What we think but cannot yet show
Everything on this page is explicitly not established. Each entry names what would have to be true, what result would kill it, and either a bench experiment or an analysis we can actually run.
Two properties are left off the chart on purpose. Every one of these entries carries published counter-evidence, and 1334 of 1338 records its own open questions — both are requirements of the schema rather than findings, so a bar for either could not move, and a statistic that cannot move cannot tell you anything. 1097 of 1338 additionally carry a bite_back note, which is required only where an entry touches an intervention.
Each carries a falsifier: the result that would end it.
A KMT2A fusion is necessary but not sufficient: the cell it lands in decides whether there is a disease and which one
The same KMT2A fusion can produce childhood leukaemia, an adult thymus tumour, or nothing at all for thirteen years. If one genetic change gives three different answers, the change cannot be the whole story - the cell it happens in is doing much of the work.
A menin inhibitor is a partial LXR agonist, and no trial in the class has looked
Menin switches off the genes that build fat and move cholesterol. If a drug removes menin, those genes should switch on. Every child on a menin inhibitor is running that experiment and nobody is measuring it.
A menin inhibitor may release RELA from menin's restraint — and released RELA is what keeps MLL fusions on chromatin. Prior 0.20
Menin holds down NF-kB, and NF-kB turns out to be needed to keep the KMT2A fusion protein attached to the genes it controls. So a drug that knocks menin off could accidentally free NF-kB and help the fusion hold on — the drug partly undoing its own job. Nobody has looked, and there is one good structural reason it might not happen at all.
A menin inhibitor's chromatin signature is organised by how much fusion it removed from each site, so every mark's response is a GRADIENT in eviction depth - and any readout that averages over a fusion-occupancy set is underpowered by construction, sometimes to exactly 1.000
The drug's effect on chromatin is not one number per mark - it is a slope across sites, steepest where the drug removed the most fusion protein. Any measurement that averages across all the sites at once will understate the effect, and can report exactly nothing when a great deal has happened.
A published MLL leukaemia stem-cell programme immortalises progenitors WITHOUT Hoxa or Meis - Myb, Hmgb3 and Cbx5 - and nobody has ever asked what a menin inhibitor does to it, although this graph reads the drug's effect almost entirely through HOXA9 and MEIS1
The only targeted drugs for this leukaemia work by switching off two genes, HOXA9 and MEIS1. A 2009 paper found three other genes that together keep MLL leukaemia stem cells self-renewing without needing HOXA9 or MEIS1 at all. If that route is active in patients, it is a way for the leukaemia to survive the drug - and it has never been tested.
AFF1 reaches P-TEFb only through AFF4, and the index fusion's elongation phenotype is therefore a claim about AFF4 availability
Everyone says the AF4 protein in the infant-leukaemia fusion grabs the machine that switches genes on. Nobody has ever shown AF4 touching that machine directly - and its close relative AF5q31, which has been crystallised holding it, may be the one doing the grabbing for both.
AFF1- and AFF4-containing super elongation complexes occupy physically distinct neighbourhoods in living cells, and KMT2A::AFF1 inherits the AFF1 one
AFF1 and AFF4 are interchangeable parts of the same machine, but in living cells they may sit in different places with different company. If so, the fusion built from AFF1 inherits AFF1's company, and treatments aimed at the machine as a whole would miss.
Against a correctly measured same-condition floor, a menin inhibitor's effect on H3K4me3, H3K79me2 and H3K27me3 is a genome-wide redistribution - up in the decile below the peaks, down monotonically to the sparsest decile - while its effect on the FUSION stays confined to the top 0.1%. The drug displaces the complex at the peaks and moves the marks everywhere
A menin inhibitor knocks the cancer-driving protein complex off a small number of very strong sites. This proposes that its effect on the chemical tags on DNA packaging is much broader - shifting across nearly the whole genome - which would mean the drug's consequences are far less targeted than the protein-displacement picture suggests.
AKAP95's requirement in KMT2A-rearranged leukaemia runs entirely through IDR co-condensation with the retained fragment, and not at all through DPY30 - so a DPY30-binding-dead AKAP95 will fully rescue
AKAP95 touches the KMT2A machinery in two ways. We propose that only one of them - the droplet-forming one - matters for the leukaemia, and that an AKAP95 which cannot grip DPY30 will work perfectly well in a leukaemia cell.
An ENCODE-only archive bound systematically undercounts chromatin-modifying enzymes, because ENCODE's target catalogue is transcription-factor-shaped - measured 3-for-3 on the H3K9 machinery, with ratios of infinity, 2.8x and 7.3x
The big public atlas this project checks first was built mainly to map transcription factors, so enzymes that modify chromatin are thinly represented in it. Every time we checked one of the H3K9 enzymes, other archives had more data than the atlas did - which means an absence declared from the atlas alone is likely to be wrong in one direction.
Atlas-only interactome partners are predicted by how abundant the prey is, not by which bait was pulled - and if so this graph's interactome denominator is inflated
TESTED AND IT DID NOT HOLD. The idea was that the weakest partners found by huge screens are mostly just abundant proteins that stick to everything, which would have made the list still to write up shorter than it looks - but across all 31,944 results of the screen, how common a protein is barely affects whether it lands in the weak group.
Both MBTD1 fusions end at H3K27me3-poor chromatin, one by excluding PRC2 and one by poisoning it - so the shared endpoint is the phenotype and MBTD1 is the coincidence
MBTD1 is part of two different cancer fusion genes, and both of them appear to end up with the same missing chemical tag on DNA - the tag that keeps genes switched off - but they get there in opposite ways. One builds a chemical environment the silencing machine cannot invade; the other carries a decoy that jams the machine directly.
CALM-AF10 was meant to separate the menin axis from the HOXA programme - and the indirect evidence has now turned against separability, because KMT2A-free HOXA lesions keep proving menin-DEPENDENT
Two ideas have been quietly treated as one: that leukaemias need the menin protein, and that they run on the HOXA growth genes. CALM-AF10 turns on the same HOXA genes and needs the same DOT1L enzyme as the KMT2A leukaemias, but it contains no KMT2A at all - so if it does not need menin, the two ideas come apart.
Cell state in KMT2A-rearranged leukaemia is organised by lineage, not by the fusion — five independent axes split the same way, and every one of them splits AML from ALL rather than partner from partner
Across five completely different kinds of measurement — energy use, fat requirements, oxygen response, how cells respond to the newest drug, and which survival protein they depend on — the myeloid and lymphoid forms of this leukaemia behave differently, while different fusion partners within one form behave alike. That pattern suggests the cell's identity, not the fusion, is what sets these states.
Chromatin-bound XPO1 as a shared, druggable address for HOXA-driven leukaemia - and the objection that it may be a general chromatin cofactor rather than an address at all
The idea is that a single protein parked on the HOX genes acts as a docking station that many different leukaemia mutations grab onto, and that blocking it could treat all of them at once. The uncomfortable extension is that KMT2A-rearranged leukaemia, which was supposed to use a completely different route, might be using it too.
Competing model 3: reciprocal status is a proxy for where the break fell, and the thing that changed is the DIRECT product
A third explanation: the mirror-image gene is not doing anything. Whether a patient has it depends on exactly where the chromosome broke - and that same break also decides how much of KMT2A ends up in the famous fusion protein. The break, not the mirror image, may be what matters.
CTCF loss on KDM3B depletion is a graded property of anchor strength, not a set of targeted events: decile medians fall monotonically 0.78 to 0.62 across 12,443 RPE anchors, and if that gradient is biological then every strong anchor in the genome is a candidate instability site
When one enzyme is removed, loop-anchor protein falls off sites all over the genome - and the stronger a site was to begin with, the larger the proportional loss. If that is real biology rather than a statistical artefact, then the KMT2A gene is not being singled out; it is simply one of the strongest anchors, and every other strong anchor is at the same kind of risk.
CTCF/cohesin occupancy at the KMT2A exon-11 anchor is both the cause of the break and the chaperone of its correct repair - so losing it gives FEWER breaks and MORE rearrangements
The same protein clamp that invites the cutting enzyme to the KMT2A gene may also be what holds the cut ends together so they can be sewn back correctly. If so, taking the clamp away produces fewer cuts, but each cut is far more likely to end up as a cancer-causing rearrangement.
CXCR4 function in KMT2A-rearranged blasts is an output of the fusion's own HOXA/MEIS1 program, so a menin inhibitor should switch off the niche axis without a CXCR4 antagonist
The leukaemia's central switch, which the drugs in this project turn off, also seems to control how well the leukaemia cell uses the marrow's hold-still signal. If that is right, those drugs are already doing part of the job a second niche-targeting drug is being proposed for.
Does KMT2A::AFF1 change histone turnover at its target genes, or does it read a chromatin state it did not make?
A KMT2A fusion forces genes to be read continuously, and reading a gene knocks its histones out. So the fusion may be speeding up histone replacement at the genes it controls - or it may just be sitting on genes that were already turning over fast. The two possibilities predict different things about how quickly a drug's effect should appear and disappear.
DOT1L was not refuted - the wrong thing was blocked, at a dose nobody finished escalating
The usual story is that DOT1L was tested in patients and failed, so the target is dead. This hypothesis says the drug blocked only the enzyme's chemistry while leaving its other job intact, and was stopped before anyone found the highest safe dose - so the target was never really tested.
DPY30 aims the H3K4 machinery while ASH2L powers it, which is why only DPY30 tracks menin
Four proteins form the shared core of the enzyme that marks active genes, and only one of them, DPY30, is needed in the same cells that need menin. We think that is because DPY30's job is aiming the enzyme at the right genes, which is exactly what a KMT2A fusion held in place by menin requires.
Every chromatin readout of a menin inhibitor measured in this corpus is graded by how much fusion the drug removed from a site, so a within-sample average over the fusion's occupancy set is the wrong statistic and will report near-null in a responding cell - the signal is the SLOPE against eviction depth, not the level
The drug's effect on the chromatin is strong where the drug works hardest and absent where it barely works, so averaging over all the sites cancels it out. The proposal is that the useful measurement is how steeply the effect changes across sites, not its average.
Every PI-dimension coverage line in this project is two-valued, and a third of what it counts as missing is a state the metric has no way to express - so the epigenetics figure will not improve by filling fields, and should not be made to
The score for each topic has only two settings: recorded, or not recorded. But there are three real situations - we know it, nobody has measured it, and there is nothing to know because the protein never touches chromatin. Lumping the last two together is why the chromatin score looks so bad, and filling in fields would hide the problem rather than fix it.
Fifth explanation for the Fanconi signal: it runs through menin binding FANCD2, not crosslink repair — tested, refuted, prior 0.07
This graph has found that KMT2A-rearranged leukaemia cells depend unusually heavily on the Fanconi DNA-repair genes, and four explanations for why have all failed. A fifth was proposed — that the link runs through menin physically holding one Fanconi protein, FANCD2 — and this entry states it precisely enough to be killed, and then reports that it has been.
H2B ubiquitination is upstream of the fusion's defining mark and is still not a target, because the fusion uses a general mechanism rather than a special one
KMT2A-fusion leukaemia needs the H2B ubiquitin tag, but so does almost every other dividing cell, because the fusion is exploiting the ordinary transcription machinery rather than a special version of it. If that is right, no drug aimed at the tag-writing enzymes will ever have a safety margin.
H3K79me2 reports only the DOT1L arm of a menin inhibitor's action and is blind to its PRC1.1 arm - so the mark's response is unchanged in a PCGF1 knockout even though PCGF1 loss is a published resistance mechanism, and a cell resistant through PRC1.1 loss would look fully drug-engaged by this readout
The menin drug does at least two things to chromatin: it takes the fusion off the DNA, and a repressive complex moves into the vacated spots. The H3K79 tag tracks the first of these and not the second. That matters because losing a component of that repressive complex is a published way for the drug to stop working - and cells that fail that way would still look like the drug was engaged if you measured the tag.
Half the menin at fusion-occupied promoters is held by a tether that is not the fusion
When the fusion protein is destroyed, half the menin at its genes stays behind. We propose that this leftover menin is being held by a different partner - most likely LEDGF, or ordinary MLL1 - and that this leftover pool is what a menin-blocking drug fails to reach.
HP1-gamma holds a constitutive, sub-threshold foothold on the KMT2A exon-11 CTCF anchor in undamaged cells - the CTCF-primed model's missing resting state - and the reason nobody has seen it is that it sits below the reproducibility filter every public peak set applies
One model of how the KMT2A gene breaks says a repair-licensing protein is only recruited after damage. A faint but reproducible signal for that protein sits on the exact spot in undamaged cells, which would mean the licence is partly pre-loaded rather than built from scratch - but the signal is faint enough that it might be nothing.
Hypothesis — forced-oligomerisation interfaces are the mutation-resistant drug targets the fusion field has been ignoring
Several fusion proteins only work because the partner glues copies of them together, and the idea here is that a drug that unglues them could not be escaped by a simple mutation. One case has now been checked against patients and it went the other way: arsenic, which acts on exactly such a glue in a different leukaemia, was defeated by a single amino-acid change while the cancer protein kept working.
Hypothesis — the elongation hijack is how one partner class satisfies a more general requirement, and the general requirement is what should be drugged
This graph explains KMT2A leukaemia by saying the partner brings in the machinery that reads genes too fast. That cannot be the whole story, because some partners are cytoskeleton proteins with no such machinery. The proposal here is that what all partners really supply is a way of sticking copies of the KMT2A fragment together, and that the gene-reading machinery is just one way of cashing that in.
Hypothesis — the menin lesson generalises: every non-enzymatic fusion has an obligate partner interface that is more druggable than the fusion
Menin is a protein that KMT2A fusions must grab hold of, and blocking that grip became a drug that works for two different leukaemias at once. We propose that most other pocketless fusion proteins also have one partner they cannot do without, and that this is where their drugs will come from.
Hypothesis (first clause REFUTED 2026-09-21): the fusion's core dependencies may be cell-type-invariant, but the cell type itself is not fixed
The idea is that what kind of cell the leukaemia came from is decided first, and the fusion then runs the same core machinery regardless. If true, the central drug targets would work the same way in myeloid and B-cell versions of the disease.
Hypothesis: a defined subset of fusion binding sites survives saturating menin inhibition on menin-independent tethering, and that residual set is where bypass resistance comes from
The drug works by prising menin off the DNA and taking the fusion protein with it. But the fusion has other ways of holding on, so some sites should stay occupied no matter how much drug you give. We think those leftover sites are exactly the ones that let the leukemia escape.
Hypothesis: a menin inhibitor must come BEFORE an antigen-directed immunotherapy, not with it or after it - and every published experiment has already assumed this without anyone testing it
If a menin inhibitor works by making leukaemia cells easier for the immune system to see, then giving it first and the immune therapy afterwards should work much better than the other way round. Every laboratory experiment has quietly done it that way, and no trial has ever asked.
Hypothesis: a single dormant leukemia stem cell state confers cross-resistance to menin inhibition, venetoclax and T-cell engagement at once — the common-cause failure that orthogonal combination therapy assumes away
Adding drugs that work in completely different ways is supposed to close every escape route. Engineers who build backup systems know they usually fail together for one shared reason. There may be one state these cells can enter that defeats all three approaches simultaneously.
Hypothesis: adaptive scheduling of menin inhibitors will fail, because MEN1 resistance mutations probably cost the cell nothing
There is a treatment strategy that deliberately leaves some drug-sensitive leukemia alive to keep the resistant cells in check. It only works if the resistant cells are worse off when the drug is away. For this drug, nobody has checked, and the closest comparison suggests they are not.
Hypothesis: CLEC12A is a cell-intrinsic dependency in KMT2A-rearranged leukaemia, not only a surface marker - and if so, antigen-negative escape is self-defeating
In one other kind of leukaemia, deleting CLEC12A killed the cells, so CLEC12A may be something the cancer needs rather than just a label on its surface. If the same is true in KMT2A leukaemia, then the way a CLEC12A therapy usually fails - the cancer drops the target - would instead make the cancer weaker.
Hypothesis: genetic and non-genetic escape from menin inhibition are substitutes, so MEN1 mutations will be rare in infants and the field's resistance biomarker will fail in children
A leukemia cell escapes a drug by whichever route is cheapest. If it can simply change its behaviour, it never needs to mutate. Infant leukemias carry almost no mutations, so they should escape by changing behaviour, and the genetic test the field relies on will come back clean while the child relapses.
Hypothesis: H3K79me2 at a KMT2A fusion's own sites survives menin inhibition when the C-terminal partner contacts DOT1L directly (MLLT3/AF9) and decays when it does not (AFF1) - so the mark's response is a property of the partner and not of menin, which predicts the opposite of the direct-contact intuition and is testable in one line
The histone tag disappeared in one leukaemia cell line and stayed put in another. The two lines differ in which partner gene is fused to KMT2A, and the one where the tag survives is the one whose partner grips the tag-writing enzyme directly - which is backwards from what you would guess.
Hypothesis: IGF2BP2 and IGF2BP3 buffer each other in KMT2A::AFF1 leukaemia, so single-gene CRISPR under-reports both
If two of the three IGF2BP siblings can cover for each other in MLL/AF4 leukaemia, then deleting either one alone would look almost harmless even though the pair is essential. That would explain why mouse experiments say IGF2BP3 matters a lot and cell-line screens say it barely matters.
Hypothesis: KMT2A and AFF1 break where they do because both breakpoint regions are CTCF/cohesin loop anchors sitting in the same nuclear hub
Chromosomes are held in loops, pinned at anchor points, and those pins are where DNA most often snaps. This hypothesis says KMT2A and its commonest partner are both pinned, both pinned in the same part of the nucleus, and that is the whole explanation for the fusion.
Hypothesis: KMT2A-rearranged infant ALL is two populations, not one — a mixture-cure model will fit better than proportional hazards, and combination therapy only needs to be given during the first year
Nearly all relapses happen in the first year and then almost stop. That pattern usually means you are looking at two kinds of patient mixed together, not one kind whose risk fades. If so, the job is to find out at diagnosis which child is which, and to add extra drugs only during the dangerous window.
Hypothesis: KMT2A::AFF1 does not require phase separation — recruitment biochemistry is sufficient, and the condensate model does not transfer from NUP98/ENL
We predict that the KMT2A-AFF1 fusion works by simple recruitment — grabbing the right machinery at the right genes — and does not need to form liquid droplets, unlike some other fusion proteins that genuinely do. This is worth settling because a wrong answer would send the field chasing droplet drugs.
Hypothesis: KMT2A::AFF1 is continuously required, not a hit-and-run initiator whose job is done once the epigenome is set
One worry about drugging the fusion is that it might only be needed at the very beginning, like a match that lights a fire and is then thrown away. If that were true, removing it later would not put the fire out. The evidence says the fusion is more like the gas supply than the match.
Hypothesis: KMT2A::AFF1 mislocalises to or disrupts nuclear speckles, and paraspeckle biology (NEAT1) is a fusion-repressed differentiation brake
Cell nuclei contain droplet-like compartments that concentrate the machinery for processing RNA. The partner half of this fusion belongs to a protein family that lives in those droplets and can disturb them, and in a different fusion leukemia a related compartment controls whether cells mature. Neither has been checked here.
Hypothesis: KMT2A::AFF1 rewires splicing, because AFF1's own protein family are G-quadruplex RNA-binding splicing modulators and the fusion keeps AFF1
The second half of this fusion protein comes from a family of proteins whose day job includes deciding how RNA gets cut and stitched together. The fusion keeps that half, so it should change RNA stitching - but the RNA-stitching changes that have actually been found in this family of leukemias are explained by other proteins entirely, and nobody has tested whether the fusion's own half does anything.
Hypothesis: KMT2A's AT-hooks are a DNA-topology reader, not an AT-sequence reader — which would link the fusion's address to the scaffold-attachment architecture that also makes its own gene fragile
The hooks on KMT2A are usually described as gripping AT-rich DNA, but the only experiment on KMT2A's own hooks found they recognise a DNA shape instead — the four-way cross that forms where DNA is twisted. If that is right, the protein is drawn to the same kind of structurally strained DNA that sits around the point where its own gene breaks in this disease.
Hypothesis: LEDGF licenses elongation at fusion target genes rather than anchoring the fusion, which predicts that a LEDGF agent is epistatic with BET and CDK9 inhibition and invisible in a fusion-ChIP assay
We propose that LEDGF does not hold the leukaemia protein onto DNA but instead lets it switch genes on. If that is right, a LEDGF drug would look like it is doing nothing in the standard test, and would not add much to drugs that block the same switch.
Hypothesis: menin inhibition does not clear the fusion from chromatin but redistributes it onto the four fifths of the unmethylated CpG address it normally ignores
The fusion protein can stick to a particular chemical pattern in DNA, and there is five times more of that pattern than it actually uses. Menin seems to pick which fifth. Take menin away and the fusion may not fall off at all - it may spread out over the rest, doing nothing useful at first, but sitting somewhere new.
Hypothesis: menin inhibition does not merely displace KMT2A::AFF1 from chromatin, it destabilises it — the evicted fusion is degraded, not parked
Menin inhibitors knock the fusion protein off DNA, and everyone assumes it then sits in the nucleus waiting to come back. This proposes the opposite: stripped of its chromatin anchor the fusion becomes unstable and the cell destroys it, which would mean the drug is a partial degrader and that intermittent dosing could work.
Hypothesis: menin inhibition pushes cells toward the GMP-like state, and the GMP-like state is where menin inhibition works best — so a lineage switch on therapy should predict response, not failure
Two separate studies say things that do not fit together: leukaemias that change lineage under treatment end up looking like a particular kind of myeloid cell, and that same kind of cell is the one a menin inhibitor kills best. If both are true, a lineage switch on the drug should be good news, yet the one patient profiled switched and did badly.
Hypothesis: menin inhibitors differ in how much JUND they liberate, and the least JUND-liberating compound will be the most durable
Menin holds a second protein, JUND, in check. The drugs that displace the fusion also let JUND go, and free JUND helps the leukemia survive. If different drugs release different amounts of JUND, the best drug is the one that releases the least.
Hypothesis: menin-inhibitor effect is hysteretic — the program collapses faster than it recovers — so intermittent dosing can hold efficacy at much lower cumulative exposure
Switching the fusion protein off happens within minutes. Nobody has ever measured how long it takes to switch back on after the drug is stopped. If switching back on is slow, a child could be given the drug in pulses instead of continuously, with the same effect and far less toxicity.
Hypothesis: menin-inhibitor sensitivity falls with cell-cycle exit, so the compartment that causes relapse is the one the drug reaches least — and nobody has sorted by cycle state to check
Menin inhibitors work by letting a leukaemia cell mature instead of staying stuck as a blast. Maturing is something a cell has to actively do, and the cells that survive treatment are the ones that have stopped doing almost anything. The prediction is that these drugs are weakest against exactly the cells that come back, and the experiment to check it has never been run.
Hypothesis: only a mechanistically orthogonal partner pre-empts menin-inhibitor resistance; a second chromatin agent will not
Every way this leukemia escapes a menin inhibitor works by rearranging how genes are switched on. So adding a second gene-switching drug should fail, while adding something that kills by a completely different route — an immune therapy or an apoptosis drug — should work.
Hypothesis: prior therapy itself installs menin-inhibitor resistance, and the dose of prior therapy predicts it better than any mutation does
The treatment a child has already had may be what makes the next drug stop working. If so, you could predict who will not respond just by counting what they have already received, and the right answer is to give the drug earlier.
Hypothesis: releasing the KMT2A::AFF1 differentiation block sends cells toward myeloid, not mature B, in a measurable fraction of infant cases
Menin inhibitors work by unfreezing leukaemia cells so they can finish maturing. But these particular cells were never fully committed to being B cells, so unfreezing them might turn them into myeloid cells instead of healthy B cells. That would look like a response followed by a different leukaemia.
Hypothesis: reserving menin inhibitors for relapse in infants is actively harmful, because pretreated disease is already resistant
Menin inhibitors are currently given only after other treatments fail. But there is evidence that leukemia which has already been through heavy chemotherapy stops responding to menin inhibitors even before it has ever seen one. If that is right, saving the drug for later means giving it when it can no longer work.
Hypothesis: SET::NUP214 activates HOXA partly through recruited wild-type MLL1, which would make a KMT2A-unrearranged leukaemia menin-inhibitor-sensitive
If SET::NUP214 switches on HOXA by borrowing the normal, unbroken KMT2A protein, then the menin drugs that pull KMT2A off chromatin should work in this leukaemia too - even though it has no KMT2A rearrangement at all.
Hypothesis: t(4;11) infant ALL is a hypermethylated genome whose transposable elements can be reawakened, and that reawakening is an untested immunotherapeutic lever
Our DNA is full of ancient virus-like sequences that are normally kept switched off by chemical tags. t(4;11) infant leukemia is unusually heavily tagged, and drugs that strip the tags wake these sequences up in other cancers and make the cell look infected to the immune system. A trial has now shown those tags can be stripped safely in these infants - but doing it alongside chemotherapy did not help them live longer, and nobody has yet checked whether the immune alarm even goes off.
Hypothesis: the ceiling on menin-interface disruption is cell-state plasticity, not target mutation
We think the main reason blocking menin will fail is not that the drug stops fitting — it is that the leukemia cell can shift into a different state where it no longer needs the fusion sitting on its genes. If true, the drug needs a partner that locks the cell state.
Hypothesis: the cell-of-origin methylome, not the breakpoint, decides whether a KMT2A::AFF1 leukemia is HOXA-high or HOXA-low
The fusion protein finds its genes by reading which DNA sites are chemically untagged. So the genes it switches on should depend on the chemical state of the cell it landed in — which would mean an infant's fetal cells and an older child's cells give different diseases from the same fusion.
Hypothesis: the fusion is exempt from KMT2A's regulation, not free of it - the brakes still operate at the same loci, on the other allele
Instead of saying the leukemic protein escaped its controls by deleting them, this says the controls are still running at the same genes on the healthy copy of the gene. For one molecule that is just another way of saying the leukemic protein has nothing to grab onto - but for the gene it sits on, it predicts something, and in one case a drug has already proved it.
Hypothesis: the fusion output program fails at a threshold on total cut-set suppression, so three nodes at 40% beat one node at 90%
No single gene in the leukemia program is essential on its own; knocking one down only partly works. If the program collapses when enough of a small group of genes is suppressed together, then partly suppressing three of them could work better, and be far less toxic, than nearly eliminating one.
Hypothesis: the haem and iron dependency of KMT2A-rearranged lines is monocytic physiology, not fusion biology — and the panel cannot tell us either way
This project found that fusion-carrying leukaemia cells need iron-and-haem genes more than other leukaemia cells do. Almost all those cells are monocytes, whose normal job is recycling iron, so the finding may be about the cell rather than the fusion — and the comparison group that was supposed to settle it turns out to contain the wrong cells.
Hypothesis: the HOXA-negative, IRX1-negative KMT2A-rearranged blast is a low-oncoprotein-dosage state, not a state whose oncoprotein has been redirected to another programme
When the fusion protein is missing from its usual genes, the simplest explanation is that there is not much of it on the DNA anywhere - not that it has gone somewhere else. That is a testable difference, and it predicts a drug that targets the fusion will do less in those patients.
Hypothesis: the KMT2A break can be prevented rather than the fusion treated - stated so it can be killed, and it probably should be
If we knew what cuts KMT2A, could we stop the cut instead of treating the leukaemia it causes? The honest answer is that we know at least one molecule whose blockade protects the right human cells, and that almost everything else about the idea is wrong: the cut happens before birth, it happens in healthy people all the time, and the step that is actually rare is not the cut.
Hypothesis: the KMT2A breakpoint map is shaped by product viability, not only by DNA fragility, and the signature is that breaks respect inter-domain linkers
Chromosome breaks in this gene are not scattered randomly; they cluster in one small region. The usual explanation is that the DNA there is fragile, but there may be a second reason: those are the only places you can cut the protein and still get something that works.
Hypothesis: the KMT2A genomic breakpoint carries prognostic information that the field has been sitting on for twenty years
The exact place where the chromosome breaks differs between patients, and we think that difference changes how the disease behaves. It was tested once in thirty patients and never again, while two groups hold everything needed to test it properly.
Hypothesis: the menin inhibitor's effect is inactivation in place, not eviction - the fusion keeps most of its sites and loses its output
There is a third possibility between the fusion being thrown off the DNA and it moving somewhere new: it stays exactly where it was and simply stops working, because the machinery that made it productive has gone. The published data fit this better than they fit either of the other two.
Hypothesis: the menin-MLL complex writes the same mark onto two address lists with opposite cell-state outputs, and which one a tissue reads decides whether blocking menin is a therapy or a lesion
The same molecular machine switches on growth-blocking genes in hormone glands and identity-freezing genes in blood. If that is right, a drug that jams the machine helps in one tissue and harms in the other, for exactly the same reason.
Hypothesis: the myeloid clone that appears after CD19-directed therapy was already there at diagnosis, and its abundance is a pre-treatment biomarker
When a B-cell leukemia turns into a myeloid one after immunotherapy, the myeloid cells may not be newly converted; they may have been present in small numbers all along and simply grown once their competitors were killed. If so, you could count them before treatment starts.
Hypothesis: the paediatric gap is set by the population ratio, not by paediatric biology
The reason babies get left out of new cancer drug trials may have nothing to do with how hard babies are to treat, and everything to do with what fraction of the eligible patients they are. If so, you can predict in advance which new drugs will reach children and which will not, just by counting patients.
Hypothesis: the PTH rise on menin inhibitors is on-target and exposure-dependent, not renal
We think the rise in parathyroid hormone seen in a third of patients on revumenib is caused by the drug hitting menin in the parathyroid gland itself, not by the kidneys. If so, patients with higher drug levels should show bigger rises, and the effect should fade after the drug stops.
Hypothesis: the R-loop landscape of a KMT2A::AFF1 blast is remodelled by the fusion and is both a vulnerability and a source of ongoing genome instability
When a gene is read, the RNA can stay stuck to the DNA and form a three-stranded knot called an R-loop. Fusion proteins force genes to be read very hard, so these knots should pile up - but they pile up in any hard-read gene in any cell, the main way of measuring them is disputed, and nobody has measured them in the infant B-cell form of this leukemia at all.
Hypothesis: the reciprocal AFF1::KMT2A product protects only when it is structurally complete, so its prognostic effect is a breakpoint class in disguise
In half of these leukemias the swap produces a second, mirror-image gene as well as the cancer-causing one, and children whose leukemia makes it do far better. We think that is not luck: the mirror gene only helps when the break left it intact, which means the real prognostic variable is where the break fell.
Hypothesis: the same drug is protective in the islet and harmful in the parathyroid and the growth plate
We think blocking menin helps insulin-making cells while harming the parathyroid glands and the growing skeleton, because those tissues use menin for different jobs on different timescales. If that is right, the same child can benefit and be harmed at the same time, in different organs.
Hypothesis: the same-condition floor measured this way is largely a property of where the EXTERNAL stratifier puts its top bins relative to each library's own coverage, not of how well the two replicates agree - which would explain a twentyfold arm span sitting on top of rank concordances that barely differ.
A puzzle: repeat measurements that agree about equally well produce wildly different reliability numbers. The proposed explanation is that the number depends less on the repeats agreeing with each other than on how well a third, separate dataset used for sorting happens to line up with each of them.
Hypothesis: the second hit in infant KMT2A-rearranged leukaemia is a fetal developmental window, not a mutation
The fusion cannot turn ordinary newborn blood stem cells into leukaemia in the laboratory, yet in a fetus it apparently can, almost on its own. The proposal is that the missing ingredient is not another broken gene but the particular state of a cell that only exists for a window during pregnancy.
Hypothesis: the spread in infant presenting-feature frequencies is mostly measurement, not biology
Reported rates for the same sign in infant leukaemia differ enormously between studies - CNS involvement from 14% to 34%, enlarged liver and spleen from 65% to 88%. The hypothesis is that most of that spread comes from how and when each study looked, not from the babies being different.
Hypothesis: the survival benefit of menin inhibition is governed by the rate at which it delivers patients to a curative consolidation, not by its remission rate — and the field is measuring the wrong thing
A remission lasting under five months only helps if you can use it to get the child to something that cures. So the number that matters is how many treated children actually reach a transplant or cell therapy while still in remission, and almost no trial reports it.
Hypothesis: the three quarters of H3K4me3 that survive menin displacement at fusion target sites are held by a menin-independent complex, and identifying it names the combination partner for a menin inhibitor
When the drug pries menin off, most of the chemical mark at the leukaemia genes stays put. Something else must be holding it there. Finding out what would point at a second drug to give alongside the first.
Hypothesis: transfer-RNA and ribosome biology is a genuinely empty cell in KMT2A-rearranged leukemia, and translational capacity may be the unmeasured dependency
Cells need a supply of protein-building machines (ribosomes) and RNA delivery adaptors (transfer RNAs) to grow. This matters in other blood diseases, and in KMT2A-fusion leukemia it has been looked at only once, as the side effect of a drug rather than as a property of the leukemia itself - but every fast-growing cancer cell needs more of these machines, so finding that this leukemia does too would not by itself mean anything, and the main drug used to test the idea turns out to work a completely different way.
Hypothesis: treatment-remodelled niche, not niche-imposed dormancy, is what residual KMT2A-rearranged disease exploits
The usual story is that leukaemia cells hide in a protected corner of the bone marrow and sleep through chemotherapy. The evidence assembled in this beat points somewhere else: the leukaemia and the chemotherapy both damage the marrow's support structures, and it is the damaged, rebuilt marrow — not a pristine hiding place — that the surviving cells grow back in.
Hypothesis: tubulin/microtubule association is a recurrent, unnamed property of the KMT2A partners whose transforming module has no other assignment
Two KMT2A partners that supposedly have nothing in common turn out to share one thing: they both stick to the cell's microtubule tracks. The idea here is that this is not a coincidence — that binding to a long polymer, rather than simply pairing up, may be how some of these partners force KMT2A into the clusters that cause leukaemia.
Hypothesis: what breaks KMT2A in a fetus is TOP2B doing its normal transcription-coupled job in a non-cycling progenitor, with no poison involved and repair capacity as the rate-limiting variable
Therapy-related KMT2A leukaemia happens because a drug jams a DNA-cutting enzyme. Infants get the same disease before birth with no drug anywhere, and this hypothesis says the answer is the same enzyme doing its ordinary job in a cell that is not dividing - so the thing that differs in the rare fetus is not the cutting, but how well the cut is cleaned up.
Hypothesis: what the KMT2A half loses is control of its own abundance, not control of its activity
Everything the broken half of KMT2A throws away that anyone has actually tested turns out to be a piece that tells the cell when to destroy the protein. So the leukemic protein's main advantage may not be that it does something new - it may simply be that it never goes away.
Hypothesis: where a blast goes when the fusion program is released is written in its methylome at diagnosis, not in its transcriptome
When a drug switches off the leukemia program, the cell goes somewhere. We think it goes wherever its original identity was set before it became cancerous, and that identity is recorded in a chemical marking on the DNA that the cancer never erased.
Hypothesis: where a KMT2A fusion sits in the nucleus is set entirely by retained KMT2A modules, and no partner changes the compartment - so every published 'the fusion is here' result is a wild-type KMT2A result
Every experiment that has looked at where an MLL fusion protein sits inside the nucleus found it in the same place as the normal protein, whichever partner was attached. This proposes that is not a coincidence or a limitation of the pictures: the address really is written entirely in the MLL half, and the partner changes what the protein does once it arrives, not where it arrives.
Hypothesis: YEATS-pocket occupancy disables the reader module of all three SEC-family complexes in proportion to how much of the reader pool each holds — so the drug's footprint is set by an unmeasured partition, not by the fusion
The idea to test: a drug that plugs the reading pocket cannot tell which of three machines it is sitting in, so it switches off reading in all three. How much that matters depends entirely on one number nobody has measured — what fraction of the reader proteins live in each machine.
If HOXA9 substitutes for BMI1 at CDKN2A, then a menin inhibitor should CREATE a BMI1 dependency that does not exist at baseline
BMI1 looks useless as a target in this leukemia: knocking it out does nothing, because the fusion's own product HOXA9 does BMI1's job of silencing the brake gene. But that is exactly the setup for a two-drug trick - take HOXA9 away with a menin inhibitor and BMI1 should suddenly become essential.
If inhibiting one methyltransferase failed, constraining the methyl donor may not
Blocking a single methyl-writing enzyme did not help patients. This asks whether limiting the shared chemical those enzymes all consume would do better, and what it would cost.
If RPA loading at the KMT2A breakpoint cluster region is the licensing step for the etoposide-induced translocation, then the nuclear RPA reserve is a modifier of therapy-related KMT2A-rearranged leukaemia risk - and an ATR inhibitor, which kills by exhausting RPA, would raise that risk rather than lower it
After chemotherapy with etoposide, the DNA-coating protein RPA piles up on the exact stretch of the KMT2A gene where the cancer-causing break happens - and it piles up more, and for longer, when the damage-sensing kinase ATM is missing. If that piling-up is what allows the translocation to form, then how much RPA a cell has spare becomes a risk factor, and drugs that use up a cell's RPA could make therapy-caused leukaemia more likely, not less.
If the fusion's address requires unmethylated CpG, a hypermethylating co-mutation should be selected AGAINST - and IDH2 is selected FOR, which means the address is not the limiting resource
The KMT2A fusion can only land on DNA that has no methyl marks. A mutation that adds methyl marks should therefore make life harder for it, and should be rarer in those leukaemias - but the data show it is commoner. Something about the simple picture is wrong, and this entry says which part.
Ig-based MRD fails specifically in the KMT2A-rearranged patients whose IGH repertoire is proximally biased and still recombining
The idea is that the leftover-disease test built on antibody genes fails in a predictable subset of babies - the ones whose leukaemia is still rearranging those genes and using the nearest available segment - rather than failing at random. If true, you could tell in advance which children need a different test.
IKK inhibition and menin inhibition evict the fusion from chromatin by routes that do not share a node, so they should be additive — and the same reasoning predicts that a menin inhibitor alone partially defeats itself. Prior 0.30
Two completely different drugs both end up knocking the leukaemia fusion protein off the genes it controls: one by removing its anchor protein menin, the other by blocking an immune kinase cascade. Because they work through different molecules, using both should knock it off harder — but there is a catch, because removing menin may switch the kinase pathway on.
In a cell with no fusion, a menin inhibitor is substantially a KMT2B drug
In a leukaemia cell, a menin inhibitor mostly attacks the fusion protein. In every other cell in the child's body there is no fusion, so the drug can only be acting on menin's two normal partners - and one of those, KMT2B, has never been measured under the drug at all.
In KMT2A::AFF1 there is no oncogenic DOT1L unit: all H3K79 methylation is the endogenous MLLT10-directed activity, working on genes the fusion has opened
If the two-machine picture of the H3K79 tag is right, and if the infant fusion has no hook for the tag-writing enzyme, then the tag on its target genes must be put there by the cell's normal machinery. That predicts a menin drug should hardly change the tag at all in infant leukaemia - the opposite of what the project currently expects.
In NUP98-rearranged leukaemia, who responds to a menin inhibitor is decided by the co-mutations, not by the fusion
Every NUP98 patient is currently treated as one group for menin drugs. The evidence suggests the fusion only makes the cell eligible to respond, and whether it actually responds is decided by the other mutations it happens to carry.
In the human index fusion the menin inhibitor subtracts MENIN and leaves MLL1-N-reactive material in place at the same sites - so the drug's chromatin action is decoupling rather than eviction, and what this literature calls fusion eviction is a menin measurement
The drug takes menin off the DNA almost completely but leaves the MLL1 protein where it was, at the very same spots. If that surviving material includes the cancer fusion protein, the drug is separating two partners rather than clearing the fusion away, and a different kind of drug would be needed to remove it.
Induction chemotherapy phospho-inactivates KMT2A fusions through PIKK sites the partner retains - and a DNA-PK or ATM inhibitor would block that
Three of the four partner proteins carry an off-switch operated by the kinases that respond to broken DNA, and in each case the switch sits on the part of the protein the cancer fusion keeps. Since standard chemotherapy breaks DNA, it may be turning the fusion off for a while without anyone realising - and a class of drug now being trialled alongside chemotherapy would stop that from happening.
IRX1 is a second address for the KMT2A::AFF1 oncoprotein, not merely the gene that turns HOXA off
Either the leukaemia protein has moved house from HOXA to IRX1 in half these patients, or IRX1 is something the cell switched on that then silenced HOXA. Both fit what we can see, and telling them apart changes which drugs should work.
IRX1 occupancy is a lineage-switch cell state, and in GSE252378 it is confounded with the AFF1 partner and one institution
The three samples where the fusion sits on IRX1 share three things, not one: the same fusion partner, the same hospital, and the same unusual disease behaviour where the leukaemia changes lineage. Any of the three could be the reason.
IRX1 status is set by the cell of origin, and the fusion partner predicts it only because different partners transform different cells - so the partner supplies no IRX1 address
Maybe IRX1 is not switched on by the fusion at all. Maybe it is already on or off in whichever blood cell the fusion happens to land in, and the partner matters only because different partners tend to arise in different cells.
Is the CpG address a contested resource? On current evidence, no - and the beat that went looking says so
The idea tested here is that the dozen human proteins reading the same unmethylated CG addresses get in each other's way, so the KMT2A fusion's address depends on who else is in the room. After checking every family member, the evidence does not support it - and the one direct test showed the fusion staying put while a rival pushed the normal protein off.
KMT2A is a 14-3-3 client by proxy: the contact is on an AFF-family scaffold, not on KMT2A
KMT2A keeps showing up near 14-3-3 proteins, but it has no good docking site of its own. Its partner AF4 does have one, at a spot that is already known and already phosphorylated - so KMT2A may simply be travelling with a partner that 14-3-3 holds.
KMT2A rearrangement creates a requirement for the FA ubiquitin-transfer module — the claim, its prior, and the four results that would kill it
The claim in one sentence: leukaemia cells driven by a KMT2A fusion need a specific part of the Fanconi DNA-repair machinery more than other leukaemia cells do. The evidence is a computer analysis of cell-line data that survives every statistical check this project could run — and no experiment at all.
KMT2A-rearranged cells need the factors retained through elongation MORE and the factors ejected at pause release LESS — a six-gene asymmetry nobody has reported
In an unbiased scan of every human gene, the two proteins that stay with RNA polymerase once it gets moving are needed more by KMT2A-rearranged leukaemia cells, and two of the four proteins that have to be thrown off first are needed much less. If that is real rather than noise, it means these cells run with the brake already released.
KMT2A::AFF1 cells run with more free P-TEFb than other leukaemias - and if they do not, this graph's CDK9 story is about recruitment only
If the fusion makes the cell hold less of its elongation enzyme in reserve, then drugs that block the enzyme should work differently in these cells than in others. Nobody has ever measured the reserve in a KMT2A-rearranged leukaemia, so the question is completely open.
KMT2A::AFF1 has no measured fusion-specific interactome: the one dataset that appears to supply one is a rank-selected tail in the wrong lineage, and the question is open rather than answered badly
It looks as though we know which proteins bind only our fusion — a published table lists 110 of them. On inspection that list is the weak end of a fixed-size ranking, from a T-cell line that does not have our fusion, and it excludes the proteins that are genuinely specific to it. The honest position is that nobody has measured this yet.
KMT2A::AFF1 oligomerises through the retained AFF1 CHD, putting the commonest fusion in the oligomerisation class
The bit of AFF1 kept by the fusion includes a module that, in its closest relative, makes two copies of the protein stick together. If the fusion does the same, then the most common childhood leukemia fusion works partly by the same trick as the rare ones, and it has a well-shaped surface that a drug could grip.
KMT2A's proximity promiscuity is BAIT-CLASS-SPECIFIC: it is recovered at twice its size-and-abundance prediction by RNA-binding-protein baits in a leukaemia cell and at exactly its prediction by transcription-factor baits in a kidney cell
KMT2A is found next to half of the RNA-handling proteins tested in a leukaemia cell, but only next to the expected share of transcription factors tested in a kidney cell. If that difference is real, it says KMT2A spends its time in the RNA-processing part of the nucleus - which is a claim about where the protein lives, not about which proteins it binds.
Menin dependency tracks the availability of an intact menin-binding motif, in cis or in trans — not the fusion family
Menin drugs work when the leukaemia's key protein complex is held onto DNA by menin. KMT2A fusions carry that grip themselves; NUP98 fusions borrow it from the cell's normal copy — and the one NUP98 fusion that destroys the grip is the one the drugs fail against.
Menin inhibition collapses the serine/one-carbon/ASNS programme in KMT2A::AFF1 cells - PHGDH -3.24, PSAT1 -2.40, ASNS -1.72, flat in a wild-type line - so it should re-sensitise KMT2A-rearranged ALL to asparaginase, a combination PubMed has zero records for
A menin-blocking drug switches off the machinery these leukaemia cells use to make their own amino acids, including the one gene that lets them survive the standard drug asparaginase. That predicts the two drugs should work better together, and nobody appears to have tried it.
Menin inhibition recruits PRC1 to the BENC MYC enhancer cluster and evicts BRG1, exactly as CBFbeta-SMMHC inhibition does in inv(16)
In one leukaemia, a drug that breaks a fusion protein causes a silencing machine to replace an activating machine on the DNA switches that control MYC. This entry proposes that the menin inhibitors used in this project do the same thing at the same switches, and sets out exactly what result would prove it wrong.
Menin inhibition removes the fusion from its strongest sites without relocating it to a new programme, and PRC1.1 fills in only where the eviction is complete - so the drug's chromatin action is subtraction plus local substitution, not redistribution
The menin drug appears to take the fusion protein off the DNA and leave it nowhere in particular, while a repressive complex moves into the vacated spots. If that is right, the drug is simpler and safer than the alternative, in which the fusion would move somewhere new that nobody is watching.
Menin inhibition restores graft-versus-leukaemia only where MHC class II is limiting - so the post-transplant maintenance rationale may hold in myeloid disease and fail in B-precursor infant ALL
A 2026 study showed that menin inhibitors make leukaemia cells visible to donor immune cells by switching a display molecule back on. But that molecule is already switched on in the type of cell that infant leukaemia comes from, so the same drug may do nothing for visibility in babies - even though several trials are already testing it after transplant.
Menin inhibitors should leave half the H3K79 methylation standing - and if they do not, the menin-interface model is wrong somewhere
A screen found that the H3K79 tag is put down by two separate machines: one steered by the leukaemia fusion and needing menin, and one the normal cell runs through a protein called AF10. If that is right, a menin drug should wipe out the first and leave the second - and measuring exactly how much tag is left is a direct test of the project's central idea.
Menin is the best handle on KMT2A::AFF1 because of where the break falls, not because menin is biologically special
Menin may not be a uniquely good drug target so much as the only well-studied protein that still has somewhere to hold on to after the KMT2A gene is broken. If that is right, then finding more targets means looking at the small piece that survives the break, not at the famous partners that are cut away.
Menin persists on chromatin after the fusion is degraded because LEDGF/PSIP1 supplies a fusion-independent tether - the authors' own proposal, never tested, and the single cheapest experiment that would change how menin displacement is interpreted as a drug readout
When the cancer fusion protein is destroyed menin stays stuck to DNA, and the likely reason is a second anchor - a protein called LEDGF - that holds menin independently of the fusion. If that is right, a drug that only breaks the menin-fusion contact is attacking the more replaceable of two attachments, which matters both for how well it can work and for how easily cancers escape it.
Menin primacy was established on the der(11) product alone and has never been tested against the der(4) product
The argument that menin is the fusion's one indispensable handle was built by studying the famous half of the swap. The other half is present in roughly half of these children, keeps the entire enzyme machinery, and has no menin handle at all - so the drug cannot reach it, and nobody has checked whether that matters.
MLL fusions engage NuRD's remodelling arm without its deacetylase arm - and if so, every 'NuRD is repressive' inference in this graph is pointing the wrong way
NuRD has two halves that work independently: a motor that slides nucleosomes and a set of scissors that removes acetyl marks. The published fishing experiment recovers the motor half and none of the scissors half, which raises the possibility that MLL fusions borrow the motor and leave the scissors behind - the opposite of treating NuRD as a silencing machine.
Most gene-body 'AFF1-exclusive' calls in GSE252378 are a fold-change threshold cutting an oncoprotein-level gradient - and IRX1 is the one member where the non-AFF1 samples are genuinely at background
Because patients' leukaemia cells carry different amounts of the fusion protein, a test that asks whether a gene has twice the average signal will pass in the high patients and just fail in the low ones, which looks like the gene being specific to the high group. Five of the six genes that look AFF1-specific behave exactly like that, and IRX1 does not - in the other eighteen patients its signal is essentially zero rather than just under the bar.
NG2 is absent from the candidate leukaemia stem-cell populations in t(4;11) ALL and 11q23 AML, so an NG2-directed therapy would debulk the disease and leave the fraction that initiates relapse
NG2 marks the bulk of this leukaemia but not the small population of cells thought to restart the disease after treatment. A therapy aimed at NG2 would therefore clear the visible disease and leave behind the cells that cause relapse.
NG2 reads the fusion's transcriptional programme, not the DNA lesion - so a menin inhibitor should erase the graph's best diagnostic marker while the rearrangement stays put
NG2 is the protein used to spot a broken KMT2A gene by flow cytometry, but NG2 is switched on by the broken gene's activity rather than by the break itself. If that is right, a drug that stops the broken gene from working should make NG2 disappear while the break is still there - which would quietly break the test doctors use to track the disease.
Nobody knows which way to push H2Bub in this disease, and the only available compound pushes it upward
Two good arguments point in opposite directions: the ubiquitin tag feeds the leukaemia's key mark, so less of it should help — but the same tag is what lets a blocked cell finish maturing, so more of it should help too. The only drug-like molecules that exist against this pathway raise the tag, so the question is not academic.
NUP98 fusions share the menin dependency (confirmed preclinically) — and the partner-supplied reader/writer module is not why
KMT2A-rearranged and NPM1-mutant leukaemias already share one drug because both need the same helper protein, menin. NUP98-rearranged leukaemia needs it too — this was shown in mice in 2022 — but through borrowed normal MLL1, not through the NUP98 partner module this entry originally blamed.
One demethylase family sits at both ends of this disease - KDM3B gates whether KMT2A rearranges, JMJD1C is what the resulting fusion needs to survive - and a single compound hits both, so its therapeutic index runs in opposite directions in the two cell populations it reaches
Two closely related enzymes bracket this leukaemia: losing one makes the KMT2A gene break in the first place, and the leukaemia that results then depends on the other to stay alive. The problem is that the only drug-like molecule available hits both of them, so it might treat the leukaemia in a patient while creating new ones in their healthy stem cells.
Open ground: the questions in this graph that nobody in the landscape appears to be working on
We compared the 777 open questions recorded in this knowledge graph against what every major group in the field is actually publishing. Six areas came back with real questions and no visible competitor, and that is where a small lab should spend its time.
Partner-protein occupancy is not fusion occupancy: every KMT2A::AFF1 chromatin claim resting on an AFF1, AFF4 or SEC-subunit ChIP is an upper bound that includes free complex
Many statements about where the fusion protein sits are actually statements about where one of its partner proteins sits - and that partner is also present, doing a normal job, on many of the same genes. The two are not the same measurement.
PRC1.1 is not both an effector and a resistance gene by coincidence: the two published readings are one activity measured from opposite ends
Two recent papers say opposite-sounding things about the same Polycomb machine: one says a menin drug switches it on to silence the cancer's genes, the other says losing it makes the drug stop working. We propose these are the same fact seen from either side - the drug needs that machine to do its job, so cells without it are resistant.
PRC2 is the fusion's repressive arm, recruited through menin, and the index disease's version of it runs to HDAC7
The usual story is that this leukaemia's fusion protein switches genes ON. This hypothesis says it also switches genes OFF, using the Polycomb machinery, and that it reaches that machinery through the same menin anchor the whole field is already drugging - which would make menin inhibitors partly Polycomb drugs without anyone having noticed.
REFUTED IN THE CYCLE IT WAS WRITTEN - the proposal that fusion-driven and normal-tissue-clean surface antigens are disjoint in KMT2A::AFF1 because of what the fusion's output programme is; an expression-matched permutation says it is the base rate
This entry proposed that the surface proteins this leukaemia makes because of its fusion gene are unusually likely to also appear in healthy organs. A fair comparison against other proteins made at the same level in the same cells found no difference at all, so the proposal is recorded here as tested and wrong.
REFUTED THE SAME CYCLE: window displacement against the current annotation is 1.83%, not systematic - and the real blind spot it hid is one row per symbol at 14.69%
The guess was that this project's list of gene start sites had gone badly out of date. It has not: 82% of the entries sit on exactly the same spot as the current annotation. The audit that checked it found a different and bigger problem instead - a gene can have several start sites and the list keeps only one.
Subtraction is a partner class, not a curiosity — and the class is defined by compartment, not by sequence
We think losing something matters in several KMT2A leukemias, not just one. But the losses do not all look alike: sometimes a piece of the protein is missing, sometimes a normal protein is dragged to the wrong place, and sometimes a healthy gene copy is switched off. Only the first kind shows up if you look at the sequence.
The 2.03-fold over-representation of shared windows among fusion-occupied promoters is generic promoter activity, not a fusion preference - and if it is not, the rule is finding something different from what it claims
Windows that cover two genes at once turn up in nearly half the positive results, though they are only a fifth of all windows. The boring explanation is that those spots are busy in every cell for reasons that have nothing to do with leukaemia. If the boring explanation is wrong, then the test is measuring something other than what it is named after.
The 41 residues after the CXXC fold (1204-1244) carry a substantial part of the fusion's DNA grip, and removing them is why the structural literature's affinity is so low
The piece of the fusion used for all the published structures stops just after the zinc-finger fold. The original 2002 work used a longer piece with a positively charged tail. If that tail does much of the DNA gripping, then the structures and the low affinity number are both describing a truncated version of the real thing.
The ATF4/ISR collapse under menin inhibitors reports loss of the KMT2A-fusion transcriptional programme by ANY route, not engagement of the menin interface - and the loss is of ATF4 activity, not ATF4
This graph has been treating a stress-response shutdown as the signature of blocking menin. The same shutdown happens when you block a completely different enzyme in the same complex, which means it reports the fusion protein losing its grip on the genome rather than anything specific to menin - and the master switch itself is untouched, so what is lost is the switch being read, not the switch.
The bait/prey grade this project uses to triage the interactome systematically discards its strongest evidence, because the strongest methods have no bait
We sort protein partners by whether our protein was the hook in the experiment, and treat the rest as low quality. But the best experiments - measuring two purified proteins binding directly - have no hook at all, so they land in the reject pile.
The clinical outcome data do not support a single KMT2A class, and they do not refute the menin interface either
Patients with different partner genes really do have different outcomes, but only in some age groups and some disease types, and the reasons may have nothing to do with menin. So the outcome data cannot confirm or kill the idea that one drug works for all of them.
The DAXX requirement in B-lineage KMT2A-rearranged lines is a telomere phenotype, not a fusion-target phenotype - and the two non-KMT2A lines in the deep group are the test
Cells with this fusion need DAXX, but so do two cells without it - and nobody can explain why those two are in the same group. The guess is that it has nothing to do with the fusion and everything to do with how these cells keep the ends of their chromosomes from fraying.
The dependencies this project keeps failing to see are the buffered ones: paralogue redundancy and guide placement, not absence of biology
When we test whether a leukemia needs a gene, we delete one gene at a time and see if the cells die. That test is blind in two specific ways: it misses genes that have near-identical backups, and it misses genes where the cutting tool happened to land in a part that does not matter. Two candidates this cycle are invisible for exactly those reasons.
The druggable node in the homeodomain code is the MEIS1-PBX3 interface, not the HOX-PBX one - and it is the only one nobody has attacked
The only drug-like molecule aimed at this biology pulls HOXA9 away from its partner PBX. But the experiment that compared the combinations found that the pairing which actually reproduces the disease is PBX3 with MEIS1, not HOXA9 with PBX. If that is right, the existing molecule is aimed at the wrong handshake.
The element carries an equally strong CTCF/cohesin peak in HCT116 and in a B-lymphoblastoid line, yet breaks in one and not the other - so occupancy is necessary and not sufficient, and transcription of KMT2A is the switch
The protein complex that marks this spot is just as strongly attached in bowel-cancer cells as in blood cells, but the DNA only snaps there in blood cells. The difference is probably whether the gene is switched on and being read.
The fetal state is PERMISSIVE for lymphoid KMT2A-fusion transformation and PROTECTIVE against myeloid - which is why LIN28B has opposite signs in the two literatures
Two sets of careful experiments say opposite things about the same fetal gene: in human fetal B-cell precursors it is needed for leukaemia to start, and in mouse fetal blood cells destined to become myeloid leukaemia it actually holds the cancer back. This entry proposed that being a fetal cell helps one lineage become leukaemic and protects the other - and then found a paper showing the fetal state helping a myeloid leukaemia too, which cuts the idea down to a claim about that one gene.
The fusion may reach the H2B ubiquitin mark through MOF-MSL rather than through RNF20/RNF40 — the one configuration in which this layer could be selective
The KMT2A fragment kept in the fusion is reported to contact MOF, and the MOF complex contains a second enzyme that puts a ubiquitin tag on H2B at a different spot — a tag that also switches on DOT1L, and that was shown to matter at HOXA9 and MEIS1. If the fusion is pulling that enzyme to its target genes, the ubiquitin layer would be fusion-specific after all.
The H3K9 machinery's address at KMT2A is the PROMOTER, not the exon-11 breakpoint anchor: G9a, HP1a and KAP1 all read the top decile of active promoters there in MCF-7 and KAP1's peak replicates a K562 peak to 267 bp, while five proteins across five cell types are ORDINARY or absent at the anchor - so every H3K9-directed intervention this graph has discussed is aimed 52 kb from where the machinery is
Every drug idea this project has raised for the silencing machinery at KMT2A assumed the machinery sits where the gene breaks. It does not seem to; it sits on the gene's on-switch, 52,000 bases away. If that holds, those drugs would change how much KMT2A is made rather than how stable its breakpoint is.
The H3K9me3 this graph argues about at the KMT2A exon-11 anchor is not deposited by SUV39H1: in SU-DIPG13 the mark sits at the 57.8th percentile of class-matched CTCF anchors while the writer reads an EXPLICIT ZERO in the same nuclei, and genome-wide the two barely track each other (Spearman 0.148, 0.078)
At the exact spot this project cares about, the silencing tag is present at an ordinary level and the enzyme usually credited with making it is completely absent - in the same cells, measured the same day. If that holds up, the tag there is somebody else's work, and any drug aimed at this enzyme would miss.
The H3K9me3 writer reaches the KMT2A anchor only through CTCF and only after the break - which would make this graph's CTCF arm and its H3K9 arm one arm, with CTCF upstream, and would make every resting-state H3K9me3 measurement at that element the wrong assay
Two of this graph's arguments about the KMT2A break may be the same argument. A 2026 paper shows that CTCF is what brings the H3K9-silencing enzymes to a DNA break - and CTCF sitting on the KMT2A gene is already the centre of our story. If that is what happens here, the silencing mark appears because the DNA broke, not before it.
The HDAC7 direction conflict is scaffold versus catalysis: removing the protein and blocking its catalysis are opposite interventions, because class IIa HDACs barely have any catalysis
Two groups report opposite things about the same enzyme in the same infant leukaemia: one says put HDAC7 back, the other says take it away. The proposed explanation is that class IIa enzymes like HDAC7 barely work as enzymes at all - their real job is to hold other proteins in place - so deleting the protein and blocking its chemistry do completely different things, and the two experiments were never measuring the same intervention.
The HP1 paralogue asymmetry is a guide-cross-reactivity artefact until somebody shows it is not - and if it is real, HP1-beta is doing a job the other two cannot, which is the only HP1 question in this graph with a fitness phenotype attached
Three nearly identical proteins, and cancer cells only really need one of them. Either that is a genuine division of labour nobody has explained, or the experiment that measured it was confused by how similar the three genes are - and the cheap test tells you which.
The HSPC-like blast fraction is the compartment that survives menin inhibition, and its size at diagnosis predicts depth and durability of response
Menin inhibitors produce deep remissions in this leukaemia that usually do not last. We think the cells that survive the drug are the small stem-cell-like group already present at diagnosis, and that counting that group before treatment would predict which children relapse. The prediction is falsifiable in both directions and the data to test it largely already exists.
The IRX1 address requires the AFF1 partner; the menin interface explains the shared HOXA address and not this one
Every KMT2A fusion keeps the same front half of the protein, the part that binds menin - and every one of them sits on HOXA. Only the AFF1 version sits on IRX1, so whatever puts it there is not the menin-binding part.
The KMT2A exon-11 anchor is not uniquely TARGETED by the KDM3B-G9a axis - it is unusually SENSITIVE to an H3K9me1 rise that happens almost everywhere, and the sensitivity is a property of the element
Turning off one enzyme raises a silencing tag across the whole genome, yet only the KMT2A gene loses its anchoring protein and only that gene breaks. The proposal is that KMT2A is not singled out by the enzyme - its anchor is just unusually easy to knock off, so a general change lands on it hardest.
The KMT2A::AFF1 oncoprotein-target promoter set is a fusion-level property, and patient identity adds nothing measurable on top of it
Where this cancer protein lands on DNA may be decided by the fusion itself rather than by the individual child. If so, two samples from the same child are no more alike than samples from two different children - which is what the data show.
The menin arm and the CTCF arm are not two stories: menin/SETD2/H3K36me3/LEDGF licenses CtIP short-range resection and CTCF/H3K9me/HP1-gamma licenses BRCA1/BARD1-EXO1/DNA2 long-range resection, which are the two consecutive halves of ONE reaction - so a menin inhibitor and an H3K9-axis inhibitor are acting on the same pathway at different steps
Two separate 2026 papers each describe how chromatin tells the cell to trim a broken DNA end, one driven by menin and one by CTCF - and neither paper mentions the other. The trimming happens in two stages, and each paper turns out to describe one of them, which would mean the drug these children take and the chromatin at the gene that breaks are acting on a single pathway.
The menin interface is a requirement of 5'-KMT2A fusions in haematopoiesis, not of KMT2A rearrangement as such - and a solid tumour has been quietly proving it for six years
This project assumes that a KMT2A rearrangement causes cancer by keeping its grip on menin. In a rare sarcoma, the rearrangement throws that grip away and still causes cancer, which suggests the menin requirement belongs to the leukaemia version of the molecule rather than to KMT2A rearrangement in general.
The menin that matters in an infant KMT2A-rearranged leukaemia is the menin on the fusion, and the N-minus-C antibody difference can measure it in any patient sample without a tag
We think the menin worth targeting is the menin attached to the broken fusion protein, not the menin attached to the normal version of the same gene. And we think there is a cheap way to tell them apart in real patient samples: compare an antibody that sees the front half of the protein with one that sees the back half.
The menin-inhibitor class splits at Val371: three of four clinical compounds contact the first residue of the gamma-H2A.X motif and bleximenib does not, so DNA-repair effects should separate from KMT2A-displacement effects within the class
Three of the four menin drugs that have been crystallised touch a menin amino acid that is also the first letter of the short stretch menin uses to recognise damaged DNA; the fourth, bleximenib, stops about ten angstroms short of it. If that contact matters, the three should interfere with DNA repair in a way bleximenib does not, at doses that block the leukaemia target equally.
The myeloid/lymphoid metabolic split in KMT2A-rearranged leukaemia is set by CELL OF ORIGIN (fetal versus adult), not by lineage or by the fusion partner
Myeloid and lymphoid KMT2A leukaemias want opposite things from their mitochondria, and everyone assumes that is because one is myeloid and one is lymphoid. There is a second explanation: the lymphoid ones start before birth from fetal blood cells and the myeloid ones mostly do not, so the difference may be about when the leukaemia began rather than what kind of cell it became.
The N-over-C oncoprotein score is not fusion-specific by construction, because taspase1 cleaves wild-type KMT2A into an N320 and a C180 fragment that need not occupy the same chromatin - so a fusion-negative cell has a real N/C signal
The test for the cancer fusion protein compares antibodies against the front and back of a normal protein, assuming both halves sit in the same places. But the normal protein is cut in two inside the cell, and the two halves can end up in different places, which would make the test fire even where there is no fusion.
The necessary lesion in ZNF384-rearranged leukaemia is LOSS of the ZNF384 C-terminus, not GAIN of a fusion partner
Everyone assumes the fusion partner is what makes this leukaemia dangerous. But three children have had the same disease with no partner at all - just a small change at the far end of ZNF384. The alternative idea is that losing that far end is the thing that matters, and the partner is a modifier rather than the cause.
The order in which the fusion's programme collapses is a property of the partner, not of KMT2A - so there is no single causal chain
When the cancer protein is removed, one part of its machinery fails first. We propose that which part fails first depends on which partner gene is fused - so the single flow-chart the field draws is really several different flow-charts wearing one name.
The partner, not the site, is what could make a retained KMT2A modification fusion-specific - the AML1-ETO mechanism asked of KMT2A::AFF1
A chemical tag on a part of MLL1 that the fusion keeps is, by itself, nothing special - the normal protein has it too. But in a different leukaemia fusion, the attached partner drags the tagging enzyme into place, so the fusion creates a tag the normal protein does not carry. The question is whether MLL1's partner does the same thing.
The partner's deubiquitinase domain is what lets a PHD-retaining KMT2A fusion survive and transform
KMT2A fusions that keep a set of protein fingers called PHD domains should not cause cancer, because those fingers normally trigger the protein's destruction. The idea here is that when the partner gene is a de-ubiquitinating enzyme, it removes the destruction tag and rescues the fusion, which would explain why the rare minor-breakpoint fusions work at all.
The PBX requirement in KMT2A-rearranged leukaemia is a family property, and every experiment that shaped this graph's view of it removed one paralogue at a time
Three almost identical PBX proteins can each partner the HOX factors that drive this leukaemia, and every experiment so far has removed just one. If they substitute for each other, single-gene results would look weak even if the family as a whole were essential.
The RAS-mutant subclone is where the leukaemia goes, not what it is - a dissemination modifier that a marrow aspirate systematically under-counts
In infant KMT2A leukaemia the RAS mutation is usually in a minority of cells, yet it predicts a worse outcome - which does not fit a passenger mutation. The proposal here is that those cells are the ones that leave the bone marrow and spread, including into the brain and spinal fluid, so a bone-marrow sample sees only the fraction that stayed behind.
The second handle on the retained fragment is CK2-dependent, not 14-3-3-dependent
This project has assumed menin is essentially the only grip on the part of KMT2A that leukaemia fusions keep. We think there is a second grip - the one LEDGF holds - and that it can be loosened by blocking the enzyme CK2 rather than by blocking menin.
The separation MARGIN, not the recurrence count, is the statistic that identifies a partner-specific promoter
Counting how many patients a gene switch is 'on' in turns out not to distinguish a real partner-specific gene from noise. A different number might: how far above the comparison patients the signal sits. In this dataset the three genes with the biggest gap include the one gene with independent published evidence, which is suggestive and needs a proper test.
The serine/one-carbon programme has a SECOND, chromatin-proximal menin input that is not ATF4 - which is why PHGDH is menin-selective where CHAC1 is not, and why the arm is absent in KMT2A::MLLT3
This graph has assumed that a menin inhibitor switches off serine manufacture because it switches off the stress factor ATF4. Two measurements say there must be a second, separate route: a DOT1L inhibitor copies the ATF4 output genes almost exactly but only half-copies the serine genes, and in a different kind of KMT2A leukaemia the ATF4 genes still fall while the serine genes do not move at all. The second route is most likely the menin complex acting directly on the serine genes' promoters, which has already been shown in a different cancer.
The size correction that dissolved KMT2A's proximity promiscuity should run the OTHER WAY for PSIP1: at 530 aa, recovery by 9 of 50 BioID2 baits may be MORE than its length predicts, which would make PSIP1 rather than KMT2A the promiscuous prey in this project's core pair
A survey found the big protein this project studies near half of fifty tagged baits, and a previous analysis showed that is mostly because it is enormous and easy to bump into. Its much smaller partner LEDGF was found near nine baits, which for a protein that size may actually be more than expected.
The SWI/SNF dependency in KMT2A-rearranged leukaemia belongs to ncBAF, not to canonical BAF - and the drug in the clinic is aimed at the wrong module
All the classic experiments on SWI/SNF in MLL-rearranged leukaemia were done before anyone knew the machine comes in three flavours. When you sort the modern dependency data by flavour, only the third flavour - the one containing BRD9 - shows a signal specific to KMT2A-rearranged cells. If that is right, the drug now in trials targets the wrong part.
The transcriptional CDK with a real KMT2A-rearranged selectivity signal is CDK13, not CDK9 - and nobody has looked
Screens say KMT2A-rearranged leukaemia cells need CDK13 more than other leukaemia cells do, and unusually the effect is the same in both the lymphoid and myeloid halves of the panel. CDK9, the kinase everyone studies, shows no such pattern. Almost no one has studied CDK13 in leukaemia at all.
The YEATS asymmetry is breakage geography, not selection: the reader's fate is decided by where topoisomerase II cuts, before any leukaemia exists
The two near-identical partner genes lose or keep their reader module because of where in the DNA the chromosome happens to break, not because one version makes better cancer. If that is right, the difference is a fact about chromosome breakage and not about what the leukaemia needs.
Three of menin's nine declined partners are E3 ligase substrate receptors - either menin turnover is ubiquitin-controlled, or substrate receptors are this beat's artefact signature
Three of the nine weak menin partners are proteins whose job is to pick out other proteins for destruction. Either menin is one of the things they destroy, or these proteins simply stick to everything when they are artificially overexpressed.
Two inhibitors of the same DOT1L active site leave different residual H3K79me2 landscapes, and the difference sits at the KMT2A-fusion target loci - so 'how much H3K79me2 is left' is a compound-specific readout, not a generic measure of DOT1L inhibition
Two drugs that block the same enzyme leave different leftovers behind, and the leftovers are concentrated at the very genes the leukaemia depends on. If that is real, a chromatin test used to decide whether one drug is working cannot be used to decide whether another is.
Under all three DOT1L inhibitors, DOT1L's fold-enrichment at canonical target loci is HIGHER than DMSO (PBX3 4.85 to 13.92, SENP6 3.37 to 9.68) - either catalytically dead DOT1L is retained or trapped on chromatin, or the reading is an artefact of a 1.35-fold spread in MACS fragment-size estimates, and this deposit cannot tell which
When the DOT1L enzyme is switched off by a drug, it seems to stay stuck on the leukaemia genes even harder than before. That would be interesting, but the experiment has a technical quirk that could produce the same appearance, so it is written down as a question rather than a result.
WDR5 is partitioned, not shared: NSL and SET1/MLL complexes compete for one pool, so perturbing NSL changes how much WDR5 reaches the wild-type KMT2A complex the fusion depends on
WDR5 is a small hub protein that several big chromatin machines need. A structural study showed two of those machines cannot hold it at the same time. If that is true inside a leukaemia cell, then interfering with one machine should change how much WDR5 is available to the other - including the normal KMT2A complex that the leukaemia fusion relies on.
What could refute the menin-interface bet, restated after the provenance audit: three tests, none of which is about menin binding
This project bets that the sticky point between menin and the fusion protein is the thing to attack. That bet is safe in one way - every version of the fusion keeps the menin-binding piece - and exposed in another, because the evidence that attacking it removes the disease's chromatin programme came from a different fusion in mice.
WRAD reaches KMT2A fusions only as a passenger on a recruited wild-type complex
The WRAD proteins may well be found sitting next to the leukemia fusion - not because they grip it, but because they are already attached to the normal KMT2A the fusion is known to need. If that is right, a test that only asks 'is WRAD nearby?' will say yes and mean nothing.
ZEB2 level sets how easily a KMT2A-rearranged blast can change lineage under CD19-directed pressure
When children with KMT2A-fusion leukemia are treated with therapies aimed at a B-cell marker, the leukemia sometimes escapes by becoming a different kind of blood cell entirely. ZEB2 is a protein whose known job is to stop blood cells from switching on the wrong identity, so this asks whether how much ZEB2 a blast has decides how easily it can make that escape.
ZMYND11::MBTD1 AML is menin-inhibitor sensitive despite having no KMT2A rearrangement, because the translocation deletes ZMYND11's MYND domain and with it a physical brake on wild-type KMT2A
A rare leukaemia is caused by joining two genes, neither of which is the gene the approved drug revumenib targets - so nobody expects the drug to work on it. This entry argues the opposite: the gene at the front of that fusion normally holds back revumenib's target, and the join throws away exactly the piece that does the holding, which should leave the target running free and the drug useful.
A methionine rescue arm, which separates SAM limitation from ordinary cytotoxicity
If a drug works by starving cells of a methyl donor, then adding the donor back should undo it. That single extra arm tells you whether the mechanism is what it is claimed to be.
A pre-registered, institution-matched serial-pair replication: does a patient's own target set predict itself better than another child's does?
To settle whether one child's leukaemia keeps its own set of target genes, you need several children profiled twice at the same hospital with the same method - including children whose disease did not change type. The public record has one such pair, which is not enough.
A time-resolved BRG1/RING1B/H2AK119ub CUT&RUN at BENC under menin inhibition - the inv(16) experiment, run on the right disease
Treat KMT2A-rearranged leukaemia cells with a menin inhibitor, then at a series of short time points measure which proteins are sitting on the DNA switches that control MYC. If a silencing protein arrives before an activating protein leaves, the drug is working the way it does in a different leukaemia where this has already been shown.
Ask whether the childhood AML in the KDM3B-haploinsufficiency syndrome cohort carried a KMT2A rearrangement - the one human test this mechanism has, and it is a records request
Fourteen people born with only one working copy of the KDM3B gene were described in 2019, and two of them developed cancer as children - one of them leukaemia. If the mechanism is right, that leukaemia should carry a broken or duplicated KMT2A gene, and finding out means asking the authors for the karyotype.
Assay telomere maintenance across the thirteen B-lineage ALL DepMap lines, and see whether it predicts the DAXX dependency split
Grow thirteen leukaemia cell lines that are already sitting in freezers worldwide, and run one cheap test on each that says how they keep their chromosome ends intact. If the six cells that need DAXX are also the six with the unusual telomeres, we have an explanation; if not, we have eliminated the only one we had.
Bin the deposited H2AK119ub NTvKO 2 x 3 design: twelve files, three replicate conventions, public since 2025-11-23 and never analysed - it is the only route to the silencing clause of hyp-q117's prediction 5, it is the internal validation of the PCGF1 knockout this project's null depends on, and it carries a spread the H3K79me2 arm cannot
A dozen public files measure the repressive chromatin mark in exactly the control-versus-knockout, three-drug design this cycle analysed for a different mark. Nobody has looked at them. They would test whether the vacated sites really are being switched off, and they would also check, from its expected effect, that the knockout worked at all.
Blot all three PBX proteins in KMT2A::AFF1 B-lineage cells, then delete them singly and in combination - the family has never been tested together in any human leukaemia line
Three near-identical proteins can each partner the HOX factors that drive this leukaemia, and every experiment so far has removed only one of them at a time. This proposes measuring all three in the right cells and then deleting them in every combination.
Catalytic-dead UCH test: does KMT2A::USP2 need its deubiquitinase activity to transform?
Build the leukaemia fusion protein twice, once normal and once with its enzyme switched off by a single amino-acid change, and see whether the broken version still turns blood progenitors into a cancer-like state. If it does, the leading explanation for this whole class of fusions is wrong.
Catalytic-dead versus degrader: a four-arm test that resolves the HDAC7 direction conflict and supplies the pharmacodynamic marker this drug class has never had
Two groups disagree about whether to add or remove HDAC7 in infant leukaemia. This experiment settles it by separating two things that are always confused: whether the enzyme's chemistry matters, or whether the protein just needs to be physically present. It also adds the one measurement that every HDAC drug study in this disease has skipped - checking whether the drug actually did anything to the chromatin it was supposed to.
Cell-surface capture proteomics on an isogenic human KMT2A::AFF1 model against its matched normal progenitor - the experiment an mRNA screen cannot replace
Take human blood stem cells, put the fusion gene into half of them, and then chemically label and identify only the proteins sitting on the outside of each cell. Comparing the two lists gives the surface proteins the fusion itself creates, measured as protein rather than inferred from gene activity.
Combine an IKK inhibitor with a clinical menin inhibitor in a KMT2A::AFF1 B-ALL line, and read out fusion OCCUPANCY rather than viability
Treat infant-type leukaemia cells with a menin drug, an IKK drug, or both, and then measure how much of the fusion protein is still sitting on its target genes — not just whether the cells died. Measuring the chromatin is the only way to tell whether the two drugs are doing the same job or two different ones.
Decompose the CBX1 DepMap screen to single guides before anyone quotes -0.349 as biology: a no-wet-work, one-afternoon check that would either kill the HP1 paralogue asymmetry or make it the only HP1 question in this graph with a phenotype
Before believing that cancer cells specifically need one of three near-identical proteins, check whether the gene-editing tool used to test it was accidentally cutting the other two as well. The data to check are already public.
Degrade DOT1L, inhibit DOT1L, or leave a dead enzyme in place - run all three against the same cells and the same marrow
The argument about whether the DOT1L drug failed because it was the wrong kind of drug has never actually been tested in leukaemia cells. This experiment does the three comparisons side by side: remove the protein, block its chemistry, or replace it with a chemically dead copy, and see which one the leukaemia minds most - and which one healthy bone marrow minds most.
Degrade one CXXC reader, ChIP the fusion: the one experiment that settles whether the address is contested
Take a leukaemia cell whose KMT2A fusion is doing the damage, rapidly remove one of the other proteins that reads the same DNA addresses, and see whether the fusion spreads into the vacated spots. Nobody has done this, and it would answer in one go a question the whole field has only guessed at.
Does a menin inhibitor break homologous recombination? Test the menin - SETD2 - LEDGF - CtIP chain in a blast, before the next genotoxic combination is designed
A menin inhibitor might block the DNA repair that keeps a cell alive, and nobody has checked that with the drugs children actually receive. One study has checked a tool compound given together with a second drug, and found the repair was disrupted.
Does a menin inhibitor create a BMI1 dependency? A two-arm knockout-under-drug experiment in KMT2A::AFF1 cells, with the CDKN2A readout that decides the mechanism
Knock out BMI1 in KMT2A-fusion leukemia cells and nothing happens. The test is whether that changes once a menin inhibitor has switched off HOXA9, which is thought to be doing BMI1's job. Measure the brake gene directly, so that a positive result comes with its explanation rather than without one.
Does KMT2A::AFF1 bind the histone chaperone SET/TAF-I beta, and does that contact overlap the menin-binding motifs? A co-IP and a peptide competition
In 2005 somebody showed that the front end of the MLL protein - the piece every fusion keeps - sticks to a histone chaperone called SET, and that the two together switch on HOXA9. Nobody checked whether an actual fusion protein does the same, or whether the drug this project cares about would break that contact.
Does KMT2A::AFF1 deliver DOT1L to the AHD or displace it? A competition experiment in t(4;11) cells
The infant fusion sticks to the same patch of AF9 that the H3K79-writing enzyme uses. This experiment asks whether the fusion brings the enzyme along or pushes it off, which are opposite predictions nobody has tested.
Does menin hold FANCD2 in a KMT2A-rearranged cell, and does revumenib let go? Nobody has asked — zero records over every menin-inhibitor name
The single experiment that would decide whether menin's Fanconi connection matters in leukaemia has never been done: pull menin out of a KMT2A-rearranged leukaemia cell, see whether FANCD2 comes with it, then repeat with the drug on board. This entry writes down the experiment and records, with the exact queries, that the literature contains nothing.
Does menin inhibition have any MHC class II headroom to work with in B-precursor infant ALL? A flow cytometry experiment on samples that already exist
A 2026 paper showed menin inhibitors make myeloid leukaemia cells visible to donor immune cells by switching on a display molecule. This experiment asks whether that molecule is already switched on in infant leukaemia cells, using a standard laboratory stain on samples that are already stored in freezers.
Does revumenib displace JunD, and at what concentration relative to KMT2A? (The yes/no half is already answered)
A menin-blocking drug does knock JunD off menin - that has now been shown with laboratory tool compounds and in leukaemia cells. What is still missing is the same measurement for the three drugs actually given to patients, in an endocrine cell, with the concentration needed for JunD displacement set against the concentration needed to displace KMT2A.
Does the diagnostic marker survive the drug? Surface NG2, CD133, CD10 and CD33 under menin and DOT1L inhibition, with the rearrangement held fixed
This experiment treats KMT2A-rearranged leukaemia cells with each of two drugs that interrupt the broken gene's activity in different ways, and measures whether the surface markers doctors use to recognise the disease are still there afterwards. The point is to find out whether the diagnostic test survives the treatment.
Does the fusion carry an H2B ubiquitin ligase? A co-IP and an H2BK34ub CUT&RUN in KMT2A::AFF1 cells
Pull the fusion protein out of a leukaemia cell and look for MSL2, then map where the K34 ubiquitin tag sits across the genome. Two standard experiments that together decide whether the ubiquitin layer is something the fusion controls or something it merely borrows.
Does the fusion's programme fall apart in the same order in KMT2A::AFF1 as in MLL::AF9? The chain's only ordering measurement, repeated on the index fusion
There is exactly one experiment that shows which step of this leukaemia's machinery fails first when the cancer protein is removed, and it was done on a different fusion protein in a different kind of leukaemia. This is that experiment, done on the one that causes infant leukaemia.
Does the sarcoma fusion bind menin? One co-immunoprecipitation with the right control settles six years of the question
Express the sarcoma fusion protein and the leukaemia fusion protein in the same cells, then pull on menin and see which one comes with it. The prediction is that the leukaemia one does and the sarcoma one does not, and nobody has ever checked.
Does WDR5 move between complexes? An acute KANSL1 degron with WDR5 co-IP, WDR5 CUT&RUN and H4K5/K8/K16ac readout in a KMT2A-rearranged line
Tag the NSL scaffold so it can be destroyed within an hour, then ask whether the shared hub protein WDR5 moves over to the KMT2A-type machinery - in the test tube, on the DNA, and in the chemical marks left on histones.
Download and bin GSE127507's nine RS4;11 tracks (1.99 GB), rebuild a human binner, and score the eviction gradient in the index fusion KMT2A::AFF1 - the one measurement that would move this finding from KMT2A::MLLT3 to the disease this graph is about
Download nine public files from a 2019 study of the exact fusion this project studies, count the signal in fixed-size windows, and repeat the analysis that was just done in two other cell types. No laboratory work is needed.
Dynamic BH3 profiling before and after menin inhibition — a one-week assay on primary cells that would choose the BH3-mimetic partner, and has never been run in lymphoid disease
There is a quick laboratory test that says how close a cell is to dying and which survival protein is holding it back. Several trials are already combining a menin inhibitor with a drug that targets one of those survival proteins, and nobody has run the test to check the combination makes sense.
EWSR1::ZNF384 versus EWSR1::FLI1 - the same EWSR1 exon-7 module on two different DNA-binding proteins, and nobody has compared them
Two cancers - a bone tumour and a childhood leukaemia - appear to use the exact same piece of the EWSR1 protein, glued to two different DNA-binding proteins. Putting them side by side in one experiment would show whether that sticky piece does the same thing wherever it lands, or whether what it does depends entirely on where it is sent.
Experiment — is a KMT2A-septin fusion menin-dependent? The one-plate version
A small, cheap experiment: grow mouse blood progenitors carrying a KMT2A-septin fusion, add a menin-blocking drug, and see whether the colonies stop forming and the HOXA9 signal drops. Nobody has done it.
Experiment (proposed) — cross-family interface point-mutant panel: is the obligate interface real in each fusion?
For each fusion protein we build versions with tiny, precise changes that break just one contact point, then ask whether the cancer program still works. Knock-outs tell you a protein matters; this tells you which handshake matters, which is what a drug has to break.
Experiment: a 6 / 24 / 72 hour / 7 day VTP-50469 time course in SEM reading menin occupancy and H3K4me3 together, to decide whether the 23% is a partial effect or a slow one
Measure the same thing at four time points instead of one. If the effect keeps growing, the mark is just slow to change; if it stops at a quarter, something is genuinely holding the rest in place. These are different biology and currently nobody can tell them apart.
Experiment: a topology-controlled binding survey of the MLL AT-hook region, plus a triple AT-hook alanine mutant in an endogenously tagged fusion
Two experiments that settle a thirty-two-year-old ambiguity. First, purify the hooks and offer them DNA of the same base composition in different shapes, to see whether they care about the letters or the shape. Second, mutate all three hooks in a real leukemia cell and see whether the fusion still sticks to the nuclear skeleton and still finds its genes.
Experiment: apply all three pressures and ask whether the survivors are the same cell
Expose the same leukemia to three completely different kinds of treatment, one after another, and look closely at the cells that survive each one. If they all look the same, there is a single hiding place that defeats every approach at once.
Experiment: bait the KMT2A breakpoint cluster region and sequence everything it joins to, in a fetal-like human progenitor
Instead of asking which partners are found in patients, put a break in KMT2A on purpose and sequence every other place in the genome it sticks to. The clinical partner list then becomes a prediction you can check rather than the only thing you know.
Experiment: block the beta-2 adrenergic receptor in a KMT2A-rearranged model — the leukaemogenic nerve signal runs through a receptor an old, cheap, infant-licensed drug already hits
Leukaemia damages the nerves that control the bone marrow, and the harmful signal travels through one specific receptor — beta-2. Drugs that block beta receptors are among the oldest and cheapest in medicine and one of them is routinely given to babies. Nobody has tested whether blocking that receptor slows this leukaemia.
Experiment: blot menin in a SET-NUP214 immunoprecipitate, then run a menin inhibitor and selinexor alone and combined in LOUCY and MEGAL
One extra antibody on an experiment that has already been done would show whether the menin protein travels with SET::NUP214. If it does, the next step is to give the cells a menin drug and see whether the HOXA genes go down.
Experiment: build the missing KMT2A::AFF1 degron and titrate it — what fractional occupancy does the program actually need?
Nobody has made a version of the KMT2A-AFF1 fusion that can be switched off on demand in a real leukemia cell. Building one and removing it in controlled amounts would tell us how much of the fusion a drug actually has to block.
Experiment: combinatorial dose-titration of the HOXA9/MEIS1 program to find its minimal cut set
Use a gene-dimmer-switch technique to turn down several leukemia genes at once, by controlled amounts, in every combination, and find the smallest group that has to be dimmed together before the leukemia program switches off.
Experiment: CRISPR knockout of CLEC12A in KMT2A-rearranged models, with and without a cooperating RAS lesion
Delete CLEC12A from KMT2A leukaemia cells and see whether they still grow, copying exactly what was done in a different leukaemia where deleting it worked. Do it with and without an extra RAS mutation, because the original experiment had one and we do not know whether that mattered.
Experiment: degrade TOP2A and TOP2B separately in a fetal-like human progenitor and map DNA breaks at the KMT2A locus without a poison
Nobody has ever asked which of the two nearly identical DNA-cutting enzymes breaks the KMT2A gene when no drug is present, because the standard way to see the enzyme working is to jam it with a drug. This experiment removes each enzyme on demand and counts the breaks directly, so the drug is never needed.
Experiment: does a MEN1 resistance mutant pay a price when the drug is not there? The competition assay nobody has run
Mix cells carrying the drug-resistance mutation with ordinary leukemia cells and watch which wins when there is no drug. If the resistant ones lose, then giving the drug in pulses could keep them suppressed; if they do not, that whole idea collapses.
Experiment: does AF1q self-associate, and does KMT2A::MLLT11 transform? The two measurements nobody has made on the 10th most common KMT2A fusion
Two straightforward experiments on a leukaemia fusion found in 28 children and adults that nobody has ever put into a cell: measure whether the partner protein sticks to copies of itself, and test whether the fusion makes blood progenitors grow without stopping. Either answer would change how this disease is explained.
Experiment: does chemotherapy itself install menin-inhibitor resistance? The two-arm controlled test, plus the serial-sample ladder
Split one leukemia sample in two, give one half chemotherapy in mice, leave the other alone for the same length of time, then test both with a menin inhibitor. If only the treated half has become resistant, treatment is the cause.
Experiment: does KMT2A::AFF1 form condensates at endogenous levels, and does it need them? The separation-of-function and IDR-swap test
Tag the fusion protein in a real leukemia cell, look at whether it forms droplets, then delete the floppy region that would make droplets while checking the protein still binds its partners. If the cell stays leukemic, droplets are not the mechanism.
Experiment: does marrow stroma protect KMT2A-rearranged cells from a menin inhibitor? Nobody has run the week-long experiment that would say
Grow this leukaemia on a layer of bone-marrow support cells, add a menin inhibitor, and see whether the support cells keep the leukaemia alive. It is a standard laboratory test, it takes about a week, and despite menin inhibitors being the most important new drug class in this disease, nobody has published it.
Experiment: does MLL-AF4 transform a staged fetal progenitor when it cannot transform cord blood?
Put the same fusion gene into blood stem cells taken from different stages of development and see which ones turn into leukaemia. If only the earliest ones do, the missing ingredient is the age of the cell, not a second broken gene.
Experiment: endogenously tagged KMT2A::AFF1 imaged through a menin-inhibitor time course, with a five-compartment segmentation and a half-life arm
Put a fluorescent tag on the fusion protein at its own gene in a leukemia cell line, then film the cells while the menin inhibitor is applied and score where the protein goes and whether it disappears. Nobody has filmed it. One study has measured the amount of fusion protein left in a treated patient — it was undetectable — but that patient's leukaemia had also switched lineage, which lowers the protein independently, so what the drug did to it is still open.
Experiment: five-fraction subcellular blot of a t(4;11) line, with a menin-inhibitor arm - converting thirty years of unattributable images into attributable numbers using only existing antibodies
Instead of photographing the leukemia fusion protein, break the cells into five layers - cytoplasm, soluble nuclear liquid, loosely bound chromatin, tightly bound chromatin and the insoluble scaffold - and run each layer on a gel. The fusion and the normal protein separate by size in every lane, so for the first time you can say how much of each is where, with antibodies that already exist.
Experiment: give an antioxidant with the CXCR4 antagonist — if the drug stops working, the niche is a redox buffer and not a hiding place
There are two competing explanations for why blocking the CXCR4 receptor hurts this leukaemia: either it evicts the cells from a protective spot in the marrow, or it removes a chemical support that keeps the cells from being damaged by their own metabolism. One cheap experiment tells them apart — add an antioxidant, and if the drug stops working, it was the chemical support.
Experiment: go back to the diagnostic sample and look for the myeloid clone before it existed
Take stored samples from the day of diagnosis, from children who later switched to a myeloid leukemia and from matched children who did not, and search very hard for a few myeloid-looking leukemia cells hiding among millions.
Experiment: induce CyP33 in a KMT2A-rearranged cell and ask, allele by allele, whether the brake still works at HOXA9
Turn up the enzyme that normally makes KMT2A switch genes off, in a leukemia cell, and watch separately what happens to the broken copy of the protein and the healthy one.
Experiment: is the primed state in KMT2A::AFF1 actually bivalent? Multi-mark CUT&Tag on primary infant blasts
Everyone says these leukaemia cells keep two sets of instructions half-switched-on at once, but nobody has actually shown the chemical marks that would prove it in this disease. This experiment measures them directly on real patient cells.
Experiment: long-hold degron depletion of endogenous KMT2A::AFF1, to ask whether the leukaemic program recovers without the fusion
Tag the fusion protein so it can be destroyed on command, destroy it, and then keep it destroyed for a month while watching whether the leukaemia programme comes back by itself. If it does, drugging the fusion will not work.
Experiment: map what stays on chromatin when the menin inhibitor has done everything it can do
Flood the cells with menin inhibitor until no more can bind, then ask where the fusion protein is still sitting on the DNA. Anything still there is a place the drug can never reach.
Experiment: measure KMT2A-AFF1 spatial proximity in the human fetal progenitor, in the cell and at the time the translocation actually happens
The idea that KMT2A joins AFF1 because the two genes sit near each other in the nucleus has only ever been measured in cancer cell lines. Nobody has looked in the cell type where the fusion actually forms, and the technology to do it on tiny numbers of cells now exists.
Experiment: measure the two missing time constants — how fast the KMT2A::AFF1 program collapses on drug, and how fast it comes back off drug
Add a menin inhibitor, watch how quickly the leukemia program shuts down, then wash the drug out and watch how quickly it comes back. Nobody has ever done the second half.
Experiment: put four residues back - the minimal add-back control the PHD3 result has never had
Add just four amino acids back into the leukemic protein to restore one lost contact, and see whether that alone cripples it. If four residues are enough to break it, the famous experiment that added a whole finger was probably measuring the insertion, not the biology.
Experiment: put the resistant and the sensitive fusion side by side - if a MEN1 interface mutant keeps the fusion in place while the wild type moves it, relocation is real and menin is what chooses the address
Some patients relapse because menin changes shape so the drug cannot grip it. Those cells are a perfect built-in comparison: same cells, same drug, one with menin still working. If the fusion moves to new places only where the drug works, the moving is caused by the drug and not by anything else.
Experiment: read the endocrine limb and the marrow limb of menin-MLL in the SAME animals at the SAME exposure
Give mice a menin-blocking drug at a dose that works on leukaemia, then measure the drug's effect in the hormone glands and the bone marrow of the same animals. Nobody has done both at once, so nobody knows whether the helpful and the harmful effects happen at the same dose.
Experiment: spike-in-calibrated, fusion-attributable occupancy of KMT2A::AFF1 in the first 24 hours of menin inhibition, scored for gains as well as losses
Tag the fusion protein so we can see it specifically, give the drug, and take chromatin snapshots every few hours instead of waiting three days. Add a known amount of foreign chromatin to every sample so we can tell a real global change from a scaling artefact, and look for new sites as carefully as for lost ones.
Experiment: spike-in-normalised H3K4me3, menin and IgG CUT&Tag in SEM under VTP-50469 and EPZ-5676 - the one missing experiment that turns every ratio in this beat into an absolute number
Repeat the key experiment in infant-leukaemia cells with a measuring stick added: a fixed amount of DNA from another species mixed into every sample. That converts 'the mark moved around' into 'the mark went up or down by this much', which is the question nobody can currently answer.
Experiment: time-to-resistance, not synergy index, as the readout for menin-inhibitor combinations
Almost every study of drug combinations in this leukemia measures how many cells die in a few days. This experiment instead measures how long it takes the leukemia to come back, which is the thing patients actually care about.
Experiment: titrate wild-type KMT2A up and down in one leukemia cell and watch what happens to the fusion's grip on chromatin
Two good papers disagree: one says the leukemia needs the healthy copy of KMT2A, the other says adding more of it pushes the leukemic protein off the genes it controls. Dial the healthy copy up and down in one cell and measure both, and you find out whether there is an amount that is bad for the leukemia.
Experiment: where do KMT2A::AFF1 blasts go when a menin inhibitor releases the block? Single-cell lineage destination under drug
Treat infant leukaemia cells with a menin inhibitor and watch, one cell at a time, what they turn into. The question is whether they mature into healthy B cells or drift toward myeloid cells, because only one of those outcomes is a cure.
Export or docking? A locus-resolved H3K79me2 ChIP under an XPO1 inhibitor separates the two CALM-AF10 models, and it has never been run
Two explanations for how CALM-AF10 causes leukaemia make opposite predictions about what happens to a specific chemical mark on DNA-packaging proteins when you block nuclear export. Measuring that mark gene by gene, with and without the drug, tells you which explanation is right.
Express a partner-free ZNF384 C-terminal truncation in HSPCs alongside TCF3::ZNF384 - the one experiment that decides whether the fusion partner is necessary
Build the version of ZNF384 that one child actually had - the protein chopped off near its end, with no partner attached - and put it into blood stem cells next to the normal fusion. If it causes the same leukaemia, the partner is not the cause. If it does nothing, the partner is.
Four hours of MS67 in MV4;11, two KMT2A antibodies: the cheapest test of whether WRAD reaches the fusion or only wild-type KMT2A
A commercially characterised compound strips one protein out of leukemia cells within two hours. Running it in the index leukemia line and then blotting with two antibodies - one that sees the fusion protein and one that cannot - would finally say whether that protein machine touches the cancer-causing fusion or only the normal version of the gene. It takes an afternoon.
GSE139196: SUV39H1 ChIP-Seq in MCF-7 with and without 40 minutes of estradiol - a ligand-inducible SUV39H1 occupancy experiment, filed under the retired symbol KMT1A in its own sample titles
In 2021 a group measured where the H3K9 silencing enzyme sits on DNA in breast cancer cells, before and forty minutes after a hormone signal. Because the enzyme is the same one this project argues about, the data are usable here even though the cells are not blood cells.
GSE225028: SUV39H1 and H3K9me3 CUT&RUN in SU-DIPG13 and SU-DIPG36 - the only published human SUV39H1 occupancy map deposited as a readable hg38 bigWig, and the enzyme and its own mark were profiled in the same cells
A 2024 study measured where the H3K9 silencing enzyme sits on DNA in two brain-tumour cell lines taken from patients, and measured the chemical mark that enzyme writes in the very same cells. The processed files are public, aligned to the standard human genome, and can be read at any gene without doing a new experiment.
Half the 'free' PRC1.1 intersection really was free: Zhou's locus set is a publisher table, Neville's does not exist, and only that half costs 25 GB
CORRECTED 2026-09-20: two papers disagree about where a protein complex acts after a menin inhibitor, and this project spent three cycles saying the disagreement could be settled free by comparing the location lists each published. One of the two did publish its list at the journal and the other did not, so half the comparison costs nothing and the other half still needs a real pipeline run on 25 gigabytes.
Immunoprecipitate the KMT2A fusion and map its phosphosites - the experiment this whole field has never done
Pull the actual cancer fusion protein out of leukaemia cells, break it into pieces and ask a mass spectrometer where the phosphates are. Nobody has done this, so every claim about which switches the fusion keeps is currently an argument from where the chromosome broke rather than a measurement.
IRX1 CUT&RUN in primary t(4;11) blasts — testing Kühn's HEK293T mechanism where it matters
The idea that IRX1 shuts down HOXA by sitting on it comes entirely from kidney cells in a dish. Nobody has checked whether IRX1 sits there in an actual patient's leukaemia.
IRX1 occupancy under PARTIAL menin inhibition in KMT2A::AFF1, with a spike-in, because complete eviction destroys the readout
To see whether the fusion protein moves to IRX1 when a menin-blocking drug pushes it off HOXA, you need a dose that only partly works. At a full dose the protein leaves the DNA entirely and there is nothing left to follow.
Is BENC even on in the disease this project is about? A module-resolved accessibility and deletion map in KMT2A::AFF1 B-lineage cells
Everything known about this MYC control region comes from myeloid leukaemia. The disease at the centre of this project is a B-cell leukaemia of infants, and nobody has checked whether the region is even switched on there. This is the experiment that checks.
Is menin an SCF-FBXW7 substrate? A one-week immunoblot and a free desk scan that settle Arm A
A simple test of whether menin is destroyed by a specific cellular disposal system: remove the disposal component, and see if menin protein goes up. If it does not, one whole idea about menin can be closed cheaply.
Is the menin interface a property of the fusion or of the blood? Test KMT2A::MAML2 thymoma
The drug that works against KMT2A leukaemias works by breaking the fusion protein's grip on a partner called menin. The same fusion also causes a thymus tumour, so testing whether menin still matters there tells you whether that grip is a property of the fusion itself or of blood cells in particular.
KMT2A-N CUT&RUN in an untreated non-AFF1 lineage-switching blast: the one experiment that separates partner from cell state
All three leukaemias where the fusion sits on IRX1 are the same fusion type AND the same kind of sample from the same hospital. One untreated sample of a different fusion type but the same kind would tell you which of those two things matters.
KMT2A-N/C CUT&RUN on one untreated KMT2A::AFF1 blast from outside Texas Children's, scored at the six AFF1-exclusive gene bodies with a pre-declared margin column
Every patient in whom the fusion protein is found on IRX1 comes from one hospital, and two of the three are the same child sampled twice. One more untreated AFF1 patient from a different hospital, profiled the same way, would settle whether this is biology or a local effect.
Knock a tag-plus-degron cassette into the endogenous KMT2A::AFF1 allele in SEM - one construct that makes five claim classes attributable
Put a small molecular handle, plus a switch that lets you destroy the protein on command, onto the fusion gene in a patient-derived leukaemia cell line - on the real gene, not an extra copy. Every measurement made through that handle can only be seeing the fusion, which fixes five different kinds of claim at once.
Knock down CDK13 and CDK12 side by side in lineage-matched KMT2A and non-KMT2A lines, and ask whether the DepMap signal is real
A public screen says KMT2A-rearranged cells need CDK13 more than other leukaemia cells do. This is the cheap experiment that would confirm or kill that: switch the gene off in matched pairs of cell lines and see whether the difference reproduces.
Knock S395A into the MLLT3 half of a KMT2A::MLLT3 model and ask whether it becomes chemotherapy-resistant
Change one letter in the cancer fusion so the DNA-damage enzyme can no longer tag it, then give the mice chemotherapy. If the untagged version survives treatment better, then part of what chemotherapy does in this leukaemia is switch the fusion off, which nobody currently believes.
Measure ABCA1, ABCG1 and SREBF1 before and during menin inhibition - a three-transcript panel on samples that already exist
Take blood samples already stored from patients before and after they started a menin inhibitor, and measure three genes that menin is supposed to keep switched off. If they go up, the drug has an effect on fat and cholesterol handling that nobody has recorded.
Measure all three H3K79 states on a real scale, across a dose series and after a menin inhibitor - the measurement this whole field has been doing by proxy
Nobody has measured the three versions of the H3K79 tag separately, on an absolute scale, in this leukaemia. Doing so across a range of drug doses would answer whether a DOT1L drug leaves a partial tag behind, and whether a menin drug removes only half of it - two predictions the project's central ideas rest on.
Measure H3K27me3 after a menin inhibitor - the free pharmacodynamic nobody has read, and the one experiment that separates two mechanisms
Two different explanations exist for why menin inhibitors switch silenced genes back on, and they predict opposite things about the Polycomb 'off' tag. The early part of that measurement has now been found in public data for a different fusion, so what remains is to do it in the fusion this project is about, and for longer.
Measure H3R2 and H4R3 methylation states at KMT2A-fusion targets by isomer-resolving mass spectrometry, and test the gate by writer perturbation read out on the enzyme's product - the experiment that has never been done in any fusion
Nobody has ever measured these arginine tags in a cell carrying a KMT2A fusion, in any fusion, in any species. This is the cheapest way to find out whether they are there, and a second step that tests what they do without relying on antibodies that cannot tell the two versions apart.
Measure IRX1 transcript and KMT2A-N oncoprotein occupancy in the SAME KMT2A::EPS15 blast - the one cell where expression-without-occupancy is most likely and has never been paired
Take one child's leukaemia cells with a KMT2A-EPS15 fusion and, from the same sample, measure both whether IRX1 is switched on and whether the fusion protein is sitting on it. Every existing measurement does one or the other in different children, which is why the question is still open.
Measure replication stress directly in KMT2A-rearranged cells instead of inferring it from a dependency ranking
The whole Fanconi story rests on an argument that KMT2A leukaemia cells must be struggling to copy their DNA, inferred from which genes they need. Nobody has watched them copy their DNA. This is how you would: stretch out single DNA molecules from the cells, measure how fast and how evenly the copying machinery moves, and compare against leukaemia cells without the fusion.
Measure the clearance clock directly: SNAP-tag H3.3 pulse-chase at KMT2A::AFF1 target genes, with and without DOT1L inhibition, in cycling and arrested cells
Tag the replacement histone so you can tell new copies from old, then watch how fast new ones appear at the genes this fusion controls - in dividing cells and in cells forced to stop dividing. That number is the one the whole drug-timing argument depends on, and nobody has measured it.
Measure the free-versus-7SK-bound P-TEFb split in KMT2A::AFF1 cells - the experiment this whole beat turns on, which has never been done
Spin the contents of leukaemia cell nuclei through a sugar gradient so that heavy particles and light ones separate, then look at which fractions the elongation enzyme is in. It is an old, cheap, unglamorous experiment, and nobody has ever run it on a KMT2A-rearranged leukaemia.
Measure the menin-KMT2B affinity nobody has measured, and the four clinical inhibitors' IC50 against it
Nobody has ever measured how tightly menin grips KMT2B, or how well the leukaemia drugs block that grip. The measurement is a standard one and would take weeks.
Measure the one number the whole YEATS-selectivity argument turns on: what fraction of ENL and of AF9 is in SEC, in SEC-L2, in SEC-L3 and free — with an exchange control, because the question may not be well posed
Take a leukaemia cell, pull out each version of the machine one at a time, and count how much of the two reader proteins came out with each. Nobody has ever done this, and every argument about whether a reader-blocking drug is specific depends on the answer.
Measure the pausing index at KMT2A::AFF1 target genes — the one number the whole pausing argument turns on, and it has never been reported
Run a nascent-RNA sequencing experiment in an infant-leukaemia cell line carrying the KMT2A::AFF1 fusion, and count how much polymerase is stalled at the start of the fusion's target genes compared with other genes. One dataset either supports or kills the idea that these cells run with the transcriptional brake off.
Measure the RAS allele fraction in CSF blasts and in paired marrow blasts from the same infant on the same day
Every genetic study of this leukaemia has sampled the bone marrow. This experiment asks whether the cells carrying the RAS mutation are actually somewhere else - in the spinal fluid - by genotyping both compartments from the same child at the same moment.
Menin CUT&RUN in an IRX1-high KMT2A::AFF1 blast — the missing track in every public deposit
Nobody has ever looked at where menin, the drug target, sits on the DNA in the half of these infant leukaemias that has switched off HOXA. Until someone does, we cannot say whether the drug has anything to grip.
Move the cohesin degron into a cell where the element actually breaks: RAD21-AID in Nalm6 or CD34+, sBLISS plus both END-seq variants, with the fusion readout in the same cells
Repeat the one cohesin-removal experiment that exists, but in a blood-precursor cell rather than a bowel-cancer cell, because that is where the KMT2A spot actually breaks. Count both the breaks and the rearrangements in the same cells, at several timepoints.
Nobody has ever run proximity labelling on a KMT2A fusion, and the interaction records that look like they did are RBP-atlas preys from HEK293
There is a standard modern method for mapping everything that comes near a protein inside a living cell, and it has never been applied to a KMT2A fusion. The partner database looks as though it has been, but those entries come from a study of unrelated proteins in which our protein merely happened to get tagged.
Occupancy-matched three-arm comparison of ziftomenib, revumenib and bleximenib on gamma-H2A.X resolution and HR efficiency in a KMT2A-rearranged B-lineage line, with the peptide competition as the biochemical pre-test
Three menin drugs touch a menin amino acid that starts the protein's DNA-damage recognition patch and a fourth does not, so this experiment gives the same cells the same amount of each drug and asks whether DNA-break repair suffers differently. The drug that does not touch that spot is the built-in control, which is what makes the comparison interpretable at all.
One afternoon of titrations that closes a 1,300-fold question: tagless 1147-1203 against tagless 1147-1244, matched in every other respect
Nobody has ever compared the short and long versions of the fusion's DNA-gripping module in the same tube on the same day. Making both proteins and titrating them against the same DNA at the same salt would settle the disagreement that has stood since 2002.
One blot: phospho-eIF2alpha and ATF4 protein in RS4;11 under VTP-50469, with a dose-response and a DOT1L comparator arm
Every claim this project makes about menin-blocking drugs switching off a stress programme rests on gene-activity measurements that cannot see the step where the switching actually happens. One western blot, on cells two laboratories already grow, would settle it.
One experiment that decides whether KDM4C is a target or the KDM4 family is: paralogue rescue of SD70 sensitivity in SEM
A drug that blocks three near-identical enzymes kills this leukemia, but deleting any one of them does nothing. Put each one back, one at a time, in cells treated with the drug: whichever ones rescue the cells tell you which enzyme the drug actually needs to hit.
One immunoprecipitation that decides whether 14-3-3 touches KMT2A at all
Pull down 14-3-3 from leukaemia cells and see whether KMT2A comes with it. Then remove the partner protein AF4 and do it again. If KMT2A stops coming, it was never holding on itself.
One stain, in the right order: H3K27me3 across a low-grade endometrial stromal sarcoma series first, and the three MBTD1::EZHIP blocks second
There is a routine hospital stain that shows whether a tumour has lost a particular chemical tag on its DNA, and nobody has ever run it on the rare uterine cancer whose fusion gene should destroy that tag. The trick is to stain the ordinary cases first, because if they have lost the tag too then the rare cases would prove nothing.
Paired IRX1 RNA and KMT2A-N occupancy in the same non-AFF1 blasts: does the expression subgroup exist where the occupancy subgroup does not?
We know the fusion protein does not sit on IRX1 outside one fusion type. We do not know whether IRX1 is switched on there anyway. Measuring both in the same cells answers it in one go.
Point the EED inhibitor at the index fusion - a clinical-stage drug, a subunit with no paralogue, and zero leukaemia experiments
There is a class of drug that blocks Polycomb by binding a part of the machine that has no back-up copy, so the redundancy problem does not apply. One of them has already been through a 139-patient trial. Nobody has ever tested one in leukaemia.
Prime, wash, divide, rechallenge - the one experiment that would turn 'DOT1L provides transcriptional memory' into a measurement, and it doubles as a dosing-schedule study
A 2026 paper says DOT1L gives leukaemia cells a memory of which genes should be switched on, and testing memory means treating cells, taking the drug away, letting them divide, and treating again. That paper did the take-the-drug-away part - for a different drug, with a timecourse - and never did the treat-again part, which is the step that makes it a memory experiment.
PROPOSED: a side-by-side CHD4 co-IP in a t(4;11) B-ALL line and a KMT2A-wildtype B-ALL line, probing one subunit from each NuRD module - the first experiment that would ask whether a KMT2A fusion sees the whole complex
Pull down the NuRD motor protein from a leukaemia cell that carries the KMT2A::AFF1 fusion and from one that does not, run them side by side on the same blot, and ask whether the deacetylase half of the machine comes along in both. Nobody has ever done this in a KMT2A-rearranged cell.
PROPOSED: measure let-7 and restore it in the human fetal-liver KMT2A::AFF1 model - an axis measured in mouse MLL-fusion AML but never in the index fusion
LIN28B is famous for holding down a growth brake called let-7. If that is how it lets infant leukaemia start, then putting the brake back should stop the leukaemia just as removing LIN28B does. In mouse MLL-fusion leukaemia that is exactly what putting the brake back does - and nobody has tried it in the human infant fusion.
PROPOSED: run a second oncogene in LIN28B-null human fetal liver CD34+ cells - the control that decides whether the KMT2A::AFF1 transformation block is fusion-specific or just a broken progenitor
Removing LIN28B stops the infant-leukaemia fusion from turning fetal blood cells cancerous - but it also stops those cells from growing normally. So the honest question is whether LIN28B is special for this fusion, or whether a damaged cell simply cannot be turned cancerous by anything. Putting a different cancer gene into the same cells answers it.
PROPOSED: swap the partner, hold everything else - KMT2A::MLLT3 versus KMT2A::AFF1 in LIN28B-null human fetal liver CD34+ cells
The fetal gene LIN28B is needed for one MLL fusion to cause leukaemia in human fetal cells, but in mice a different MLL fusion is actually held back by the same gene. Putting both fusions into the same human fetal cells, with and without LIN28B, would tell us whether the difference is the fusion partner or the species and lineage.
Proximity-labelling map of KMT2A::AFF1 — the experiment nobody has run in twenty years of interactome work
Tag the fusion protein with an enzyme that sprays a chemical label on anything that comes near it, then collect everything that got labelled. It would give the first list of what is actually next to the fusion inside a living leukemia cell, which no one has ever produced.
PUBLISHED, DONE: human fetal liver CD34+ cells fail to transform on KMT2A::AFF1 induction without LIN28B (Ling et al., Blood 2026) - a dedicated human index-fusion experiment, read from the abstract only
Scientists put the exact infant-leukaemia chromosome swap into human fetal liver blood stem cells with gene editing, and the cells only turned cancerous when a fetal-only gene called LIN28B was present. Switching LIN28B off in leukaemias that had already formed took away their growth advantage, in dishes and in mice.
Put mutant IDH2 into SEM and ask three questions at once: does the unmethylated-CpG pool shrink, does the fusion move, and does its competitor move
Take the standard human infant-leukaemia cell line, switch on the mutant IDH2 enzyme that drives DNA over-methylation, and then map three things: where methyl-free DNA is, where the fusion sits, and where its main rival sits. Three readouts, one experiment, and it distinguishes three competing explanations that argument alone cannot.
Put the same fusion in two cell types and degrade it, rather than cutting the gene
Two experiments in one: grow the identical fusion in a myeloid and a B-cell background, and switch the fusion protein off directly instead of deleting the whole gene. Together they would answer whether cell type changes what the fusion needs.
Raise a KMT2A::AFF1 junction-specific monoclonal by the published TEL 6F2 protocol - rat immunisation with a junction peptide, ELISA and western screen, then a ChIP test against the wild-type-parent controls that the ETV6::RUNX1 campaign never ran
An antibody that recognises only the join between the two fused proteins would see the cancer protein and nothing else, and the recipe for making one against a childhood leukaemia fusion was published in 2013. This is that recipe applied to the fusion this project is about, with the specificity controls the original campaign left out.
Raise H2Bub in KMT2A-rearranged cells with the existing USP22 macrocycles and see which way the disease moves
Selective molecules that raise the H2B ubiquitin tag already exist and get into cells. Put them on KMT2A-fusion leukaemia lines, measure the tag, the downstream mark, the leukaemia genes and whether the cells grow or mature, and the direction question is answered.
Re-grade the interactome on PSI-MI method class instead of bait role - a dry experiment, and an EXECUTION of a prescription data/symbol-collisions.json already wrote down in cycle 57 and nobody has run
Sort the protein partners by what kind of experiment found them rather than by which protein carried the tag, then read the two dozen the old sorting threw away, and see whether the order changes.
Re-score all nine GSE260742 same-condition floors under two further external stratifiers already binned on disk - KDM2B DMSO and the Oct-2024 input - to decide whether the twentyfold arm span is a property of the replicate pairs or of the stratum definition. Zero downloads, one re-run of an existing script.
The cheapest way to check the surprising result in this analysis is to redo it using a different dataset to sort the genome. Everything needed is already on the machine and it takes one run.
Re-score every committed within-sample share ratio at S_FLAG against Q117's eviction quartiles - the bins are already on disk in five scratch directories, no download and no wet work - and start with the H3K79me2 PCGF1 arm, because cycle 83's central null is the one most likely to change
Every number of this kind the project has published averages over sites where the drug works well and sites where it barely works. Splitting them the way this cycle did needs no new data - the files are already on this machine - and could change what several earlier conclusions mean.
Re-supply the AF10 PZP module in cis to KMT2A::MLLT10 and ask whether transformation is suppressed
Put the missing sensor back into the fusion protein and see whether the leukemia stops. This has already worked for a related fusion; the point is to test it on the one that matters here.
Read all 76 archived CBX3 and BRCA1 tracks at the KMT2A exon-11 anchor at SIGNAL level against a CTCF-anchor-matched null, and score the anchor's rank among its own class in each - no wet work, no new data, and it decides whether the sub-IDR peak is a foothold or a threshold artefact
Three public datasets show a faint signal for two repair proteins at one small spot in the KMT2A gene, faint enough that the standard filter throws it away. This is the computer-only experiment that would decide whether that faint signal is real, by reading all seventy-six available datasets and comparing the spot to thousands of similar spots in the same cells.
Recalibrate the N/C oncoprotein threshold on the fusion-negative CD34+ control at the interval scale it will actually be used at, and report the false-positive count at every scale rather than one
A cutoff that makes almost no mistakes on one kind of genomic region can make hundreds on another. The fix is to measure the mistake rate on a healthy control at exactly the region size you intend to use, and to publish that number beside the results.
Redesign: the proximity-labelling map of KMT2A::AFF1 cannot be controlled with the wild-type allele - it has to be controlled with time
The plan was to tag the leukemic protein with a labelling enzyme and compare it with the same tag on the healthy copy of the gene. That comparison cannot work, because the healthy copy sits on exactly the same genes - so both tags spray the same neighbourhood. The fix is to destroy the leukemic protein for half an hour and label again: whoever leaves with it was there because of it.
Regress prey abundance against interactor confidence class across all 1,330 baits of the Hein atlas - the data are already downloaded and the run is an afternoon
A computer-only experiment: take the whole published screen, and ask whether the proteins it calls weak hits are simply the most common proteins in the cell. Nothing has to be done at a bench.
Reopen PXD009338 at peptide level — the search results are already deposited, so no re-search is needed, but the deposit carries no sample-to-file mapping and that is the real cost
The raw data behind the only MLL-fusion interaction map is public, and so are the search results, so checking how solid each protein identification is does not need the analysis to be redone from scratch. The catch is that the deposited files are not labelled with which fusion each one came from, and working that out is the actual work.
Repeat the H3K79 partition in a human KMT2A::AFF1 line: menin inhibition and MLLT10 knockout, side by side with a KMT2A::MLLT3 line
Do the key experiment again, but in human infant-type leukaemia cells instead of mouse myeloid ones, with a mouse-type comparison running alongside. Measure how much of the H3K79 tag a menin drug removes in each.
Repeat the one published RPA-at-the-KMT2A-breakpoint experiment in a haematopoietic cell, with an RPA-dose arm - the protocol is fully specified in a 2010 paper whose every arm used fibroblasts and a CML line, and whose primer sets, dose and controls are all published
One paper has ever measured the DNA-coating protein RPA sitting on the part of the KMT2A gene that breaks to cause leukaemia. It used skin-derived cells and a chronic-leukaemia line, and it was published in 2010. The experiment should be repeated in the blood cells this disease actually arises in, with an extra arm that raises and lowers RPA to see whether the amount of RPA changes how often the break turns into a translocation.
Rerun the 2002 SPR with the two controls it did not have: cleave the GST, and put the protein on the chip instead of the DNA
The 33 nM measurement used a two-headed protein flowed over a surface densely coated with DNA, which is the standard way to get an artificially tight number. Two straightforward repeats - remove the second head, and swap which partner is stuck down - would show how much of the number is real.
Resolve acetyl-K1130/K1133 occupancy on the fusion versus wild-type KMT2A in the same cell - the one experiment that makes a shared-sequence site interpretable
The chemical tag site sits on a stretch of protein that both the normal MLL1 and the cancer fusion carry, so an ordinary measurement cannot tell which protein was tagged. This experiment separates the two proteins first, then measures the tag on each, which is the only way the question means anything.
Restore the AF9 YEATS reader in cis to MLL-AF9, and read out the ENL patients who already did the experiment for us
Nearly all MLL-AF9 leukemias have thrown away a reader module that the near-identical ENL fusion keeps. Put it back and see whether the leukemia gets weaker — and in parallel, find the rare patients whose ENL fusion threw it away and see whether their disease behaves like AF9.
Retrospective paired-target MRD on stored Interfant marrows, stratified by IGHV usage
Use marrow samples that were already collected and already tested two ways, and ask whether the antibody-gene test failed more often in the babies whose leukaemia used the restricted antibody gene segment. No new patients and no new samples are needed.
Run a fusion-specific and a total-AFF1 readout side by side in the same KMT2A::AFF1 nuclei, and report the two target sets rather than one
Measure where the fusion protein is using a method that can only see the fusion, and at the same time with the usual antibody that also sees the normal partner protein, in the same cells. The difference between the two answers is the thing nobody has measured.
Run the menin inhibitor against CALM-AF10 next to a KMT2A-rearranged control - the missing two-week experiment
Take leukaemia cells carrying the CALM-AF10 fusion, treat them with a menin-blocking drug alongside cells carrying a KMT2A fusion, and see whether the HOXA genes switch off. Nobody has published this, and it would take about two weeks.
Score the HSPC-like fraction on paired pre-treatment and on-treatment samples from KMT2A-rearranged patients receiving a menin inhibitor — a flow-first design with a single-cell confirmatory arm
Take marrow or blood from children on a menin inhibitor, before treatment and again a few weeks in, and count what fraction of the remaining leukaemia cells look like blood stem cells rather than B cells. If that fraction grows, the drug is leaving behind the cells that cause relapse; if it shrinks, the drug is clearing exactly the right cells. Either answer changes how the drug should be used.
Separate AKAP95's two routes into the KMT2A axis - starting with one transfection that re-reads a thirteen-year-old control
A staged experiment to find out which of AKAP95's two grips on the KMT2A machinery the leukaemia actually needs. The first stage is a single cheap test using a construct that already exists.
Separate ncBAF from canonical BAF in KMT2A-rearranged cells: paired degron, paralogue-pair CRISPR, and complex-resolved CUT&RUN
Three experiments that would settle whether MLL-rearranged leukaemia cells need the minority flavour of the SWI/SNF machine or the common one: degrade each part quickly and watch what happens, knock out the interchangeable pairs together, and map where each flavour sits on the genome.
Separate positioning from catalysis: an ASH2L mutant that keeps RBBP5 and loses DPY30
Build a version of ASH2L that still assembles the H3K4 enzyme but cannot hold DPY30, put it into a KMT2A::AFF1 leukemia line, and see whether the cells behave as if DPY30 were gone. If they do, DPY30's job really is aiming the enzyme.
Separating CK2's two routes into KMT2A leukaemia, which every published experiment has confounded
Blocking CK2 helps in this leukaemia for two possible reasons, and no experiment has ever told them apart. A short time-course with the right readouts would.
Separation-of-function RTF1 rescue in a KMT2A-fusion line: does the fusion need PAF1C for elongation, or for H2B ubiquitination?
Knock out RTF1 in a KMT2A-fusion leukaemia cell, then put back either the normal protein or a version that can still help transcription but cannot trigger the ubiquitin tag, and see which one rescues the leukaemia programme. It separates two explanations that the field currently runs together.
Separation-of-function test: does KMT2A::AFF1 need its C-terminal homology domain to dimerise?
Make three versions of the fusion protein — normal, one with the suspected sticky module cut off, and one with just a few surface letters changed so it keeps its shape but cannot stick — and see which ones can still turn blood cells cancerous. If only the normal one works, the sticky module is doing real work.
Settle the monocytic confound isogenically: degrade the fusion, read haem-gene dependency before lineage has time to move
The project found that fusion-carrying leukaemia cells need certain iron-and-haem genes more than other cells do, but almost all those cells are monocytes, whose normal job is handling iron. The clean way to tell the two explanations apart is to switch the fusion off inside one and the same cell and see whether the iron dependency goes away before the cell has had time to change into a different kind of cell.
Settle the WRAD question: tagged co-IP plus degron-timed proximity across three fusion classes
One set of experiments answers a twenty-two-year-old question. Tag each kind of KMT2A fusion protein, pull it down, and look for the four WRAD proteins - then repeat with the normal KMT2A switched off, so a positive result cannot be explained away as WRAD hitching a ride.
Single-cell CUT&Tag on p279 — is the double-positive blast one cell state or two clones?
One patient's leukaemia appeared to have the driver protein on both HOXA and IRX1 at once. That could mean two different groups of cells mixed together, or one cell doing both. Only looking cell by cell can tell.
Six to twenty-four libraries in RS4;11: two DOT1L inhibitors titrated to equal spike-in-measured global H3K79me2 loss, three replicates per arm, plus the no-primary control that no public deposit contains - the smallest experiment that would convert four moderate entries in this beat into testable ones
Run the same chromatin experiment three times instead of once, with two different drugs given at doses that remove the same total amount of the mark, and include a control that shows how much signal the antibody produces when there is nothing to find. That last control does not exist in any public dataset we could find, and without it we cannot tell real leftover mark from background.
Sort KMT2A-rearranged blasts by cell-cycle state, then measure menin-inhibitor response — the one-week experiment nobody has run
Separate leukaemia cells into the ones that are dividing and the ones that are resting, then give each group the same drug and see which responds. It is a standard sorting technique and a standard drug assay, and putting the two together would say whether this drug class can touch the cells that cause relapse.
Spike-in-normalised menin CUT&RUN in a KMT2A::AFF1 degron line, with replicates and a PSIP1 arm
The experiment that would settle this: measure menin on chromatin in the fusion this project actually cares about, with a built-in ruler so the amounts are real, repeated enough times to trust, and with LEDGF knocked down in one arm to see whether that is what holds the leftover menin.
Spike-in-normalised menin CUT&RUN with paired N- and C-terminal KMT2A antibodies in an infant KMT2A::AFF1 sample, on and off drug, in the same disease state
The experiment that would settle this: take one baby's KMT2A::AFF1 leukaemia cells, split them, treat half with a menin inhibitor, and measure menin and both ends of the KMT2A protein with a counting standard added to every tube - so the numbers are amounts, not proportions.
Spike-in-normalised paired KMT2A N- and C-terminal CUT&RUN, with a fusion-junction readout, in a HOXA-negative KMT2A-rearranged blast beside a HOXA-positive one from the same institution and batch
Repeat the measurement with an internal ruler added to every tube, so that 'there is less of it' can be told apart from 'it is somewhere else'. Run the odd patient and a normal one side by side.
Strand-resolved transcription across the AFF1 / AFF1-AS1 / C4orf36 promoter - the only measurement that can say which gene the signal belongs to
An occupancy measurement looks at protein sitting on DNA, and a piece of DNA shared by three genes has no owner. To find out which gene is actually being switched on, you have to measure the RNA and which direction it is going. Nobody has.
Stratify an existing menin-inhibitor cohort by reciprocal fusion status
Roughly half of these children carry a second fusion protein that the menin drug cannot touch. Nobody records which half a patient is in. The test is one cheap PCR on samples that already exist.
Swap the AFF1 C-terminus onto a KMT2A-AF9 backbone and score IRX1 occupancy: the direct test of partner causality
Every KMT2A fusion keeps the same front half of the protein; only the back half differs. Building a fusion with an AF9 back half and an AFF1 back half in the same cells, and asking which one puts the protein on IRX1, tests the partner explanation directly.
Tag endogenous KMT2B in an index-fusion line and measure what revumenib does to it - the experiment the reagents have blocked
Nobody has ever watched what a menin inhibitor does to KMT2B on the DNA, because the antibodies do not work. Putting a small tag onto the gene itself in a leukaemia cell line removes that obstacle entirely.
Tag the endogenous KMT2A::AFF1 allele in RS4;11 and ChIP the tag with and without VTP-50469 alongside MENIN and MLL1-N in the same chromatin - the one experiment that converts every occupancy claim in this literature from a reagent statement into a fusion statement
Put a small molecular handle on the cancer fusion protein in the standard human cell line for this disease, then pull it down with and without the drug. That single experiment would say whether the drug actually takes the fusion off the DNA, which nobody currently knows for this fusion.
Targeted PRM for the KMT2A::AFF1 junction peptide - the experiment that makes every other attribution claim testable
Chop the proteins from a leukemia cell into pieces and look, with a mass spectrometer, for the one piece that can only come from the fusion protein. We have computed exactly which piece that is and what it weighs, so the experiment is a targeted search rather than a fishing trip.
The 2x2 paralogue swap: separate 'is the reader there' from 'which backbone is it on', in one background with matched expression
Build four versions of the cancer protein in the same cells - AF9 with and without its reader, ENL with and without its reader - and measure them the same way. Nobody has ever run these two fusions side by side in one experiment.
The analysis that would settle it: a partner-by-lineage interaction term in pooled individual-patient data
Every study so far has looked at partners inside one disease or one age group. The decisive test is to put them all together and ask directly whether the partner effect changes depending on the disease and the age.
The binary ladder for AFF1: does the index fusion's partner touch P-TEFb without AFF4, and does it touch AFF4 at all?
Buy or make the pieces, mix two at a time, and measure whether they stick. The whole field's picture of how the AF4 protein plugs into the gene-switching machine comes from experiments that pulled everything out at once, so nobody knows which piece it actually holds on to.
The cheap experiment nobody has run: an MLL1 C-TERMINAL ChIP beside the standard N-terminal one in a single human KMT2A-rearranged line under a menin inhibitor - the C-terminus is lost in the fusion, so the difference of the two tracks is the fusion, with no tag and no knock-in
The fusion protein keeps the front half of KMT2A and loses the back half. So an antibody against the back half sees only the normal protein, and subtracting it from the usual antibody's signal leaves the fusion - one extra experiment in a run people already do.
The cheapest missing experiment in this beat: put revumenib on a parathyroid cell
Nobody has ever put a menin-blocking drug onto parathyroid cells. Take cells from a routine parathyroid operation, add the drug, and see whether the cells stop obeying the signal that normally tells them to make less parathyroid hormone.
The cheapest study that would settle this: nine assays and a hand radiograph, on children already being followed
Add a short list of blood tests and one hand X-ray to the follow-up that children treated with menin inhibitors already receive, at fixed times during and after treatment. It uses blood that is already being drawn and would answer within two years whether the drug leaves a hormonal mark.
The days-long experiment that would close or confirm the imprinting liability: allele-specific XIST and MEG3 readout in cells treated with a YEATS antagonist
Treat cells with the reader-blocking drug and check whether genes that are supposed to be silent on one copy stay silent. The same readout was used to show that removing AFF3 breaks this, so the assay already exists and has never been pointed at a drug.
The decisive experiment: acute degron removal of IGF2BP3, IGF2BP2 and both, read on self-renewal and xenograft initiation rather than bulk dropout
Instead of switching off IGF2BP3 for two weeks and counting cells, this experiment removes the protein within hours and asks whether the leukaemia can still start fresh disease in a mouse. It also removes the sibling protein, alone and together, because a gene that has a backup will look unimportant when tested by itself.
The decisive test that needs no new data: overlay the 12,443-anchor CTCF-loss ranking on the copy-number changes Gray 2023 reports, and ask whether the 280 anchors that lose CTCF as hard as KMT2A also amplify
We now know that thousands of loop anchors lose their anchor protein when this enzyme is removed, and that the KMT2A gene is one of the harder-hit ones but far from unique. The question that settles the mechanism is whether all those other hard-hit anchors also get copied and broken. If they do not, then losing the anchor protein is not enough on its own, and something else about KMT2A is doing the work.
The deposit this graph filed as 'four SUV39H1 ChIP-seqs' is actually 77 samples in MCF-7 carrying the ENTIRE H3K9 axis in one cell type and one pipeline - writer SUV39H1, writer G9a/EHMT2, reader HP1-alpha/CBX5 (wild-type AND two phospho-mutants), co-repressor KAP1/TRIM28, erasers KDM4B and KDM4C, the mark H3K9me3, and an shKDM4B+shKDM4C eraser-knockdown arm
A dataset this project had catalogued as four experiments on one enzyme turns out to contain seventy-seven, covering every protein that writes, reads, removes or is the silencing tag - all in the same cells, on the same day, with the same method. That is the comparison this project has never been able to make, because until now every piece came from a different cell type.
The experiment that closes the LEDGF hole: doxycycline-inducible PSIP1 knockdown switched on AFTER a human KMT2A::AFF1 xenograft is established
Put a switch on the LEDGF gene in human leukaemia cells, let the leukaemia take hold in a mouse, then flip the switch off and see whether the leukaemia dies. That one experiment would tell us whether a LEDGF drug could treat a patient, rather than only prevent a disease that has already happened.
The experiment that separates lineage from ontogeny: one KMT2A fusion, two cell-of-origin contexts, three metabolic perturbations, and a limiting-dilution readout alongside the viability one
Put the same leukaemia gene into blood cells from before birth and into blood cells from an adult, then measure how each one burns fuel and how well each one can restart the disease in a mouse. If the two behave differently, the difference in metabolism that everyone attributes to cell type is really about when the cell was born.
The experiment the public data cannot supply: H3K9me1/2/3 and CTCF at the KMT2A exon-11 anchor in a cord-blood CD34+ progenitor, with and without KDM3B depletion
Everything known about this mechanism was measured in eye cells. The experiment that matters is the same measurement in the blood stem cells where infant leukaemia actually starts - and those cells are already the ones the original paper used for its one blood-cell result.
The experiment this beat could not do: an epitope-tagged KMT2A::AFF1 ChIP with a Drosophila spike-in, in a human infant-derived B-lineage line, at three replicates and two menin-inhibitor exposures - the design that would turn a within-sample ranking into an absolute occupancy measurement
Do the same experiment in the human infant leukaemia the lab actually studies, with a measuring stick added to every sample so that losses and gains can be compared properly, and with three repeats so the numbers have error bars.
The experiment this beat could not do: G9a, HP1a, KAP1 and H3K9me3 at the KMT2A promoter and exon-11 anchor in a KMT2A-rearranged B-lineage cell and a CD34+ progenitor, WITH an input arm, an IgG arm and a same-cell CTCF arm
Everything this project knows about these silencing proteins at the KMT2A gene comes from breast and blood-cancer cell lines that never get this leukaemia. The experiment that would make it relevant is a standard one and nobody has done it.
The experiment this deposit is one arm short of: MEN1 ChIP after fusion degradation with a PSIP1-depletion arm, a spike-in, replicates, and timepoints past three hours - four additions to an existing protocol that would settle why menin stays
We found that menin stays stuck to DNA after the cancer fusion protein is destroyed, and the likely explanation is a second anchor holding it there. The experiment that would prove it is small: repeat the existing one, but also remove the suspected anchor, add an internal measuring standard, do it more than once, and look at later timepoints.
The experiment: cross-tabulate MLLT1 breakpoint class against age, lineage and outcome in the Frankfurt archive, then confirm reciprocal status by RNA in whoever has it
Three steps, all on data that already exists. First check that the two patient groups are alike in age and disease type; then link them to how they did; then confirm on a subset that the mirror-image gene really is absent in one group and present in the other.
The measurement that settles the CTCF dispute: degron-grade CTCF removal versus cohesin removal, with DNA breaks AND rearrangement outcome read out at the same 1 kb of KMT2A in the same cells, reported as rearrangements per break
Two sets of papers disagree about whether the CTCF protein makes the KMT2A gene break or stops it rearranging, and neither set measured both things. This experiment removes the protein cleanly and counts both, in the same cells, at the same spot.
The missing measurement: is MHC class I bivalent, does a menin inhibitor raise it, and does raising it cost NK control - in a KMT2A-rearranged cell, for the first time
Nobody has ever checked whether the immune identity-card genes are held in the half-on state in this particular leukaemia, whether a menin inhibitor turns them up, or whether turning them up makes a different kind of immune killer worse at its job. One set of experiments on cells this project already uses would answer all three.
The missing reagent is a spike-in, not a better statistic: a spike-in-normalised H3K4me3, H3K79me2 and H3K27me3 ChIP-seq at 72 h of a clinical menin inhibitor in a KMT2A-rearranged human line, which is the only design that can tell a genuine genome-wide mark redistribution from the compositional shadow of a peak-localised change
Every chromatin result in this thread is a percentage of that sample's own total, which means a change that happens everywhere at once is completely invisible. Adding a fixed amount of foreign chromatin to each sample before the measurement gives an absolute yardstick and separates 'the marks moved from here to there' from 'there is simply less of the mark'.
The missing reagent: a KMT2A-fusion or wild-type-KMT2A proximity bait in a KMT2A-rearranged leukaemia line - IntAct has curated no KMT2A-bait interaction experiment since April 2015 and no PSIP1-bait since October 2012
Nobody has tagged the protein this project is about and asked what sits next to it in a leukaemia cell for over a decade. Everything currently known about its neighbours comes from other people tagging other proteins and finding it by accident.
The noncoding dependency screen five hypotheses in this graph are waiting on - and the KMT2A-rearranged screen that already exists and bounds it
Most of the genome is not genes, and almost every dependency screen in this leukaemia has only tested genes. This is the screen that would test the rest of it - the switches, the repeats and the structural boundaries - in the fusion that causes infant leukaemia.
The one experiment that separates fusion partner from lineage: an H3K79me2 ChIP under a menin inhibitor in ML-2 or any KMT2A::AFDN line, alongside the MLL1n and MENIN samples that already exist for that line and can be re-binned for free
Two cell lines gave opposite answers about the histone tag and they differ in two ways at once. A third line, already used in the same published dataset, would separate the two explanations - but the tag was never measured in it.
The one experiment that settles this beat: dose-matched menin and DOT1L arms in RS4;11 with H3K4me3 ChIP at PHGDH, plus 13C-serine flux - and it answers the cycle-58 question at the same time
One experiment in cells two laboratories already grow would decide whether the serine result in this beat is real biology or an artefact of comparing two drugs at arbitrary doses. Match the doses so both drugs hit the leukaemia's own marker genes equally hard, then look at the serine genes, at the chromatin mark on the PHGDH gene, and at whether the cells actually make less serine.
The one experiment that would close the SUV39H1 arm of this graph's H3K9 argument: CST #8729S CUT&RUN in a CD34+ progenitor and in SEM or RS4;11, with an IgG arm and a SUV39H1-degron arm that neither existing deposit has - because after this cycle the enzyme has been measured in exactly two human cell types and neither is a blood cell
Everything public about where this enzyme sits on human chromosomes comes from two cell types: a children's brain tumour and a breast cancer line. One experiment in blood-forming cells and one in a leukaemia line with the relevant gene fusion would settle the question this project keeps having to answer by substitution.
The one experiment that would close this beat: MEN1 and DOT1L ChIP in RS4;11 (KMT2A::AFF1) under DMSO, a menin inhibitor and a target-engaging DOT1L inhibitor, with Drosophila spike-in and two replicates per arm - about 24 libraries, and the depositing lab already runs every component of it
Everything found here was measured in the wrong leukaemia and without the internal ruler that would let anyone say how much protein was really lost. The same lab already has all the pieces; running them together in the right cell line would settle it.
The reciprocal ChIP: is chromatin-bound XPO1 present in a KMT2A-rearranged cell, and does a menin inhibitor move it?
Nobody has ever checked whether the export protein XPO1 sits on the HOXA genes in a KMT2A-rearranged leukaemia cell - the disease this whole project is about. One standard experiment, run with and without a menin inhibitor, answers that and also tests whether the two drug targets are on one assembly or two.
The test bed nobody names: KMT2A rearrangement can be induced and counted in primary human CD34+ and fetal liver stem cells, and it happens without any drug at all
You can take real human blood stem cells, treat them, and count how often the KMT2A gene breaks and rejoins wrongly. The most uncomfortable result from that assay is that the breaks also appear in untreated cells that are simply being told to divide.
The two co-IPs that settle whether menin's SUV39H1 window and its gamma-H2A.X motif are one surface: swap each laboratory's construct into the other's assay
Two labs made the constructs needed to answer this and neither tested the other's. Put the deletion mutant from the 2026 repair paper into the 2013 enzyme-binding assay and the two patient mutants from 2013 into the 2026 repair assay, and the question closes.
Three blots and one drug curve: does the ZMYND11::MBTD1 fusion protein still bind KMT2A, and does a menin inhibitor work in a leukaemia with no KMT2A rearrangement
Two experiments, run in order. First a single protein-pulldown asking whether the leukaemia fusion protein still grips the enzyme that its normal parent gene holds back - if it does, the idea is dead and nothing else needs doing. If it does not, then treat fusion-driven leukaemia cells with the approved drug revumenib and see whether they die, in a leukaemia the drug was never meant for.
Three extra tubes: the glucose study that answers the inversion in children
Add fasting glucose, insulin, C-peptide and glucagon to blood that is already being drawn from children on menin inhibitors, and take the same tests again a few months after the drug stops. It would show whether the effect seen in diabetic mice happens in a child, in either direction.
Three routes to one observation: the CUT&RUN panel that says whether menin eviction, H3K36me3 loss or MLL loss takes LEDGF off the fusion's genes
We know LEDGF leaves the leukaemia's genes when the menin drug is given, but not why. Three different explanations predict different results in one experiment that uses drugs and cell lines that already exist.
Tile chr5:2,750,677-2,752,892 at 100 bp from the GSE252378 bigWigs and ask which TSS the p454 signal sits on
The data needed to tell IRX2's signal from its neighbour's already exists; it was thrown away when the signal was averaged over two thousand letters. Reading the same files in hundred-letter steps instead would show where the signal actually sits, at no extra experimental cost.
Titrate PRC1.1 capacity against menin-inhibitor response, and measure H2AK119ub on the same timecourse as H3K79me2
Two papers disagree about whether one Polycomb machine is the menin drug's helper or its off-switch. This experiment dials that machine down a bit at a time instead of deleting it, and watches the drug's effect change - which tells you which story is right.
Track the sub-1% cooperating-lesion clones through menin inhibition, because nobody has measured what the drug selects for
A menin inhibitor is now approved for this leukaemia, and drug-resistant cells making up less than one cell in three hundred at diagnosis are known to grow back after chemotherapy. Nobody has checked which of those hidden cells a menin inhibitor spares. This experiment does that by sequencing the same patient's samples very deeply, before, during and after treatment.
Treat a KMT2A::AFF1 patient-derived xenograft with an NG2-directed agent and read out the CD34+CD19-CD117+ fraction at nadir - does the reservoir the 2006 flow data predicts actually survive?
Give a mouse carrying a child's leukaemia a drug aimed at NG2, then look at what survives. If the surviving cells are the immature ones that lack NG2, the therapy is a debulking agent that needs a partner drug rather than a cure.
Turn menin off for a defined window, then back on: the missing experiment that would bound reversibility
Mice that lose menin in their insulin cells get overgrowth in two months and tumours by six. Nobody has tried switching menin back on after one, two or three months to see whether the damage undoes itself - which is exactly the question for a child who takes the drug for a while and stops.
Two constructs that would decide whether menin dependency is about the menin-binding motif or about HOXA level
Two engineered versions of these leukaemia proteins would settle a question the whole field is arguing around: add the missing menin grip back to the one drug-resistant NUP98 fusion, and separately delete the grip from the cell's own normal copy in the drug-sensitive ones.
Use centrifugal elutriation, not a chemical block: the one separation method demonstrated to work in KMT2A-rearranged cells, applied to both the cell-cycle question and the small-persister question
Standard drugs that freeze cells at one stage of division do not work on these leukaemia cells, but spinning them apart by size does. The same spin could separate the rare small quiet cells that survive treatment, which nobody has tried.
Which marker moves first? A re-analysis of flow files that already exist, to put an order on lineage switch
Every hospital that treats this leukaemia keeps the raw data files from every flow cytometry test it runs, including the ones taken between diagnosis and relapse. This experiment proposes going back through the files of patients who switched from a B-cell to a myeloid leukaemia and asking, marker by marker, which one changed first.
Reproducible scripts against real public data.
'AFF1-only' is a property of mid-activity promoters: zero in the bottom three deciles, zero in the top decile, and 81% of the 332 in deciles 6-8
Sorting all nineteen thousand gene switches by how much oncoprotein signal they carry on average, the 'only in KMT2A::AFF1' ones are all in the middle of the range. The quietest thirty per cent never pass the test at all, and the loudest ten per cent pass it in every patient, so neither can ever look partner-specific. Partner-specificity is only visible in the middle, which is where a threshold decision is fragile.
'Four consecutive beats found 40-50% of their genes anti-selective' was the null expectation, not a pattern - and the real test shows the curated set IS enriched for selectivity, p = 0.0016
Across four rounds of work this project kept noticing that about half the genes it examined were *less* needed in this leukaemia than in others, and treated that as a meaningful signal. About half of all genes are in that group by definition, so the observation carried no information. Testing it properly shows the opposite: the genes this project has chosen to write about really are more selectively required than genes picked at random.
"Diverse models of MLL-rearranged leukaemia" is one mouse cell line: the RNF20 H2Bub requirement resolves to murine RN2 (MLL-AF9/NrasG12D)
A paper this graph cites for the claim that the H2B-tagging enzyme RNF20 is needed by MLL-rearranged leukaemia never says which leukaemia it used. Its deposited sequencing is three samples from a single mouse cell line, RN2, which carries MLL-AF9 - not the infant fusion this project studies.
"ENL" returns 1,109 PubMed records of which about 30% are the protein: a measured alias-precision failure
The short name ENL is used for at least three unrelated things in the medical literature - the leukaemia protein, a skin reaction in leprosy, and a plant compound. Searching for it returns mostly the wrong ones, so any count built on that word is wrong by roughly threefold unless you check every hit.
"Seen in MLL-AF4 and no other bait" excludes AFF1, AFF4 and ELL — the three most AFF1-specific proteins in the experiment — and keeps ACADM and GLUD1
A published table lists 110 proteins that stuck to the MLL-AF4 fusion and to none of the six other fusions tested, and those 110 have been treated as the partners unique to our fusion. But the rule that built that list throws out AFF1 itself — the very protein our fusion is half made of — because a trace of it also turned up in one other pulldown, while it keeps two mitochondrial enzymes.
[CORRECTED cycle 110 - F2 not F1, and the floor spans ~20x across arms] Measuring the same-condition histone-mark noise floor from replicates already binned on local disk, zero downloads: 15.97% at the peaks and 3.03% at D1, so the borrowed 10% floor was 3.3x too generous in the bulk - independently reproducing, in a different deposit, cell system, mark and antibody, what the GSE299100 beat measured at 9.4% and 2.4%
To tell a real change from noise you need to know how much two repeats of the same experiment differ. This project had been using a rough figure borrowed from a different kind of measurement. Repeats of the right kind turned out to be sitting on the local disk already. They show the noise is large where the signal is concentrated and about three times smaller everywhere else - so the borrowed figure was far too forgiving in exactly the region where a recent conclusion was drawn.
[CORRECTED cycle 117 - the columns were not scaled] [CORRECTED cycle 110 - the gradient is NOT an artefact] Grading this project's own stratified-share statistic on data where the truth is known: it is not blind to a global collapse, it is DISTORTING - a pure 5.84x genome-wide loss with no redistribution whatsoever is rendered as a monotone gradient rising from the peaks into the sparse deciles, reporting a 9.70x INCREASE at D10 where the absolute change is 1.66x, and 62 of 66 strata are confidently outside the measured noise floor
This project measures chromatin as proportions - what fraction of a signal sits where. A public dataset now provides a case where the true answer is known independently: a drug wipes out a chemical mark across the whole genome. Re-running this project's own method on that dataset shows it does not simply miss the change. It reports a large change of the wrong kind - it says the signal moved out of the busy regions and into the quiet ones, by up to ten-fold, when in reality nothing moved anywhere and everything just got smaller. The shape that comes out looks exactly like a result this project published three cycles ago.
[CORRECTED cycle 117 - the columns were not scaled] Correcting yesterday's correction: share(s) = absolute(s)/G is an exact identity, so a within-sample share reproduces the true SHAPE faithfully and misstates only the LEVEL - but that is the worse failure, because 34 of 66 strata lost up to 68% of their signal while reading as gains of up to +440%
Yesterday this project concluded that its proportion-based method invents a false pattern out of nothing. That was wrong, and the arithmetic shows why: the pattern the method reports is exactly the true pattern, just rescaled. The genuine problem is different and more dangerous. A proportion only ever says how one region did compared with the average of everything. When everything is falling, a region that fell by half can still look like it doubled. In this dataset that happened to more than half of the regions examined.
`absence_class` is one token per ENTRY while the entries carry seven to eleven distinct gaps each, so it is a summary label - and inside gene-nbn it hid a verifiably false ENCODE claim for three cycles
We label each entry with one word describing the kind of gap in it, but a typical entry contains seven to eleven different gaps of different kinds. Inside one of them that single word was hiding a statement that four public experiments do not exist, when they do.
124 entries already write NOT APPLICABLE as a category statement on the localization dimension, and the epigenetics dimension has no equivalent - so a correctly-annotated Golgi enzyme and an unwritten chromatin protein are indistinguishable in the epigenetics figure
When a protein sits in the Golgi and has no address inside the nucleus, this project's localization entries say so explicitly - 'NOT APPLICABLE, and this is a category statement, not a gap'. Nothing like that exists for the chromatin topic, so a protein that correctly has no chromatin biology looks exactly like one nobody has got round to.
148752 and 152985 are one patient before and during revumenib, not two treated patients - so the drug-free arm this corpus said did not exist is sample 148752
The two infant samples this project believed were both taken while a menin-blocking drug was being given are in fact one child, sampled twice: once at diagnosis before the drug, and once at relapse while on it. That means the before-and-after comparison the project had written off as impossible is available.
18% of this graph's evidence is the same finding said again
Every cycle this project reports how many pieces of evidence it holds, and the number keeps going up. It turns out about one in six of those pieces is the same result from the same paper, written out again in a different entry.
21.6% of this project's 19,404 promoter windows contain another gene's TSS, so an attribution caveat is corpus-wide and not an IRX2 peculiarity
This project measures each gene by taking the average signal in a 2,000-letter stretch around its start. For one gene in five, a different gene starts inside that same stretch, so the number belongs to two genes at once. IRX2 is not a special case - it is an ordinary member of a large group.
22 of the 37 promoters in the published B-ALL-specific peak class are the clustered protocadherin gamma locus - one 23.8 kb interval at chr5:141,473,710-141,497,491 - and it reads below an average promoter in all 27 samples
A gene list that looks like 37 genes is really 16 places in the genome, because 22 of the entries are neighbouring copies of one repeated gene family sitting under a single peak. That one place is quiet in every patient, so it drags the whole list down.
3,288 papers are 2,025 research groups, the per-entry overcount is 11.7% and barely moved while the graph doubled - and TWO entries now rest on three papers from a single laboratory, where the first run of this analysis found none
Counting papers overstates how many separate teams stand behind a claim, and here it overstates it by about a eighth. That fraction has stayed the same as the graph doubled in size.
4,589 genes with no prior, and menin comes first (run)
**Extended, not superseded:** `ana-depmap-kmt2a-selective-scan-full` re-ran this over the whole protein-coding genome rather than the 4,589 genes here, and **menin still came first** — so this scan's headline survives the wider search. The -full run additionally surfaced a haem-biosynthesis cluster that this one could not see. Until now every dependency analysis here started from a shortlist taken from the literature, which can confirm but never discover. This one scanned nearly five thousand genes with no shortlist and asked which are most specifically needed by leukemia cells carrying the fusion. Menin came first.
42.5% of every node in this corpus names a histone mark, so 'names a mark' is a weak instrument - but it is not noise, and the two types where it collapses are the ones that tell you it is measuring something
This project's coverage tool gives an entry credit for epigenetics if a histone mark is named anywhere in it. Almost half of everything in the whole knowledge graph names a histone mark, because the whole project is about a chromatin disease - so that credit is worth much less per entry than it looks.
42% of every gene-body occupancy call is at a locus where wild-type MLL1 is also bound - but the AFF1-exclusive class is not, HOXA9 and MEIS1 are not, and SPAG7 is
The antibody this project relies on recognises both the fusion protein and the normal version of the same protein. Comparing it with a second antibody that sees only the normal version shows that about four in ten of the sites counted as 'occupied' have the normal protein there too. The handful of genes the project cares about are not among them - with one exception, which should be dropped.
43 of the 62 discussed-but-unwritten interactome partners (69%) are genuinely draftable from the committed corpus, so cycle 93's decision to work them before the 34 never-read ones has a real cost basis - but the clause that decided it was the citation clause I thought was safe, not the proxy I flagged as weak, and the 12 RNA-processing failures turn out to be a promiscuity census that should NOT be promoted partner by partner
The previous cycle decided to finish the partner proteins the project has already read about before starting the ones it has never read, on the grounds that the first group is cheaper, and this checked whether that word was true: for forty-three of the sixty-two the project's own files really do contain enough to write a page, so the decision holds. The nineteen that fail are mostly RNA-processing proteins, and looking at where they came from showed they should not simply be written up - they come from a survey in which the main protein was picked up by half of all the experiments, which may say more about its size and abundance than about any real partnership.
43.9% of every occupancy call this project's rule makes is at a shared window it cannot attribute - twice the 21.6% base rate, in 27 of 27 samples
One promoter window in five in this project overlaps another gene's. But among the windows the fusion-occupancy test actually calls positive, it is nearly one in two. So the calls the test makes are twice as likely as chance to belong to two genes at once.
49 of 91 proteins in this graph have never had a drug-like molecule bound to them
For every protein in the graph we asked two questions that need no new experiment: how much of it has anyone ever seen at atomic resolution, and has anyone ever made a small molecule that sticks to it and solved the picture. Slightly more than half have never had such a molecule, and the ones that have are mostly not the ones the fusion keeps.
56 of KMT2A's 130 observed phosphosites survive the translocation, and the retained fragment is phospho-denser than the part that is lost
We counted every place on KMT2A where a phosphate has actually been seen, then asked how many of those places are on the piece that leukaemia fusions keep. Forty-three percent of them are, and per unit length the kept piece is more heavily phosphorylated than the piece that is thrown away.
577 citations carried a bare PMID with no title, so the audit built to catch a fabricated title could not see them: the third dropped denominator in this project
An audit built to catch invented citations came back reassuring - four suspicious out of roughly 8,700 - but it could only grade a citation that had a title stored beside it. 577 citations stored nothing but a bare reference number, so for those the checker confirmed only that the number pointed at a paper, never that it pointed at the right one.
78% of the noncoding beat rests on a claim that nobody has done something
We went through every entry about the non-protein-coding side of this disease and counted how often it argues from something nobody has done yet. Eighteen of twenty-three entries do, and most of those statements carry no date and no search behind them — which is exactly the kind that was proved wrong three times in a single sweep.
A 109-transcription-factor BioID network puts KMT2A next to MYC, YY1 and the ETS and KLF families - and AF10 next to ERG and KLF5
A 2022 study tagged 109 human transcription factors and listed everything each one touched. KMT2A turned up beside MYC, YY1 and several other well-known factors; AF10, a KMT2A fusion partner, turned up beside ERG and KLF5.
A caveat that weakened the index fusion, wrongly: the MLL-AF4 spreading deposition is 30 SEM samples plus six other lines, all human
Two entries hedge the strongest human t(4;11) evidence in this graph with the note that the paper does not name its cell lines. It deposited fifty-nine human samples and they are named in detail - thirty of them are SEM, the infant pro-B line carrying exactly the fusion this project studies.
A checker for the error the citation checker cannot see
Our existing checker catches a citation pointing at the wrong paper. It cannot catch the opposite mistake, where the paper is right and the sentence attached to it says something the paper does not. This scores every claim against the abstract it cites.
A ChIP-seq series of 158 samples that names no antibody anywhere in its metadata: the string 'antibod' occurs ZERO times in GSE260742's 419,824-character SOFT dump, and the reagents are in a supplementary spreadsheet the full text never quotes
A public dataset of 158 chromatin experiments does not record which antibody was used for any of them. The information exists, but only in a spreadsheet attached to the paper, which no database search can reach.
A congress abstract book is one indexed document containing thousands of unrelated abstracts, so it matches almost any conjunction: across 18 Europe PMC full-text intersections read to completion, 83 of 321 records (25.9%) are proceedings, ranging 0% to 50% per query
Searching the whole text of papers is more powerful than searching summaries, but it has a new way of going wrong: a conference programme is filed as a single document containing thousands of separate short reports, so it matches nearly any combination of search words. A quarter of the results this beat got were conference programmes.
A detector for cross-paper contradictions, and the measurement that says it does not work
The graph was holding two opposite answers to the same question in different places, and nothing noticed. So we built something to notice. Then we tested it on the case we already knew the answer to, and it failed: the real contradiction came 76th in a list of 200, below two hundred pairs that were not contradictions at all.
A DOT1L inhibitor DOUBLES H3K79me2's mid-occupancy share (1.89-2.01) and a menin inhibitor CUTS it (0.69-0.78), in THREE murine samples one of which is a replicate (corrected cycle 97; the id still says four) with no overlap between the ranges, while the FLAG and MLL1-N channels sit at 0.905-0.929 under both drugs - so the two drugs do not act on one axis and 'a fraction of the achievable effect' presumes they do
Both drugs reduce the H3K79 tag where the fusion protein sits, but the DOT1L drug spreads it out of its peaks into the middle of the genome while the menin drug lowers it in the peaks and the middle alike. Expressing one drug's effect as a percentage of the other's assumes they are the same kind of change, and in four separate cultures they are not.
A DOT1L inhibitor, which does not touch the menin-KMT2A interface, removes 70% of the fusion's enrichment from its own strongest sites - two thirds of what a menin inhibitor removes - so fusion chromatin residency is partly downstream of H3K79 methylation rather than only of menin
A drug aimed at a completely different enzyme knocks the fusion protein off its favourite DNA sites almost as effectively as the drug designed to do exactly that. Whatever holds the fusion on the chromatin, it is not only its grip on menin.
A failure-mode atlas for drugging the KMT2A::AFF1 axis: 90 modes over a 10 x 11 grid of layers and mechanisms, 46 cells argued not applicable, and eight modes the corpus has no entry for
We made a checklist of the ways treating this leukaemia can go wrong, built so that it cannot quietly skip a category: ten places something can fail (the drug, the target protein, the proteins around it, the DNA packaging, the RNA, the cell, the cancer cell population, the body, the clinic, and our own knowledge) crossed with eleven ways it can fail (a relative steps in, another route takes over, the target changes, the drug lets go, it creates something new, it harms something else, the population evolves, the cell changes identity, it springs back, other drugs interfere, or we mismeasure). That gives 90 specific failure modes. Eight have nothing about them anywhere in our existing entries. The list is only as complete as the grid, and the 46 cells we ruled out are claims that could be wrong.
A falsifier written and scored inside Q118 last cycle, cashed again in the second cell line one cycle later: H3K79me2 at the MLL1-N held bins is 0.7008 in RS4;11 (does not fire, threshold 0.75) and 0.9980 in MOLM13 (FIRES) - the same pre-registered clause, opposite verdicts, and nothing in the clause names a cell line
Last cycle a prediction was written down in advance: if the histone mark stays above a stated level after the drug, then saying the protein complex is 'decoupled' understates what survives. In the first cell line the mark fell and the prediction did not trigger. In a second line the mark does not move at all and it triggers outright.
A filter-free promoter ranking with no peak caller and no target list recovers 40 of the top 50 published oncoprotein-target genes in the drug-free infant sample, and 0 of 50 in the drug-exposed one
Before believing anything else, we checked that our way of measuring 'where the fusion protein sits' actually finds the genes the original authors said it sits on. In the sample taken before the drug it found 40 of its top 50 on their list. In the sample taken during the drug it found none - which is itself the result.
A full 2x2 on the MLL1 channel in one run - DMSO, EPZ5676 1 uM, VTP-50469 100 nM and both - closes cycle 83's thirty-cycle-old threat: at the fusion's strongest MLL1 sites the menin inhibitor removes 58.3% against the DOT1L inhibitor's 17.5%, a 3.33-fold separation at TEN TIMES the DOT1L dose Q136 could test, with sequencing depth running opposite to the effect in every arm
A long-standing worry about this project's central idea was that a drug aimed at a different target might disturb the same machinery just as much, which would mean the intended drug is not doing anything special. Nobody had ever run both drugs side by side on the relevant measurement. A public dataset turns out to contain exactly that experiment. The intended drug removes more than three times as much, and a technical artefact that could have faked the result runs the wrong way in every arm, so it cannot be the explanation.
A genome-wide DAMAGE-STATE occupancy dataset for CtIP and MRE11 does exist and is public - GSE218590, 58 samples, +/- 4-OHT and a light-activated Cas9 time course - and it cannot answer the KMT2A question, because there is not one AsiSI site in the 90,377 bp of the gene
Last cycle this graph said nobody has ever measured repair proteins on damaged DNA anywhere, for any of these proteins. That is wrong - a 58-sample public dataset does exactly that - but the breaks in it are made at specific sequences, and none of those sequences occurs anywhere inside the KMT2A gene, so it still cannot answer our question.
A menin peak that keeps its call has still lost a sixth of its enrichment - and only where the fusion was
Looking only at the menin sites that were still detected after the fusion was destroyed, the amount of menin at them had fallen by about a sixth if the fusion had been there, and not at all if it had not. So even the sites that 'kept their peak' lost menin.
A null that costs one command, run first rather than afterwards - and SSRP1 is genuinely LESS required in KMT2A-rearranged lines, while the rest of the chaperone layer is not much of anything
We built a quick way to ask whether a group of genes is really needed more by this cancer than by other leukaemias, so the check can be run before writing rather than after. Run on the histone-packaging machinery, almost nothing stands out - and one protein stands out for being needed LESS.
A partial resolution: the leptin-LEPR chemoresistance model is murine MLL-AF9, but the two AML arms were never separated and the proteomics deposition is unreachable
An entry on marrow fat cells notes that a paper mixes two different mouse leukaemia models and never separates their results. Reading the full text confirms the leukaemia model is MLL-AF9 and that the arms stay mixed; the study's protein dataset could not be retrieved at all, so the separation the entry needs is still missing.
A per-sample parameter nobody reads: MACS's fragment-size estimate varies 822-1112 bp across ten ChIPs from one sequencing run, and within the DOT1L arm peak count correlates +0.7 with fragment size and -0.8 with library depth - so the obvious confound is not the operating one, and the non-obvious one is
When comparing these experiments people usually worry about how deeply each was sequenced. In this dataset that worry points the wrong way - the number of hits tracks a different, rarely-checked setting that varies between samples for reasons nobody reports.
A quarter of this graph's gene entries have a UniProt localization that nobody measured - and the subnuclear vocabulary UniProt does carry avoids the genes we care about most
Every gene entry here that names a protein can be checked against UniProt for where that protein lives. Doing that for all 299 of them shows that for about one in four the stated location was never measured in a human protein - it was copied from another species, inferred by a curator, predicted from the sequence, or left with no source at all. All four look identical if you just read the text.
A quarter of this graph's load-bearing evidence is under twelve months old, and the year field is right in a way a checker would call wrong
About a quarter of the evidence carrying the most weight here was published within the last year, and no check this project runs would flag that - a paper from last month passes every one. It is partly unavoidable, because the drugs this graph is about reached patients while it was being written.
A ratio-of-ratios combines four independent tracks and a pair ratio combines two, so its noise floor is wider by a factor of root two in log space - and cycles 101 and 102 scored two ratio-of-ratios against a pair floor. Applying the correction to the measured 9.4% peak floor gives 13.5%: VTP's 1.3% stays comfortably inside, and SGC0946's 22.8% stays outside but its margin falls from '2-6x outside' to 1.7x
Comparing two measurements doubles the random error; comparing two comparisons doubles it again. Two of this project's results compared two comparisons but judged them against a yardstick built from a single comparison, which made them look more solid than they are. Both conclusions still hold, but one of them by a much smaller margin than was reported.
A real defect in a committed scan, computed three days ago and never written up: `depmap-kmt2a-selective-scan-full` fixes both comparison arms from the FIRST gene that returns data, so five of sixteen eligible KMT2A-rearranged lines are excluded from all 17,728 gene comparisons - including the genes where they were screened
A genome-wide screen comparing leukaemia cell lines with and without the fusion picked which lines go in each group once, using the first gene it happened to look at, and then reused that grouping for every other gene. Five eligible lines were left out of every comparison, including ones where they had actually been tested.
A residue axis for menin, derived from coordinates: three spatial lobes, not four named ones, and the sharpest seam in the protein runs through the drug pocket
Everyone describes menin as a curved left hand with a thumb, a palm and fingers, but nobody writes down which amino acid each part starts and ends at, so no claim about them can be checked. Measuring the actual atomic coordinates of 69 menin structures gives numbers for the first time in this graph - and shows the protein really has three lumps, not four, with the cleanest division falling right inside the pocket that every menin drug binds.
A sharper test than complex enrichment: is the claimed partner's OBLIGATE partner also there? — and its fourth failure is the one protein the source paper named as its own non-selectivity control
When a protein turns up in a pull-down, the usual defence of it is that it belongs to a machine that genuinely binds the bait. A harder question is whether the one partner it cannot work without also turned up. Testing thirteen such pairs, four failed — and one of the four is the protein the original authors had already singled out as a false lead, which nobody told the test.
A surfaceome-first screen for KMT2A::AFF1: 2,404 surface proteins, 169 expressed in the fusion's own cell lines, 15 fusion-enriched, 3 clean in normal tissue - and the two sets do not intersect
We asked whether KMT2A::AFF1 leukaemia carries any protein on its outer surface that a CAR-T cell could attack without also attacking healthy organs, checking every surface protein in the human proteome rather than the handful already in use. The proteins that are specific to this leukaemia turned out to be present in normal blood vessels or brain, and the proteins absent from normal organs turned out not to be specific to this leukaemia.
A sweep for the failure that cost two cycles: of 234 analysis directories holding a non-empty results file, only 5 are cited by no entry at all - so the backlog is real but small, and one of the five contains a defect in a committed scan
After twice discovering finished calculations that nobody had written down, this checks all of them at once. Almost everything has been written up - five directories out of two hundred and thirty-four have not, and one of those five records a flaw in a screen the project still relies on.
A tempting second finding that does NOT survive its control: about half of each treated sample's peaks are absent from DMSO, but the two independent VTP-50469 libraries share only 10.5% of their gained sites - against 7.7%, 8.1% and 10.6% for the DOT1L-inhibitor arms - so the gained peaks are threshold noise and there is no evidence of redistribution
After treatment, roughly half the spots the experiment picked up were new ones. It would be exciting if the protein had moved somewhere specific - but the two independent experiments on the same drug barely agree about where the new spots are, which means they are just measurement noise.
A third of the papers we said we could not read were free to read in full
Many entries here say we only read a paper's summary, which sounds like the paper was locked away. For a third of them the whole paper was free the entire time and nobody checked.
A three-question triage to detect subtraction mechanisms, applied to every KMT2A partner in the graph
A short checklist that can be run on every fusion partner already in the database, to catch the kinds of hidden losses that the project's current way of looking would miss.
A tighter CXXC domain fails and a weaker one works: affinity rank does not predict function in the MLL-AF9 swap series
Four different CpG-gripping modules were put into the fusion in place of its own. The one that gripped DNA more tightly could not make leukaemia and the one that gripped more weakly could. How tightly the module holds DNA is not what decides whether the fusion works.
A73051 has the lowest oncoprotein share of the 25 untreated samples, and its sequencing depth is mid-range (37th-39th percentile), so the reading is not a depth artefact - the coordinator's cycle-135 qualification misidentified enrichment as coverage and was reversed in cycle 138
One patient's leukaemia looked interesting because the fusion protein was missing from the genes it usually sits on. It turns out to be missing from essentially everywhere in that sample - the profile looks like a patient on the drug that strips the fusion off chromatin, except this patient never had it.
A73051 is the one HOXA-negative non-AFF1 sample in GSE252378, and it is IRX1-negative too
One paediatric leukaemia in this dataset has almost no fusion protein on the HOXA genes - the same state that in t(4;11) goes with IRX1 being switched on. In this one, IRX1 is switched off as well.
Absence claims: 93% are unbounded in prose, but the evidence claims are 71% bounded - and the query that tested them was itself broken
This graph says 'nobody has tested this' about 1,878 times, and almost none of those sentences in ordinary prose come with the date, search and number that would let anyone check them - though the ones carrying a formal confidence rating mostly do, with the bound written in the caveat beside the claim rather than inside it. Worse, the literature search I wrote to test the nine riskiest claims was itself broken and returned 'zero papers' for every one, which looked like confirmation until the same database returned 535 papers for one half of that search and 7,161 for the other half.
Absence: the tissue with the clinical signal is the one tissue nobody has put under a menin inhibitor and measured - one record in the whole literature, and it is a review
A third of patients on the approved menin drug show a rise in parathyroid hormone. Nobody has ever taken parathyroid tissue, given it the drug, and measured what happens inside the cells. The only paper that mentions both the drug class and the gland is a review article.
Across 21 stopped trials of menin, DOT1L, BET, CDK9 and LSD1 inhibitors in acute leukaemia, 12 (57%) were stopped for sponsor or enrolment reasons, and the only two risk-benefit stops are one molecule - failure here has mostly been structural
When trials of these epigenetic leukaemia drugs have ended early, it was mostly because a company changed course or too few patients signed up, not because of a stated safety problem. For a disease as rare as infant leukaemia, running out of patients and sponsors is the failure to plan for.
Across all seven MLL-fusion baits, how deep a bait's fixed top-300 list runs into non-significance predicts how many "specific" partners it appears to have — Spearman +0.821, exact p = 0.034, and MLL-AF4 is the extreme on both axes
The published table took a fixed number of top-ranked proteins from each of seven pulldowns rather than everything above a quality bar. Baits whose fixed quota reached further down into weak, borderline hits ended up with more proteins that no other bait saw — and MLL-AF4's quota reached furthest of all. So "MLL-AF4 has the most unique partners" is mostly a statement about where its list stopped.
Across eight murine contrasts and two human stratifications, the H3K79me2 eviction span separates a menin inhibitor from a DOT1L inhibitor 22-fold with no overlap AND detects menin engagement in a human cell whose aggregate reads 0.998 - two properties the aggregate has one of and one of
Two different measurements of the same chemical tag behave differently. The simple average can tell the two drugs apart but in human cells it cannot tell the menin drug from no drug at all. The pattern across sites does both.
Across every same-condition H2AK119ub replicate pair in GSE260742 - nine pairs, three arms, both pairing shapes - rank concordance over the shared bin universe runs 0.6481 to 0.8796 and never reaches 0.88. The floors this thread scores against are all computed from pairs that are only moderately concordant.
Repeat measurements in this dataset rank the genome's regions only moderately alike - better than chance, well short of a close match. Every reliability yardstick this project uses is built from such pairs.
Across one infant's B-ALL to AML lineage switch, HOXA occupancy collapses ~28-fold and IRX1 falls only ~2.2-fold
One baby's leukaemia changed from a B-cell type to a myeloid type under treatment. The HOXA genes lost the driver protein almost entirely; IRX1 kept most of it.
Adjudicating the unwritten-gene queue by claim membership diverges from it at Spearman 0.36 - and then puts four acronym collisions in its own top five, so divergence was not improvement
A list of genes the project has mentioned but never written up is ranked by how often each is mentioned. Ranking it instead by whether anyone made a claim about the gene gave a different order - but its top entries turned out to be abbreviations that only look like gene names.
ADJUDICATION: the SvID and the gamma-H2A.X motif are OVERLAPPING BUT NOT THE SAME SURFACE - and the premise that made the question urgent, that the SvID is distinct from the MLL cavity, is FALSE at the SvID's own boundary
Two papers described two jobs for nearly the same stretch of menin, and neither knew about the other's result. Measuring the actual three-dimensional positions shows the two jobs live at opposite ends of that stretch, about 16 to 26 angstroms apart - but the stretch itself starts inside the pocket the drugs occupy, which is the opposite of what the graph had been assuming.
AFF1 and fusion occupancy transfer between infants at Spearman 0.82-0.92, but H3K4me3 transfers better at 0.93-0.97 - so the transferable part is mostly promoter structure, not fusion targeting
Where the fusion sits ranks similarly from one child's leukaemia to another's. But a general marker of switched-on genes ranks even more similarly, so most of that agreement is just the fact that active genes are active everywhere.
AFF1 is NOT peptide-poor - it yields 49 detectable tryptic peptides against menin's 28, and menin cleared the identity threshold while AFF1 scored zero - but the two groups do not separate on peptide yield at all, so the driver is abundance
Last cycle found that the fusion's own partner protein scores zero on a confidence test that other proteins pass, and concluded the test cannot be trusted here. The obvious objection, which that entry raised against itself, is that this particular protein might simply be hard to detect - it is unusually rich in two amino acids that can make protein fragments awkward to measure. Cutting all ten proteins up in software and counting the measurable fragments: it is not hard to detect. It yields more measurable fragments than menin, which passes the test easily. But the proteins that pass and the proteins that fail are indistinguishable on this measure, so fragment yield explains nothing either way.
AFF1 occupancy does not distinguish fusion-occupied promoters from wild-type KMT2A promoters in primary infant KMT2A::AFF1 - reproduced at promoter resolution, and the corpus should stop reading a partner ChIP as a fusion readout
Finding the AFF1 protein sitting on a gene in a patient's leukaemia cells does not tell you the cancer-causing fusion protein is there, because normal AFF1 sits on many of the same genes anyway. A different protein, ENL, does tell you.
AFF1-exclusivity describes 378 of 719 genes and 72% of them rest on one patient - and the MEIS1 comparison I used to call IRX1 indistinguishable was an artefact of scoring MEIS1 with the weaker instrument
Last cycle I said IRX1 could not be told apart from MEIS1, a gene nobody doubts. That was wrong: I had measured MEIS1 with this project's own weak method, which found it in 2 patients when the published method finds it in 6. But the bigger finding is that being 'unique to this fusion' is not special - more than half the genes in the published table have it, and most of those were seen in a single patient.
AFF1, the fusion's 3-prime partner, carries 126 genome-wide-significant in-gene SNPs across 505 traits in one contiguous 201 kb block (98.9th length-matched percentile), 631 of 849 rows lipid-related and 68 red-cell indices
The gene that supplies half of the leukaemia fusion protein also sits in one stretch of inherited DNA that influences hundreds of blood-fat and red-cell measurements. Most of that is probably one signal repeated, so it is a lead and not a finding.
AFF1's modern proximity neighbourhood is four RNA-binding proteins - NPM1, RBM15, SBDS, ZRANB2 - and it barely overlaps AFF4's
The first modern map of what AFF1 sits next to inside a living human cell finds four neighbours, all of them proteins that handle RNA. Its close relative AFF4 has four neighbours too, and only one of them is the same.
AFF1's own promoter window is occupied in 20 of 27 samples - top 4% of the whole table - and the call cannot be attributed to AFF1
The measurement everybody assumed had never been made was sitting in a table nobody read. Under this project's own rule, the window used for AFF1 - the gene the index fusion is named after - is called occupied in 20 of 27 patient samples, which puts it in the most consistently occupied 4% of all 19,404 windows. And 92% of that window belongs to two other genes, so the call cannot be assigned to AFF1.
AFF1's promoter window is 92.5% AFF1-AS1's - worse than IRX2's 89.2% - and the project's promoter table has no row for the neighbour at all
The 3' partner of this project's index fusion, AFF1, has another gene starting 151 letters away pointing the other direction. The window used to measure AFF1's on-switch is 92.5% that other gene's window - worse than the IRX2 case this beat was sent to investigate - and the other gene is not in the project's table at all, so nobody could have noticed.
AFF4 has never been placed in a haematopoietic cell state — every SEC cell-state result is in another tissue
AFF4 is the scaffold of the elongation complex this disease runs through, and nobody has shown what cell condition it maintains in blood. The results that do exist are in brain tumours, bladder cancer and skin.
Against a chromatin-machinery name list written before the intersection, only 8 of 151 symbols have an entry with no mark, no epigenetics field and no chromatin concept at all - and 7 of the 8 are the P-TEFb brake and the AFF paralogues
To find the real gaps without relying on sampling, we wrote out 151 proteins whose job is to write, read, erase or move histone marks - before looking at the data - and then asked which of their entries say nothing about chromatin in any wording. Eight do. Seven of those eight are the machinery that holds the transcription accelerator in reserve.
All 39 bounded absences in this graph, classified by WHAT BOUNDS THEM: 23 rest on an abstract-level literature search, 7 on a threshold or a query filter, and only 2 on reading the primary source
This project writes down things nobody has measured, which is useful only if the search that found nothing was capable of finding something. Sorting all thirty-nine of those write-ups by the tool that produced them shows that nearly six in ten rest on a search of titles and summaries only, and just two rest on somebody reading the actual papers.
All 46 comparator lines screened: the two we tripped over are the only two we can find
Last cycle we found that two of the 'healthy comparison' leukaemia cell lines used in every calculation here secretly run the same biology as the disease we are studying. The obvious next question was how many more there are. We checked all of them: the same two, and no others we can detect — though four lines could not be checked at all, and the method can only find what it is looking for.
All 46 patient phosphosites sit outside folded domains, none of LEDGF's twelve is in the clamp that holds KMT2A, and six of KMT2A's eleven are not on the fusion protein
We took every phosphorylation site measured in patients with this leukaemia and asked where on each protein it sits. Not one is inside a folded, structured part - they are all in the floppy stretches between the working parts.
All 56 core-complex partners that anyone deliberately fished for already have an entry: the unwritten interactome is 100% by-product
This project carried a figure saying only about a third of the fusion's known partners were written up, which sounded like a large research gap. Splitting them by whose experiment found them changes the picture completely: every partner somebody deliberately went looking for is already written, and everything still missing turned up incidentally in somebody else's experiment.
All four input samples explicitly titled to the MLLAF9 ChIP arm of GSE260742 carry supplementary_file = NONE, so no input-referenced peak call is possible from the processed deposit - and the only deposited batch-1 inputs are two BAMs filed under a different arm's title with the same cell line and the same F3 filename prefix
Every control sample that belongs to the main experiment has its description in the database but no data file attached, so nobody can do the standard background subtraction using the public processed files. The raw reads are there, so it is a compute job rather than a lost experiment.
All four murine VTP-50469 H3K79me2 contrasts in GSE260742 give Spearman -1.000 against eviction depth, and all four SGC0946 contrasts are flat: the sign reproduces four times out of four, the span only to within a factor of 1.99, and the replicate pair that establishes this was already binned on disk from cycle 82
A second copy of one of the experiments was already sitting on this machine from an earlier round of work. Running it gives the same direction of effect - four times out of four across all the available experiments - but the size of the effect varies by about a factor of two between copies.
All nine unwritten MEN1 partners trace to exactly two publications, neither of which mentions menin outside an antibody catalogue row
Nine famous cancer genes appear in a database as menin's binding partners, and every one of them comes from just two big screens that were never about menin. Seven come from a screen that deliberately hunted for interactions made by MUTANT proteins, and that paper never discusses menin in its text at all.
All three withdrawn mechanism-targeted trials admit infants and none of the 52 trials that do not has been withdrawn (exact p = 0.0032, noticed after looking), but the stated reasons are business and administrative and all three are 2025-2026 registrations
Three menin-drug trials were cancelled before enrolling anyone, and all three would have been open to babies. It is probably a coincidence of timing and company decisions, but it is worth watching.
aml_stjude_2024 - 887 paediatric AML cases with per-sample KMT2A fusion-partner structural-variant calls for all 180 KMT2A-rearranged cases, plus mRNA-seq and FAB morphology
A public dataset of 887 children with myeloid leukaemia. For every one of the 180 with a KMT2A fusion it records which partner gene was involved, plus gene activity and the leukaemia's cell type - which is what let us ask whether IRX1 tracks the partner or the cell type.
An activity-stratified permutation null for promoter-set overlap, and the check that its hypergeometric shortcut is not an instrument bug
To know whether two lists of genes overlap more than expected, you need to know what 'expected' means. Because busy genes are busy in every sample, the simple textbook expectation is far too low, so we built a fairer one and then proved our fast version of it gives the same answer as the slow honest version.
An independence null OVERSTATES partner-exclusive recurrence 16-fold here - it is not a lower bound, and the brief that dispatched this beat said it was
A standard shortcut is to pretend each gene switch behaves independently of the others, on the assumption that this under-counts how often a pattern appears by chance. In this dataset the shortcut over-counts by sixteen times, so using it would have made a real pattern look impressive when it is not.
An occupancy-matched floor built from ~8,600 mid-occupancy promoter windows per track closes a gap Q117 named in its own memo - and it was necessary, because GAPDH and ACTB GAIN share in two of three human lines while Q117 measured them LOSING share in mouse
Comparing treated and untreated samples needs a reference point that does not move, and everyone uses one or two housekeeping genes. Those genes move - upward here, downward in last cycle's mouse data - so this beat built a reference from about eight thousand ordinary promoters instead.
Analysis (planned) — comparative druggability scan of every obligate fusion-partner interface in the atlas
Before doing any laboratory work, we can rank every known fusion-partner contact point by how drug-friendly it looks and how safely the partner could be blocked. The point is to find the next menin on a computer, cheaply, rather than by luck.
Analysis (planned): is CLEC12A a selective dependency in KMT2A-rearranged lines in DepMap, lineage-split and reported as raw effect?
Before doing any bench work, look in a public database that has already deleted every gene in hundreds of cancer cell lines, and ask whether deleting CLEC12A hurts KMT2A leukaemia cells in particular. It costs nothing and could settle the question.
Analysis (runnable): build a methylome cell-of-origin score for infant KMT2A-rearranged ALL and test it as a predictor of lineage destination
Take the chemical marks on the DNA of these leukemias at diagnosis, work out which normal blood cell each one came from, and see whether that tells you where the cell goes when treatment releases it.
Analysis (runnable): build the first correlation matrix of menin-inhibitor resistance routes, from screens and from patient cohorts
Collect every published case and every screen hit where a leukemia escaped a menin inhibitor, record which escape route it used, and count how often each pair of routes occurs together. That table has never been made.
Analysis (runnable): can a menin ligand tell KMT2A from JUND? A structural and co-folding selectivity profile of the whole inhibitor class
A computer study comparing how menin grips the fusion protein and how it grips JUND, then asking whether any of the existing drugs, or any possible drug, could block one grip without breaking the other.
Analysis (runnable): do KMT2A partners rank by nuclear contact frequency in the same order as they rank in patients?
KMT2A fuses to some partners far more often than others, and the usual explanation is that those partners sit closest to it in the nucleus. That correlation has a published clinical ranking on one side and available contact data on the other, and nobody has plotted one against the other.
Analysis (runnable): does product viability leave a signature in the KMT2A breakpoint map once DNA fragility is controlled for?
A computer analysis that takes every published KMT2A break position, maps it onto the protein, and asks whether breaks avoid the folded parts of the protein even after you account for the parts of the DNA that are known to be fragile.
Analysis (runnable): does the infant methylome predict HOXA status once age and breakpoint are controlled?
Using data that already exists, we can test whether a leukemia's HOXA gene activity is better explained by the chemical state of the cell it started in or by exactly where the chromosome broke. It is a computation, not an experiment.
Analysis (runnable): fit the relapse hazard function in KMT2A-rearranged infant ALL and test whether the cohort is a mixture
Nobody has drawn the actual risk-over-time curve for these children, only the survival curve. This analysis draws it, and tests whether the shape means there are two kinds of patient rather than one.
Analysis (runnable): is anything actually enriched at KMT2A breakpoints? The test has never been run against a background that could say no
People have said for thirty years that the place KMT2A breaks is full of repeated DNA and other special features. Nobody has ever compared it to a fair stretch of DNA chosen the same way, so we do not know whether those features are enriched or simply present.
Analysis (runnable): is there a shared cell-state signature behind non-genetic menin-inhibitor and BET-inhibitor resistance?
Two different drug classes fail in this leukemia for reasons that are not mutations. We can check, using published data, whether the surviving cells look the same in both cases — and if they do, that shared state becomes the real target.
Analysis (runnable): mine the KMT2A junction sequences for repair-pathway signature, and count the one junction class with a genetic control behind it
Every KMT2A fusion leaves a scar where the two chromosomes were glued, and the shape of that scar says which repair crew did the gluing. Thousands of these scars have been sequenced for diagnostic purposes and nobody has ever tallied their shapes.
Analysis (runnable): value a menin inhibitor as a short-duration instrument — cohort meta-regression plus a discrete-event model of remission months
A modelling study that takes the published trial numbers and asks what really drives survival: how many patients get into remission, or how many of them get to a transplant or cell therapy while they are there.
Analysis: cross-validating the single-group non-menin KMT2A-fusion dependencies in one framework
Several labs have each reported a different molecule that KMT2A-fusion leukemia cannot live without. Nobody has checked them against each other or against the big public screening data, so we do not know which ones are real.
Analysis: does the fusion's target set sit at lamina-associated domain borders? A free overlay of published LAD maps on published fusion occupancy
LAD borders are marked by CG islands — the same feature the fusion reads. This checks, using only data already published, whether the fusion's landing sites sit at the edges of the silenced blocks that touch the nuclear envelope, and whether the HOXA gene cluster is one of them.
Analysis: is the KMT2A genomic breakpoint a prognostic variable? A reanalysis of data Interfant and Frankfurt already hold
Two groups already hold everything needed to answer whether the exact position of the DNA break changes how a child does. One holds the sequences, the other holds the outcomes, and nobody has put them together.
Analysis: score fetal, myeloid-priming and differentiation-stage signatures across public KMT2A-r datasets
Before running any new experiment, take the leukaemia gene-activity data that is already public, and score every case for how fetal it looks, how myeloid-primed it looks, and how immature it looks. Then check whether those scores track with which patients do badly.
Analysis: separate loss-of-protein from loss-of-occupancy by comparing degron-depletion and menin-inhibition transcriptional time courses
Two ways of shutting the fusion down — destroying it with a degron, or knocking it off DNA with a drug — should give different shaped response curves if the drug also destroys it. This analysis compares the shapes of published and in-house time courses to infer whether menin inhibition is partly a degrader, without needing a new experiment first.
Analysis: test whether published KMT2A::AFF1 occupancy is enriched at scaffold-attachment and cruciform-prone DNA beyond what CpG islands explain
Take the existing maps of where the fusion sits on the genome and ask whether it prefers stretches of DNA that are structurally strained — the kind that forms cross-shaped structures — over and above its known preference for unmethylated CG islands. No new experiment, only existing public data.
Answering a question three cycles filed and none ran: H3K79me2 also has two channels in GSE260742 - rep1 and rep2, 199 numeric leaves each - so cycle 104 quoted one of two available numbers without saying so. Unlike H2AK119ub they never materially disagree, 0 of 18 comparable ratios, and H3K27me3 and H3K4me3 were never scored under any second channel at all
One chemical mark in a public dataset was known to exist in two differently processed versions that sometimes disagree. Nobody had checked whether the other three marks in the same dataset have the same problem. One of them does have two versions - but they are two repeats of the experiment rather than two processing recipes, and they never disagree by enough to matter. The remaining two marks were only ever computed one way, so the check cannot say whether a second version exists, only that none was used.
Applied symmetrically, the criteria that were never applied make IRX1 the weakest member of the class it won - and MEIS1, a canonical fusion target, is indistinguishable from it, so the deposit cannot separate the two
Over several cycles a six-gene list was cut to one - IRX1 - by removing members one at a time. Each removal used a different test. Running the tests that were never used shows IRX1 would have been the first to go on one of them. But a gene nobody doubts scores exactly the same, so the dataset is simply too small to tell them apart.
Are the menin partners outliers, or did we pick them? (run)
Every co-dependency result here so far tested only genes we already believed were partners, which guarantees that partners come out on top. We repeated the test against 377 randomly chosen genes to find out what an unremarkable result looks like, and the four strongest partners beat every single random gene.
ARID1B is not dispensable - it is paralogue-buffered. In ARID1A-mutant lines its CRISPR dependency goes from -0.0724 to -0.3709 (p=0.0002) while ARID1A's own falls away (+0.1925, p=0.044) and ARID2/SMARCA4 stay flat. Cycle 80's 'dispensable everywhere' framing was wrong and the caveat that entry carried was right
Last cycle I reported that switching off the gene ARID1B barely affects any leukaemia cell line, and I flagged one alternative explanation I could not rule out: ARID1B has a near-identical partner gene, ARID1A, that may simply cover for it, so a screen that switches off only one would see nothing either way. This is that test, and the alternative explanation wins: in cell lines whose ARID1A is already broken by mutation, switching off ARID1B becomes about five times more damaging, while switching off ARID1A itself stops mattering in those same lines - which is the internal check that the mutation records mean what they say. Two related genes show no such shift, so the effect is specific to the pair that substitute for each other; the honest correction is that ARID1B is not unnecessary, it is covered for.
ARID1B's buffering effect survives leave-one-out in 11 of 11 - even dropping the line whose own effect is -1.2269 leaves -0.2129 at p = 0.00465. But the lineage-concentration test says there was LITTLE confound to remove, cycle 84's headline confirms an absence rather than defeating a threat, and my own pre-registered TVD threshold fired on a measure that is large by construction at n=11
Two checks I had promised and not done, both capable of undermining my own earlier conclusion: dropping each of the eleven relevant cell lines in turn leaves the finding intact every time, including when the most extreme line is removed, so it is not one cell line wearing a statistic. But the second check shows the earlier tissue-of-origin correction was not correcting much - the eleven lines are no more concentrated in particular tissues than a random eleven would be - and one of my own pre-set thresholds turned out to be automatically satisfied at this sample size, so it decided the answer while carrying no information.
ASH2L and RBBP5 rank 17,248 and 14,622 of 17,728 for KMT2A-selectivity, which is what this graph's own SET-domain-is-dispensable thread predicts - the audit found agreement where it looked for contradiction
This project has argued from mouse genetics that the enzyme part of KMT2A does not matter for the leukaemia, and that the protein complex whose job is to switch that enzyme on is therefore beside the point. A large independent gene-deletion screen agrees: the two complex members turn out to be among the least specifically required genes measured.
ASH2L is studied in leukaemia and has never been placed in a cell state there — and the KMT2A-rearranged arm needs it LESS than average
ASH2L is part of the machine at the centre of this disease, and there are leukaemia papers about it, but none asks what condition it holds a cell in. The big cancer screen also finds that leukaemias with this fusion need ASH2L slightly less than other cancers do.
Asked for the first time whether a KMT2A fusion GAINS chromatin sites under a menin inhibitor: all three pre-registered criteria fire, and the set they fire on is 52 low-amplitude bins reaching 6x the genome median, 16 of them with no gene within 20 kb and 1 of 52 on a CpG island where 66% of the fusion's real sites are
We looked for places on the DNA where the leukaemia fusion protein shows up only after the drug is given, and we wrote down the rules for what would count before looking. A small scattered set passed the rules, but one of its fifty-two sites is a CpG island where two thirds of the fusion's real sites are, so the drug knocks the fusion off without moving it anywhere it would normally bind.
Asking IntAct for its own fusion constructs found a sixth, exposed a duplicate record that would have inflated the count by three, and failed its own control 4 of 5 - so the floor is lower, not closed
A list of laboratory-made fusion proteins in a public database had been built up by hand, and nobody knew if it was complete. Asking the database to list them instead found one more, caught a duplicate entry that would have made the data look bigger than it is, and then failed a test using the items already known - so the list is still not proven complete.
At gene resolution the menin inhibitor is the stronger single agent, and only the combination approaches the degron
Comparing the two drugs head to head on how much fusion protein they knock off its target genes: the menin inhibitor beats the DOT1L inhibitor, and only using both together comes close to simply destroying the fusion outright.
At gene-body scale the AFF1-exclusive recurrent class is 6 against an exact institution-stratified null of 0.86 - but the singleton class is 63 against 74, so recurrence carries the signal and exclusivity alone does not
Six genes are hit by the fusion protein in three or more of the eight AFF1 leukaemias and in none of the other eighteen patients. Shuffling which patients count as the AFF1 group gives about one such gene, so six is about seven times more than chance. But genes hit in only ONE AFF1 patient and nobody else number sixty-three, and shuffling gives seventy-four - so that weaker pattern is exactly what chance produces.
At promoter resolution the published N/C oncoprotein criterion calls 1043 promoters in fusion-negative CD34+ HSPCs - above the median across 27 KMT2A-rearranged patients - while the same-reagent controls in that sample call 66 and 11
Healthy blood stem cells have no fusion protein at all, yet the published test says hundreds of their gene promoters are fusion-bound when the test is run on promoters instead of on the regions it was designed for. The cutoff does not travel between region sizes.
At the 2 kb bins a menin inhibitor strips the most fusion from, both repressive marks GAIN and the active mark is LOST, monotonically in eviction depth (Spearman +1.000, +1.000 and -1.000) against a baseline-matched null that is flat to within 0.007 - the silencing clause of hyp-q117's prediction 5, measured for the first time
When a drug pulls the cancer fusion protein off the DNA, the spots it leaves behind get chemically switched off - and the more of the fusion that leaves a spot, the more strongly that spot is switched off. Two different 'off' marks go up together and the 'on' mark goes down, in exactly the same places, which is why this is not a measurement artefact.
At the 692 RS4;11 bins where an MLL1 N-terminus and menin co-occupy, VTP-50469 takes menin to 0.20 and DOT1L to 0.20, drops H3K79me2 to 0.66, and leaves MLL1-N-reactive material at 0.83 - four factors, one drug, one chromatin, and a clean ordering
Measured side by side at the same places in the same cells, the drug removes menin and DOT1L almost completely, cuts the histone mark they produce by a third, and leaves the MLL1 protein itself essentially where it was. Whatever stays behind has lost its partners and most of its output.
At the MLL1-N HELD SET, floor-normalised, the MLL1-N channel is inert to three decimals in both human lines (1.036 and 1.0359) - AND THAT IS A STATEMENT ABOUT THAT POPULATION, NOT ABOUT THE CHANNEL: at MLL1's own top sites in mouse the same channel reads 0.3012 under a DOT1L inhibitor and 0.3298 under a menin inhibitor
The antibody most papers use to ask whether the fusion protein is still on DNA gives the same non-answer in two different leukaemia cell lines, to three decimal places. When a measurement is that reproducible and that flat, the thing being reproduced is usually the tool.
At the same bins in the same cells, menin's share ratio is 0.20-0.26 and the MLL1 N-terminus antibody's is 0.80-0.91 - so a shared-epitope MLL1-N ChIP under-reports the menin inhibitor's occupancy effect by 3.5 to 4.0-fold, and the number is now measured rather than asserted
The antibody most papers use to ask whether the fusion protein is still on the DNA also detects the normal, non-cancerous version of the same protein. Measured side by side at the same places in the same cells, that antibody shows a four times smaller drug effect than a menin antibody does, so it hides most of what the drug is doing.
ATF4's own transcript does NOT move under either menin inhibitor - and the whole ISR arm is reproduced by a DOT1L inhibitor, so it is not a menin-interface readout
This graph concluded that menin-blocking drugs switch off a stress programme by watching the genes the programme controls, never the master switch itself. Measuring the switch directly shows it barely moves, and - more awkwardly - a completely different drug reproduces the whole effect almost exactly.
Atlas confidence class tracks the prey:bait ratio, not abundance - so the proposed copy-number triage of the atlas-only partners does not exist
When a large survey reports that two proteins stick together, one worry is that it only found the protein because there is a lot of it lying around. We tested that on 31,944 measurements and it is not what is happening: what marks a confident result is the ratio of the two proteins pulled out together, not how abundant the partner is.
Attribution audit 1 of 5 - who co-purifies with KMT2A::AFF1: the inventory does not survive, the individual partner blots mostly do
Lists of the proteins that stick to the KMT2A::AFF1 fusion were made with antibodies that also stick to the ordinary versions of those same proteins, so a mass-spectrometry list cannot say which proteins were on the fusion and which were on the normal copy. The one thing that does work is a Western blot, because there the fusion is a different size and can be seen on its own.
Attribution audit 2 of 5 - post-translational modifications on KMT2A::AFF1: nothing survives at peptide level, and the one strong PTM result is explicitly about the wild-type protein
Every chemical tag that gets added to a protein is identified by the short stretch of sequence it sits on, and every such stretch in the KMT2A::AFF1 fusion - except the one that spans the join - is letter-for-letter identical to a stretch in one of the two normal proteins. So a modification found by mass spectrometry cannot be said to be on the fusion rather than on the normal copy.
Attribution audit 3 of 5 - where KMT2A::AFF1 sits in the nucleus: the endogenous image does not survive, the overexpression images do and say something different
Pictures of where the fusion protein sits inside the nucleus were taken with antibodies that also light up the normal protein, and in a microscope image there is no way to tell the two apart. The pictures that can be trusted were taken in cells given an extra, artificially high copy of the fusion - which is a different question.
Attribution audit 4 of 5 - the half-life and turnover of KMT2A::AFF1: this class SURVIVES, and it is the only endogenous fusion-attributable protein measurement this beat found
How long the fusion protein lasts inside a cell was measured in real patient-derived leukaemia cells, and it counts, because the fusion is a different size from the normal protein and both were visible in the same lane of the same gel. It is the one thing in this whole audit that was measured on the real fusion in a real leukaemia cell.
Attribution audit 5 of 5 - where KMT2A::AFF1 sits on chromatin: the field's standard attribution heuristic is invalidated by the paper that founded it, but two real survivors exist
Maps of where the fusion protein binds DNA were made with antibodies that also bind the normal protein, so the field works out which peaks belong to the fusion by clever subtraction rather than by seeing it directly. One of the two subtraction tricks is sound and the other is undermined by the very paper that established the biology behind it.
Attribution is asymmetric: q151's IRX2 negative survives a window that is 89.2% someone else's gene, and a positive in the same window would not have
Finding nothing in a stretch of DNA that contains IRX2's on-switch really does mean nothing was at IRX2's on-switch, no matter which other genes share that stretch. Finding something there would not have told you which gene it belonged to. So the original negative result stands, and the near-miss that came with it does not.
Audit of the infant presenting-feature evidence: what exists, what does not, what was rejected, what could not be read
This is the working record behind the rebuilt infant clinical picture: every search that was run with its record count, every paper that looked relevant and was not used with the reason, and every paper that exists and could not be read. It is written so that someone repeating this work next year repeats the decisions, not just the searches.
Audit, CORRECTED: the PHD add-back result was replicated in 2008 by a second group in a different fusion, and this graph already cited it
The graph's most exciting claim about what KMT2A fusions throw away was recorded as resting on a single 2008 experiment that nobody repeated. That is wrong. A second group did the same thing the same year in a different fusion and got the same answer — and this graph was already citing their paper in four other entries.
Auditing all 2,476 committed entries for single-convention H2AK119ub: 122 mention the mark, 10 carry both conventions, 2 name the one they use, and 2 quoted a paired-convention number without saying so. Both are amended - and the more useful result is that the two conventions disagree PER STATISTIC rather than globally, so declaring the mark unusable would have been wrong
One chemical mark in a public dataset was deposited under two different processing versions, and those two versions have twice given opposite answers to the same question. That raised the worry that pages in this project might be resting on whichever version happened to be used. Checking all two and a half thousand pages found two that quoted one version without saying which - both now fixed - and, more importantly, that the two versions actually agree on some measurements and disagree on others, so the mark is usable as long as every number says which version it came from.
Auditing the graph against the PI's brief, not against its own ledger (run)
For thirty cycles this project has measured itself against a list of research beats it wrote for itself, and nobody had checked that list against what the PI actually asked for. Most of the brief is covered well, but one thing the PI asked for by name — real patients and their stories — has no entries at all.
Auditing the seven strikes before applying them: nine of ten decisions stand, four for different reasons than were given, and one clinical argument fails
Another team argued that seven candidate drug targets should be dropped. Before dropping them we went back and read every paper they cited. Almost all the decisions were right, but several were right for the wrong reason, and one of the arguments does not survive reading the trial it rests on.
BARD1 and EXO1 are the only two proteins in the resection layer with ZERO chromatin-occupancy datasets anywhere - ENCODE, ChIP-Atlas, ReMap and a hand-checked GEO title list all return nothing - and BARD1's silence is the loud one, because its obligate partner BRCA1 has thirty-one
Of the twelve DNA-repair proteins this graph tracks around the KMT2A gene, ten have been mapped onto chromatin somewhere in a public dataset. Two have never been mapped at all, and one of them is the obligate partner of a protein that has been mapped thirty-one times, which points at a missing reagent rather than a missing idea.
Before quoting anything, this beat re-derived six figures that two earlier beats had stored - three human, three mouse - from the deposited files end to end, and all six came back identical, which is the control that licenses every cross-beat comparison in the entry above it
Numbers in this project are often quoted from earlier work rather than recalculated. Before comparing anything, this beat recalculated six of them from the original public files, and every one matched.
Between-patient variation in oncoprotein occupancy is not a fusion-partner effect: across six partners the within-partner range of the amplitude statistic is as wide as the between-partner range of its medians
Which partner gene the fusion has does not explain how much chromatin it occupies. Two children with the same partner differ about as much as two children with different partners.
Boltz co-folding plan: six complexes worth folding, three that must not be, and the positive controls that decide whether to believe any of it
The lab has credits for Boltz, a tool that predicts how two molecules fit together and how tightly they stick. We list exactly which pairs are worth predicting, which would produce confident-looking nonsense, and the known-answer cases we should run first so we can tell the difference.
Both datasets this beat used carry an exposure mismatch that could have produced its results, and both were measured rather than argued away: MOLM13's MENIN pair is 4 days against everything else's 3, and GSE260742's VTP arms are 48 h against 72 h DMSO controls - the first moves a criterion by 0.024 and the second is matched between the two arms being compared
In both datasets some samples were treated for longer than others, which could easily have created the differences we found. Checking each one shows neither explains the result.
Bounded absence: every chromatin measurement under a menin inhibitor is in an adult-derived line, and none is in an infant KMT2A::AFF1 model
The drug is being given to babies with a specific leukaemia. Every experiment that maps what the drug does to DNA-bound proteins was done in cells taken from adults, and mostly from a different kind of leukaemia. The standard infant cell line has never been used for this.
Bounded absence: in GSE159608's SEM line, VTP-50469 was profiled for H3K4me3 only - no menin, no H3K27ac, no H3K79me2, no RNAP2 and no IgG under the drug - and no library in the series carries a spike-in normalisation, so an absolute change in any mark is unmeasurable here
The dataset that contains the menin inhibitor experiment in infant-leukaemia cells only measured one chemical mark under the drug, and measured nothing that would let anyone tell whether the mark went down everywhere or just moved around. Both limits are worth writing down, because they bound what anybody can conclude from this dataset - including us.
Bounded absence: no menin-inhibitor trial reports response rate by 3-prime fusion partner in the primary literature
We searched for any published menin-inhibitor trial that reports how well patients responded broken down by which partner gene they had. There are none. The one trial that says it looked put the numbers in a supplement that is not publicly retrievable.
Bounded absence: no published study has tested whether a KMT2A fusion GAINS chromatin sites under a menin inhibitor - and the same pipelines found gains for two other proteins when they looked
If the drug pushed the fusion protein onto new places on the DNA, we would only know if somebody had looked for new places, and nobody has. Every study asked what was lost - and the two that did ask about gains, for other proteins, both found them.
Bounded failure with a near-miss worth recording: a GEO series matching the topic exactly belongs to a different paper
A paper on DOT1L being sugar-modified never names the MLL fusions it used, and it deposited nothing readable. A search of the data repository turned up exactly one dataset on precisely that topic - and it belongs to a different research group and a different paper. Attributing it would have produced a confident wrong answer.
Bounded failure: 'menin inhibitor-resistant cells' is a drug entry's central claim and neither the paper nor any repository defines it
A drug entry in this graph says a DOT1L degrader works in cells resistant to menin inhibitors. Nobody has said what made those cells resistant, and the paper is unreadable with no deposited data - so the graph cannot tell whether the finding would apply to a real resistant patient.
Bounded failure: a PMCID that exists and a body that does not - the MLL amino-terminus transformation paper is unreadable by both PMC routes
A paper this graph cites for menin binding through the front end of MLL does not say which fusion partners it used. It has a PMC record, so it looks readable, but both ways of asking PMC for the text return only the title page - and it deposited no data.
Bounded failure: the CDK4/6 cell lines behind two therapeutic entries are unrecoverable although the article is marked free access
Two entries recommending CDK4/6 inhibitors in KMT2A-rearranged leukaemia rest on a paper that never names the cell lines it tested them in. The article is tagged as free to read, but both ways of fetching its text return only the title page, and it deposited nothing.
Bounded failure: the founding menin-cofactor paper names no fusion, has no PMC record and deposited nothing - and three entries depend on it
The 2005 paper that established menin as an essential partner for MLL leukaemia never says which fusion proteins it used. It is behind a paywall with no free version anywhere, and it deposited no data, so there is nothing left to ask. Three entries in this graph carry that unknown.
Bounded failure: the one unresolved source that is actually about the index fusion
A 2026 paper on how the protein IGF2BP3 affects the amount of t(4;11) fusion message and protein is the only unreadable source in this beat that studies this project's own fusion. Its model systems are described only as model systems, it deposited nothing, and it cannot be read.
Bounded unreachable: the Blood supplemental table that three entries read their NLRC5 and CIITA group assignments from cannot be retrieved by ANY programmatic route - PMC serves a landing page and the publisher 403s this client - so those claims are re-readable but not re-verifiable without a human's institutional access
Three entries in this graph rest on numbers read out of a spreadsheet attached to a published paper. This entry exists so a future cycle does not waste time trying to download that spreadsheet: every route tried returns either a missing-file error or a web page pretending to be the file.
Bounded unreadable: the only affinity purification of a SET-retaining KMT2A-derived fusion complex fails all four routes, and the molecule is the reciprocal product
One 2011 paper is the only source for the claim that the mirror-image fusion protein sits in a sixteen-piece machine, and it cannot be read by any free route - only its summary is public. Its summary also shows the protein in question is the mirror-image product, not the main fusion this graph is about.
Bounded with every archive and every route named: after both public SUV39H1 deposits were read this cycle, there is still NO SUV39H1 occupancy measurement in any haematopoietic cell, any KMT2A-rearranged cell, or any damage state - in ENCODE (three routes, zero), ChIP-Atlas (one deposit, MCF-7 only) or GEO (two deposits, glioma and breast)
Now that both public datasets have been read, the gap is easy to state: nobody has ever measured where this enzyme sits in a blood cell. Every measurement that exists comes from a brain tumour or a breast tumour, and the disease this project is about is a leukaemia.
Bounded, with every archive named: H3K9me1 and H3K9me2 in a CD34+ progenitor are absent from BLUEPRINT, EGA, Roadmap/Human Epigenome Atlas, the IHEC Data Portal and ENCODE - 49,663 IHEC datasets carry 178 assay labels and H3K9me2 is not one of them - and the absence is a consortium artefact, because GEO has had the CD34+ H3K9me2 data since 2017
The big public epigenome projects only ever measured six standard chemical tags on DNA-packaging proteins, and the two tags this project cares about are not among them. So none of the big archives has them for blood stem cells. A single 2017 paper in a different archive does have one of them, which shows the gap is about what the consortia chose to measure, not about what is knowable.
Bounding an absence the graph already asserts: the H3K79me2 decay half-life, counted, dated and controlled - and why three of the six hits are about the drug's half-life
Three entries in this graph say nobody has measured how fast the H3K79 tag fades once the enzyme that writes it is blocked, and none of them ran a search to check. A dated, controlled search confirms it - with a trap worth carrying away: in a DOT1L literature the word 'half-life' almost always means the drug's, not the tag's.
BRD9 survives the null; CEBPA sits at the median of random genes
Last cycle we found that some genes only look important to this leukaemia because of which cell types were tested. This puts the survivors through a proper statistical check against seven hundred random genes. The menin group comes out as the most extreme in the genome, one new gene holds up, and the two rejected ones score no better than chance.
Building the instrument that asks whether this corpus has already measured something: it fails, 85% of comparable pairs flag as disagreements and the one real conflict never surfaces - because measurements live in prose, so the fix is a typed reporting field and it is now in the standing brief rather than in an open question
This project has three tools for checking whether the outside world has already done an experiment, and none for checking whether the project itself has already measured the same thing — which has now caused two beats to contradict each other without noticing. Building that missing tool did not work: the numbers are written into sentences rather than into fields, so nothing can reliably tell which number describes what. The remedy is to write them down properly from now on, which has been added to the instructions.
Canela 2017 deposited NO cohesin-depletion experiment - and it did deposit human END-seq in a B-cell-precursor leukaemia line, which nobody here had looked for
This project believed the 2017 paper had removed a cohesin ring subunit and mapped the breaks. It did not - that experiment is in the 2019 paper. What the 2017 paper does have, and nobody here had noticed, is a break map in a human B-cell leukaemia line, which is much closer to the disease this lab studies.
Canela 2019's full text never names MLL or KMT2A, and its HCT116 cohesin arm is lovastatin-arrested in G1 - so the only directional cohesin experiment was run in a cell where transcription-coupled fragility should be lowest
The one experiment that removes the cohesin ring and then counts DNA breaks was done in cells deliberately frozen before they divide, using a drug that stops them growing. That is the state in which this stretch of DNA is least likely to break in the first place.
Canela 2019's RAD21-degron END-seq IS deposited and IS readable today - one sample per arm, no replicates, no minus-auxin control, hg19, bigWig coverage only
The one experiment that removes the cohesin ring from human cells and then maps DNA breaks genome-wide is public, and its processed files can be read over the internet without downloading anything big. But the deposit contains a single sample for each condition, with no repeat and no matching untreated control for the depleted cells.
CD19 looks lower in KMT2A-rearranged patients in the RNA-seq arm of TARGET and identical in the microarray arm of the same study - so this graph should not record a CD19 deficit
There is a widespread belief that KMT2A-rearranged leukaemia carries less CD19 than other B-cell leukaemias, which matters because CD19 is what most immune therapy targets. Two measurements of the same patient group point in opposite directions, so this project records the question as open rather than picking the answer it expected.
CDK9 peak widths more than double after fusion degradation, firing the pre-registered re-score trigger
For one of the measured proteins, CDK9, the detected regions became two and a half times wider after treatment - a known sign that the signal has collapsed and the detection software is guessing. Any number computed from CDK9's regions has to be reported with that warning attached.
Cell identities considered and not written: what the search actually returned for basophils, mast cells, eosinophils, NK cells and the rest
Before writing entries for the cell types this disease talks about, every plausible candidate was searched in PubMed and the results counted. This records the ones that were deliberately not written and why, so that nobody spends a day rediscovering that there are zero papers connecting this leukaemia to mast cells.
Cell type or the fusion? MV4-11 LOOSENS the confound the graph's own lineage split was built on - but its founding paper calls it biphenotypic
To ask whether a gene matters because of the leukaemia's cell type or because of its fusion gene, you need the same fusion growing in two different kinds of cell. One cell line, MV4-11, was meant to be that, and it was missing from this project's earlier comparison. Adding it suggests menin and DOT1L are needed in both cell types equally - but the paper that first described MV4-11 calls it a mixed B-and-myeloid line, so the two cell types are not as separate as the design needs.
Cell-state enumeration: 23 queries, what was written, what was rejected and why — including an empirical confirmation that PubMed reads SEM as two journal names
This is the working record of how the cell-state entries were found: every search run, how many records each returned, which papers were read and set aside, and why. It exists so that someone repeating this work next year repeats the decisions as well as the searches.
Census of every published chromatin-occupancy measurement made under a menin inhibitor: eight datasets, every one TSS-anchored, every one tabulating losses only
We listed every experiment anyone has published that maps where proteins sit on DNA after a menin inhibitor is given. There are only about eight, they all measure signal in a window around the start of genes, and every one of them reports a table of what was lost. None reports a table of what was gained.
check-pi-dimensions.py selected fusion entries by id prefix and swept in an analogy node, reporting 50/51 where the truth is 49/50 - identical to one decimal, which is why it survived - and the one genuinely unlinked fusion has no analogue in this corpus, so no edge was written
A script that measures how complete the knowledge graph is was counting fusion records by looking at whether their name started with "fusion-", and one record that starts that way is not a fusion at all - it is a note about a family resemblance between fusions - so it was being counted as both a fusion and a completed one, inflating the top and bottom of the fraction equally. The reported score and the true score were both 98.0% after rounding, so nothing looked wrong until a different tool counted 50 fusions where this one counted 51; the fix changes no conclusion, and the single fusion that genuinely lacks a link still lacks one because there is nothing in this graph it is honestly analogous to.
Chromatin accessibility is named in the PI's own dimension and the detector has no token for it - nor for enhancer, nucleosome, Polycomb or remodelling - so 121 of the 195 entries scored as having nothing do name a chromatin concept
The tool looks for the names of specific chemical tags on histones, like H3K4me3. It has no word for 'enhancer', 'nucleosome' or 'open chromatin' - even though the project's own definition of this topic includes chromatin accessibility. So entries that describe a protein's chromatin biology without naming a tag score as empty.
Co-dependency scan: what else does a menin-dependent cell need? (run, not planned)
If a cancer cell cannot survive without menin, what else can it not survive without? We compared how badly 36 genes are needed across about 1,200 cancer cell lines and asked which ones track menin. The answer put the two proteins menin physically holds on to at the top, which is a good sign the method works.
Comparing every citation's stored author list against the paper its DOI resolves to - 15,219 comparable blocks, 99.6% agreeing, and all 60 disagreements are placeholder strings, one Crossref metadata error and one Scandinavian letter - which closes the year check's known blind spot and finds no wrong-paper citation on either axis
The dangerous citation error here is a working link to the wrong paper. A check built two cycles ago compared publication years and caught about seven in ten of the mistakes this project has recorded — but it was blind to the commonest slip, a mistyped digit that lands on a neighbouring paper in the same journal and year. This compares the authors instead. Across fifteen thousand citations it found none pointing at the wrong paper; every disagreement was a formatting difference or a placeholder.
Comparing every citation's stored year and journal against the record its DOI actually points at - 16,420 comparable blocks, 92.0% exact, and all 22 that disagree by two years or more resolve on reading to six DOIs with online-versus-print date conventions, so the wrong-paper failure mode is bounded from above for the first time
The dangerous kind of citation error in this project is not a broken link — it is a working link to the wrong paper, which has happened twice and both times a human caught it by reading. A previous attempt to catch that automatically failed completely. This one compares the year and journal each note records against the year and journal of the paper its link actually opens. Sixteen thousand notes agree; the twenty-two that did not turned out to be papers published online in one year and in print in another.
Confidence in this graph carries no information about whether anyone read the paper
We count a paper as read if any entry cites it, which treats a passing mention like a careful reading. Nearly a third of our citations say in their own notes that only the abstract was seen, and our confidence labels do not track that at all.
CORRECTED SAME CYCLE: the replication test was UNINFORMATIVE, not negative. Eleven of the twelve proximity panels do not sample the nucleus - they are centrosome, RTK, kinase, ion-channel, CFTR and mitochondrial baits - so KMT2A's 0th percentile in three of them measures panel composition, not promiscuity. In the one transcription-factor panel KMT2A is at the 71.7th and menin is absent, which is DIRECTIONALLY CONSISTENT with the K562 result at n=1
Two earlier cycles argued that one protein is picked up by proximity experiments far less than expected, and this set out to check whether that holds in every other such experiment available. The first answer was that it does not hold - but looking at what those experiments were actually about showed that eleven of the twelve were mapping things outside the cell nucleus, where neither protein belongs, so they cannot speak to the question at all. The one experiment that did look at nuclear proteins points the same way as the original, which means the question is still open rather than settled either way.
CORRECTED: the AF4-specific interactome gap is ZERO, not one - RPA2 was declined on purpose in cycle 159 and my no-match read was the tool warning I ignored
A list of 110 proteins that looked like they stick to this cancer protein and only it has been treated as the biggest hole in the project. Filtering out the ones the data cannot support leaves six, and five of those were already dealt with. One protein is left.
CORRECTED: two of eleven CXXC readers HAVE been measured alongside a KMT2A fusion, and the result supports the menin-interface hypothesis
CORRECTED 2026-09-21: this entry originally said nobody had measured two of these DNA-reading proteins in the same cells as the fusion, and a 2017 paper we already cite 29 times had done exactly that in human t(4;11) cells. Its answer is the interesting part - the fusion sits on only a fifth of the available addresses while the other readers sit on half and nearly all, so the fusion is being placed somewhere specific rather than competing for space.
Corrections owed to other beats: nine entries whose stated provenance is now resolvable, and one unflagged import
While auditing, we found places where other entries say a fact is unknown that we were able to look up, and one place where an entry does not say which leukaemia its evidence came from. None of these is a wrong result; they are provenance statements that can now be made specific.
CORRECTS A DISPATCH PREMISE: GSE149183's menin-versus-DOT1L-inhibitor head-to-head exists in MOLM13 (KMT2A::MLLT3) ONLY. RS4;11 - the index fusion - is in the series but carries H3K79me2 alone, with no VTP-50469 arm, no MEN1 ChIP and no DOT1L ChIP, and the two arms are on different genome builds
This dataset was expected to contain the key comparison in the exact leukaemia subtype the lab studies. It does not - that cell line was only used for a different measurement, so the main result has to be read as coming from a related but different leukaemia.
Counted absence: AFF1/AF4 - the partner in the commonest infant fusion - has no phase-separation literature under any alias, and the two hits that look like it are token collisions
The protein fused to MLL in the commonest infant leukemia has never been shown to form liquid droplets by anyone. Searching every name it has ever published under returns two papers, and one is about plants while the other is about a lab technique that happens to have 'droplet' in its name.
Counted absences keyed on a complex subunit miss the papers that name only the complex - demonstrated on our own SSRP1 entry, which declared no KMT2A paper about FACT while this graph cites one about the index fusion in four other entries
When we record that no paper exists on a protein, we search for that protein's name. If the paper that did the experiment only names the machine the protein belongs to, our search cannot see it. One of our own entries says no study links this machine to this leukaemia; a study does, on exactly the right fusion, and we cite it elsewhere.
Counter-argument reuse across 8,080 notes is 1.04x and only 1.6% are class-level objections that ignore their own claim: the ledger's standing quality worry was measured on 53 notes and does not generalise
This project requires every claim to carry a published argument against it, and reached 100% coverage a while ago - but immediately warned that a high score might just mean the same generic objection pasted everywhere. Measured across all 8,080 of them, that turns out not to be what happened: 95% are one-off texts written for their own claim, and the generic ones are mostly generic for a good reason.
Counting laboratories instead of papers demoted nothing, because BioPlex 2.0 and 3.0 have different last authors - and the 3' partners' own neighbourhoods are the emptiest part of the graph
We tried to fix a counting error by counting research groups instead of papers, and the fix did nothing because the two papers we were trying to merge have different senior authors. Fixing that properly changed the answer, and also showed that the partner proteins' own neighbourhoods are the least written part of this project.
CSPG4/NG2 is 0.0 nTPM in normal bone marrow, so the ratio metric is undefined for it - and that zero is the favourable half of its profile, which a ratio-ranked table hides entirely
NG2 is the surface marker this project calls its clearest example of the leukaemia switching on a gene the body uses elsewhere. It is undetectable in normal bone marrow, which is good news the usual scoring method cannot express.
CURATED from committed-but-unread results: murine MLL-AF9 (MLLT3) retains H3K79me2 like human MOLM13 (MLLT3) and unlike RS4;11 (AFF1) - four confirmations fire, two refutations do not, the cross-partner delta is 3.2-3.6x the between-culture spread, and the flatness survives an amplitude stratification in 8 of 8 measurements
Two leukaemia cell lines carrying different fusion partners had given very different answers about whether a chemical mark on DNA packaging survives a menin inhibitor, and the obvious worry was that the difference was just noise between cell cultures rather than anything to do with the partner. This checks that, and the difference between partners turns out to be three to four times larger than the variation between two cultures of the same cell type - and a mouse line carrying the same partner as one of the human lines behaves like that human line rather than the other, in a completely separate experiment. The result also survives being recomputed on only the strongest sites, in all eight measurements where that was tested, so it is not an artefact of averaging weak signal with strong.
CURATED: Q120 called the MLL1-N channel 'inert to three decimals'. At MLL1's OWN top sites in mouse it falls to 0.3012 under a DOT1L inhibitor and 0.3298 under a menin inhibitor - so 'inert' was population-specific, and a drug that does not touch menin moves the MLL1 channel as much as one that does
An earlier analysis reported that the signal from an antibody against MLL1 barely moved under a menin inhibitor, and described it as inert to three decimal places. Looking instead at the places where that antibody's signal is strongest, it drops to about a third of its untreated level - and it drops by almost exactly the same amount under a completely different drug that blocks an enzyme rather than the menin interaction. So the earlier 'inert' description was true only of the particular set of sites it was measured on, and the fact that both drugs move this signal equally argues that neither is moving it through the mechanism it was chosen to test.
Cycle 102 declared before it ran that an input pair 'bounds library and coverage variation but NOT ChIP efficiency' and is therefore a LOWER BOUND - the size of that underestimate is now measured in human KMT2A::AFF1 MV4;11 and it is sixteen-fold at the peaks: 0.6% for an input pair against 9.4% for a menin ChIP pair over the same bins, in the same cells, on the same day
When scientists want to know how noisy a measurement is, they often use a simpler control sample that skips the antibody step. This compares the two directly in identical cells: at the sites that matter most, the simple control looks sixteen times steadier than the real measurement does, so using it as the yardstick makes small changes look much more convincing than they are.
Cycle 104's M1 - 'no scoreable mark moves at D1, so the menin inhibitor's chromatin effect is proportionally a peak phenomenon' - DOES NOT SURVIVE. The same-condition floor at D1 is 2.4%, not the 10% borrowed from the top stratum, and all three scoreable marks exceed it (H3K4me3 9.8%, H3K79me2 7.8%, H3K27me3 4.8%). Two of the three exceed even GSE260742's OWN stratum-matched input floor, which was on disk the whole time. Cycle 104's alternative M2 fires instead
Last cycle concluded that a menin inhibitor changes the chromatin marks only at a tiny set of the very strongest sites, and not in the much larger region just below them. That conclusion used a 'how much do identical samples differ' yardstick measured at the strongest sites and applied to the larger region, where the true yardstick is about five times tighter. With the right yardstick, all three measurable marks do change in the larger region - and that region holds twenty-three times more signal.
Cycle 104's open question - 'why do two marks break the floor at D10 and not at D1?' - has a false premise. Against the measured same-condition floor of 2.3% at D10, THREE marks break it there (H3K79me2 16.2%, H3K4me3 12.6%, H3K27me3 6.3%), and the same three break it at D1 too. There is no D1-versus-D10 contrast to explain; there is a continuous bulk gradient that a single borrowed 10% figure had chopped in two
Last cycle noticed that two chromatin marks looked changed in the emptiest part of the genome but not in the busier part just below the peaks, and asked why. With a correctly measured yardstick the marks are changed in both places, and a third mark joins them - so the puzzle was created by the yardstick, not by the biology.
Cycle 107's headline same-condition floor of 3.03% at D1 is a median over a MIXED population of two genotypes. Restricted to the non-targeting control genotype it is 10.49%, so the borrowed 10% floor cycle 107 declared 'too generous by 3.3x' was in fact about right there - cycle 107's own F2, not its F1.
A previous analysis averaged the repeat-to-repeat wobble across two different cell lines and got a small number, then used that small number to declare an older, larger estimate too lax. Splitting the average back into its two cell lines shows the older estimate was close to right for the normal cells, and the small number came almost entirely from the gene-knockout cells.
Cycle 110's 'D10 GAINED 66%' rescored on the absolute scale: D10 lost 76.5% of its H3K79me2, no stratum gained anything, and on the genuine RS4;11 rows the gradient spans 41-fold against the contaminated table's 30-fold
An earlier cycle of this project concluded that the least-marked tenth of genes gained chromatin signal while the most-marked lost it; with the fruit-fly yardstick properly applied, that tenth had lost three-quarters of its signal and nothing gained anything. The shape of the earlier result was right and the line between gaining and losing was in the wrong place.
Cycle 76 warned that count thresholds in falsifiers may not transfer across set sizes. Sweeping all 2,301 entries found 5 scored criteria with a count threshold and exactly ONE that is both fragile and live - hyp-q117's uncashed PEAK ARM, now amended before it runs
Last cycle I found a pre-registered test whose pass/fail answer flipped depending on whether you read its "20 out of 52" as a raw number or as a percentage, so this time I searched all 2,301 records for tests written the same way and found that almost none are: of roughly 7,600 places a numeric condition could hide, only five are real scored tests containing a count, and four of those are already safe. Exactly one had the flaw, and because it has not been run yet it could simply be rewritten to state the number both ways at once - so my alarm was much bigger than the problem, and the useful result is one repair made before anything depended on it.
Cycle 77 published epigenetics as this graph's weakest PI dimension at 35.9% by dedicated field. Reading the entries that have no field shows 10 of 10 carry genuine, specific, cited epigenetic annotation in prose - so the dimension is 56.2% (at least 50.4% on a Wilson bound) and about 20 points of the 'gap' was a formatting artefact
Last cycle I reported that the knowledge graph's weakest area was epigenetics - the chemical marks on DNA packaging - because only about a third of gene records had a dedicated field filled in for it, which looked bad in a disease defined by exactly such a mark. So I pulled out the actual sentences from records that have no such field but mention a mark somewhere in their text, and read ten of them at random: all ten turn out to state real, specific, referenced epigenetic biology about that very gene, not passing mentions of the disease. The information was already there and simply not filed in the expected place, so the true coverage is about 56% rather than 36%, roughly 20 points of what I called a research gap was a paperwork gap, and the genuine gap is the 44% of records that say nothing about marks at all.
Cycle 81 named lineage as the single biggest hole in the ARID1B paralogue-buffering result. Permuting ARID1A status only WITHIN lineage strata moves the estimate from -0.2985 to -0.2961 - a change of 0.0024 - with p = 0.00140, the instrument check surviving, both specificity controls flat and zero mutant lines dropped. Lineage explains none of it
The previous finding was that switching off the gene ARID1B matters much more in cells whose partner gene ARID1A is already broken - but ARID1A tends to be broken in particular cancer types, so the whole effect might have been about which tissue the cells came from rather than about the gene. This repeats the comparison in a way that can only ever compare cells of the same cancer type with each other, so tissue of origin is held fixed by construction. The answer barely moves: the effect size changes by about two parts in a thousand and remains statistically clear, while the internal checks that validate the comparison all still behave, so tissue of origin explains essentially none of the original result.
Cycle 81's pre-registered PRIMARY stratum was underpowered at n=4 for five cycles. A different source gives n=124: ARID1B buffering is -0.2475 (p=5e-05) in truncating-ARID1A lines and +0.0117 (p=0.66) in missense-only ones - a complete dose-response. My own specificity control failed, and it failed because I chose it from inside the complex I was studying
For five cycles the sharpest version of this question could not be asked, because the records I was using identified only four cell lines whose ARID1A gene was clearly broken - too few to test. A different public database holds the same cell lines' mutations with a field saying what each mutation does, which raises that number to over a hundred, and the answer is clean: the effect appears only in lines whose ARID1A is truncated and is completely absent in lines carrying a single amino-acid change. One of my own control genes also moved, which by my stated rule made the result uninterpretable - but I had picked controls from inside the very protein complex under study, and five genuinely unrelated genes all stayed flat, so the control was measuring biology rather than a fault.
Dataset: cBioPortal - open, easy, and containing exactly two KMT2A::AFF1 fusion records
cBioPortal is the friendliest cancer-genomics database there is: no login, no forms, a simple web interface. We checked what it actually holds for our disease, and the answer is sobering - only two records of the exact fusion this project is about.
Dataset: ChEMBL chemical matter - open, quantitative, and easy to misread as a druggability score
ChEMBL is a free catalogue of every published measurement of a drug-like molecule binding a protein. It tells us instantly whether anyone has ever made chemistry against a target. The catch is that a big number can mean the chemistry was aimed at a different part of the protein entirely.
Dataset: DepMap CRISPR essentiality - the bulk download is now behind a captcha, and this changes how phase 2 must be built
DepMap tells us which genes cancer cells cannot live without, and it is the backbone of every target decision in this project. As of now you can no longer download the big files with a script - a robot check blocks it - so we have to go through a different door or use an older copy.
Dataset: gnomAD constraint - an open, instant tolerability prior that cannot discriminate between our candidates
gnomAD collects the genetic variation of hundreds of thousands of people and tells us which genes almost never tolerate being broken. It is free and instant. The awkward finding is that every single target we are considering is in that intolerant category, so the measure cannot help us choose between them.
Dataset: GTEx normal-tissue expression - fully open, and the version-string trap that silently returns nothing
GTEx measures how much of each gene healthy adult tissues make. It tells us which organs would be exposed if we blocked a target. It is completely free to use, but there is a trap: if you get the gene's version number slightly wrong, the service returns an empty answer instead of an error.
Dataset: Human Protein Atlas - the protein-level and single-cell check that GTEx cannot give
GTEx tells us how much of a gene's message a tissue makes. The Human Protein Atlas goes further: it stains actual tissue to show where the protein really is, and it breaks blood down into individual cell types. That matters when the tissue we must protect is bone marrow.
Dataset: PDB and AlphaFold - the structural census, including the two absences that shape every plan
Before designing a drug you need to know the shape of what you are aiming at. We checked every structure this project relies on against the official databases, and also checked what is missing. Two things are missing that matter: there is no computed model of the full KMT2A protein, and no experimental structure of any intact fusion protein.
Dataset: St. Jude Cloud - the right cohort for this disease, behind a Data Access Agreement we have not applied for
St. Jude Cloud holds the largest collection of childhood cancer genomes in the world, including the infant leukaemias this project is about, but the raw data needs a signed agreement reviewed by a committee. Nobody in this project has applied yet, and that free application is the single cheapest thing we could do.
Dataset: TARGET via the GDC - open expression and mutations, but the fusion calls themselves are controlled
TARGET is a large US programme that sequenced childhood cancers, and its data now lives in the NIH's Genomic Data Commons. Much of it is free to download. The one thing a fusion project most wants - the file listing which fusion each patient has - is locked behind an application.
DAXX and ATRX are absent from the only MLL-fusion AP-MS interactome - and PML is in it against all seven baits, alone, without a single one of its other body components
There is exactly one experiment that asked what sticks to seven different versions of this fusion protein. The two chaperones this beat is about are not in its results at all - but PML, the scaffold of the nuclear compartment they normally live in, is there against every single bait, and none of PML's usual partners came with it.
Day 3 of a DOT1L inhibitor has already removed 85-90% of H3K79me2 absolutely, and the collapse becomes progressively LESS uniform: the p10-p90 spread of per-gene retention widens from 2.4-fold at d3 to 8-44-fold at d9 in all four arms
Three days of the drug already strips most of the chromatin mark, and the loss is very even across genes at that point. By day nine there is far less left, and what is left is spread wildly unevenly - some genes keep thirty times more of it than others. The collapse does not flatten out; it fans out.
DepMap across the whole CTCF-BRCA1 licensing cascade in one call: CTCF is a near-total pan-essential at -1.83, BARD1 is deeper than BRCA1, the HP1 paralogue that matters is flat while the one that does not is four times deeper - and DNA2's row has 80 screens where everything else has 1,258
We looked up how much each gene in this repair chain matters for cancer-cell survival, all in one query so the numbers are comparable. The protein at the top of the chain is essential in almost every cell line tested, which means it can only ever be damaged by a drug, never helpfully tuned.
DepMap agrees with the mouse genetics: ARID1B is not a FUSION-SELECTIVE dependency (+0.0145, p=0.80) and reads -0.08 in both groups - but CYCLE 81 SHOWED THAT FLATNESS IS PARALOGUE BUFFERING, not dispensability. ARID1A and ARID2 ARE dependencies and are equally so in both groups - so no BAF subunit tested is fusion-selective. And MEN1 (-1.26) is a harder and more selective dependency than KMT2A itself (-0.81)
A 2022 mouse study found that deleting the gene ARID1B makes MLL-AF9 leukaemia worse rather than better, which is the opposite of a drug target, so this analysis asked whether large-scale human gene-knockout screens agree - and they do: switching ARID1B off barely affects any leukaemia cell line, whether or not it carries a KMT2A rearrangement. Two related genes, ARID1A and ARID2, are genuinely needed for these cells to grow, but they are needed just as much in leukaemias without a KMT2A rearrangement, so none of them is a target specific to this disease; the unplanned finding is that MEN1, the protein the current drugs actually block, scores as a harder and more selective requirement than KMT2A itself does.
DepMap cannot answer whether MEN1 is selectively required in NUP98-rearranged lines, because it contains none — and it takes only ~3 to answer it
We tried to check in a big public database of cancer cell lines whether NUP98-driven leukaemias need menin as much as KMT2A-driven ones do. There are no NUP98 cell lines in the database at all — and the effect is so strong that only about three would be needed to settle it.
DepMap co-dependency cannot rank the 97 unwritten interactome partners either, pan-cancer OR restricted to leukaemia lines: mean r 0.008 rising to 0.023 under restriction while the leukaemia-matched null rises with it, giving 1.08x the chance-expected count. Recomputing the null on the restricted lines is the only thing that stopped this being reported as a discovery, and my first pass criterion called a below-chance result a PASS
The project's first priority is an entry for every documented partner of the fusion, and ninety-seven documented partners still have none - but the previous cycle showed they are all backed by exactly one experiment each, so nothing about the published evidence can say which to do first. This tried a completely different ranking: whether switching off each partner harms the same cell lines that switching off its bait harms, which would suggest a real functional link. It does not work either - the correlations are indistinguishable from randomly paired genes, and slightly tighter than random - so there is no principled order in which to work through the remaining list, and saying so is more useful than inventing one.
Depth matching changes this pair's call counts by under one percent, so library depth is not what separated the two beats - a candidate explanation cleared
We suspected that differences in how deeply each sample was sequenced might explain the disagreement. It does not: equalising them changes the numbers by less than one percent for this pair.
Diffing the core scaffold's curated interactome against the graph: 46 of 217 partners are things here, CDK9 is not one of them, and the filter I built to rank the rest was counting one platform twice
We asked an outside database which proteins are known to stick to the machine this cancer builds, then checked how many of them we had actually written down as separate things. One in five. The single best-documented one - the enzyme at the centre of our own explanation of the disease - had no page of its own.
Do NUP98's seven fusions resemble each other more than they resemble their partners' other fusions? The test cannot be run for six of them, and the seventh resembles neither side
NUP98 is stuck onto seven different partner genes in seven different leukaemia fusions, which makes it the best case for asking whether a fusion family is really about the shared piece. The answer turned out to be that the question cannot be asked properly here - six of the seven partners appear in no other fusion this project has written - and the one partner that can be compared behaves unlike both groups.
Do the other WRAD subunits behave like DPY30? (run, not planned)
Four proteins form the shared core of every complex that puts the H3K4 mark on DNA, and only one of them, DPY30, tracks menin in leukemia cells. We tested all four the same way to check that this was real and not an artefact of which genes happened to be in the earlier panel.
Does the menin co-dependency survive lineage adjustment? (run)
Our earlier scan found that cells needing menin also need KMT2A and LEDGF, but that could have been a trick of the data: two genes can look related simply because both matter in the same tissue. We removed that effect and the result held almost perfectly, then got noticeably stronger when we looked only at leukemia.
DOT1L carries 52 in-gene genome-wide-significant SNPs at the 97.8th length-matched percentile, led by height, pulse pressure and blood-pressure traits - a speculative cardiovascular lead for DOT1L-directed drugs
Inherited variation in DOT1L, the enzyme several leukaemia drugs target, is linked to height and blood pressure. That is a faint hint, not a warning, that blocking it might touch the circulation.
DOT1L sits in a B-lineage proximity map, and we nearly did not look
The enzyme at the centre of this project's story turns up next to the master switch of B-cell identity, in the right kind of leukaemia cell. The result was sitting in a spreadsheet attached to a paper we had set aside.
Drug exposure and clonal replacement are completely confounded across this pair - the relapse sample is a different clone in a different lineage, and the paper's own reading is that the drug killed the cells the diagnostic sample was made of
The two samples are not the same cells before and after a drug. The drug wiped out the B-cell leukaemia; what grew back was a myeloid leukaemia from the same child. So anything that looks different between the two could be the drug, or could simply be that these are different cells.
Eight claims this corpus carries from PMID 23995757, checked one at a time against the full text: five confirmed, one scope-limited, one materially different, one not in the paper at all
We re-read the 2013 AKAP95 paper line by line to check what our own notes say it contains. Most of it was right. Two things were not: a binding experiment was described more precisely than it was actually done, and a localisation fact came from a database, not from that paper.
Eight of 27 GSE252378 samples fail a pre-registered depth or readability gate, and two of them are the only HOXA-negative non-AFF1 candidates
Some samples in this public dataset were not sequenced deeply enough to say that a gene is unoccupied - a blank there means 'not measured', not 'not there'. Eight of the 27 samples are in that category, and they were set aside before the main question was asked.
Eleven abstract-level bounded absences re-tested with the instrument that overturned the HDAC6 one - ten survive, one is not checkable this way, none overturned
A previous cycle found one of this project's no-one-has-looked claims to be wrong, by searching for drug names and cell-line names instead of disease names. Doing the same thing to eleven more of those claims did not break any of them: ten held up under the better search and one asks a question no literature search can answer.
ELL has never been placed in a cell state — its entire leukaemia state literature is its own fusion transcript used as a residual-disease marker
Nobody has ever shown what condition ELL holds a cell in. Every paper that pairs ELL with a cell-state word in leukaemia is using the MLL-ELL fusion as a laboratory test for leftover disease, not studying what ELL does.
ELL S309 is the most frequently observed phosphosite across public datasets - and it is a proline-directed site with no kinase, not an ATM site
Someone pooled all the public data on where phosphates land on ELL and found one spot that shows up far more than any other. It is not one of the spots the DNA-damage kinase uses, nobody knows which enzyme makes it, and the technique cannot cleanly tell it apart from the serine right next door.
ENL is the only elongation cofactor that separates fusion-occupied promoters from wild-type KMT2A promoters by more than the H3K4me3 negative control does, in primary infant KMT2A::AFF1
Of the proteins that travel with the fusion, only ENL reliably marks the genes the fusion sits on in real patient samples. DOT1L does so only just, and no better than a general marker of active genes in two of the three samples.
ENL redistributes rather than persists: it is the only channel that holds more of its signal at the myeloid-program peaks than at the B-lineage ones, by 2.06x in my instrument and 1.26x in the authors' own - and in absolute terms it falls further than menin does
The published headline is that ENL stayed on the fusion's target genes while menin and the fusion left, but ENL actually fell at least as much as menin did. What it did do, uniquely among the five proteins measured, is hold on at the genes of the new myeloid identity while letting go of the B-cell ones.
Enumerating every H2AK119ub number in 168 analysis result files instead of hand-picking fourteen: the two conventions flip sign on 35 of 68 comparable ratios but MATERIALLY on only 2, the exhaustive list finds exactly one real single-convention quotation against two arithmetic coincidences, and the statistic cycle 105 cited as proof they agree is one of the two places they materially disagree
A chemical mark in a public dataset was processed two different ways, and earlier work worried the two versions might contradict each other. Checking every number the project has ever computed from that mark shows the two versions do land on opposite sides of no-change about half the time, but nearly always by a hair - only two out of sixty-eight comparisons differ by enough to matter. Awkwardly, one of those two is the very comparison the previous cycle used as its example of the two versions agreeing.
Enumerating the WRAD-versus-KMT2A-fusion literature: what was searched, found, rejected and could not be read
This is the audit trail for the WRAD question: every search that was run with its exact wording and how many papers it returned, every paper that looked relevant and was set aside with the reason, and the one paper that matters and could not be read.
Enumeration for the index-fusion model beat: what was searched, found, rejected and could not be read
A record of everything this piece of work looked at, including the things it decided not to use and why, so that someone repeating it next year repeats the decision and not just the search.
Enumeration for the partner-outcome beat: what was searched, what was used, what was rejected, what could not be read
This is the record of the search itself: which questions were typed into PubMed, how many papers came back, which ones were used, which ones were deliberately set aside and why, and which ones could not be opened at all.
Enumeration for the subnuclear beat: every compartment queried, every count, and the papers that looked on-topic and were not used
A record of exactly what was searched for this beat, what each search returned, and which papers were read and then set aside - so that somebody repeating the work next year repeats the decisions, not just the searches.
Enumeration for the USP2 / MAML2 / second-BCR beat: what was searched, what was counted, what was rejected
Before writing anything, every relevant search was run and every hit was checked by reading its title and abstract. This records the numbers, including the searches that found nothing and the papers that looked relevant and were not.
Enumeration of candidate cooperating lesions in KMT2A-rearranged leukaemia - what was written, what was recorded as a gap, and what was read and rejected
This is the working record of a search: every gene that was considered as a partner of the KMT2A fusion, what was found for each, and - importantly - which ones were looked at and deliberately not written up, with the reason and the exact searches used, so that someone repeating this in a year can tell a real gap from a gap nobody checked.
Enumeration of the reciprocal KMT2A fusion literature: what was searched, found, rejected and could not be read
A record of exactly which searches were run for this topic, what came back, which papers were looked at and set aside and why, and which could not be read at all - so that anyone repeating this in a year repeats the decisions and not just the searches.
Enumeration: every phosphorylation-dependent and kinase-regulated contact on KMT2A and its fusions that this beat could find
A list of everything we searched for, what we found, what we looked at and rejected, and what we could not read. The point is that someone repeating this next year can tell our decisions from our blind spots.
EPHA7 is the one core member with a published KMT2A::AFF1 promoter ChIP - and this beat's own data put it second on separation while denying it to the AF9 samples
Of the seventeen genes, one - EPHA7 - was already shown in 2007 to have the MLL-AF4 fusion protein sitting on its promoter, and switching that fusion off in a t(4;11) cell line turns EPHA7 down. Our data agree for the AFF1 patients. They disagree with the same paper about MLL-AF9, where we see nothing.
Epigenetics reads 39.2% and looks like the thinnest PI dimension. Roughly half of the shortfall is a missing FIELD NAME over content that is already there, a third is proteins with no chromatin context to record, and the real content gap is about a ninth of the corpus
A tool that scores how well this project records the chromatin biology of each protein says only 39% of entries have it, which made epigenetics look like the worst-covered topic here. Looking entry by entry, most of the missing 61% is either content that IS written down but under a field name the tool does not recognise, or proteins like cell-surface antigens and Golgi enzymes that have no chromatin biology to record at all.
Estradiol does not redistribute SUV39H1 - because the replicates do not agree well enough to see it: over 56,171 CTCF anchors the ACROSS-condition Spearman (0.358) is HIGHER than the within-condition replicate correlations (0.326 for -E2, 0.207 for +E2), and the same 147 bp anchor reads 51.8, 69.1, 90.8 and 96.3 in the four samples
The one dataset that measured this enzyme before and after a hormone signal cannot show a change, because two repeats of the same condition agree with each other no better than a before agrees with an after. The repeats disagree so much that a single measurement from this kind of experiment cannot be trusted to tens of percent.
Every age-restricted target nominated by the model is pan-essential or invisible
We checked the four targets this literature nominates against a large cell-line dependency dataset. Three are needed by almost every cell alive and the fourth shows nothing, which does not disprove the claims but does say where they rest.
Every DOT1L dependency number this project has computed, in one place, lineage-split and in raw Chronos units
This project has measured how much cancer cell lines need DOT1L about a dozen times, in different runs, and the answers look different depending on which statistic you read. Put side by side and split by whether the leukaemia is myeloid or lymphoid, the picture is a real but moderate dependency that the available cell lines are too few to prove.
Every gene in the H2B ubiquitination axis, ranked against the whole genome for KMT2A selectivity — and the writers are not selective
We looked up every enzyme that writes, carries or erases the H2B ubiquitin tag in a ready-made scan of 17,728 genes that ranks how much more KMT2A-rearranged leukaemia cells need a gene than other leukaemia cells do. The enzymes that write the tag came out flat or backwards; only the erasing side produced anything.
Every human t(11;19) KMT2A::MLLT1 cell line, with accessions fetched and read back - and not one of them is known to be reciprocal-negative
There are seven human cell lines with this fusion. All but one come from babies under a year old, six are B-cell leukaemias and one is a therapy-related myeloid leukaemia from an adult. None has been shown to lack the mirror-image gene, which is the one property this project most needs a line to have.
Every human t(9;11) cell line, with every accession fetched and read back: eleven lines, ten patients, and two of them are B-ALL
Eleven human cell lines carry this translocation, but two of them come from the same man, so there are ten patients. Nine are monocytic leukaemias and two are infant B-cell leukaemias that almost nobody uses.
Every IntAct number in this project is drawn through a query that misses half the record - and what it misses is the taspase1 fragments
Every count this project has taken from the IntAct interaction database was fetched with a query that silently skips any record filed under a protein fragment rather than the whole protein. For MLL, whose database records are mostly filed under its two cut pieces, that means half the record has never been seen here.
Every MS-evidenced KMT2A modification site, flagged retained or lost against the breakpoint - and the retained portion is NOT enriched for acetylation once you use the right null
We listed all 174 chemical tags that mass spectrometers have ever found on the MLL1 protein and asked, for each one, whether it survives into the cancer fusion protein or is thrown away. The surprise is a negative: none of the four SUMO tags survives, while the tags that do survive are simply the ones that happen to sit on the part of the protein that is kept - there is nothing special about them.
Every multi-entry dangling reference in this graph has exactly one author: demand counts entries, and entries are not independent
Yesterday's audit ranked the graph's missing entries by how many other entries asked for them, and one was asked for seventeen times. Checking who wrote those seventeen: one agent, one day, one folder - so the number seventeen measures one decision, not seventeen.
Every partner this project imports from the only MLL-fusion AP-MS interactome was pulled from a Strep-HA transgene in Jurkat T-ALL cells — so the host's own lineage factors are in the prey pool, and one of them is LEF1
The one experiment that has ever mapped what MLL fusion proteins bind was done by putting a tagged copy of each fusion into Jurkat cells — a T-cell leukaemia line that does not have the fusion and is the wrong lineage for the infant B-cell disease we study. Anything the host cell is full of will show up in the pulldown, and for a T-cell line that includes T-cell transcription factors.
Every peak file of all 13 ENCODE BRCA1, CBX3 and CBX1 experiments opened at the exon-11 element, not just the two Q37 reached: BRCA1 is there in BOTH pre-IDR replicates of THREE experiments - HeLa-S3, H1 and HepG2, global IDR 0.059, 0.577 and 0.482 - and it is a clean zero in both K562 experiments and in GM12878 against 300,000-peak pre-IDR lists
Public datasets usually publish only the peaks that survive a strict reproducibility filter. Opening every unfiltered list instead shows the repair protein BRCA1 sitting on this one spot in both repeats of three different cell types - two more than anyone here had checked - and genuinely absent from both blood-cell datasets even in lists three hundred thousand entries long.
Every published KMT2A partner in sarcoma, counted - plus the papers this beat read and did not use, so the next beat does not repeat the search without the decisions
This is the full list of genes reported joined to KMT2A in soft-tissue sarcoma, how many cases each has, and what each contributes. It also records which papers were read and judged off-beat, so that a future reader can tell a paper that was dismissed from a paper nobody looked for.
Every upstream NF-kB component this beat enumerated, with its query, its count, and why it did or did not get an entry
This is the working record of which NF-kB proteins were checked, what each search returned, and which ones were dropped. Writing down what was looked at and rejected is as useful as writing down what was found, because otherwise the next person cannot tell a gap from a decision.
Every WRAD subunit, DOT1L, ENL, AF9 and LEDGF carries a cited histone-mark finding and none carries an epigenetics field - 12 of the 32 index-fusion core proteins are pure notation gaps, 9 are covered, and the 11 with nothing are all on the negative-regulatory side of elongation
For the 32 proteins at the heart of this project's disease, nine record their chromatin biology in a field a reader can filter on, twelve describe it only in the text, and eleven say nothing at all. Those eleven are not scattered - they are the brake that holds transcription in reserve, not the accelerator the fusion hijacks.
EWSR1::FLI1 has no residue-resolved, fusion-demonstrated PTM in its retained portion - every mapped site falls past the breakpoint, and the one site shown on the fusion has no residue
For the Ewing sarcoma fusion protein, every chemical modification anyone has pinned to a specific amino acid sits in the part the fusion throws away. The only modification actually demonstrated on the fusion itself has never been traced to a particular spot.
Exactly one gene in 8,248 beats menin, and it is FLT3
Last cycle we checked two thousand genes and reported that none scored better than menin for this leukaemia. Checking eight thousand finds one that does: FLT3, a gene that already has approved drugs. The claim was true of the sample and not of the genome.
Excluding spleen, lymph node and thymus moves 9 of 29 normal-tissue safety ratios and not one of the worst cases - the artefact is real, concentrated and confined to the middle of the table, and it takes the ACTB reference down with it
A gene can look dangerously active in an organ simply because that organ is full of immune cells. Redoing the safety ranking with the three most lymphocyte-rich organs taken out changes nine genes out of twenty-nine, but leaves every one of the worst offenders exactly where it was.
FAERS shows no endocrine signal for either menin inhibitor - and proves it could not have
We searched the FDA's public database of side-effect reports for every hormone- and bone-related problem we could think of, in every report naming a menin inhibitor. Not one of twenty terms appeared, while the two known side effects of the drug appeared hundreds of times more often than chance.
Fifty-three 2 kb bins gain MLL1-N-reactive signal under VTP-50469 in RS4;11, and the same 53 coordinates gain 1.77-fold in MOLM13 against a class-matched null of 1.01 with the MOLM13 input flat at 1.01 - so there IS a small, reproducible, input-controlled redistribution, and it carries no CpG-island signature
Fifty-three small stretches of DNA pick up more of the MLL1 signal when the drug is added, and the same stretches pick it up in a second, unrelated leukaemia line while the background control does not move at all. So something really does move - but it is about thirty times smaller than what stays put, and it is not where the CpG-island theory predicted.
First deposit to put both drug classes on the MEN1 channel in one experiment: VTP-50469 leaves ZERO of the 19 DMSO menin peaks at 15 canonical menin-MLL target loci, while EPZ5676 keeps 14, compound 10 keeps 14 and compound 11 keeps 8 - strength-matched permutation p = 0.0003, and the zero is unchanged at every flank from 0 to 100 kb
Two kinds of leukaemia drug were tested side by side in the same experiment, and we asked where the protein menin still sits on the DNA afterwards. The menin-blocking drug removed menin completely from all fifteen of the genes this cancer depends on, while all three of the enzyme-blocking drugs left menin sitting on most of them.
First systematic enumeration of the KMT2A fusion interactome, and the diff against this graph
Somebody finally went through the main papers that list what sticks to KMT2A and its fusions, wrote down every protein they name, and checked that list against what this graph already had. The surprising result is that most of the famous partners attach to the part of KMT2A the cancer throws away.
Five new arginine-axis entries in one day, and not one of their genes is a selective dependency - in an instrument that recovers MEN1 at p=5e-05 and TDRD3 is not a dependency at all
In a single day this project wrote up an arginine-tagging control point on the machine that drives this leukaemia. Asking whether any of the genes involved is actually needed more by these leukaemia cells than by other leukaemia cells, the answer is no for every one of them - measured with a method that correctly finds the two genes already known to matter.
For AFF1 in primary infant material, which antibody you use moves the answer more than which patient, which lineage, or cell line versus patient - a pre-registered falsifier that fired
Two different antibodies aimed at the same protein in the same patient sample gave more different answers than two different patients gave with the same antibody. That means cross-sample comparisons of this protein are partly measuring the reagent.
Forty-one SEM libraries in GSE159608 sort themselves by antibody from the signal alone, across two protocols, two operators and ten months - so the deposit's own labels are confirmed by an instrument that does not share the labels' failure mode
A public dataset labels each experiment with which protein or chemical tag it measured; we ignored those labels, downloaded the raw data for 41 experiments on the same infant-leukaemia cell line, and asked the data to group itself. It grouped exactly the way the labels said it should, even matching experiments done ten months apart by different people, so the labels are right.
Four beats in a row were picked by noticing the same shape by hand; this is the instrument that finds it - 171 gene symbols named in six or more entries with no entry of their own
We kept noticing that this project writes up a mechanism and never writes down the proteins inside it. That had been spotted four times by hand. This finds them all at once, and the top of the list is the cell-surface markers the whole disease is diagnosed and treated by.
Four counted absences re-run against a second index, and all four hold
When this graph says nobody has studied something, that claim was usually checked against one literature database, and four such claims in this project have previously turned out to be wrong because the search was narrower than the sentence it was used to support. I re-ran four of them against a different database, and this time all four held up - the one apparent exception was a brain-tumour review that happened to mention both search terms for unrelated reasons.
Four days after the fusion is destroyed, menin signal at non-fusion promoters is 27% higher than control
Destroying the fusion for four days does not just remove menin from the fusion's genes - menin shows up more strongly at other genes. It looks like menin is relocating rather than being cleared, although the way the data were scaled makes that hard to prove.
Four matched input tracks, used for the two things an input is actually for: no chromosome-scale copy-number difference between the CRISPR clone and its control (max 1.0988, min 0.9470 in both processing states), and the knockout's H2AK119ub depletion from CpG islands is REAL because its own genomic DNA is enriched there, not depleted
A CRISPR-edited cell clone can quietly gain or lose whole chromosomes, which would fake any comparison with its parent. Sequencing the raw DNA of both shows they match. The same raw-DNA tracks also show that a difference seen in the edited cells is genuine rather than an artefact of how the DNA was sequenced.
Four regulatory elements on AF9, and the fusion keeps exactly one - the phospho-switch that controls SEC binding
AF9 has four known control knobs, and the KMT2A::MLLT3 cancer protein throws away three of them and keeps one. The one it keeps is the switch that decides whether AF9 grabs the gene-reading machine, which is the opposite of what you would expect if the discarded parts were the important ones.
Four sources give ZNF384 three different zinc-finger counts - 8, 6 and 'five to eight' - and the paralogue's feature table explains where the 6 comes from
Nobody agrees how many zinc fingers ZNF384 has. The database and the paper that discovered the fusion family both say eight; the more recent leukaemia papers draw six; one says between five and eight. Looking at ZNF384's closest relative explains it: the relative has exactly six, and the newer papers are drawing the part the two share.
Four YEATS antagonists exist, four public transcriptome datasets of YEATS inhibition exist, and not one paper or dataset has ever asked what happens to the AFF2- or AFF3-scaffolded complexes — and the analysis needs no wet work
Four different drug-like molecules that block the histone-reading pocket have been published, and their effects on gene activity are freely downloadable. Nobody has checked what they do to the two sister machines that contain the same reader — even though the data to check it are already online.
Fractional occupancy needs a free concentration as well as a Kd, and nobody has measured either for a KMT2A fusion
You cannot say how much of the time the fusion is sitting on a CpG site from a binding constant alone - you also need to know how much fusion protein is floating free in the nucleus, and nobody has measured that. The published binding constants alone already change the answer ten-thousand-fold.
Fusion-driven surface antigens are NOT more exposed in normal tissue than expression-matched controls - ratio 0.91, permutation p = 0.79 - so the screen's empty intersection is the base rate and not a property of the fusion
This beat proposed that the surface proteins this leukaemia makes because of its fusion gene are unusually likely to also appear in healthy organs. Testing that against a fair comparison - other proteins expressed at the same level in the same cells - shows no difference at all. The reason no good target was found is simply that almost no surface protein is absent from healthy tissue, whatever switches it on.
GEO labels both aliquots of the lineage-switching patient 'treated with the Menin inhibitor', but the paper says 148752 is the B-ALL at diagnosis and only 152985 arose during treatment - the deposit's drug flag is a case label wearing an aliquot label's clothes
A public dataset marks two samples from one child as having been taken during drug treatment. The paper says the first was taken at diagnosis, before any drug. Anyone comparing the two as treated-versus-treated would be comparing before with after and not know it.
Giving the reopening conditions an owner, and the first run changes two decisions (run)
When this project rules a drug target out, it writes down what would change its mind. Nobody was checking whether that had happened. A script now runs those checks against the medical literature, and on its first run it found new work that revives one ruled-out target and confirms a change made to another.
Gothe 2019's deposit is all-human, has base-resolution break lists, includes primary CD34+ marrow and a transcription-inhibition arm - and contains no CTCF or cohesin perturbation of any kind
The Gothe paper's data is entirely human, includes real bone-marrow progenitor cells, and lists every single break it found at single-base resolution - the richest deposit of the three. What it does not contain, anywhere, is an experiment that removes CTCF or cohesin.
Gothe 2019's full text IS reachable - not through PMC, which has no deposit, but through an author search that surfaces the bioRxiv preprint; fourteen routes, and the two Europe PMC id links that look like a PMC record are both traps
A previous cycle concluded this paper's full text could not be reached and gave up. It can: search the first author's name in Europe PMC and the free preprint version appears, with the complete methods. Two of the obvious shortcuts return long lists of identifiers that look like the paper's own record and are actually the papers that cite it.
GSE127507 deposits a matched input for BOTH conditions in MOLM13 only - RS4;11 and ML-2 have a DMSO input and no VTP-50469 input - so no input-corrected cross-condition occupancy statement is possible in the human index-fusion line from this series, and none exists anywhere
To be sure a change in a treated sample is real you need the matching background control for the treated cells, and this dataset only has it for one of its three cell lines - not the one that matters most here. Every number for the index-fusion line therefore rests on a statistic that does not need it, which is a real constraint and not a formality.
GSE127507 ran a Drosophila spike-in, measured it, and deliberately did not use it - so the VTP-50469 founding dataset carries no recoverable scale factor either, and the depositors say so in the sample record
The team that made the key menin-drug dataset did add a fly-DNA yardstick to their experiment, but then decided the yardstick showed nothing interesting and threw the measurement away. That means nobody can compare the raw signal between treated and untreated samples, so every comparison in this re-analysis is a ranking inside one sample rather than a before-and-after amount.
GSE159608 — 419 automated CUT&Tag libraries across 7 KMT2A-rearranged primary patient samples, 5 lines and a normal CD34+ donor, including KMT2A fusion occupancy, a menin-inhibitor arm in SEM, and three single-cell CUT&Tag libraries
This is a large, freely downloadable public dataset that maps where the KMT2A fusion protein and nine chromatin marks sit on the genome, in seven real patients' leukaemia cells and in the standard laboratory cell lines, with healthy donor stem cells as a control. Three of its libraries read the marks one cell at a time. This project has cited the paper twelve times and has never written down the accession number, so nobody could reuse the data.
GSE173599 deposits menin ChIP at three hours and nowhere else - no 24 h, no 4 d - so the central question this beat raises, whether menin's persistence is fusion-independence or merely a slow off-rate, cannot be answered from the only degron deposit that exists, and the GEO record advertises four-day menin ChIP that was never deposited
Menin stays on DNA three hours after the cancer protein is destroyed, and the obvious follow-up is to look a day later - but nobody deposited that measurement. The dataset's own description says four-day menin measurements were made, and they are not in the files.
GSE252378 carries an institution-level batch effect: unrelated pairs from the same hospital overlap more than cross-hospital pairs, and a cohort-median comparator absorbs that into the same-patient claim
Samples processed at the same hospital look more like each other than samples from different hospitals, whatever patient they came from. Because the one patient with two samples was from a single hospital, comparing them to a whole-cohort average made them look unusually similar when they were not.
GSE252378 contains menin CUT&RUN in exactly one patient - a KMT2A::ENL lineage-switching pair under a menin inhibitor - and none in any of its eight KMT2A::AFF1 patients
The dataset that finally provides fusion-protein maps in real infant patients does not include the menin protein for the version of the fusion this project studies. Menin was measured in only one patient, who has a different fusion partner.
GSE252378 contains no megakaryoblastic sample, so the 0-of-18 non-AFF1 IRX1 occupancy bound does not cover the one subgroup where IRX1 is actually expressed
The study that found the fusion protein never sits on IRX1 outside one partner did not include any child with the kind of leukaemia in which IRX1 is switched on. So it could not have seen the thing we now know to look for.
GSE252378 fragment BEDs: a complete KMT2A N-terminal / C-terminal antibody set in 27 patient samples across six fusion partners, deposited and until now unopened by this corpus
A public dataset holds raw fragment positions from 27 children's leukaemia samples, each profiled with antibodies against both ends of the KMT2A protein. Because only the fusion keeps the front end, comparing the two gives a per-gene readout of where the cancer-driving fusion protein sits.
GSE252378 holds eight KMT2A::AFF1 oncoprotein panels from two centres, not the three this corpus had used
We thought this public dataset had three infant leukaemia samples to work with. Reading the sample list properly, it has eight — from two different hospitals. The extra five more than doubled what could be measured.
GSE252378 holds the only AFF1 CUT&RUN in primary KMT2A::AFF1 patient material - four libraries across two patients, three disease states and the SEM cell line, all hg38
A public dataset contains the only measurements of where the AFF1 protein sits on the DNA of real infant leukaemia patients who carry the KMT2A-AFF1 fusion. There are fewer independent patients in it than the sample names suggest, because two of the three patient samples come from the same child at two different stages of their illness.
GSE252378 labels one of its two SEM AFF1 libraries KMT2A-AFF2, a fusion that does not exist in the deposit and does not exist in SEM - a filter on that field silently halves the SEM data
One sample in a public dataset is tagged with the wrong fusion name because of a single-character typing error. Anyone filtering the data by that tag would lose half the cell-line measurements without noticing.
GSE252378's 27 KMT2A-fusion samples are 24 independent patients: three pairs are one patient sampled twice
A public dataset's 27 leukaemia samples come from 24 children, because three pairs are the same child sampled twice as the disease changed. Counting them as 27 independent patients overstates the evidence.
GSE260742 declares an H3K27me3 second channel in its MLLAF9 Control arm, and no floor can be computed from it: all three rep1/rep2 pairs agree about only 86.31-88.90% of the shallower track's covered bins, so 11-14% of the SHALLOWER track's coverage falls where the DEEPER one has none. Depth cannot explain that.
A dataset says it repeated one of its measurements, and it did - but the two repeats do not even agree about which parts of the genome carry a signal. More than a tenth of the shallower repeat's signal sits where the deeper one sees nothing at all, which is not something you can blame on it being shallower.
GSE260742's murine H3K79me2 samples come from FOUR distinct cultures - F3, F3_NT, NT and the PCGF1-KO clone C1 - and the titles name only two of the distinctions; 'Timecourse Control' is the F3 line, and 'NTvKO Control' covers two different filename families that are not the same material
In this dataset the sample titles and the file names disagree about which batch of cells each experiment used. Reading the file names shows four separate cultures where the titles suggest two, and it settles a question the previous cycle had to leave open about whether one experiment used the same cells as another.
GSE260742's PCGF1-knockout samples carry no deposited measurement of PCGF1 - no protein, no transcript, no sequencing of the edited locus - and the knockout rests on one guide and one clone with no rescue, so every null result from that arm is a null in an unverified genotype
The dataset labels six samples as coming from cells with PCGF1 deleted, but nothing in the dataset shows that PCGF1 was actually lost in those cells. There is one guide RNA, one clone, and no rescue experiment. So a result saying 'deleting PCGF1 changes nothing' cannot be separated from 'PCGF1 was not deleted'.
GSE260742's VTP-50469 kinetic series carries no SGC0946 sample at any timepoint, so the denominator of F cannot be measured inside that batch and the fraction cannot be recomputed at any time - the only SGC0946 arms in the series are in different experiments at a single 72-hour or 7-day exposure
The project expresses the menin drug's effect on a histone tag as a fraction of what the DOT1L drug achieves. The dataset follows the menin drug over time but never the DOT1L drug, so the fraction cannot be tracked over time at all - and the two measurements that make it come from different experiments.
GSE299100 supplies the vehicle-versus-vehicle pair that neither GSE260742 nor GSE127507 could - two independent control hairpins, both DMSO, same date, uniformly processed to hg38_rmdup.tdf, for MEN1, MLL1 and input - in MV4-11, which Cellosaurus confirms carries KMT2A::AFF1 though it presents as monocytic rather than the infant B-ALL this project studies. It also holds MEN1 and MLL1 ChIP under five different menin inhibitors in that line, and the six control files total 1.87 GB
Four recent results depend on sorting genomic regions by how much a drug removes a protein, then measuring a second change in the same regions - a design that can invent a pattern from nothing. Checking it needs two untreated samples that differ only by chance, and the two datasets this project had been using could not provide such a pair. A search of public deposits found one that can, and it is better than expected: the samples come from the exact leukaemia subtype this project studies rather than a mouse model of a related one, and the same deposit also compares five different drugs of the class in that cell line.
GSE68720 - 97 infant ALL patients from COG P9407 with the KMT2A partner named per sample (mll_par) and a GPL570 expression matrix, 74 partner-named KMT2A-r cases and 17 germline controls
A public dataset of 97 babies with leukaemia that records, for each child, both which partner gene the KMT2A fusion joined and the activity of every gene. It is the only such dataset we could find in this disease, and it is downloadable in one 8 MB file.
GSM8006584 names its antibody KMT2A-C1 in its title, its characteristics field and its own filename, and KMT2A-C2 in its description - so keying the antibody off the free-text description gives patient TB-92-0488 two C2 libraries and no C1
One sample record in this public dataset contradicts itself about which antibody was used. Three of its four fields say one thing and the free-text sentence says another, and trusting the sentence would quietly delete half of that patient's control.
GSM8123377 and GSM8123403 carry the same supplementary filename but for the `.filt.bam` marker, yet GEO titles one 'MLLAF9 Control - DMSO H2AK119ub rep2' and the other 'NTvKO Control - DMSO H2AK119ub paired'. Scored directly, they are DIFFERENT libraries (Spearman 0.83), so the collision is a naming accident and both arm titles stand.
Two samples in a public dataset have essentially the same filename but are labelled as belonging to two different experiments. Comparing them directly shows they really are two different samples, so the labels can be trusted and the matching names are just a slip.
GSM8123386/387/388 carry the GEO title 'MLLAF9 Control - ... H3K79me2 rep2' and the supplementary filename 'ChIP_K79_F3_NT_*'. Two committed analyses treat them as a separate culture and one asserts 'FOUR INDEPENDENT MURINE CULTURES'. If the titles are right there are three, one of them replicated. REPORTED AS A DISPUTE, NOT APPLIED.
Three deposited samples are labelled as repeats of an existing experiment in one field and as a different experiment in another. Two earlier analyses followed the filename; the dataset's own title field says repeat. The difference changes how independent a published comparison is.
GSM8123405 is mm10 DATA in an hg38-DECLARED TDF - the exact mirror of Q122's mm10-declared / hg38-coordinate defect, in the same series - and the container silently clipped 72.8 Mb (2.76%) of chr13-chr19 and chrX at deposition; the declared extents name the container and max_end names the data
One of the deposited mouse files was packaged as if it were human. The mouse data inside is genuine, but wherever a human chromosome is shorter than the mouse one, the packaging cut the end off - about 3% of the genome, silently. A check that reads only the label calls the file human; a check that reads where the data actually stops calls it mouse.
H3C1 stays rejected, and now with the two numbers the registry row lacked: its accession carries TEN gene names, and removing every substrate and ribosome symbol moves interactome coverage 43.8% -> 46.1%
Histone H3 sits at the top of this project's list of unwritten interaction partners on excellent evidence, and it should not get an entry because it is the thing the machinery acts on, not a partner of it. Two numbers settle it: the protein record for H3C1 covers ten different genes, so an entry would have to pick one arbitrarily, and counting substrates as partners only distorts the coverage figure by two percentage points.
H3K4me3 and wild-type MLL1 are the negative controls for the partition, and both come out flat
If splitting genes into 'fusion sat here' and 'fusion did not' made every measurement look different, the split would be telling us about the split and not about the fusion. Two things that should not change when the fusion is destroyed - a histone mark laid down by the normal protein, and the normal protein itself - do not change, which is what makes the menin result mean something.
H3K4me3 is by far the most reproducible track in primary infant leukaemia CUT&Tag - replicate rho 0.97-0.98 - which makes it both the assay's proof of competence and the benchmark that every fusion-site enrichment should be quoted against
A general marker of switched-on genes is measured almost identically twice in the same patient sample, which shows the technique works on this difficult material. It also sets the bar that any claim about the fusion has to beat.
H3K4me3 is the one mark in GSE260742 with no declared second channel at all - six samples, no replicate suffix anywhere - so no same-condition pairing exists and its noise floor remains a transfer from another mark whatever else is measured.
Of the four chemical tags this dataset measures, one was never repeated at all. Its reliability therefore has to be guessed from the others, and no amount of reanalysis can change that.
H3K9me1 depletion over the KMT2A breakpoint cluster replicates in a leukaemia line - as a broad-domain property of the 7.5 kb BCR, not of the 1 kb CTCF element, where the two tracks disagree in sign
The one part of this chain that is a property of a resting cell - that the mark is unusually scarce across the stretch of DNA where KMT2A breaks - holds up in a leukaemia cell as well as in the skin-type cell it was found in. But it holds for the whole stretch, not for the small element inside it, where the two available datasets point opposite ways.
Half the KDM4C deposition is not a KMT2A fusion at all: MOZ-TIF2 8, MLL-GAS7 6, MLL-AF9 2, and the human xenograft deposited nothing
Two entries cite a survival result in a human leukaemia sample called MLL3 whose fusion partner is never named. The paper's deposited data contains no human sample at all - it is sixteen mouse samples, and half of them carry MOZ-TIF2, which is not a KMT2A fusion, while the KMT2A arm is mostly the very rare partner GAS7.
Half the PRC1.1 intersection was free all along: Zhou's locus set is a publisher supplementary table, and 159 of this graph's gene entries are in it
I reported that settling a disagreement between two papers would need a large computing job, having checked the public data archive but not the journals' own supplementary files, and said so. Checking them now, one of the two published its list there - 9,292 rows, free - so half the job was already done and only the other half is still expensive.
Half this graph's claims are unopposed; these are the ones that matter
About half the individual statements in this project have nothing published arguing against them. Working through them in file order treats them all as equally important. This ranks them by how much else would move if one turned out to be wrong.
Has anyone tested an XPO1 inhibitor in a KMT2A-rearranged system? Yes - four preclinical studies and zero clinical ones, counted with the queries and the controls
We asked whether anyone had ever put an XPO1-blocking drug into KMT2A-rearranged leukaemia, and counted the answer instead of guessing. Four preclinical studies exist, all combinations, all in myeloid models; no clinical trial has ever selected patients by this genotype.
Holding mark, arm and drug conditions fixed and varying only which confound the replicate pair carries: the floor ranges are DISJOINT at D1 and D10, a processing-confounded pair inflates D1 2.20x while a batch-confounded pair inflates D10 2.21x, and the direction reverses by stratum - so the floor is a property of the pair, not of the mark or the arm, and every floor this project has published came from the pair type that is wrong for D1
To say a measurement changed you need to know how much two repeats of the same experiment differ. This project has now measured that several times and got wildly different answers. Comparing repeats that differ only in how the files were processed against repeats that differ only in when they were sequenced shows the answer depends almost entirely on which kind of repeat you happen to have - and in opposite directions in different parts of the genome. Every figure this project has published came from the kind that is wrong for the region three rounds of argument were about.
How much of this graph's mechanism is somebody else's fusion? A counted answer: 54% of fusion-bearing evidence, and 0% from a dedicated human KMT2A::AFF1 experiment
We took a random sample of the entries that carry mechanism in this graph and asked, for each one, which leukaemia fusion the evidence actually came from. Of the entries whose key experiment used a fusion at all, more than half used a fusion that is not the infant one this project is about, and not a single one rested on a dedicated experiment in human infant-type cells.
How the three occupancy archives spell zero: ENCODE 404, ReMap 401, ChIP-Atlas 404 - and ChIP-Atlas spells a one-character URL typo exactly the same way, so a whole sweep can come back cleanly and be entirely false
None of the three big protein-on-DNA archives answers an empty question with an empty answer; all three answer with what looks like a broken connection. Worse, one of them gives the identical broken-connection answer when you mistype one character in the address, so a search that finds nothing anywhere can be a typo rather than a fact.
HOXA is occupied as a 107 kb cluster; IRX1 is occupied as one gene, with both IRX clusters cold
When the leukaemia protein switches on HOXA, it lights up the whole neighbourhood of genes at once. When it switches on IRX1, it lights up that one gene and nothing around it.
HOXA occupancy is the partner-invariant: occupied in 17 of 18 non-AFF1 KMT2A-r samples, 2.94x to 29.1x
Whichever gene KMT2A is fused to, the fusion protein sits on the HOXA9/HOXA10 genes. That part of the disease is shared across all six partners tested.
HOXA9 is where menin leaves most completely (share ratio 0.14-0.23 over floor) and MLL1-N-reactive material stays most completely (0.74-1.02 over floor), in all three MLL-rearranged lines - so the menin inhibitor decouples menin from the HOXA9 locus without clearing the MLL1 N-terminus from it
At HOXA9 - the gene most people think this leukaemia lives on - the drug takes menin off almost completely while the MLL1 signal stays where it was. Whatever is still sitting there after treatment, it is not being removed by a menin inhibitor.
HOXA9/HOXA10 is unoccupied and unmethylated at H3K4 in two of the three primary infant KMT2A::AFF1 states, while the SEM cell line carries the strongest AFF1 signal over that locus of any sample in the deposit
HOXA9 is the textbook target gene of this fusion, but in two of the three real infant patient samples the fusion is not on it at all and the gene is switched off. The cell line that laboratories use as a stand-in for these patients has the strongest signal there of anything measured.
HOXB4, the one direct IRX1 target with ChIP evidence, carries no oncoprotein in any of seven infant patients
IRX1's best-established direct target gene is HOXB4. In seven babies' leukaemias the cancer-driving protein was not on HOXB4 at all — consistent with what the original paper said, that HOXB4 belongs to IRX1 and not to the fusion.
HPA's per-cell-line RNA file is not dead - it lives under /download/tsv/ while proteinatlas.tsv lives under /download/, and the wrong prefix returns HTTP 200 with a 117,090-byte HTML page
A previous cycle concluded that the Human Protein Atlas cell-line data file could not be downloaded any more. It can - the web address just needs an extra folder in it. The dangerous part is that the wrong address does not return an error; it returns a normal-looking success and a web page pretending to be a data file.
HR genes are menin-bound at 1.76x and so is ribosome biogenesis at 1.78x: the pathway-specific reading does not survive its own control
RETRACTED IN PART, within the hour, by the control this entry itself named as required: the claim was that menin specifically binds the DNA-repair machinery. Every highly-expressed housekeeping gene set turns out to be menin-enriched to the same degree, so repair genes are part of the cell's busy machinery rather than singled out.
HTAN KMT2A-r ALL single-cell multiome — 25 patients (18 infants), scRNA-seq + scATAC-seq + snmC-seq2, with 2 paired pre/post lineage-switch cases and analysis code on GitHub
This is the single most useful public dataset for the questions this project asks about which cells a KMT2A leukaemia is made of: twenty-five patients, eighteen of them babies, with three different measurements made on the same individual cells — which genes are on, which bits of DNA are open, and which bits are chemically silenced. Two of the patients were sampled before and after their leukaemia changed lineage. The analysis code is public too.
Hunting the one reagent that would unbind every share-based result in this thread: the fly/human ratio is in no supplementary table and the words dm6 and scale factor appear zero times in the full text - and the 181116 batch the processing line names is in NONE of the 78 samples across all four subseries of GSE127508, so no deposited sample is established to carry the spike-in at all
Every measurement in this line of work is a proportion, which means it cannot tell a real global change from a bookkeeping effect. The fix is a "spike-in": a known amount of fruit-fly material added to each sample as a ruler. One key dataset's notes say a fly spike-in was added and then judged unnecessary. Chasing it down shows the number was never published, and - more awkwardly - the specific batch of samples the note refers to is not in the dataset at all. It can still be recovered by re-reading the raw sequencing files, and this works out what that would cost.
hyp-q117's prediction 5 is stored TWICE in one entry, in sections[0].body_md and in data.predictions[4], and the two strings differ - here is the scored outcome and the ALL-CAPS cashed text proposed for both locations, with the clause-by-clause verdict a one-line 'confirmed' would have hidden
A prediction written two years of cycles ago has now been tested, but only one of its three parts could be tested, and that part came out half right. This entry writes down exactly what was and was not answered, in the form the project uses for a settled prediction.
hyp-q117's prediction 5 says 'in a PCGF1 OR BCOR knockout' and GSE260742's only genetic perturbation is the PCGF1 knockout: across all 158 sample records the strings BCOR, BCORL1, KDM2B, RING1 and RNF2 appear zero times as a perturbation, so a verdict on PCGF1 is not a verdict on the prediction as written
The prediction named two genes that could be removed. Only one of them was, in the only dataset that could test it. So half the prediction is answered and half has never been looked at.
I audited whether a valid absence token is the RIGHT token, expecting the gene-nbn error to be widespread. It is not: 3 of 4 read were correctly scoped, the scoping is done in prose, and 39 of 40 flagged entries carry no machine-readable scope
Last cycle we found one entry whose gap label was wrong, so this checked whether that was common. It is not - the labels read correctly, and where an entry means something narrow it says so clearly in its own text, just not in a form a program can read.
I withheld these four numbers last cycle to keep an underpowered contrast out of a narrative, and they point the OTHER WAY: the four truncating-ARID1A leukaemia lines mean -0.0401 against a leukaemia wild-type of -0.0923, a naive +0.0523 where pan-cancer gives -0.2475. And NONE of the four carries a KMT2A fusion, so the intersection this whole thread needs is EMPTY, not merely small
Last cycle I noted that only four blood-cancer cell lines have a clearly broken copy of the partner gene, decided that was too few to test, and deliberately did not look at their numbers so that a weak comparison could not creep into my write-up - while recording that this also meant any contradictory signal in them would be invisible to me. This cycle I looked. They point the opposite way to the main finding, and more decisively, not one of the four carries the gene rearrangement this project exists to study, so the specific combination needed to test the idea in this disease does not exist in the dataset at all. The main pan-cancer result is unaffected, but carrying it across into this disease now has four data points against it and no route to a direct test.
In 18 of 27 patient samples, comparing two antibodies raised against the SAME end of KMT2A calls at least as many oncoprotein-target promoters as the real N-over-C comparison does, and swapping the antibody pair moves a sample's own call count by a median of 2.57x
The test for fusion binding compares antibodies against the front and back of a protein. If you instead compare two antibodies against the same end, where there is nothing to find, the test still fires - in many patients as often as the real comparison does.
In a competitive menin-binding scaffold with assumed rates, the drug exposure needed scales with menin relative to the fusion-menin Kd, and whether recovery after a lapse lags collapse is governed by the drug's residence time - a conditional counter-lead to the pulsed-dosing premise, not a prediction
We built a small computer model of the loop that keeps this leukaemia's growth program switched on, and of a drug that competes with the fusion protein for the same grip on menin. Almost every number in it is a guess within a wide range, so it does not predict any patient. What it does show is which guesses matter. The more menin the cell holds relative to how tightly the fusion grips it, the more drug is needed, and how fast the program comes back after a missed dose depends mostly on how long the drug stays stuck to menin once it lands.
In a PCGF1-knockout clone the menin inhibitor still evicts and the vacated sites are still silenced, but LESS: H2AK119ub at the most-evicted quartile falls from 1.4450 to 1.1325 and from 1.0151 to 0.8842 in two replicate conventions, and H3K27me3 from 2.0007 to 1.0134 - so prediction 5's silencing clause is directionally right and quantitatively overstated
Knocking out one component of the 'switch-off' machinery makes the switching-off weaker but does not stop it. That is what you would expect when the component has five close relatives that can do part of the same job, and it means a prediction that said the sites would simply NOT be silenced was too strong.
In all 24 read-queue rows inspected, the candidate is the BAIT and this project's protein is the prey - nobody has ever tagged KMT2A, AFF4, menin or PSIP1 and found any of them; plus seven bounded rejections
Fifteen proteins are queued as unwritten interaction partners of the leukaemia machinery. Reading the raw records one by one shows that in every single case somebody was studying the other protein and ours happened to turn up - no experiment has ever started from our proteins and found any of these - and seven of the fifteen should not get entries at all, for reasons recorded here with their numbers.
In an unsupervised 5 kb correlation of 41 SEM libraries, both KMT2A N-terminal AutoCUT&RUN libraries have menin as their single nearest neighbour - the menin-interface hypothesis appearing as the top-level structure of an analysis that was not looking for it
We grouped 41 experiments on infant-leukaemia cells purely by how similar their data looked, with no gene lists and no assumptions. The experiments that tracked the front half of the leukaemia-driving protein came out closest to the experiments that tracked menin, the partner it is supposed to hold on to - which is what the leading hypothesis predicts, arrived at without asking for it.
In drug-naive primary infant KMT2A::MLLT1 B-ALL, menin is as enriched as the fusion N-terminus at the top fusion promoters (2.73x vs 3.12x over an accessibility reference) while the wild-type C-terminal reporter sits at background (1.10x)
In a baby's leukaemia cells taken before any menin-blocking drug, menin is sitting exactly where the fusion protein is - and it is not sitting where the normal, unbroken version of the same gene is. That is the clearest direct evidence this project has that the menin the drug targets is the menin attached to the fusion.
In human KMT2A::AFF1 MV4;11 (monocytic, not B-lineage), the vehicle-versus-vehicle noise floor is 9.4% at the top 0.1% of menin/MLL1 sites and 0.1-2.6% in every other stratum - so cycle 102's borrowed 4-10% band is about right AT THE PEAKS and four to five times too generous everywhere else, and it was applied everywhere
To tell whether a drug really changed something in a cell, you need to know how much two identical, untreated samples differ just by chance - and for four recent results this project had to guess that number from the wrong kind of comparison. Two genuinely identical control samples have now been measured three times over, and the answer is that the guess was roughly right at the handful of strongest sites and about five times too forgiving everywhere else.
In human RS4;11 (KMT2A::AFF1), menin occupancy at its own top 0.1% of 2 kb bins falls 14.9x to 2.0x over the track median and its share of total coverage falls to 0.200 - reproducing Q117's murine FLAG-MLLAF9 figures (15.9x to 3.0x, ratio 0.179) to within 1.3-fold on all five statistics
In cells from the leukaemia this lab studies, the menin drug strips about four fifths of menin off the places on DNA where menin normally sits hardest. The same measurement made last cycle in a mouse leukaemia, using a completely different tag and dataset, gave almost exactly the same number.
In infant ALL, IRX1 transcript is enriched in KMT2A::AFF1 (21 of 48) and present but rare outside it (3 of 26) - so the occupancy zero outside AFF1 is not an expression zero
In 97 infant leukaemia patients, the IRX1 gene is switched on in about 44% of those whose fusion partner is AFF1, but in only about 12% of those with a different partner. It is not completely off outside AFF1 - a few children with other partners have IRX1 as high as anyone.
In MCF-7 the two signals point in opposite directions over the same 90 kb: H3K9me3 sits at the 15.8th and 10.3rd percentile of 56,171 same-cell CTCF anchors across the KMT2A gene body (0.82x and 0.65x the random-tile mean) while SUV39H1 sits at the 81.7th-86.8th in all four of its tracks - and the 147 bp anchor carries ZERO H3K9me3 in the minus-estradiol track
In breast cancer cells the silencing tag is unusually scarce right across the KMT2A gene - at the exact 147-letter spot this project cares about there is none at all in one sample - while the enzyme that is supposed to make that tag looks, if anything, plentiful there. Enzyme and product disagree about the same piece of DNA.
In MOLM13, human KMT2A::MLLT3, from the VTP-50469 founding paper's own series GSE127507, H3K79me2's aggregate response at MLL1-N's top sites is 0.998 - and the same sites split by eviction depth run 0.7746 / 1.0601 / 1.0995 / 1.1421 at Spearman -1.000, with the INPUT stratified identically staying flat
In human leukaemia cells from a completely different published dataset, averaging says the menin inhibitor does nothing at all to the activating tag - the number is 0.998 out of 1. Splitting the same sites by how much fusion protein the drug removed shows a clean trend from a 22 percent loss to a 14 percent gain.
In murine MLL-AF9, H3K79me2's response to a menin inhibitor runs 0.586 / 0.819 / 0.928 / 1.187 across the four eviction-depth quartiles of the fusion's own top-0.1% sites - Spearman -1.000, span 0.6005, the largest of any mark measured in these bins except H3K27me3 - while the aggregate over the same 1,228 bins is a single unremarkable 0.8156
When a menin inhibitor pulls the leukaemia fusion protein off a stretch of DNA, the activating chemical tag DOT1L leaves behind disappears from that stretch in proportion to how much of the fusion left. Averaging over all the stretches at once hides this completely, because the places that lose the most and the places that gain are counted together.
In murine MLL-AF9, knocking out PCGF1 changes the menin inhibitor's effect on H3K79me2 by 0.0166 of a share ratio and the DOT1L inhibitor's by 0.0101 - both smaller than the 0.0584 two control cultures of the same line already differ by, so the mark's drug response runs outside PRC1.1
Deleting PCGF1 - a component of the repressive complex that moves into the spots the fusion protein vacates - does not change how either drug rearranges the H3K79 chemical tag on histones. The effect of removing the gene is smaller than the difference between two untouched cultures of the same cells, so whatever the drugs do to this tag, they do not do it through that complex.
In patient-derived H3K27M glioma, baseline H3K9me3 over the 7,466 bp KMT2A breakpoint cluster core sits at the 7.8th and 28.5th percentile of 59,821 width-matched CTCF-anchor windows (0.56x and 0.90x the random-tile median) - the methyl-poor-BCR pattern this graph found in epithelium and in a blood line, now in a third lineage and for the trimethyl state
The stretch of DNA where the KMT2A gene usually breaks carries unusually little of the silencing tag in two children's brain-tumour cell lines - the same pattern this project already found in eye cells and in a leukaemia line. The pattern now shows up in a third kind of tissue, which makes it look like a property of the DNA rather than of one cell type.
In SEM, the infant-derived KMT2A::AFF1 line, VTP-50469 lowers H3K4me3 at menin-bound sites by 23% relative to H3K4me3-matched menin-free sites (95% CI 13-33%), while the DOT1L inhibitor EPZ-5676 changes it by 1% (CI -4 to +6%)
A menin inhibitor is a drug that pries a leukaemia-driving protein off DNA, and we asked what it does to a chemical mark called H3K4me3 in cells from an infant leukaemia. Comparing the places menin sits against places with the same amount of the mark but no menin - inside the same sample, so the measurement itself cannot be blamed - the mark falls by about a quarter where menin was, while a drug aimed at a different enzyme does nothing to it.
In St. Jude paediatric AML, 2 of the 6 ARID1A-mutant cases are KMT2A-rearranged and BOTH ARID1A mutations are TRUNCATING frameshifts in KMT2A::MLLT10 cases - so the exact combination cycle 86 showed the ARID1B dependency is confined to DOES occur in patients, and cycle 88's empty DepMap intersection is a cell-line sampling artefact rather than a fact about the disease. The paediatric ALL cohort returned zero ARID1A mutations, which is panel coverage and not biology
Eight cycles of work showed that one chromatin gene becomes important when its partner gene is broken, and then found that no laboratory cell line carries both that broken partner and the gene rearrangement this project studies - which left the whole line of work stranded. This asks the version of the question that matters: do actual patients have both? In a study of nearly nine hundred children with acute myeloid leukaemia, six had the broken partner gene and two of those six also carried the rearrangement, so the combination is real in people even though no cell line models it. A second, larger children's study of a different leukaemia reported no cases at all - but it also reported none anywhere in the cohort, which means its sequencing panel did not cover the gene, so that zero is about the test and not about patients.
In the deposit paper's own target-gene table, the AFF1-exclusive class is 64 genes against a permutation null of 2.3, and it survives call-rate matching, institution stratification, B-lineage restriction and patient collapse
Sixty-four of the fusion protein's target genes are hit only in the leukaemias whose partner gene is AFF1 and in none of the other eighteen patients. If you shuffle which patients you call the AFF1 group and count again, you get about two such genes. That gap does not go away when you correct for the AFF1 patients having more target genes overall, for them coming from one hospital, or for them being the B-lineage ones.
In the HOXA-dead infant states IRX1 carries H3K4me3 and ENL while HOXA9/HOXA10 carry neither
In two infant leukaemia samples the HOXA genes are completely shut — no active mark, no machinery. In the very same cells IRX1 has both. The cells did not lose the programme; they moved it.
In the only system where the fusion can be measured separately, a menin inhibitor removes MLL1 from wild-type-dominated sites harder than it removes the tagged fusion from its own sites - 0.1069 against 0.1793 - which is an on-target effect on normal KMT2A-menin complexes and an argument about toxicity nobody in this project has made from occupancy data
The drug is meant to peel the cancer protein off DNA. In the one experiment where the two can be told apart, it peels the normal version off even more thoroughly at the places where only the normal version sits.
In the revumenib-exposed relapse sample nothing is enriched above a non-specific reference - not menin, not the fusion N-terminus, not ENL - and the sample's own top-ranked promoters are mitochondrial pseudogenes
In the cells taken while the child was on the drug, none of the five proteins measured is sitting anywhere in particular. The genes that score highest are the ones that score highest in a negative control. So the question 'is menin still where the fusion is' has no answer here, because the fusion is not measurably anywhere either.
In the same deposit, the MENIN inhibitor removes DOT1L from all five DMSO DOT1L peaks at canonical target loci (p = 0.0432) while all three DOT1L inhibitors keep 5/5 - and the VTP DOT1L library still calls ACTB, GAPDH and EEF1A1, which is the internal positive control the MEN1 arm lacks
The enzyme DOT1L is brought to leukaemia genes by a complex that menin holds together. Blocking menin knocked DOT1L off all of those genes, while blocking DOT1L's own enzyme activity left it sitting exactly where it was.
In the VTP-50469 timecourse Q124 declared uninterpretable, the eviction-gradient SPAN grows monotonically 0.104 / 0.198 / 0.214 / 0.335 at 4/8/16/24 h with Spearman -1.000 at every timepoint - while the aggregate it was suppressed for is non-monotone and perfectly anti-correlated with library complexity. The span is better behaved, and it is NOT clean.
In a separate experiment following the drug over a day, the pattern deepens steadily the longer the drug is on. The overall average in the same experiment jumps around and was written off as unreliable. The pattern is steadier than the average, but it is not fully free of the same technical problem, and this entry says so.
In-silico digest of 81 modelled KMT2A fusion junctions: the peptide that would prove the fusion protein exists
A fusion protein shares almost all of its sequence with two normal proteins in the same cell, so the only piece that can prove the fusion itself is present is the short stretch that straddles the join. We computed that stretch for 81 possible KMT2A fusion joins and found it does exist and is measurable - but for the commonest KMT2A::AFF1 architecture the enzyme normally used to chop proteins up cuts exactly at the join, so the proof peptide only appears when that cut is missed.
Independent positive control: of 27 samples, the ONE that calls zero oncoprotein-target promoters is 152985 - the AML that arose during menin-inhibitor treatment, which the paper reports as having oncoprotein and menin reduced to undetectable levels
The published paper says one particular patient sample lost its fusion protein from chromatin entirely after drug treatment. Rebuilt from the raw data with a different pipeline and different genomic regions, that same sample is the only one of twenty-seven that scores zero.
Indexing abstracts for all 4,000 cited papers, the U2 that has bound the citation index since it was built - 98.9% retrieved, and across seven probe topics abstract search reaches 144 papers title search cannot see, including 54 single-cell papers and 25 CRISPR screens
The project keeps a searchable list of every paper it cites, used to check whether something has already been done before sending a researcher to do it. Until now that search only looked at paper titles, so a study that did the work without saying so in its title was invisible — which is exactly how the project once sent someone to run a screen that had been published three years earlier. Now it searches the summaries too, and finds 144 papers the old search could not.
Indexing every deposit this corpus cites - 504 accessions, 102 GEO series, all 102 resolving - because the citation index repaired last cycle indexes papers only, and half the known absence failures were refuted by a dataset rather than by a paper
Last cycle this project fixed the tool it uses to check whether something has already been done, which searches every paper its notes cite. That tool cannot see datasets — and one of the two mistakes that prompted the fix was refuted by a dataset, not a paper. This builds the matching index over datasets. It immediately found a public experiment the project had recorded as missing.
Indexing every PMID in every entry - CORRECTED to 4,000 distinct papers across 2,411 entries after the original regex was found to miss 73.7% of them by matching prose but not the schema's own citation field, an undercount that made this entry declare a gap the corpus refuted 14 times over
Before sending a researcher to investigate something, this project checks whether it has already been done. Three times running that check said no and was wrong, because it searched the text of the project's own notes rather than the list of papers those notes cite. Building a searchable index of all 1,908 cited papers fixes it — and incidentally shows the project's citation record is clean: every single identifier resolves to a real paper.
Instrument check before any conclusion: reading a browser-track .tdf with a hand-written binner and ranking the result blind returns the Hoxa cluster in five of the top twenty bins, plus Meis1, Jmjd1c, Six1, Arid2, Ncor1 and Tsc22d2 - the canonical MLL-AF9 programme, which is what the track claims to be
Before trusting a home-made tool that reads an unusual file format, we ranked the strongest signals it found and looked at which genes they sit on. They are the exact genes this leukaemia fusion is famous for occupying, which means the tool and the file are both doing what they claim.
Instrument check: the top thirty RS4;11 MLL1-N 2 kb bins recovered from a deposited .tdf are JMJD1C (259x the genome median), ZMYND11, BAHCC1, CDK6, MYB, RUNX2, IGF2BP2, FOXP1, SATB1 and ARID1B - canonical KMT2A::AFF1 targets, annotated only against the series' own RefGene windows
Before trusting any measurement, the strongest signals in the re-processed data were listed and checked against what is already known to be bound in this leukaemia. They are the expected genes, which means the pipeline that read the deposited files is working.
IntAct curates 47 interactions against engineered MLL fusion proteins, and every accession-keyed query in this project is blind to all of them - while its PSICQUIC mirror stops at 2024
The interaction database this project measures its coverage against contains experiments done directly on laboratory-made MLL fusion proteins - the exact kind of evidence the project says it lacks. Those entries have no standard protein identifier, so every automated query here misses them completely.
IntAct curates SEVEN MLL fusion proteins as first-class interactors - ENL, AFF4, AF9, GAS7, ELL, AF6 and CALM - and not the index one
Interaction databases can store a fusion protein as its own molecule, and they have done it for seven different MLL fusions. The one this project is about is not among them, so no recorded partnership in the database is a partnership of the actual index fusion.
IntAct holds 51 interaction records whose participant IS a KMT2A fusion across 6 constructs, our coverage instrument dropped every one of them silently, and closing the single gap they contained took the partner list to 21 of 21
This project measures how much of the fusion protein's partner list it has written up, and the program doing the measuring was silently skipping the very records where the partner was measured against the fusion itself. Reading them found fifty such records, five different fusion constructs, and exactly one protein the graph had never written down.
Interactome round 4: what was counted, what was written, what was rejected, and the two alias traps this beat fell into and caught
Before writing anything, this round counted how many published papers exist for each candidate protein in KMT2A leukaemia. Most of the proteins the brief asked about turned out to have almost no literature at all, and two of the searches produced wildly wrong counts because of two-letter gene nicknames.
Interactome round 5: 42 queries, 13 entries written, 9 candidates rejected with counts, and four alias failures of four different kinds
Before writing anything, this round counted how many published papers exist for each candidate protein in KMT2A leukaemia, and checked every dependency number twice - once in myeloid cells and once in B-cells. Most of the proteins the brief asked about had almost no literature; two of them had strong-looking dependency numbers that evaporated when the cells were split by type.
IRX1 carries the KMT2A::AFF1 oncoprotein in exactly the patients where HOXA does not — seven patients, GSE252378
In seven babies with the same leukaemia, we looked at where the cancer-driving protein sits on the DNA. In the ones where it sits on the HOXA genes it is completely absent from IRX1, and in the ones where HOXA is dead it is on IRX1 instead.
IRX1 expression in KMT2A-rearranged paediatric AML is a FAB-M7 megakaryoblastic property, not a fusion property: 13 of 20 M7 versus 2 of 96 M5, and non-KMT2A-r M7 is 54 of 97
In 887 children with AML, the IRX1 gene is on mainly in one kind of leukaemia - megakaryoblastic (M7) - whether or not the cell carries a KMT2A fusion. So IRX1 being on says which blood cell the leukaemia came from, not which fusion it carries.
IRX1 fusion occupancy is confined to KMT2A::AFF1: 0 of 18 samples across five other partners
In leukaemias driven by a broken KMT2A gene, the fusion protein sits on the HOXA genes whichever partner it is joined to. It sits on IRX1 only when the partner is AFF1 - in 18 samples carrying five other partners it never does.
IRX1 is reported as expressed in AML subsets, so its absence from oncoprotein occupancy outside KMT2A::AFF1 is not an absence of IRX1
IRX1 is switched on in some myeloid leukaemias. Our finding is only that the KMT2A fusion protein does not sit on the IRX1 gene outside one fusion type - which is a different claim from IRX1 being off.
IRX1 sits in the 2-of-8 tier - two different patients, not the same one twice - and is not in the recurrent core
IRX1, the gene this project has been chasing for several cycles, is switched on in two of the eight KMT2A::AFF1 samples at promoter scale. Those two are two different children, which is the good news. But forty-seven other genes share the same two-of-eight pattern, so IRX1 is not among the seventeen most recurrent.
IRX1's own three-sample pattern is p = 0.022 naively and 0.077-0.176 once patients, institution or lineage are held fixed - the class it belongs to is significant, IRX1 alone is not
IRX1 has the fusion protein sitting on it in three of the eight AFF1 leukaemias and none of the other eighteen, which taken on its own looks unlikely by chance - about one in fifty. But two of those three samples are the same child sampled twice and all three come from one hospital, and correcting for that drops the odds to between one in six and one in thirteen.
IRX2 IS fusion-occupied at gene-body scale in p454 - and q151's promoter-scale zero of 27 turns on one replicate 5% below the threshold, which two other instruments put above it
An earlier round found no sign of the fusion protein on IRX2 in any of the twenty-seven patients, using a two-kilobase window at the gene's start. Measuring the whole gene instead, one patient - p454 - is clearly positive, and the original scientists' own published list also names IRX2 in that one patient. The earlier zero came down to a single measurement of 1.898 against a cut-off of 2.0.
IRX2 is not fusion-occupied at promoter scale in GSE252378: 0 of 27 samples, against a null of 2.73
The broken KMT2A protein does not sit on the switch of the IRX2 gene in any of 27 children's leukaemia samples, even though the same measurement finds it sitting on HOXA9 in 21 of them. One sample comes close and is discussed honestly, but the answer is no.
IRX2 transcript is NOT AFF1-confined: re-deriving GSE68720 finds 1 of 16 KMT2A::MLLT1 infants positive, and that case is IRX2-positive with IRX1 off
This project recorded that a gene called IRX2 is switched on only in infants whose leukaemia carries one specific fusion partner. Re-running the same data finds one child with a different partner who also has it switched on - and in that child the neighbouring gene IRX1 is off, which the record said never happens.
Is any histone chaperone selectively required in KMT2A-rearranged leukaemia? Twenty-nine genes, lineage-split - and the answer is no, with one number worth a null and a positive tail nobody expected
We asked whether cells carrying this fusion need any of the histone-packaging genes more than other leukaemia cells do. Across twenty-nine genes the answer is no - and the genes they appear to need LEAST are ones this project had assumed were central to how the fusion works.
Is IRX1 expressed in normal blood at all? Two of the three papers say no, one says exclusively yes
Before you can say a leukaemia has switched IRX1 on abnormally, you need to know whether healthy blood cells use it. The published answers flatly contradict each other and nobody has resolved it.
Is the selectivity just pan-essentiality? No: it explains 0.4% of the variance
The obvious objection to our genome-wide hit list is that genes every cell needs will look important in KMT2A-rearranged cells too, so the ranking is measuring neediness rather than anything about the fusion. We tested that directly and it is not what is happening: how essential a gene is in general predicts almost nothing about how much more the KMT2A-rearranged lines need it.
Is the YEATS asymmetry just intron geometry? Computed from hg38: the direction is explained, the magnitude is not, and both genes fail the same prediction
We measured how much DNA sits in each part of the two genes, to see whether AF9 loses its reader simply because there is far more DNA to break in the wrong place. The shape of the answer is right but the size is badly wrong, and in a way that says breaks are not scattered at random.
Join the KMT2A breakpoint archive to the Interfant MRD and outcome database, on patient
Every baby who gets the DNA-junction leftover-disease test has the exact position of their broken gene sequenced as a side effect of building the test. Nobody has ever put that position next to how the baby did, even though both pieces of information already sit in databases run by collaborating groups.
KDM2B, the PRC1.1 CpG-island recruiter, sits at the 99.7th percentile of its own distribution on the fusion's strongest sites in DMSO and does not gain share under either a menin inhibitor or a DOT1L inhibitor - it holds position while comparable bins drift down
The protein that tows the repressive PRC1.1 complex to unmethylated CpG islands is already sitting on the fusion's favourite sites before any drug is given, so it has nowhere to climb. What changes under treatment is not where it is but what it is allowed to do there.
Kerry 2017's 'sufficient but not necessary' is a figure title with no experiment behind the second half - and the paper's own knockdowns run the other way
A well-known claim says the fusion protein does not really need menin to find its place on DNA. We went to the source and read it: that claim appears only in the title of one figure. The experiments in that figure show the opposite - when menin is knocked down, the fusion binds less.
King verification of Q154: the AFF1-exclusive target class survives every stratification and a different seed, and IRX1 individually does not - the partner signal is real at the class level and gone at the gene level
An agent reported that KMT2A::AFF1 leukaemias share a set of target genes no other partner touches, and that IRX1 - the gene this project has chased for many cycles - is not individually special once you account for which hospital and which lineage the samples came from. I re-ran its statistics myself, with a different random seed, and checked its gene lists against the original paper's table. Both hold.
King verification of q181: the epigenetics field is NOT hollow (1 of 218 empty), the mark regex has three further blind spots including chromatin accessibility in 25 entries, and my own attempted correction of q181 was wrong
A check of this project's own epigenetics coverage tool. The tool was accused of under-counting, and it does - it only ever looks at the NAMES of data fields, never at what is written inside them, and it has no word for chromatin accessibility even though the project reasons about that constantly. But the opposite worry turned out to be unfounded: the entries it counts as covered really are covered, with only one of 218 empty. Along the way I nearly published a correction of the agent's report that was itself wrong, and caught it by re-checking my own pattern.
KMT2A sits in the 14-3-3 proximity neighbourhood, and the YWHAG link reproduces across two independent studies
A 2023 study tagged the 14-3-3 proteins - general-purpose molecular minders that hold other proteins in particular shapes - and found KMT2A among their neighbours. A different 2024 study, different laboratory, found the same thing for one of them.
KMT2A-rearranged AML is two diseases on the phosphorylation layer, and DOT1L phosphorylation is one of the things that separates them
When researchers measured which proteins carried phosphate in 74 poor-prognosis leukaemia samples, the KMT2A-rearranged cases fell into two groups with different drug sensitivities. One group had more phosphate on DOT1L, a protein this whole field is trying to drug.
KMT2A-rearranged or NPM1-mutant? Both, in separate arms: WT-MA9 mouse, MOLM13, OCI-AML3, and two patients on AUGMENT-101
An entry on menin-inhibitor resistance notes that the paper never says whether its leukaemias were KMT2A-rearranged, NPM1-mutant, or both. The deposited data and full text say both, in clearly separated arms: mouse MLL-AF9 cells and human MOLM13 for the KMT2A side, OCI-AML3 for the NPM1 side, and one patient of each genotype on a menin-inhibitor trial.
KMT2A::AFF1 has had a published affinity-purification interactome since 2018 — in a paper titled for SETD2, uncurated by IntAct, and cited zero times here
A 2018 paper mapped every protein touched by seven versions of this fusion, including the one we study, and deposited the raw measurements publicly - and we had never cited it. It stayed invisible because its title is about a different protein, and because the interaction database we draw from never entered it.
KMT2A::AFF1 retains THD2 by position, so the HBO1 contact can apply to it — but the paper that established that contact tested MLL-ELL, MLL-AF6 and MLL-AF10, never the index fusion
A 2021 paper found these fusion proteins grab a histone-modifying machine through one stretch of protein, and the fusion this lab studies still contains that stretch, well before the point where the protein is cut. But the paper never tested our fusion, and the one it found most dependent on that stretch was a different fusion entirely.
KMT2B has no haematopoietic cell-state result, and two of the seven records that look like one are about a different gene in muscle
Nobody has shown what cell condition KMT2B maintains in blood. Worse, a search for it returns papers about muscle stem cells that are really about a different gene, because both genes share the nickname MLL4.
KMT2B, AFF2, AFF3, AFF4 and ENL carry most of the contact residues of the two interfaces the fusion keeps (menin-KMT2A and AF9-AFF1), so a substitute is structurally possible and selectivity cannot be read from either surface; at the menin side KMT2B differs at W7
If a drug blocks the contact between menin and KMT2A, or between AF9 and AFF1, a close relative of one of the proteins could in principle step in, and the drug might hit those relatives too. We looked at the exact amino acids that touch in solved 3D structures and asked whether the relatives have the same ones. They mostly do: the menin-grabbing stretch of KMT2B shares about 11 of 18, and the surface AFF1 uses to grab AF9 is almost identical in AFF2, AFF3 and AFF4, while ENL matches AF9 at 17 of 18 spots. So these two contacts are family-wide, not special to the fusion. This is about whether a substitute could exist, not whether it binds, is made in the cell, or is used.
LAMP5 was proposed as an immunotherapy target for MLL-rearranged leukaemia without a normal-tissue filter; two atlases independently put it above 290 in the basal ganglia, and it is not partner-specific either
A 2022 paper found LAMP5 by searching every surface protein in MLL-rearranged leukaemia and proposed it as a target for immune therapy. Checking the same protein in healthy organs, which that search did not do, shows it is one of the most brain-concentrated proteins in the body - and this leukaemia is treated in babies.
LAMP5, a surface antigen this project named as a fusion-enriched candidate, is 19.6x higher in normal cerebral cortex than in bone marrow - worse than every one of the 22 chromatin targets previously scored, including MEIS1
LAMP5 was on the shortlist of surface markers this leukaemia switches on. In a normal body it is about twenty times more abundant in brain cortex than in the bone marrow being treated, the highest such ratio measured here.
Ledger: every module the KMT2A half loses, with coordinates, how each was confirmed, and whether its loss has been shown to matter
A single table of everything the broken half of KMT2A throws away, where each piece sits on the protein, and whether anyone has shown that losing it changes anything. Re-checked in September 2026: far more of these losses matter than the table first said, and most of what they do is stop the protein from being cleared away on schedule.
Lineage-split dependency for the round 5 panel: 23 genes, 2 myeloid-identity artefacts caught, and 3 survivors of which one has no literature at all
The cell lines used to find KMT2A-specific vulnerabilities are mostly myeloid, so any gene that myeloid cells happen to need looks specific for free. This analysis scores every candidate twice - once in the myeloid lines and once in the B-cell lines - and reports both numbers.
Lineage-split DepMap: neither USP2 nor MAML2 is a dependency in KMT2A-rearranged lines, and USP22 is
Before saying a gene is needed by KMT2A-rearranged leukaemia cells, you have to check both kinds of those cells - the myeloid ones and the B-cell ones - because the available cell panel is mostly myeloid and a myeloid gene looks important for free. Doing that check, neither USP2 nor MAML2 is needed by these cells, while USP22 clearly is.
Live DepMap on fourteen genes: CDK6's pan-essentiality is a lineage artefact of pooling, and not one gene of the homeodomain cofactor code is a dependency anywhere
We asked a public database, live, how badly each of these fourteen genes is needed across 1,258 cancer cell lines and across the 57 blood-cancer lines specifically. Two answers came back: the cell-cycle gene CDK6 is needed everywhere a bit and in blood cancers a lot, while its twin CDK4 is the reverse - and none of the HOX/PBX/MEIS genes this beat is about is needed anywhere at all.
Looking for a second handle on the retained fragment, and not finding one
If the piece of KMT2A that survives the translocation had a second docking site for a drug to grab, the whole strategy would have another option. We looked for the obvious kind of site and did not find a convincing one.
Measured against a technical floor of 4-10% from two independent input pairs, PCGF1 loss lowers baseline fusion occupancy at the top-FLAG sites by 16.3% and changes the DOT1L inhibitor's effect on that occupancy by 22.8% - while leaving the menin inhibitor's effect unchanged at 1.3%. In the knockout SGC0946 displaces more fusion than VTP-50469 does, which is the opposite of the control
Last cycle measured how much of the cancer-driving fusion a menin inhibitor strips off DNA, and found it identical in normal cells and in cells lacking a repressive protein. That comparison used a second drug as a yardstick for how much variation to expect, without checking the yardstick. Checked now against two sets of no-antibody control samples, the real variation is much smaller than assumed - which leaves the original conclusion intact and turns the yardstick itself into a result: the second drug behaves very differently in the two cell types, and in the mutant cells it removes more fusion than the menin inhibitor does.
Measured: the cell-type, cell-state and epigenetics literature for KMT2A and MEN1 was NOT missing from this corpus - 7 of 16 load-bearing papers were already cited, one of them 32 times, and 0 of 16 were in gene-kmt2a
We checked whether this project had simply never read the papers about how KMT2A and MEN1 behave in normal cells. It had read them, across dozens of entries, and had written none of them into the two files named for those genes.
Measuring every PI-named axis structurally and by content: the type fields understate coverage by a median of 19.5x and by up to 81x, no axis falls below 18% of the corpus, and 'which axis is thinnest' is retired as a beat selector because it reports a lower bound as though it were the truth
This project files each entry under a category label. For several rounds the coordinator chose what to work on next by counting those labels and picking the rarest. Counting the actual text instead shows the labels miss most of the relevant work - by twenty to eighty times - so every topic the lab cares about is already covered far more than the labels suggest. The method of picking work by counting labels has been retired.
MED23 confirmed, and the zero-record claim corrected: it is not literature-free in leukaemia, an shRNA screen already found it
A previous agent flagged MED23 as a promising but unfinished lead and asked for a yes or no. The answer is yes on the data — the four B-cell leukaemia lines with the KMT2A fusion need it, and the numbers are printed here line by line — but the claim that nobody has ever studied it in leukaemia is wrong. A 2016 screen in myeloid leukaemia cells found it, and that paper was missed.
Mediator's KMT2A signal is in its pan-essential core, CDK8 is expressed and not required at all, and MED13L is rank 9 while its paralogue MED13 is rank 13,569
A large protein machine scored highly for this leukaemia, so we asked which part of it mattered. Almost all of it is needed by every cell, which makes it a poor target - but one swappable piece stands out, and its near-twin does nothing.
MEF2C loses occupancy under VTP-50469 in all three human MLL-rearranged lines (MLL1-N 0.18 / 0.25 / 0.04 over floor), which resolves Q117's one locus-level disagreement with the published account in the paper's favour and localises the disagreement to the murine system
Last cycle's mouse re-analysis found one gene, MEF2C, behaving opposite to what the original paper reported. The human data in all three cell lines agrees with the paper, so the odd result was a feature of the mouse experiment rather than a problem with the published account.
MEN1 and CDK9 each carry exactly one genome-wide-significant in-gene common variant and PSIP1 two, so GWAS says almost nothing about on-target liability - yet the revumenib label lists hyperparathyroidism, the organ MEN1 loss famously affects
Inherited variation in the genes these drugs block is almost silent, so population genetics cannot warn us about on-target side effects. The revumenib label nonetheless lists raised parathyroid hormone, which fits what losing menin does in people born without it.
MEN1, DOT1L and LEDGF dependency in the only CALM-AF10 cell line DepMap has screened
We asked a large public screen whether the leukaemia cells carrying the CALM-AF10 fusion need the menin protein to survive. Only one such cell line has been screened, and in it menin barely matters while the DOT1L enzyme clearly does - but one cell line is one cell line.
Menin dependence of a non-elongation KMT2A fusion has never been demonstrated — the counted absence, and what the pivotal trial actually enrolled
Menin drugs are given to patients because their leukaemia is 'KMT2A-rearranged', on the reasoning that menin sticks to the KMT2A half, which every fusion keeps. Nobody has ever tested that in a leukaemia whose partner is a cytoskeletal protein, and the registrational trial did not knowingly enrol a single such patient.
Menin dependence varies 2.83x across KMT2A::AFF1 lines and the spread survives at three donors - but RS4;11 is last on KMT2A itself, so the heterogeneity is largely fusion-dependence heterogeneity, and DAXX is the one gene decoupled from it
Last analysis found that two cell lines with the same fusion needed menin very differently, which would be awkward for the idea that the fusion works through menin. Fetching every available line with that fusion, the difference holds. But the odd line out also barely needs the fusion gene itself - it is the least fusion-dependent line on the list - so it needs less menin because it needs less of the whole machine, not because menin works differently there. That largely rescues the idea. DAXX behaves differently: it is needed evenly, regardless of how much a line depends on the fusion.
Menin does NOT leave chromatin when the fusion is removed: three hours after MLL::AF9 is degraded to 2.5-7.1% of its own chromatin sites, menin still holds 80.0% of its strongest peak calls genome-wide - so menin occupancy is not a readout of fusion occupancy, and a menin inhibitor that displaces menin is doing something fusion loss does not do
A drug in clinical use works by prising a protein called menin off the cancer protein it partners with, so we asked the control question nobody had asked: if you simply destroy the cancer protein, does menin fall off by itself? Genome-wide it does not - menin still sits at four-fifths of its strongest spots three hours after the cancer protein has gone from its own, so something other than the fusion is holding it there.
Menin does not outlast the fusion under revumenib in primary infant material - 0.586 against 0.485 at oncoprotein peaks, a ratio of 1.21 against a pre-registered threshold of 1.5 - so the fusion-independent menin plateau seen after a degron does not reproduce here
After the fusion protein is destroyed in a lab cell line, about half the menin stays put, which is why this project believes something other than the fusion holds it there. In a real patient on a menin-blocking drug, menin and the fusion fall away together. The lab result does not reproduce - but the two experiments remove menin by completely different routes, so this is a weaker blow than it looks.
Menin has never been tested against SET::NUP214 - a bounded absence in three independent registers: literature, cell-line dependency data, and the founding paper's own blots
We checked three different places for any experiment asking whether SET::NUP214 leukaemia needs menin, and found none in any of them. That is a gap in the evidence, not a finding that it does not need menin.
Menin is the best-observed of the six core proteins and the only one whose interval separates from the rest - the interface the programme is built on is the one the record actually supports
Across the six proteins at the centre of this project, most recorded partnerships were never tested one-on-one. Menin is the exception: fewer than half of its partnerships are of that inferred kind, and it has eight partners whose contact has been measured directly - more than the other five put together on that measure.
Menin loses half its promoter signal where the fusion sat, and nothing where it did not
When the cancer fusion protein is destroyed, menin falls away from exactly the genes the fusion was sitting on - about half of it goes in three hours - while menin at every other gene does not budge. The earlier finding that menin 'stays put genome-wide' was counting mostly places the fusion had never been.
Menin-bound genes are the RNA-processing and DNA-repair machinery, H2AK119ub-only genes are neuronal and developmental, and 'menin target' is therefore a much weaker label than this graph treats it as
A published dataset sorts about nine thousand genes into menin-bound, silenced, or both, and nobody had asked what kinds of genes fall in each bin. The menin bin is full of the cell's basic housekeeping machinery, which means calling a gene a menin target says much less than it sounds like.
Menin, ENL and DOT1L are CUT&Tag and the four KMT2A antibodies are CUT&RUN, run three weeks apart - so every menin-versus-fusion comparison in this pair crosses both assay and batch
The experiment that asks 'is menin sitting where the fusion is' compares two measurements made with different chemistries on different days. That does not make the comparison wrong, but it means a difference between menin and the fusion could be a difference between two methods.
Menin's 33 curated partners, enumerated: 10 written, 10 discarded as mutation-directed or single-method, 3 that are not proteins, and what was read and rejected
This is the ledger for the whole beat: every protein a database says menin binds, what the evidence for each actually is, which ones got written up, which were rejected and why. Recording the rejections matters as much as the entries — otherwise the next person repeats the same search and cannot tell what was already considered and dismissed.
Metabolic enzyme enumeration in KMT2A-rearranged leukaemia: the instrument, its positive controls, seventeen bounded absences - and the counted zero that would have been wrong
Before saying nobody has studied an enzyme in this leukaemia, you have to prove your search can find the enzymes people HAVE studied. This entry records that test, the seventeen enzymes that came back empty, and one case where the empty answer was correct as a count and wrong as a conclusion.
MI-2-2 was doing a weak version of the same thing - 5.8x enriched overlap with the potent inhibitor, 75% concordant - and the selectivity claim it left behind gets STRONGER when low-expressed probes are filtered out, 9.6x to 22.1x
A weak and a strong version of the same kind of drug were tested on the same leukaemia cells in two separate public studies, and the genes the weak one moved are mostly a subset of the genes the strong one moved, in the same direction. A separate worry about the earlier analysis - that its headline number came from barely-detectable genes - turned out to be wrong in the reassuring direction.
ML-2 (KMT2A::AFDN, a non-elongation partner) is the line where menin eviction is weakest (retains 37.2% of its own top windows against 20.5% and 13.0%) and where the DOT1L track moves most of any track in the beat (retains 31.6%, 13.69 percentile points below a rank-matched null at p = 3e-10) - the opposite pattern to what an AF9/ENL-free fusion would predict
The third cell line carries a fusion partner that is not part of the transcription-elongation machinery, so DOT1L should have been the least affected thing in it. It is the most affected, and menin is the least - which is the reverse of the expected pattern and worth checking before it is believed.
MLLT10/AF10: twenty observed phosphosites, twelve PIKK motifs, and not one kinase-site relation in the literature - a bounded absence
Phosphates have been seen on the AF10 protein in twenty places, but nobody has ever shown which enzyme puts any of them there or what any of them does. We report the exact queries and counts, because 'nobody looked' and 'someone looked and found nothing' are different things.
Model: interface hotspot prediction - trainable in general, pure extrapolation here, because SKEMPI contains zero menin entries
There is a standard way to build a model that predicts which single amino acid changes will break a protein-protein grip. The training data exists and is large enough. We checked whether it contains any examples from our own proteins, and it contains none at all - so any prediction it made about menin would be a guess from unrelated examples.
Model: predicting who responds to a menin inhibitor - no patient-level data is public, so this cannot be trained either
We were asked whether we could build a model that predicts which patients will respond to a menin inhibitor. Hundreds of patients have now been treated, but the trials publish only group averages. The individual patient data that a model needs is held by the companies and is not released, so the answer is no.
Model: resistance-mutation prediction - cannot be trained. There are six patient-observed labels in the world, and a base-editor screen already does the job better
We were asked whether we could build a model that predicts which mutations make menin inhibitors stop working, and we counted the real examples: six mutations across about a dozen patients. That is not a training set, and an experiment that makes every possible mutation and sees which survive is both cheaper and more informative.
MOLM13 (KMT2A::MLLT3) is the only line in GSE127507 with a matched input in BOTH conditions, and it gives the one differential-input-controlled measurement in the series - MLL1-N occupancy at its own top 0.1% of bins is unchanged (share ratio 0.9438, input 0.9542, input-corrected 0.9891) with ZERO gained bins genome-wide, while menin retains only 13.0% of its own top windows
Of the three cell lines in this dataset only one has the background control for both treated and untreated samples, and it is the line carrying the same fusion as last cycle's mouse work. There the MLL1 signal does not shift at all after correcting for background, while menin keeps only one site in eight.
Neither infant sample has an IgG control, and neither of their two sequencing batches contains one - so every number from this pair is confounded with chromatin accessibility
To know that a protein is really sitting on a gene, you run the same experiment with an antibody that binds nothing and check you get less signal. For these two infant samples that control was never run, and it was not run on anything else processed at the same time either.
Neither rival mechanism leaves its fingerprint, and the best correlate of Fanconi dependency is menin
Two explanations were on the table for why Fanconi-repair genes matter in this leukaemia: a build-up of toxic aldehydes, or stalled DNA copying. Across 54 leukaemia cell lines neither leaves the pattern it should, and the gene whose requirement tracks Fanconi most closely is menin.
NEITHER, and it is the closest call this project has had: KAP1/TRIM28 reads the 90.62nd, 81.85th and 89.45th percentile of MCF-7 CTCF anchors at the exon-11 element in three of four tracks - satisfying the first clause of the pre-registered PRESENT rule - and the six-non-anchor-interval clause blocks it at 65.87
One protein came closer than any before it to being found at the exact spot where KMT2A breaks - in three of four measurements. It still fails, because everything else nearby is elevated too, so the spot is not being singled out.
NG2 in KMT2A-rearranged patients replicates in two independent TARGET profiles - 3 of 4 and 9 of 13 above the 90th percentile of everyone else - but it is bimodal, and the 552-fold median ratio I first computed is an artefact of a near-zero denominator
In children's leukaemia samples, the NG2 protein's gene is switched on far higher in KMT2A-rearranged patients than in any other kind - but only in about half to three-quarters of them, not all. A first pass produced a headline 552-fold difference, which turned out to be an accident of dividing by a number close to zero, and is not reported as a result.
Niche biology HAS been measured in KMT2A-rearranged disease - but in the index fusion itself there is essentially one marrow-niche molecular experiment, from 2011
We counted, molecule by molecule, how much of the bone-marrow niche literature has actually been done in this particular leukaemia rather than borrowed from leukaemia in general. There is real work in KMT2A-rearranged disease as a class, but for the exact infant fusion this project studies, almost none of it exists.
Niche components considered and not written: what the search actually returned for E-selectin, CAR cells, osteoclasts, marrow macrophages and the rest
Before writing an entry for each part of the bone-marrow niche, every candidate was searched in PubMed and the results counted on both sides. This records the ones deliberately not written and why, so nobody spends a day rediscovering that there is not a single paper connecting this leukaemia to E-selectin.
Niche proteins enumerated and NOT written in this tranche, and four symbol traps measured while doing it - three of them new to the registry
This records the niche molecules we deliberately did not write entries for and why, so nobody spends a day rediscovering that there is no literature on them. It also records four gene-name traps that made searches return confidently wrong answers.
Nine of the seventeen core promoters have a non-AFF1 sample within 10% of the threshold, and one has a non-AFF1 sample STRONGER than its own weakest AFF1 call
These seventeen genes are called 'on in KMT2A::AFF1 only' because they cross a 2x line in those patients and not in the others. For nine of them the best non-AFF1 patient is within a tenth of that line, and for one the best non-AFF1 patient reads higher than the weakest AFF1 patient that counted as a hit. The wall between the two groups is thin.
No acetyl-dead and no acetyl-mimetic KMT2A was ever put in a methyltransferase assay - the link from K1130/K1133 to catalysis runs entirely through SIRT1, and that is a weaker claim than the abstract makes it sound
The paper everyone cites says a chemical tag on MLL1 controls how well the enzyme works. Reading the actual experiments, the enzyme test was done by adding or removing the enzyme that strips the tag - never by making a version of MLL1 that simply cannot carry the tag. So the tag and the enzyme activity are linked by association, not by a clean direct test of the two spots.
No BRCA1, BARD1, EXO1, DNA2, HP1-gamma or KDM5A has ever been localised at the KMT2A locus, in any cell type: 21 queries, every querytranslation read back, 0 on-topic records - and the absence splits THREE ways by class, which is what decides what it costs to close
Nobody has ever measured where the DNA-repair machinery sits on the KMT2A gene, in any cell, in any published paper. For two of these proteins the public data already exist and nobody has looked; for four of them no data exist at all.
No experiment has ever ASKED whether MTA2, MBD3, GATAD2A or SIN3A binds a KMT2A fusion - four of them were nonetheless in the lysate of a seven-fusion pulldown, and the class of that absence is measured-not-reported, not never-measured
Nobody has ever done a targeted experiment to see whether these four scaffold proteins stick to a KMT2A fusion protein. But they were all present in the cells of one published fishing experiment that would have caught them, so the numbers exist in a public file and just were not written down.
No fusion oncoprotein in any family has a published replication-protein-A measurement: 23 queries across five fusion families and four spelling variants each, every querytranslation read back, zero records - so the PI's analogy dimension for RPA has no comparator at all, and the index fusion's single locus-level result is the entire field
One of this project's standing questions is whether a finding in the leukaemia fusion we study also appears in other cancer-causing fusion proteins. For the DNA-coating clamp RPA, the answer is that there is nothing to compare with: across Ewing sarcoma, chronic myeloid leukaemia, acute promyelocytic leukaemia and two other childhood leukaemia fusions, not a single paper mentions RPA and the fusion together.
No gene-body-scale IRX2 measurement exists in this project - the one scale that could attribute the signal is the one nobody computed
The only way to tell IRX2's signal apart from its back-to-back neighbour's is to measure along the length of the gene instead of just around its start, because the two genes' bodies do not overlap at all. Nobody has done that in this project, and this entry records exactly that gap rather than filling it with a guess.
No H3K9me3 writer and no H3K9me3 reader has ever been localised at the KMT2A locus IN THE PUBLISHED LITERATURE, in any cell type: 20 queries, 5 alias spellings, 0 on-topic records - and CORRECTED 2026-09-22: the SUV39H1 half is no longer unanswerable from archives. ENCODE still holds zero SUV39H1 experiments (verified by three routes with controls), but GEO holds two deposits, and Q114 read both at this locus and found the writer ABSENT
This graph has argued for several cycles about a silencing tag on the KMT2A gene. Nobody has ever measured where the enzymes that write that tag, or the protein that reads it, actually sit on that gene - in any cell, in any published paper. For the main writer there is not even a public dataset that could be checked.
No HA library exists in the AFF4 sequencing batch, so AFF4 is excluded from the peak partition by name
One protein, AFF4, could not be checked by the peak-counting method, because the experiment that would tell us where the fusion was sitting was never run in the same sequencing batch. This is written down rather than quietly left out.
No HDAC-inhibitor result in this disease attributes to an HDAC enzyme - the only per-enzyme evidence is genetic, and it says the enzyme is not fusion-specific
Every drug study of HDAC blockers in this leukaemia used a drug that blocks several of the eleven HDAC enzymes at once, so none of them can say which enzyme mattered. The one study that did answer the question used genetics instead of drugs - and its answer was that the enzyme it found, HDAC3, is needed just as much by two completely unrelated blood cancers.
No histone-mark readout has ever been measured after perturbing ANY NuRD or Sin3 subunit in a KMT2A-rearranged cell - the two complexes whose entire job is acetylation have zero acetylation data in this disease
NuRD and Sin3 exist to take acetyl marks off histones. In leukaemia driven by a KMT2A fusion, nobody has ever removed one of their parts and then measured what happened to any histone mark. The complexes are discussed as chromatin machines in a disease where their chromatin output has never been looked at.
No IDH inhibitor has ever been reported in KMT2A-rearranged leukaemia: 4 literature records, all reviews, and 1 registry record, an adult master protocol - and the first version of this query returned a false zero
Two IDH-blocking drugs are approved, and the IDH2 mutation they relate to is commoner in KMT2A-rearranged leukaemia than in other kinds. Yet on the day this was checked, nobody had published or registered a study of either drug in that group - and the first search that said 'zero' was broken.
No infant or childhood KMT2A-rearranged cohort has REPORTED IDH status — and at least one has certainly MEASURED it
A co-mutation that changes DNA methylation is more common in adult leukaemias driven by this fusion, and that could matter for how the fusion finds its targets. Nobody has published that number for infants. But a 2025 study sequenced the whole genomes of 36 infants and children with this exact disease, so the answer is almost certainly already sitting in its supplementary tables — it has been measured and not reported.
No junction-specific antibody exists for any KMT2A fusion - and synovial sarcoma shows exactly what one would buy
An antibody raised against the join itself would see the fusion protein and nothing else. One exists for a different cancer - synovial sarcoma - where it replaced genetic testing and works for mapping the fusion onto DNA. Nothing like it has ever been made for any KMT2A leukemia fusion, and our calculation says only about six versions would be needed to cover nearly every patient.
No KMT2A fusion junction peptide has ever been reported - the one measurement that would prove the protein exists has not been published
The short piece of protein that straddles the join is the only thing that can prove a fusion protein is really there, and our calculation says it is measurable. Searching the published literature for anyone reporting one in a KMT2A leukemia returns nothing at all.
No KMT2A-rearranged haematopoietic cell line in CCLE expresses IRX1 - including all four KMT2A::AFF1 lines MV4-11, NALM-19, RS4;11 and SEM at 0.0000-0.0053 RPKM
None of the eleven laboratory cell lines carrying a KMT2A fusion has the IRX1 gene switched on - not even the four t(4;11) lines that are the standard models for this disease. So there is no cell line in which the IRX1 half of this biology can be studied.
No menin inhibitor has ever enrolled a KMT2A-rearranged solid tumour, and the one solid-tumour menin trial that read out was terminated for lack of efficacy
Menin-blocking drugs are approved and in dozens of trials, but every one of those trials is in leukaemia. Only two have ever been opened in solid tumours - one in bowel cancer, which the company stopped because it was not working, and one ongoing in a stomach tumour - and none has ever enrolled a patient whose tumour carries a KMT2A rearrangement outside the blood.
No menin interaction outside the chromatin complex has ever been tested in a haematopoietic cell - 0 of 9, and the compartment is the wrong one in every case
Every experiment behind menin's nine non-chromatin partners was done in kidney, colon, bone, lung or skin cancer cells. Not one was done in a blood cell, which is the only place this project's disease exists.
No menin-inhibitor trial measures homocysteine or any amino-acid endpoint across 57 records - while a recruiting infant-ALL trial already gives PEG asparaginase and ziftomenib together
This project found that a menin inhibitor strongly lowers the gene for an enzyme whose inherited loss causes dangerous blood clots, and it proposed combining that drug with asparaginase, which also causes clots. It turns out the combination is already being given to infants in an open trial - and the simple blood test that would detect the problem is named in none of the 57 trials of this drug class.
No name-based test can clear a gene symbol, and GAPDH is the proof: 48 files, zero name support, unambiguously the real gene - so the triage is one-sided, 145 REAL and 144 ABSTAIN
This project keeps a list of genes it mentions constantly but has never written up, and the list is full of words that only look like gene names. An attempt to sort it automatically failed twice, and the reason it can never fully work is that a real gene is often only ever called by its abbreviation.
No orthogonal validation exists for any of the 110 MLL-AF4-specific partners, as far as four declared bounds can see — and the one partner the paper itself followed up, SETD2, is on the opposite list, shared by all seven fusions
The published list of 110 proteins said to bind only the MLL-AF4 fusion has never been checked by anyone, with any other method, in the eight years since it appeared. The one protein the authors did follow up, SETD2, was on the opposite list — the one shared by all seven fusions.
No preprint in this graph is carried at strong or established - and the audit that found that first misclassified Genes & Development as a preprint
Thirty-two citations here rest on work that has not been peer reviewed, none of them at the two highest confidence levels, and only one failed to say so. My first attempt reported 132, because the web-address prefix used by the main preprint server is also used by several respected journals from the same publisher.
No published biology exists for AFF1-AS1 under any name - and the 2026 review of non-coding RNAs in KMT2A-rearranged ALL does not mention it
The antisense RNA sitting 151 letters from AFF1's start has never been written about. Not in a paper title, not in an abstract, not inside the full text of any open-access paper, and not in the review published this year whose whole subject is non-coding RNAs in exactly this leukaemia.
No published experiment combines an AKAP95 perturbation with any menin or DOT1L inhibitor - bounded by an enumerated 92-record literature, not by a keyword zero
AKAP95 is now a known dependency of MLL-rearranged leukaemia, and menin inhibitors are the approved drugs for that disease. Nobody has yet put the two together in a single experiment.
No two entries name the same gene - 574 symbol-shaped entries, zero duplicate groups - and only two use a symbol HGNC does not approve: gene-kdm3c is JMJD1C and gene-bod1l is BOD1L1
Entries here are named after gene symbols, and some genes have more than one symbol, so the same protein could in principle have been written up twice under two names. It has not been: of 574 such entries, none duplicates another and only two use an out-of-date name.
No washout, withdrawal, release or drug-removal sample exists in any of GSE260742's 158 records, so the third clause of hyp-q117's prediction 5 - that the programme returns on drug withdrawal - cannot be scored from the deposit that prediction names, and closing it needs a wet experiment rather than a download
One part of a standing prediction says the disease programme comes back when the drug is taken away. The dataset that prediction points at never takes the drug away in any of its 158 samples, so that part cannot be checked from it at all.
Nobody has ever asked whether a KMT2A-rearranged sarcoma needs menin: zero records, and the reason is that no model exists
The drugs that work in KMT2A leukaemia work by breaking a grip between KMT2A and a protein called menin. Nobody has ever tested whether that grip matters in the KMT2A sarcomas - not once, in any cell, in any mouse - and the reason is that no laboratory model of these tumours exists.
Nobody has ever published a menin experiment in CALM-AF10 - counted, controlled and dated
We searched for any study that tested whether CALM-AF10 leukaemia needs menin, and found none. To make sure the search itself was not broken, we ran the same searches with the pieces separated and got thousands of hits, so the zero is real.
Nobody has ever shown 14-3-3 binding directly to KMT2A: five records across every paralogue alias, and one of them is about AF4
KMT2A turns up as a 14-3-3 partner in two big screens. We searched the published literature using every name the seven 14-3-3 proteins go by and found no paper that ever tested the pair directly - five records in total, and the only relevant one is about KMT2A's fusion partner, not KMT2A.
Nobody has reported what a menin inhibitor does to CXCR4: zero papers name menin and CXCR4 together, and the RNA-seq that would answer it is already public
The drug class this whole project is about collapses the gene programme that makes leukaemia cells respond to the marrow's hold-still signal, and no one has ever checked whether the receptor for that signal goes down too. The data to check it is already sitting in a public archive.
Nobody needs to enrol anyone: the PTH question is a regression on data that already exists
The key question - is the parathyroid hormone rise caused by the drug or by the kidneys - can be answered with the blood results already sitting in the sponsor's database, using statistics rather than new patients.
Non-elongation KMT2A partners considered and not written: what the searches actually returned, partner by partner
Before writing anything about the cytoskeletal KMT2A partners, this beat counted how much real literature each one has. Six got their own entries; the rest are listed here with the number of papers that exist, so nobody repeats the search.
None of the 166 leukaemia-family GWAS in the Catalog names KMT2A, MLL, 11q23 or infants, although ETV6::RUNX1, TCF3::PBX1 and hyperdiploid ALL each have their own scans - and a 987-case scan could only detect odds ratios of about 1.4-1.6 or more
Researchers have mapped inherited risk for several childhood leukaemia subtypes, but not for the KMT2A-rearranged leukaemia of babies. A study that size would only pick up fairly strong inherited effects.
None of the five selective unwritten interactors survives a null, and FANCD2 - the one that carried a story - is the second weakest of them at p = 0.12
Yesterday we found five proteins that looked more needed by this cancer than by other leukaemias, and built two write-ups on them. Testing whether those differences could have arisen by chance: none of them holds up once you account for having looked at 254 proteins to find five.
None of this graph's 2,196 cited papers is retracted - and I nearly reported that from a query that could not have found one
None of the 2,196 papers this graph cites has been retracted or flagged, and nobody had ever asked - a withdrawn paper would pass every existing check, because its title, DOI and authors all still match. The part worth remembering is that my first attempt would have given the same reassuring answer even if the graph had been full of retracted papers, because I tested the wrong thing.
Not one of the four studies in this dispute measured CTCF or cohesin occupancy in a human blood cell - so this beat read ENCODE instead, and the element is a STRONGER CTCF site in a B-lymphoblastoid line than in HCT116, with the summit at the same base
The whole argument turns on a protein sitting on one short piece of the KMT2A gene in a blood cell, and none of the four papers measured it in a blood cell. So we looked it up in a public atlas: in a human B-cell line the protein sits there more strongly than in the bowel-cancer cells the papers used, and at the very same base.
Not one phosphorylation site has ever been mapped on a KMT2A fusion protein itself - every site in this beat is an inference across the junction
Every phosphate spot discussed in this beat was found on a normal, unfused protein - nobody has ever taken the actual cancer fusion protein out of a cell and asked where its phosphates are. So every statement that the fusion 'keeps' a switch is an argument from where the chromosome broke, not a measurement.
Not re-bounded: an IDH-inhibitor name list crossed with KMT2A terms returns 489 full-text records and 131 even when tightened to the three approved drugs against the exact string KMT2A-rearranged, which is more than this beat could screen - the unrun step, stated with its size
Whether any patient with a KMT2A-rearranged leukaemia has ever received an IDH inhibitor is a question about who was enrolled in trials, and that detail sits in tables inside papers rather than in their summaries. Searching whole papers returns more than a hundred documents and this beat could not read them all.
NOT YET REPLICATED (corrected cycle 97). Within the K562 map KMT2A's 25-of-50 promiscuity is explained by size (97.1st percentile falling to 90.7th among proteins its own length) and menin is under-reached at the 23.5th, surviving an abundance control. The replication attempt was UNINFORMATIVE, not negative: eleven of twelve other proximity panels do not sample the nucleus, and the one that does shows the same shape
A survey using fifty different bait proteins found the gene this project studies turning up near twenty-five of them, far more than almost any other protein, which could mean it genuinely sits in a shared neighbourhood with RNA-processing machinery. It is one of the largest proteins in the human body, though, and bigger proteins are easier to bump into: comparing it only against proteins of similar size, it is no longer exceptional, and similar-sized proteins in the same survey average nearly twice as many hits as ordinary ones. So the striking number is mostly a fact about its size, and the individual proteins found near it should not each be written up as partners.
Nothing beats menin after all, and the statistic that said otherwise ranks narrowness
Last cycle we said one gene looked more important than menin for this leukaemia, because we divided each gene's effect by how variable it was elsewhere. Measured in plain units of effect, menin is first of 8,269 and a cluster of DNA-repair genes near the top turns out to be real.
Nothing connects ENL YEATS binding or ENL condensates to phosphorylation: zero records, and what that means for the phase-2 dossier
The ENL protein's reader pocket and its droplet behaviour are the two things a drug programme here would aim at, and we found no published link between either of them and phosphorylation. That matters before betting on the target: it means the question is open, not that the answer is no.
NUP98 fusions retain a 469-residue FG repeat region that carries ZERO annotated modification sites - the cleanest retained-portion PTM absence in the four fusion families checked
NUP98 fusion proteins keep the first 469 amino acids of NUP98 and throw away the rest. Every chemical modification anyone has ever catalogued on NUP98 is in the part that gets thrown away - the earliest one is at position 524. The retained part is completely blank.
NUP98 is an eligibility criterion in 22 of 57 menin-inhibitor trials including a paediatric trial for that genotype alone - and NSD1 and KDM5A appear in none of them, so trials select on the fusion and never on its partner
A hypothesis here says NUP98-rearranged leukaemia should respond to the same drug class as our index disease, and nobody had checked whether such patients are even being treated. They are, in 22 trials including one for children with that genotype alone, and one complete remission has been published.
Of 262 curated interactors this graph never wrote down, five are selectively required in KMT2A-rearranged lines - and one of them is FANCD2, which binds menin and sits inside a selectivity signal this project has failed four times to explain
We had 262 proteins known to stick to this cancer's machinery that we had never written up, so instead of reading them in order of how often they are mentioned we asked which ones the cancer cells actually need. Five, and one of them ties together two things this project had been unable to connect.
Of 62 registered menin- or DOT1L-inhibitor trials, 7 have stopped and none gave a safety reason - four were company or supply decisions - yet the same drugs carry dose-limiting differentiation syndrome and QT prolongation that the registry never mentions
Seven trials of menin or DOT1L drugs have been stopped early, and none of them said a safety problem was why - most were company decisions. But the published papers show these drugs do cause serious side effects, so the registry's silence tells you almost nothing about safety.
Of every antigen this project names, only NG2, LAMP5 and CD123 are present on all three KMT2A::AFF1 cell lines - CD19 is at 0.2 nTPM in MV4-11 and CD33 at 2.3 in RS4;11
This leukaemia can appear as a B-cell disease or a myeloid one, and the three laboratory lines that carry its fusion gene span both. No lineage marker is present on all three: CD19 is absent from one, CD33 from another. The only surface proteins present on all three are the two that the fusion gene itself switches on.
Of the 58 two-patient AFF1-exclusive genes, four - not one - are called exclusive by both instruments, and the surviving IRX1 claim does not beat a patient-level null (p=0.092) [CORRECTED by q171]
**CORRECTED (cycle 157, by q171).** The headline below is wrong. At the literal criterion there are FOUR such genes, not one, and 34 of the 42 could never have been one because the second instrument never calls them anywhere. The narrower claim that survives - IRX1 is the only gene called in two or more fusion samples and none of the others - does not beat chance once the null respects that two of IRX1's three samples are the same patient (p=0.092). The occupancy line is finished, as cycle 155 pre-registered it would be on this result. --- Last cycle showed that being 'unique to this fusion' is common - 378 genes have it and most rest on one patient. Narrowing to the 58 with two patients and asking whether a second, independent measurement agrees: only IRX1 passes both. But three genes come close, and the project has written 111 entries about IRX1 and three about all of them combined.
Of the 90 papers citing the discovery of KMT2A acetylation at K1130/K1133, the only three that mention MLL-rearranged disease or AFF1 are reviews, and none mentions menin - salvaged from an agent the session limit killed
A 2015 paper found two spots on the MLL1 protein that get chemically tagged under clock and energy control. Ninety later papers cite it. Only three of them mention leukaemia fusions at all, and all three are reviews - so no one has actually tested whether the leukaemia version of the protein carries those tags.
Of the menin co-dependencies, which are actually safe to hit? (run)
**SUPERSEDED (found cycle 158). Do not act on this entry's safety verdicts — read `ana-target-safety-triage-v2` instead.** This version passed **DPY30** as safe to hit using a constraint statistic (LOEUF) computed from too few expected loss-of-function variants to mean anything. v2 gates that statistic on gnomAD's expected-LoF count, treats a reassuring LOEUF at low power as **absence of evidence rather than evidence of absence**, and finds that only **LEDGF and FLT3** earn an unqualified pass. --- Knowing the leukemia needs a protein is only half the question; the other half is whether healthy tissue needs it too. We joined three public datasets for the ten genes that passed our earlier test, and only three came through without a warning flag.
Of three co-folding predictors run without alignments, only the single-sequence one (ESMFold2) recovers menin-MBM1 above shuffled-peptide controls, and its real peptides sit within 0.6 A (core) to 2 A of the deposited structures while shuffled peptides do not - on interfaces that were in its training data
Before trusting a computer model that predicts how two proteins fit together, we tested it on two contacts whose answer is already known, and against scrambled versions of the same protein pieces that should not fit. Of the three models, only one passed on both known contacts, and it placed the real pieces almost exactly where the lab structures put them while the scrambled pieces landed far away. Because those two contacts were part of what the model learned from, this shows the setup works, not that it can find a new contact.
On an absolute scale, H3K79me2 in the index fusion falls 52-fold (EPZ5676) to 91-fold (compound 10) by day 9 and NOT ONE of 17,341 measurable genes holds its signal - the first absolute-scale chromatin statement this project has been able to make about KMT2A::AFF1
Using a fruit-fly yardstick added to every sample, we can say for the first time how much of a chromatin mark is actually lost in the leukaemia cells this project is built around, rather than how it is redistributed. Nine days of a DOT1L inhibitor removes about 98 to 99 percent of it everywhere, and of 17,341 genes with enough signal to measure, exactly zero hold on to what they had.
On one universe with one rule and a structure-respecting null, the two lineages of one infant overlap no more than two unrelated KMT2A::AFF1 infants do
One baby's leukaemia changed from a B-cell type to a myeloid type, and we asked whether the same genes stayed switched on by the cancer protein. Re-measuring it carefully, the two versions of this child's disease share about as many target genes as two completely different children do.
On the absolute scale the H3K79me2 gradient is real and it runs the other way from the share: the most heavily marked decile retains 0.8% and the least marked retains 31%, a 41-fold gradient - and every decile, in both lines and under both drugs, still loses
Genes that carry the most of this chromatin mark lose almost all of it; genes that carry the least lose a smaller share. That gradient is real and not an accounting trick. But once a proper external yardstick is used, every group of genes loses - the ones that looked like they were gaining were simply losing less than average.
One laboratory holds up 92 entries, which counting papers could not show
We had counted how many different papers this project relies on and called that breadth. Papers are not independent if they come from the same laboratory, so we resolved every cited paper to its senior author. A thousand papers turn out to come from under seven hundred groups, and one group in Frankfurt stands behind ninety-two entries — more than any single paper does.
One paper holds up 70 entries, and one dataset holds up 167
We counted what this knowledge graph is actually resting on: one paper supports seventy entries and one dataset supports a hundred and sixty-seven. If either turned out to be misread, a great deal would move at once.
One trial, two published analyses, never crossed: gemtuzumab clearly helps KMT2A-rearranged children, and a common CD33 genotype abolishes gemtuzumab benefit — nobody has asked whether it abolishes it in them
A targeted antibody drug is one of the few clear wins in this fusion, and separately a common inherited variant in the drug's target gene wipes out its benefit across the trial as a whole. Both results come from the same 1,022-child trial and nobody has put them together, so whether children with the fusion who carry the variant benefit at all is unknown.
Only 2 of this graph's 26 bounded absences name a deposited dataset — the corpus asks whether a paper exists and almost never whether the data does
When we say nobody has done an experiment we mean nobody has published a paper, and the measurement may already sit unexamined in a public archive. Only two of our twenty-six such claims check, and the difference matters because one kind of gap needs years and funding while the other needs an afternoon.
Only 20% of this corpus's absences carry a machine-readable query, which I nearly published as an auditability crisis - and then read five of the other 80% and found every one bounded by a better instrument
Most of our claims that something has never been studied do not store a database search, and at first that looked like a large gap in how checkable our work is. Reading a sample showed the opposite: those claims are usually backed by something stronger than a search, such as reading the paper itself.
Only 4 of 43 open menin- or DOT1L-inhibitor trials admit infants under one year (9.3%), and only one (TINI-2) is infant-specific, though infants are about a third of the characterised KMT2A-rearranged population
Almost all the open trials of the new menin-blocking drugs exclude babies, though babies make up about a third of KMT2A-rearranged leukaemia in the largest patient series. Only one of the four trials that do admit infants is built around them.
Only 69 of the 529 entries carrying absence_class are absence entries, and every conceptually novel token authors reached for - partially-measured, measured-and-contradicted, not-worth-measuring - is on an interface entry, where the thing being graded is not an absence at all
This project requires every 'nobody has measured X' entry to say what kind of gap it is, from a fixed list of seven words. Eight entries used words that are not on the list. Rather than treat that as sloppiness, the question was whether the list is missing something. It is - but not where expected. The field has quietly spread far beyond the entries it was designed for, and the odd words all appear on entries describing a protein contact, where what needs recording is a disagreement between two labs, or a result that exists for a different fusion, or a decision not to pursue something. None of those is an absence.
Only BRD4 (mitotic chromosome) and CHD4 (mitotic spindle) of 31 fusion-pathway proteins carry the Human Protein Atlas cell-cycle flag (3.5% of all proteins), and the flag, defined by localisation to mitotic structures, does not mark TOP2A, a canonical G2/M gene
A public atlas flags almost none of the proteins in this leukaemia pathway as changing with the cell cycle, but its flag only tracks movement onto mitotic structures. It misses well-known cell-cycle genes such as TOP2A, so it cannot tell us whether the drugs work better in dividing cells.
Only one public B-lineage cohort annotates the KMT2A fusion PARTNER per sample beside an expression matrix - every other large paediatric leukaemia resource records only 'MLL yes/no' or the 11q23 band
To ask whether a gene behaves differently depending on which partner the KMT2A fusion has, you need a dataset that records the partner for each child. Almost none do - they only record whether a KMT2A rearrangement is present. We found exactly one that does, in the right disease.
Only one query was broken, and it was dragging the whole figure down
After one of our five literature searches turned out to be five times too pessimistic, we re-checked the other four. They were fine, and the overall figure we had published was being dragged down by the one broken search.
Outside ENCODE the resection layer is not one absence but four: 31 BRCA1, 45 CBX3, 22 KDM5A, 14 MRE11 and 8 CtIP ChIP-seq experiments exist in ChIP-Atlas, DNA2 and RAD50 have one each, and only BARD1 and EXO1 have none in any of the three archives
Last cycle this graph decided that nobody has ever measured where six DNA-repair proteins sit on the KMT2A gene, having looked in only one public archive. Looking in the other three shows that four of the six do have public data, sometimes a great deal of it, and only two of them have none anywhere.
p279 and p318 are one patient in two lineage states, and collapsing them removes exactly one gene - NOP58 - from the recurrent core
Two of the eight KMT2A::AFF1 samples come from the same child, sampled twice as the disease changed lineage. Counting them as one patient, the eight samples are seven patients, and one of the seventeen genes turns out to have been counted twice in the same child.
p279 carries the oncoprotein at HOXA AND at IRX1 — the published mutual exclusivity is not mutual exclusivity at chromatin
The textbook says infant leukaemias use either HOXA or IRX1, never both. One of these babies' leukaemias had the driver protein sitting on both at once.
PAF1 and MED29 return the identical occupancy call in all 27 GSE252378 samples, because 86.5% of their promoter windows is the same DNA
PAF1, a protein this project cares about, has another gene starting 271 letters away facing the other direction. When the fusion-protein signal is measured at each of them across 27 patient samples, the two genes give the same answer every single time - because the measurement is mostly reading the same stretch of DNA twice.
Paralogue redundancy does NOT explain anti-selective DepMap results here: clearly selective genes carry paralogues at 30.6% against a 32.7% base rate, with no depletion
When a gene this project expected to matter turned out not to matter in a large gene-deletion screen, two beats explained it the same way: the cell has a backup copy, so deleting one does nothing. Tested properly across 297 genes, that explanation fails — genes with a backup copy are just as likely to show a strong, disease-specific requirement as genes without one.
Partner-exclusivity does discriminate: of four comparator partner groups only AFDN fires, it is 4-of-4 AML, and lineage-matching collapses it to exactly chance while AFF1's class survives the same restriction
The worry was that this project's way of finding genes unique to one fusion partner might flag any group of patients at all, in which case none of its results would mean anything. Testing every other partner group shows it does not - only one rival group scored, and that one turned out to be explained by the patients all having myeloid rather than lymphoid leukaemia.
PCGF1 loss flattens the fusion's genomic distribution - the knockout-over-control share ratio rises monotonically from 0.810 at the highest-signal decile to 1.136 at the lowest while both input pairs stay flat - and the drug does something spatially different: VTP-50469 evicts almost only at the top 0.1% (0.7554 against 0.9937 at the decile below). The drug acts at the peaks, the genotype acts in the bulk
Cells lacking a particular repressive protein hold less of the cancer-driving fusion at the places it normally sits most densely. Because these measurements are proportions, that missing amount has to turn up somewhere else, and this looked across the whole genome to find out where. It is not piling up at any particular new location - it spreads smoothly into the vast low-signal majority of the genome, and the effect grows steadily the weaker a region's normal signal is. The same analysis run on control samples with no antibody shows nothing, so this is not a technical artefact of sequencing depth.
Peak-call retention tracks peak strength, not fusion occupancy, and that inverts the answer
Counting whether a 'peak' is still detected after treatment mostly measures how strong the peak was to begin with. Because the fusion sits on the strongest sites, that counting method says menin is MORE stable where the fusion was - the exact opposite of what the amount of menin actually does there.
Phosphosite counts for the five 3' partners: the atlas beats UniProt 1.4-6.0x, and the retained/lost split is completely different for each partner
We counted every place a phosphate has been seen on the five partner proteins that KMT2A fuses to, using a deep proteomics resource rather than the hand-curated database most people look at. The deep resource finds several times as many spots, and - the point of the exercise - whether those spots survive into the cancer protein differs wildly between partners: essentially all of them do for ENL, and only three of nine for AF9.
PhosphoSitePlus returns HTTP 403 to its own homepage from this environment, so the PTM-database absence named by the cycle-65 census cannot be bounded there - and the positive control is what proves the 403 is a block and not an answer
One of this project's claims is that no phosphorylation site has ever been mapped on a fusion protein, and the right place to check is a database of such sites. That database refuses every request from this machine, including its own front page, so the check could not be run and no conclusion may be drawn from the refusal.
Picking a GEO supplementary file by slot number instead of extension substitutes a depth-normalised bigWig for a raw fragment BED in 107 of this deposit's 171 records
Each sample in this dataset lists two downloadable files, but their order is not consistent between samples. Taking whichever comes first gives you the wrong kind of file about two thirds of the time.
PLANNED: margin-ranked partner-specificity scan over all 6,049 occupied-anywhere promoters, with a depth control
A plan, not a result. Rank every gene switch by how far its KMT2A::AFF1 signal sits above the best non-AFF1 patient, shuffle the patient labels to get a fair comparison, and check whether the ranking finds known target genes better than simple counting does.
PLANNED: re-search the public PXD009338 raw files for the twelve corepressor subunits the published table truncated away - one compute job that closes twelve measured-not-reported absences and adjudicates the module-split hypothesis
The published MLL-fusion pulldown only listed its top three hundred hits per experiment, so a dozen proteins this beat cares about were measured and never written down. The raw machine files are public, so the missing numbers can be recovered with computing time and no laboratory work at all.
Planned: read CXCR4 and SYTL1 out of the menin-inhibitor transcriptomes that are already public, in the same contrast that shows HOXA9 and MEIS1 falling
The measurement needed to test whether menin-blocking drugs switch off the marrow hold-still receptor has already been made and uploaded by other laboratories. This is a plan to download it and look.
PLANNED: refit prey recovery on accessible lysine count instead of protein length, which separates the two mechanisms that could produce the size effect
Bigger proteins are picked up more often in proximity screens. There are two reasons that could be true: a bigger protein has more of the chemical handles the labelling enzyme attaches to, or a bigger protein simply lingers nearby longer. Counting the handles would tell us which.
PLANNED: run the size-matched promiscuity correction for PSIP1 in the K562 BioID2 map, the way it was run for KMT2A - fully specified here so the next cycle executes rather than redesigns it
A recipe, written down so somebody can just run it: work out how many baits find each protein in one published survey, group proteins by size, and see whether LEDGF is found more often than other proteins of its size. It needs no new experiments, only public data.
Pooling the two AF4 antibodies, as the published analysis does, drops AFF1's fusion-site enrichment from 1.8-4.6x the null to 0.8-2.1x - below the H3K4me3 negative control in all three samples
The published study combined two antibodies for the same protein into one signal. Doing that makes the protein look even less like a marker of the fusion than the better of the two antibodies does on its own.
POST HOC: against institution-matched, anchored comparators the same-patient pair has the LOWEST excess in all six factorial cells - which leans to q142, but was not pre-registered
When we compare the one patient's two samples only against other children treated at the same hospital, and always keep one of that patient's own samples on the other side of the comparison, the patient's two diseases share the FEWEST target genes of any comparison we can make. This test was not planned in advance, so it is a lead and not a result.
Prediction 5 of hyp-q117 contains three separable claims - normal eviction, no silencing, and return on withdrawal - and the deposit it names supports one of them only through a clause it never states; the eviction and silencing clauses need tracks the NTvKO arm does not carry, and withdrawal is absent from all 158 samples
A standing prediction in this project says that if you delete a PRC1.1 component, the fusion protein is still removed by the drug but the empty spots are not switched off, and the disease programme comes back when the drug stops. The dataset it points at can only test a fourth thing the prediction takes for granted. Saying which of the three claims was tested is more useful than a verdict.
Premise (iii) was not falsified at this locus - Gothe 2019 measures MLL breakage and MLL-AF4 fusion in the same cells and they move TOGETHER under TDP2 loss; what TDP2 actually falsifies is a premise about break FORMATION that the triad never stated
This project believed that losing the repair enzyme TDP2 raises the number of cancer-causing rearrangements without raising the number of breaks, which would break the chain of reasoning it uses. In the one paper that counts both in the same cells at this gene, they go up together. The real problem is subtler: a break that is counted hours later is a break that failed to heal, not a break that formed.
Prey recovery in the two BioID depositions tracks SIZE and not abundance - and the interactome tail was being scored on the wrong side of the pair, because every one of the 22 partners is the BAIT and every core protein is the PREY
Two big screens are the only evidence for 22 proteins this project lists as partners of KMT2A and its complex. We asked whether those screens pick a protein up because it genuinely sticks to the bait, or just because it is big and common. Size matters a little, commonness does not matter consistently at all - and the bigger discovery was that we had been measuring the wrong protein of each pair.
PROM1 (CD133) occupancy tracks HOXA occupancy at Spearman +0.952 — a second, independent marker of the same split
A completely separate marker that doctors have proposed as a CAR-T target lines up almost perfectly with the HOXA half of this leukaemia. That makes it much less likely the split we measured is an accident.
PROM1 occupancy tracks HOXA at +0.93 in KMT2A::AFF1 and not at all elsewhere; it is near-absent in KMT2A-AF9 and -AFDN despite high HOXA
PROM1 (CD133) was the second marker that seemed to move with HOXA in t(4;11). Across the other fusion partners it does not - and in two partner groups the fusion sits heavily on HOXA while ignoring PROM1 entirely.
PROM1/CD133 is 5.27x higher in normal kidney than in bone marrow - worse than MEN1 on this project's own safety metric - and it was absent from the filter because that filter scored 22 chromatin targets and not one surface antigen
CD133 is the most promising surface handle on this leukaemia and the one an antibody drug would aim at. Asking where it sits in a normal body shows it is about five times more abundant in kidney than in the bone marrow you are trying to treat, which is a worse ratio than the drug target this project already flags for hormone-gland side effects.
PROM1/CD133 tracks the HOXA arm and ANTI-tracks the IRX1 arm in 48 KMT2A::AFF1 infant ALL cases - so the one surface target with a published human kill is depleted in the iALL-IRX1 subgroup
CD133 is the most promising surface handle for treating this leukaemia. In 48 babies with the t(4;11) fusion, CD133 is switched on almost every time HOXA is on - and it is much less often on in the babies whose leukaemia runs on IRX1 instead. A CD133-directed therapy would under-treat that second group.
Proximity-interactome beat: every query run, every paper rejected, and the one resource not reachable
A record of exactly what was searched, what was found, what was read and set aside, and what could not be reached - so that anyone repeating this next year repeats the decisions and not only the searches.
PubMed deletes zero-hit terms from an OR-group's querytranslation, so the translation is a free audit of which drug names in your list actually contributed - and five of mine contributed nothing
When you search several drug names at once, any name that matches nothing simply vanishes from the record of what was searched. That record is printed with every result, so it tells you for free which of your names were real and which were typos or wrong spellings - and one spelling of a compound can return twenty papers while another spelling of the same compound returns none.
Put three descriptions of the same address on one hg38 axis - the exon-11 CTCF site, the intron-10/exon-11 functional cut point, and the region that breaks more in patients who got a KMT2A fusion - and find out whether they are one element
Three separate papers describe something interesting at roughly the same spot in the KMT2A gene, but each uses a different way of naming the position, so nobody knows whether they mean one spot or three nearby spots. Putting all three on the same genomic ruler is an afternoon's work and would either produce a striking coincidence or dissolve one.
Q109 enumeration: the ten panel members that got no entry, the measurement that declined each, the papers read and not used, and four identifier traps measured on this panel
A record of everything this beat looked at and did not write up, with the number that decided each case. Without it a later cycle cannot tell what was considered and rejected from what nobody ever checked.
Q117 predicted an MLL1 ChIP would diverge from FLAG at MYB and agree at MEIS1 - both halves fire in RS4;11 (MYB 1.21 against mouse 0.71, MEIS1 0.25 against 0.23), but the MYB divergence does NOT reproduce in MOLM13, which carries the same fusion as Q117's mouse, so the MYB difference is a partner difference and not an antibody artefact
Last cycle's mouse work predicted that a different antibody would disagree about one gene, MYB, and agree about another, MEIS1. Both predictions came true in the human cells this lab cares about - but the MYB disagreement vanishes in a second human line carrying the mouse's own fusion, so the cause is the fusion partner, not the antibody.
Q117's CpG-island falsifier reproduces in the human index fusion almost to the decimal - gained bins 5.66% on an island against a class-matched null of 4.23%, fold 1.34, Fisher p = 0.49, where Q117 got fold 1.35 and p = 0.53 - so the CXXC-redistribution model fails in human RS4;11 as it did in mouse
One theory said the fusion protein, once knocked off its usual spots by the drug, would resettle on the unmethylated CpG islands its DNA-binding domain likes. Tested in human cells this comes out the same as it did in mouse last cycle: the new spots carry none of that signature, while the old spots are overwhelmingly islands.
Q117's prediction, run: an MLL1 N-terminal ChIP and a FLAG ChIP of the tagged fusion in the same murine cells diverge 1.83-fold at Myb (0.6645 against 0.3623) and agree to within 4% at Meis1 (0.112 against 0.1171) - the direction is exactly as predicted and the pre-registered numeric threshold for Myb still fails
Two years of argument about whether a popular antibody hides the drug's effect can be settled at two genes, because one mouse dataset ran that antibody and a clean tag on the cancer protein in the same cells. At one gene the two disagree by nearly two-fold; at the other they are the same number.
Q142 and Q144 measured the same patient's lineage switch and disagreed in direction - 0.117 'less than unrelated' against 0.400 'keeps 1,894 of 2,045 calls' - and the beat with 351 pairs and a chance floor beats the one with three pairs and none, which retracts a headline I put in a king log
Two analyses looked at the same infant, whose leukaemia changed type, and asked how much of the cancer protein's position on the DNA stayed the same. One said almost nothing stayed; the other said most of it did. The second looked at twenty-seven samples and worked out how much overlap pure chance would give; the first looked at three and did not. The second is right, and a conclusion drawn from the first has to be re-grounded.
q142's exact comparison - p279 against its own switched AML versus p279 against an unrelated infant - replicates in q144's own universe with q144's own rule, and q144 never made it
The earlier analysis that said a patient's two diseases share fewer target genes than two different children was right, and it stays right even when you use the other analysis's methods. The other analysis simply never ran that particular comparison.
q145's anti-correlation p = 0.0288 is a one-sided value; the two-sided exact p is 0.0576 and the rho reproduces exactly
A number this project has been quoting as a significance value was computed by counting results in one direction only. Counting both directions, which is the usual convention, gives 0.058 instead of 0.029 - just above the usual cut-off.
Q153's named threat, run: the reciprocal KMT2A-C arm shows the same bidirectional-promoter enrichment as the fusion-side N arm (44.86% against 43.86%), and IgG shows 33.86% - so the 2.03x is mostly accessibility, carries a real 7-point KMT2A-family component, and is not fusion-specific in any part
Nearly half of this project's promoter measurements sit at spots where two genes start back to back, far more than you would expect by chance. Running the same test on the half of the protein that is NOT part of the fusion, and on a negative control antibody, shows the effect is mostly just open chromatin and is not something the fusion does.
q172's four-line index arm is three KMT2A::AFF1 lines from TWO donors plus an ENL line; dropping the ENL line moves MEN1 7.7x more than DAXX, and menin's lost significance is variance, not effect
A dependency comparison last cycle used four leukaemia cell lines as the 'fusion' group. One of the four carries a different fusion from the one this project is about, and two of the other three come from the same child - so the real evidence is three lines from two people. Removing the odd line barely changes DAXX but changes menin a lot. And menin's apparently weaker result is not menin becoming less needed: the comparison group becomes more menin-dependent, while menin stays exactly as needed as before.
Q25 enumeration: what was searched, found, rejected and could not be read across the five 3' partners
The audit trail for this beat: which searches were run, what came back, which papers were set aside and why, which papers exist and could not be read, and what would change any of those decisions. It is here so that someone repeating this work next year repeats the decisions and not just the searches.
Q48 enumeration: every antigen, identifier and query this beat ran - including what was searched and rejected, and the three identifier traps found live
This is the bookkeeping entry for the surface-antigen beat: what was looked up, what was found, what was deliberately not written, and every search that came back empty with the proof that the search itself worked. It exists so that somebody repeating this work next year repeats the decisions and not just the searches.
Q50 enumeration: every query this beat ran, every zero it recorded with its bound, and the papers it read and did not use
A record of exactly what was searched, what came back empty, and what was deliberately left out - so that someone repeating this work next year repeats the decisions and not just the searches.
Ranked against four fusion families, the KMT2A acetylation site is the only candidate with no evidence on the fusion at all - and the strongest argument for it is one nobody has made
We compared the MLL1 tag site with the equivalent sites in four other cancer fusion proteins. Two of them have been taken all the way - one is the mechanism of a curative drug. The MLL1 site is the only one where nobody has even checked whether the fusion protein carries the tag.
Ranking bins by raw menin signal builds a class that the IgG control piles into eight to thirty-six fold - so the first version of our menin-bound contrast was measuring accessibility, and its own negative control said so before we believed it
To ask whether a drug strips a chemical mark from the places a protein sits, you first have to define those places, and we defined them by where the protein's signal was strongest - which turned out to mean 'where every experiment piles up', because the no-antibody control piled up there hardest of all. Dividing by that control before defining the places fixed it, and the answer changed from 'no effect' to a real one.
Ranking every KMT2A::AFF1 partner on dependency versus essentiality
We scored every protein the fusion touches on one question: is it needed much more by KMT2A-rearranged leukemia cells than by cells in general? Menin wins, LEDGF and a barely-studied zinc finger called ZFP64 come next, and several famous partners turn out to be needed less by these cells than by other leukemias.
Re-adjudicating the standing weak spot: proximity labelling HAS sampled AFF1, KMT2A and five leukaemia fusions - just never a KMT2A fusion
The project has carried a note saying nobody has ever used modern neighbour-tagging methods on this biology. Half of that is wrong: it has been done on AFF1, on KMT2A and on five other leukaemia fusion proteins. The half that is right is that nobody has done it on a KMT2A fusion.
Re-analyse the eight unreported H3K9me3 samples in GSE210480 at the KMT2A breakpoint cluster region - compute only, and it decides whether the rearrangement brake is mark-specific
Eight public sequencing samples measure a silencing tag on the KMT2A gene that the paper reporting them never showed. Downloading and analysing them would settle whether the gene-breaking mechanism runs on one particular tag or on silencing in general.
Re-auditing every cached GEO deposit on three axes after the parser fix: 1,597 comparable title-versus-characteristics pairs across 40 series and 2,620 samples, zero disagreements - and nine series including the one this cycle's beats rest on are abstentions rather than passes, because a 0-of-0 is not a clean bill of health
A previous entry claimed public data archives disagree with their own descriptions about one time in six. That was a bug in our own code, now fixed. Re-checking properly, and on a third kind of label as well, they agree every single time across fourteen hundred checks. The more useful finding is the opposite of a defect: five archives cannot be checked at all, including the one the project leans on most, and 'nothing found' there means 'nothing looked at'.
Re-bound at full text, where methods live: 20 records pair a WRAD subunit with a proximity-labelling method and a KMT2A term, 14 pair one with a tiling-CRISPR term, and none is a proximity, degron or tiling study of WRAD in a KMT2A-fusion line
Whether the four-protein module that partners the fusion has ever been mapped by proximity labelling, degraded on a timer, or scanned domain by domain in this disease is a question about methods, and methods are described in the body of a paper rather than its summary. Searching bodies finds no such study.
Re-bound at full text: 100 records pair an MLLT1 alias with reciprocal-fusion language and 8 with spliced-fusion language, and none stratifies a KMT2A::MLLT1 outcome, transcriptome or drug response by reciprocal or spliced status
Some patients carry the mirror-image copy of the fusion gene as well as the main one, and nobody has ever asked whether that changes anything for this particular partner. Searching whole papers confirms it.
Re-bound at full text: 40 records pair an ENL alias with condensate and phosphorylation language, 11 of them congress abstract books, and none tests whether YEATS reader binding or ENL condensate behaviour is phosphorylation-gated
Whether adding a phosphate switches the ENL protein's ability to read histone marks or to form droplets has never been tested, and reading whole papers rather than summaries does not change that. A quarter of what the search returns are conference programmes.
Re-bound at full text: 65 records name CXCR4 and a menin inhibitor in the same document and the 50 read are all reviews of AML therapy - no study measures the CXCR4 axis under menin inhibition
Nobody has measured what the menin drug does to the signal that keeps leukaemia cells anchored in the bone marrow. Searching whole papers finds sixty-five documents that mention both, and every one read is a review article listing treatments.
Re-bound at full text: AFF4 still has no haematopoietic cell-state result - the loose intersection is 265 records and the tight one 11, and every title read is a review, a complex-membership mention or a congress abstract book
Nobody has shown what cell condition AFF4 holds a blood cell in, and reading the whole text of papers rather than only their summaries does not turn anyone up. What the search returns is review articles that list AFF4 as part of a protein complex.
Re-bound at full text: ASH2L still has no haematopoietic cell-state result - the loose intersection is 429 records and the tight one 17, and every title read is a review, a complex-membership mention or a congress abstract book
Nobody has shown what cell condition ASH2L holds a blood cell in, and reading the whole text of papers rather than only their summaries does not turn anyone up. What the search returns is review articles that list ASH2L as part of a protein complex.
Re-bound at full text: ELL still has no haematopoietic cell-state result - the loose intersection is 1055 records and the tight one 17, and every title read is a review, a complex-membership mention or a congress abstract book
Nobody has shown what cell condition ELL holds a blood cell in, and reading the whole text of papers rather than only their summaries does not turn anyone up. What the search returns is review articles that list ELL as part of a protein complex.
Re-bound at full text: KMT2B still has no haematopoietic cell-state result - the loose intersection is 1065 records and the tight one 13, and every title read is a review, a complex-membership mention or a congress abstract book
Nobody has shown what cell condition KMT2B holds a blood cell in, and reading the whole text of papers rather than only their summaries does not turn anyone up. What the search returns is review articles that list KMT2B as part of a protein complex.
Re-bound at full text: no published experiment removes LEDGF from an established KMT2A-rearranged leukaemia in an animal - the tight intersection returns 12 records of which 6 are congress abstract books and none is a treatment experiment
Every published experiment that takes LEDGF away does so before or as the leukaemia forms, never after it is established in an animal, and reading whole papers rather than summaries does not change that. Half of what the search returns are conference programmes.
Re-bound at full text: no study of MLLT1/ENL function in cerebellum or developing brain exists, and 10 of the 22 full-text hits are congress abstract books - the weakest-bounded absence in the corpus is now the best-bounded of the four ENL ones
ENL is the drug target most abundant in the fast-growing part of an infant's brain, and nobody has studied what it does there. Searching the whole text of twelve million papers instead of just their summaries finds nothing, and nearly half of what it does find are conference programmes.
Re-bound at full text: still no publication puts KDM2B's rDNA repression together with a KMT2A fusion's nucleolar residence - 27 records, 10 of them congress abstract books, the rest reviews - but KDM2B ChIP-seq in MLL-AF9 cells is now deposited, so the rDNA re-map has a second and better dataset
Three separate findings place three different proteins on the DNA that builds ribosomes, and no single paper has ever looked at more than one of them. Searching whole papers rather than summaries confirms that, and turns up a 2026 dataset that would make the missing experiment a computer job rather than a bench job.
Re-bound at full text: still nobody has asked whether a KMT2A fusion GAINS chromatin sites under a menin inhibitor - but the data to ask it are now deposited, 158 ChIP-seq samples including spike-in-normalised FLAG-MLLAF9 under VTP50469, so the class moves from never-measured to measured-not-analysed
Nobody has checked whether the leukaemia fusion protein moves to new places on the DNA when the drug knocks it off its usual ones, and searching whole papers rather than summaries does not turn up anyone who has. What has changed is that somebody has now deposited exactly the right experiment publicly, without asking that question of it.
Re-bound at full text: the H3K79me2 decay half-life is still unmeasured - one paper in 12.4 million full texts contains the phrase 'half-life of H3K79me2', it is an unnumbered assertion citing a 2010 review, and the MLL::AF9 degron time-course that could have measured it uses the word zero times
How long the H3K79 chemical tag survives once the enzyme that writes it is switched off has never been put in a number, and searching the whole text of twelve million papers rather than just their summaries does not change that. The one sentence in the literature that calls the tag long-lived gives no figure and points at a 2010 review.
Re-bound at full text: the LEDGF-after-menin-failure intersection goes from 0 records at abstract level to 116 at full text and 90 with resistance terms, and the 50 read are review articles - no LEDGF-directed agent has been tested in a menin-inhibitor-resistant model
The most obvious clinical use for a LEDGF-blocking drug is after the menin drug stops working, and nobody has tested it. Searching whole papers turns a zero into a hundred hits, all of them reviews that mention both ideas without testing either together.
Re-bound in the matching instrument: the Antibody Registry holds no junction-specific antibody for any KMT2A fusion under five spellings, while the same query template returns the synovial sarcoma clone as a positive control and turns up two leukaemia-fusion junction reagents nobody here had found
The claim that no antibody exists against the join of a KMT2A fusion protein was based on a search of paper summaries, which is the wrong place to look for a laboratory reagent. Searching the catalogue of antibodies instead confirms the claim and, in the same two minutes, turns up two such antibodies for other leukaemia fusions.
Re-checking all 60 bounded-absence claims against the repaired citation index returns zero strong candidates - and that null is not reassurance, because the one absence known to be false ranks first for the wrong reason and its actual refutation is a deposit the method cannot see
Two of this project's 'nobody has done this' claims turned out to be wrong last cycle. This checked the other fifty-eight against every paper the project cites, and found no strong matches — which sounds like good news. It is not: the one claim we already know was wrong is one this method is incapable of catching, so a clean result from it means very little.
Re-measured against IntAct the interactome is 103/217 (47.5%) covered, up from 41.0%, and the read queue is now ZERO - no partner with two publications or two baits lacks an entry, so the instrument's own ranking criterion has stopped selecting anything. It was also persisting the list of partners that DO have entries while counting and discarding the 105 that do not, which is the exact defect cycle 57 fixed for the other list
The project's first priority is to give every documented protein partner of the fusion its own entry, and a tool measures how far along that is by comparing the graph against a curated public database. Re-running it shows just under half the partners now have entries, up from about two fifths, and that every partner backed by more than one study already has one - so what remains is a long tail supported by a single experiment each, and the tool can no longer tell anyone where to start. It also turned out the tool was saving the list of partners already done and throwing away the list of those still missing, which is the only list anyone would actually use.
Re-ran this corpus's literature bounds against the live index: 40 of 40 stored zeros still return zero, none has expired - and 19 stored 'queries' abbreviate their own clauses, so they were never runnable bounds at all
Many of our claims that something has never been studied rest on a literature search that came back empty, and searches go stale as new papers appear. We re-ran forty of them and every one is still empty - but nineteen others turned out to record only a summary of the search rather than the search itself, so they cannot be checked at all.
Re-scored under the measure that passes a degron positive control, cycle 117's MLL1 two-by-two verdict does not merely survive - it strengthens from a 3.33-fold to a 4.64-fold separation between the menin inhibitor and the DOT1L inhibitor, because the MOLM13 arms do not show the peak-broadening that broke the metric elsewhere
A finding published here last cycle was measured with a yardstick we have just shown can be badly wrong, so we re-measured it with the corrected one. It held up, and the gap between the two drugs got larger rather than smaller.
Read at last, against a pre-registered class-matched null: in SU-DIPG13 and SU-DIPG36 the SUV39H1 CUT&RUN mean at the KMT2A exon-11 CTCF anchor is 0.0000 and 1.6660 - the first an EXPLICIT zero record, not missing data - and no one of the seven KMT2A intervals in either line exceeds 1.21x the genome-wide random-tile mean. The pre-registered NEGATIVE prior held.
In two tumour lines taken from children with a brainstem glioma, the enzyme that writes the main silencing tag does not sit on the KMT2A gene anywhere along it - and at the exact 147-letter spot this project cares about, one of the two lines has a measured zero. It is a clean negative, and it says nothing about infant leukaemia, because neither cell is a blood cell.
Read from deposited files today: the KMT2A exon-11 CTCF/cohesin element is a top-0.05% etoposide break site on chromosome 11 in a human B-cell-precursor line and in primary CD34+ marrow - and is NOT one in HCT116, the only cell line in which cohesin was ever degraded
We read the public data files for the exact 147 bases of the KMT2A gene that this project has been arguing about, in ten different published experiments, without downloading anything large. In blood-precursor cells and in real human bone-marrow progenitors, that spot is among the most break-prone one-in-two-thousand stretches of the whole chromosome. In the one cell line where scientists actually removed the ring protein to test whether it causes the breaks, the spot barely breaks at all.
Reading the fusion interactome by COMPLEX rather than by protein — and the strongest signature in it is a known contaminant, which is what makes the rest readable
A single protein turning up in a pull-down is weak evidence, because abundant proteins turn up everywhere. Several members of the same molecular machine is a stronger claim, because a contaminant has no reason to bring its partners. Testing this on the published interactome of the fusion works — but the very strongest hit is a protein-folding machine that everyone knows is a false positive, and that sets the ceiling on how much any other result here can be trusted.
Reading the same-source counter-evidence: 77 notes adjudicated, and the defect is not the one the metric measures
Most of the graph's counter-arguments cite a different paper from the claim they oppose, which looks like a sign of health; about a sixth cite the same paper. We read 77 of those same-paper notes closely and found that almost all are legitimate, but that the few bad ones fail in a way nobody was checking for: the note states a fact from some other paper while wearing the cited paper's identifier.
Redone at 11x the power: at n=124 the truncating-ARID1A group IS materially lineage-concentrated (TVD permutation p = 0.00020, against p = 0.42 at n=11), so this time the lineage test had something real to remove - and the effect survives it at -0.2516. The missense null is tight, CI95 (-0.0242, +0.0473), excluding -0.10. And there are now 4 truncating-ARID1A leukaemia lines, one short of testable
Two earlier checks on this finding were correct pieces of work done on a badly undersized sample of eleven cell lines, and a better source of mutation records raises that to a hundred and twenty-four, so both were redone. The tissue-of-origin check now matters in a way it did not before - at the larger size the mutated lines really are concentrated in particular tissues, where at eleven they were not - and the finding survives that correction essentially unchanged. A second check settles a worry I had recorded against my own work: lines carrying only a single-amino-acid change really do show no effect, rather than the data simply being too thin to tell. And the number of blood-cancer lines with a clearly broken copy of the gene has gone from zero to four, which is one short of the minimum needed to test the question in the disease this project is actually about.
Registry sweep: is there any CXCR4-directed trial in KMT2A-rearranged or infant leukaemia? There is not, and the one paediatric attempt closed with a single patient
We searched the public register of clinical trials for any study that blocks the CXCR4 receptor in children with this genetic type of leukaemia, or in infants at all. There is none — the nearest thing is a trial that opened in 2012, enrolled one child, and closed for lack of patients.
Repairing 85 defective citation blocks, and the finding that the queue's own numbers were wrong in three directions at once
Sixty-seven counter-argument citations were reported as having titles cut short. We re-fetched every one from PubMed and found the report was wrong about most of them: 28 differed only by a full stop, four had the correct title flagged as the wrong one, and a graph-wide sweep found 76 real truncations rather than 67. All of them are now fixed against titles fetched this session.
Resolving all 3,438 cited DOIs, the last identifier class never checked here - 99.8% resolve, every apparent defect is a wrong DOI the corpus quotes in order to correct it, and the three that are genuinely unregistered turn out to carry a working PMID in the same block, because 99.7% of citations here carry two identifiers
Every paper reference in this project carries a digital identifier. The PubMed ones were checked twice; the other kind had never been checked at all, in over a hundred cycles. Checking all three and a half thousand of them found essentially none broken — and the handful that looked broken were wrong identifiers the project had deliberately written down in order to record that it had caught and fixed them.
RETRACTED: 'Sixteen percent of GEO sample labels contradict their own metadata' was an off-by-one in my own parser - GEO's brief format puts the title before the accession, so every sample was scored against the next sample's title; the true rate is ZERO, the deposit named as worst offender has none, and the same bug made me 'correct' an agent whose accessions were right
This entry previously reported that public gene-expression archives disagree with their own descriptions about one time in six. That was wrong: the fault was in the script that compared them, which lined up each experiment with the next one's name. The real disagreement rate is about one in a hundred, and the archive singled out as worst had none at all.
Revumenib does not move H3K27me3 in MLL-fusion leukaemia, in vitro or in vivo: the absence narrows a THIRD time and the experiment's prior collapses
For many cycles this project has said that its cheapest and most valuable experiment is to measure a particular repressive chromatin mark after giving a menin inhibitor, **because nobody had ever done it.** That claim had already been narrowed once. Re-checking it today it is wrong again: somebody has done it - in stem cells taken from human fat, not in leukaemia - and the drug **did not move the mark at all.** The experiment is still worth doing, but the sentence justifying it has to change, and the one existing result points away from the expected answer.
Route 5 rescued 2 of 32, and 21 of the 30 failures returned HTTP 200 with no article - each publisher failing at a constant, recognisable length
We added a new way of reaching papers we could not read - follow the permanent link to the publisher - and tried it on every such paper here, where it worked twice out of thirty-two. More usefully, the failures look identical per publisher, so we can now tell a real failure from a page that only pretends to have worked.
RPA1 and RPA2 are curated HGNC aliases of POLR1A and POLR1B, and RPA34 - a UniProt synonym of RPA2 - is an alias of POLR1G, so three of the ssDNA clamp's names point at RNA polymerase I subunits; the hazard is in symbol-based DATABASE lookups, not in PubMed, and the measured contamination says which
The three proteins of the DNA-coating clamp are called RPA1, RPA2 and RPA3. Confusingly, the subunits of the machine that makes ribosomal RNA are also called RPA1, RPA2 and similar names in older literature and in gene databases. A search on the symbol can therefore hand you a completely different protein from a different part of the nucleus.
RS4;11 - the human KMT2A::AFF1 index-fusion line - has a deposited H3K79me2 pair, a MENIN pair, an MLL1-N pair, a DOT1L pair and an input under the same menin inhibitor in GSE127507. Every track needed to run this beat's test on the index fusion exists. The download is 1.99 GB measured by HEAD request, and nobody has run it.
The disease this whole project is about is driven by a specific fusion, and a human cell line carrying it has all the measurements needed to test this beat's finding - already public since 2019. Running it needs about two gigabytes of downloads and a piece of code that was deleted.
RS4;11 retains 2.5x more H3K79me2 than MOLM13 under EPZ5676 at every one of ten baseline deciles - and 0.71x under compound 10, with the sign reversing in deciles 5-10: the between-line difference survives the baseline confound and fails the drug replication, so it cannot be attributed to the fusion partner
Comparing the two leukaemia lines, one appears to hold on to nearly three times more of the chromatin mark under one drug - a difference that survives every check for the obvious artefact. Then the second drug reverses it. A difference that flips sign when you change the drug is not a property of the cells' fusion gene.
Rule 80 measured: the abstract index missed the best KMT2A measurement for three marks out of seven, and confirmed the absence for the other four
We searched paper titles and abstracts for seven histone tags in KMT2A-fusion leukaemia, then opened nineteen of the actual papers and searched inside them. For three of the seven tags, the single best measurement was inside a paper the abstract search could not find.
RUN (by the King, cycle 159): the eight definitions span 55 points, from 36.8% to 78.8% on one population - every pre-registered reading fires, the beat is NOT refuted, and bare 'chromatin' alone moves it 25 points
**RUN.** The eight ways of counting give answers from 37% to 79% on exactly the same entries — a 55-point spread — so the headline figure was always a statement about the definition. The strictest reading, which asks not just whether a chromatin mark is named but where it acts, gives 24%. --- Before anyone changes the chromatin coverage score or starts filling in fields to raise it, run the same 541 entries through every reasonable definition of what 'covered' means and print all the answers together. If the answer swings wildly between definitions, the number was never measuring the thing people thought.
Run in coordinates: the protective CTCF site and the patient break-prone window are ONE element, it is cohesin-bound, and the 'functional cut point' is 3.75 kb away - two of the graph's three exon-11s were the same exon and the third was not
Three papers each described something at 'exon 11' of the KMT2A gene using a different numbering scheme, so nobody knew whether they meant one place or three. Putting all of them on the same genomic ruler shows that two of them mean exactly the same 1,000 letters of DNA - a spot where the clamp protein CTCF and its partner ring cohesin both sit - and the third means a spot 3,752 letters away.
Run on the reciprocal KMT2A-C antibody instead of the fusion-side KMT2A-N, the same rule gives zero AFF1-exclusive recurrent gene bodies against a null of 0.52 - the gene-scale class is fusion-arm specific
Two antibodies are used: one sees the part of KMT2A that is kept in the fusion, the other sees the part that is lost. Repeating the whole analysis with the wrong-side antibody gives no AFF1-only genes at all. Whatever the six genes are, they are a property of the fusion side of the protein, not of the KMT2A locus in general.
RUN, and the absence closes: all 13 released ENCODE BRCA1, CBX3 and CBX1 ChIP-seq tracks read at the KMT2A locus with three nulls - zero called peaks in 13 of 13, the breakpoint cluster at 0.57-1.24x the genome, and the exon-11 CTCF element BELOW the median CTCF anchor of its own cell type in 10 of 13
Public data already existed that could say whether two DNA-repair proteins sit on the KMT2A gene, and nobody had looked; we looked at all thirteen datasets and they do not. The one place that seemed to light up turned out to be ordinary background once it was compared against the right thing rather than against the whole genome.
RUN, by Q156, within the hour: the whole-table displacement audit - drift is 1.83%, the real blind spot is one row per symbol at 14.69%, and the case that motivated it is not displaced
The check that would say how often this project's gene-start list is out of date. It was specified here as a plan and the coordinator ran it immediately. The list is mostly fine - 82% of entries are on exactly the right spot - and the audit found a different problem instead.
RUN, with its positive control: knocking down KDM4B and KDM4C together raises H3K9me3 at the +1 nucleosome of 2,000 active promoters by 35.7 percentile points, moves the KMT2A promoter's +1 nucleosome by the same 35.3, and moves the 90 kb KMT2A gene body 5.05 points the WRONG way - so the methyl-poor domain is not an erasure product
Switching off both enzymes that erase the silencing tag makes the tag pile up at switched-on genes across the genome, exactly as published. It does not make it pile up over the KMT2A gene, so the tag is missing there for some other reason than being erased.
RUN: endogenous HP1a/CBX5 is DEPLETED at the exon-11 anchor (median 23.94th percentile) and has the lowest KMT2A gene-body reading of any protein in the panel (30.16) - while the phosphomimetic Bio-ChIP mutant 4E reads higher than the phospho-compromised 4A at all seven intervals AND that difference falls inside the genome-wide interquartile range of the same comparison, so it carries no information
The protein that reads the silencing tag is scarce across the whole KMT2A gene and scarcest at the spot where it breaks, agreeing with the tag itself. A version engineered to read the tag better did stick more, but by no more than two such experiments differ from each other anywhere in the genome, so that part means nothing.
RUN: G9a/EHMT2 reads the 95.8th percentile of 56,171 same-cell-type CTCF anchors at the KMT2A promoter in all four MCF-7 ChIP-seq tracks, and a median 33.42nd at the exon-11 anchor - so the cheapest test of hyp-q114's third prediction REFUTES it
The enzyme G9a sits firmly on the switch that turns the KMT2A gene on, in every one of four measurements, but it is unremarkable at the small spot 52 kb away where this gene breaks. So G9a is not the missing enzyme at the breakpoint either.
RUN: KAP1/TRIM28 has a called peak overlapping the KMT2A promoter in MCF-7 in both hormone conditions, 267 bp of it inside the K562 TRIM28 peak Q110 found - and its nearest chr11 peak to the exon-11 CTCF summit is 52,230 bp away
A scaffolding protein that normally helps silence DNA sits right on the KMT2A gene's on-switch in breast cells, at almost exactly the same spot it was already found in blood cells. It is nowhere near the place where the gene breaks.
RUN: KDM4B and KDM4C both read the 76.6th-79.7th percentile of MCF-7 CTCF anchors over the whole 90 kb KMT2A gene body - above the 25-track panel median of 61.39, LOCUS-WIDE ELEVATED - and both are ORDINARY at the exon-11 anchor with replicate spreads of 92.3 and 61.5 percentile points
The two enzymes that remove the silencing tag are spread across the whole KMT2A gene at a slightly raised level, and neither is concentrated at the spot where the gene breaks. Several unrelated proteins now show the same pattern, which usually means the pattern is about the region rather than the proteins.
RUN: ten ENCODE peak sets and fourteen signal tracks read at the KMT2A locus in two blood cells - not one called peak for SETDB1, CBX5, CBX1, CBX3 or TRIM28 anywhere in the gene, while G9a/EHMT2 and TRIM28 both sit on the KMT2A PROMOTER and neither touches the breakpoint cluster region
Nobody had ever looked in the public data to see whether the enzymes that write the H3K9 silencing tag, or the protein that reads it, actually sit on the KMT2A gene in a blood cell. They do not. But the enzyme this graph says protects the gene when you remove it, G9a, does sit there - at the gene's switch, not at the part that breaks.
RUN: the eight unreported H3K9me3 samples of GSE210480, read at the KMT2A CTCF element - H3K9me3 does not move where H3K9me1 moves 3.9-fold, and the me1/me2 positive controls passed
A 2023 paper measured a third silencing tag on the KMT2A gene and never showed the result; the files have sat in a public archive since. We read them. The tag the paper showed jumps almost four-fold at the key spot; the tag it hid does not move there at all.
RUN: the IDR-free, signal-level re-read of the KMT2A exon-11 anchor against a class-matched CTCF-anchor census - the pre-registered prior of NEGATIVE HELD, both arms score FAILS on the pre-registered rule (BRCA1 median percentile 62.8 of 6 tracks, CBX3 55.9 of 3), and the one thing the prior did not predict is that 2 of 6 BRCA1 tracks sit above the 90th percentile and both are the non-blood cells
Two repair-related proteins were measured at one small spot inside the KMT2A gene in every public dataset that allows a fair comparison, and compared not to the whole genome but to tens of thousands of similar spots in the very same cells. On that fair comparison the spot is unremarkable, which is what had been predicted in writing before the data were opened - except in two non-blood cell types, where one of the proteins really is among the top few per cent.
Running the HOXA programme is not enough to need menin
Two leukaemia cell lines switch on the same set of genes as our disease but by different routes. One of them needs menin even more than our own lines do; the other does not need it at all. So needing menin is not simply a consequence of switching those genes on.
Running the wrong-paper check over the 17,925 PMID citations - the larger half of the corpus and the half where both documented failures happened - gives 100.0% year agreement and 99.7% on authors, one genuine defect, and shows the DOI side's 92% was an artefact of comparing against a single date field
The checks built over the last few cycles all ran on one kind of paper identifier. The other kind is more common here, and is the kind behind both mistakes this project has actually made. Running the same checks over it found essentially nothing wrong — one citation that had stamped this year onto a paper from 1995 — and revealed that an earlier result looked worse than it was, because the two databases mean different things by 'the year'.
S821 is not phosphorylated in any dataset we could query: the 14-3-3 motif candidate is dead as stated
One spot on the KMT2A protein looked like it might be a second docking site, but only if the cell actually puts a phosphate there. We checked a phosphate database five times deeper than the one used before, and the spot is empty.
SALVAGED from a beat killed by a session limit: the PXD009338 spike control passes decisively, but AFF1 - the index fusion's own partner - has 33 PSMs and ZERO above the identity threshold, so that criterion cannot discriminate any of the six candidates
A beat was sent to reopen the raw evidence behind six candidate partner proteins and was cut off by a usage limit before it could write anything up. Its working files survived, and they contain the answer. The check that the method works passed easily. But the key test - whether a protein's peptides clear a confidence threshold - is failed by the fusion's own known partner, which means the test cannot tell a real partner from a contaminant here. Five of the six candidates score zero on it, and so does the protein we are certain about.
SALVAGED from a beat killed mid-write: with direct DOT1L inhibition as the positive control that bounds the achievable effect, VTP-50469 achieves only F = 0.09-0.25 of it. SGC0946 strips H3K79me2 to 6.7-14.8% of DMSO; the menin inhibitor leaves 77-91%. And the bin-wise responses do not correlate (rho = -0.026)
Two drugs can both be said to reduce a chemical mark on DNA packaging, but until you measure how much is achievable you cannot say whether one of them does much at all - so this compared the menin inhibitor against a drug that blocks the enzyme writing the mark directly, in the same cells and the same experiment. Blocking the writing enzyme removes almost all of the mark, leaving only about 7 to 15 per cent of the untreated level, whereas the menin inhibitor leaves 77 to 91 per cent, meaning it buys between a tenth and a quarter of what is actually achievable. The two drugs also do not shift the mark at the same places in the genome, which argues that whatever the menin inhibitor does to this mark does not work through the writing enzyme's pathway at all.
Salvaged from a session-limit death: the DMSO-versus-DMSO selection control cannot be run cleanly on GSE260742, because every same-condition vehicle pair is confounded - two differ about three-fold in aggregate signal and the only admissible same-mark pair crosses processing states (filt.bam.tdf versus tdf). Its automated verdict of MANUFACTURED is NOT supported, and the control remains UNRUN rather than passed or failed
Four recent results depend on sorting genomic regions by how much of a protein a drug removes, then measuring a second change in those same regions - a design that can invent a pattern out of nothing. The proper check is to run it on two untreated samples, where no drug effect can exist. That check was attempted and the agent doing it was killed by a usage limit mid-way, but its computed output survived: it turns out there is no clean pair of untreated samples to use, because every candidate pair differs either in sequencing depth or in how the depositor processed the files. So the check is still outstanding, and neither confirms nor refutes the earlier results.
SALVAGED: in K562, which carries no KMT2A fusion at all, a menin inhibitor moves H3K79me2 as much as it does in fusion-bearing murine MLL-AF9 (V 0.8291 vs 0.7734, F 0.1831 vs 0.2500). Both the confirmation and the refutation fail to fire, and the refutation misses by 0.0057 - this is at-the-edge and must not be read as a clean negative
If a drug aimed at a fusion protein changes a chemical mark on DNA packaging, the obvious question is whether it still does so in cells that have no such fusion - and this compared exactly that, using a leukaemia line with no KMT2A rearrangement. The drug moved the mark in that line by very nearly the same amount as in cells that do carry the fusion, which points away from the change being a consequence of hitting the fusion. The honest caveat is that the pre-set test for 'the effect does not need the fusion' missed its threshold by less than one part in a hundred and a second test also failed to reach its own bar, so this sits genuinely on the fence rather than settling the question either way.
Scored against the literature: the composition half of the YEATS-selectivity objection holds and the selectivity half is refuted — a scaffold-blind probe already has an in vivo window in the index fusion with no HSPC toxicity
The worry was that because the two mark-reading proteins sit in three different machines, a drug aimed at them cannot be specific to the leukaemia. The composition part of that is true. The conclusion is not: somebody already ran the drug in mice with this exact leukaemia and it killed the cancer without harming blood stem cells.
Scored against the same DMSO in the same batch, degrading the fusion leaves 1.0% of its chromatin sites, the DOT1L+menin combination leaves 10.7%, the menin inhibitor alone leaves 51.7% and the DOT1L inhibitor alone 75.5% - and in a second independent batch the menin inhibitor again beats the DOT1L inhibitor on the fusion's own channel
Destroying a leukaemia-driving protein outright clears 99% of its spots on DNA, the two available drugs together clear about 89%, and either drug alone clears far less - roughly half for one and a quarter for the other. This is the first time this project has been able to say how far short the drugs fall of simply removing the protein, because the comparison needed a yardstick nobody had.
Scored at complex level, coherence predicts nothing (rho -0.016 over 719 complexes) - but the Fanconi core complex is 2nd of 719 and the HBO1/MOZ acetyltransferase modules are 1st
We tested our own explanation for why one DNA-repair group looked important in this leukaemia, this time across 719 protein machines instead of a handful of pathways. The explanation failed, but the group itself came second out of 719, and a different set of machines came first.
Scored in the published criterion's own currency, the number of oncoprotein-target promoters runs 0 to 2282 across 27 KMT2A-rearranged patient samples - 20.01x between the 10th and 90th percentiles - and the spread widens to 61.10x after depth matching
Counting how many gene promoters pass the published test for fusion binding, one child's leukaemia gives a few dozen and another gives a couple of thousand. Making every sample the same sequencing depth does not shrink the gap.
Scoring complexes instead of genes, and the controls disagree with each other (run)
Every dependency test here has scored one gene at a time, which is the wrong unit for a disease about protein machines and is blind to genes that cover for each other. Scoring whole machines instead gives a different picture — and two control machines that should have behaved the same way behaved oppositely, which undermines an assumption this project had been making.
Scoring cycle 104's mark panel against a floor measured in its OWN arm: 9 of 10 mark-by-arm cells at D1 land inside the floor's own spread, so D1 is not resolvable from this deposit at all - and the same mark ratios give 0, 4 or 4 breaking marks depending only on which floor was imported, so cycle 104's pass, cycle 107's F1 and the GSE299100 beat's failure were all measuring the ruler rather than the marks
To say a chemical mark changed, you need to know how much two repeats of the same experiment differ - the noise floor. Three previous rounds of this project each picked a different floor and reached a different conclusion about the same measurements. Measuring the floor separately for each arm of the experiment shows why: within a single arm the floor itself varies as much as the effects being measured. The honest answer is that this dataset cannot settle the question, and the disagreement was never about biology.
scripts/is-symbol-covered.py returns 'no match' for BRAF, AKT1, EGFR, CD226 and IRF2 - all five have a declined interface entry. The `interface-` prefix in its PREFIXES list can never match anything.
The tool everyone is told to run before writing an entry, so as not to duplicate one, answers 'not covered' for proteins that already have an entry. It looks for a file named after the protein, and this project names those files after the pair.
Searching 32 GEO series and 2,312 samples for a menin readout under DOT1L inhibition: exactly ONE distinct experiment exists, MOLM13_MEN1-ChIP_EPZ5676_4d at 100 nM, and it is the sample Q136 already used - so the deflationary reading of Q136 is blocked not by a missing assay but by a ~50-fold DOSE GAP, and Q136's drug-class conclusion rests on a single unreplicated sample in a single cell line
A recent result here said a menin-targeting drug and DOT1L-targeting drugs do very different things to menin on DNA. The obvious way to attack that result is to ask whether the DOT1L drugs would look the same if given at the much higher doses used elsewhere. Searching every relevant public dataset shows that the measurement needed exists exactly once, at the low dose, and it is the very measurement the original result used. Nobody has ever measured menin under a high dose of a DOT1L drug, so the objection cannot be tested - and the original result turns out to rest on one unrepeated experiment.
Selective in the disease and quiet everywhere else: MYB comes out cleanest
A good drug target has to be needed by the cancer and not needed by the rest of the child. We had one list for each and had never put them side by side. Doing so puts MYB at the top.
SEM, the standard t(4;11) model line, has zero signal at IRX1 — every model of this disease is a model of one subgroup
Almost everything we know about this infant leukaemia comes from one cell line. That line is firmly in the HOXA half and has nothing at IRX1 — so the worse-prognosis half of patients has effectively never been modelled.
Seven KMT2A-rearranged sarcomas retain the menin-binding motif, three of them carry the complete leukaemic N-terminal module, and all sixteen partner names were in one unopened supplementary table
Last cycle this project concluded that the KMT2A sarcomas all throw away the piece of the protein that menin drugs grip, based on the three tumour types anyone had examined. The full list of sixteen was sitting in a free spreadsheet attached to the original paper, and seven of them keep that piece, three of them keep the entire machinery that drives the leukaemia.
Seventeen entries declare themselves part of a causal chain that does not exist, and the validator could not see it
Entries here point at each other through formal links, which are checked, and through fields that merely name another entry, which were not - and 61 of those names point at entries that do not exist. Most are deliberate promises to write something later, but one is named by seventeen different entries as the backbone they belong to, and that backbone was never written.
SGC0946 removes 86% of H3K79me2's share at the fusion's own sites and the four eviction quartiles are 0.1396 / 0.1453 / 0.1452 / 0.1360 - span 0.0036, Spearman +0.400, a bootstrap interval spanning zero - so a very large effect on this mark can carry no eviction-depth structure whatever
A drug that blocks the enzyme making the tag removes almost all of it everywhere, evenly. The menin inhibitor removes far less of it overall but removes it in a strict pattern that tracks how much fusion protein was pulled off. Size and shape are different measurements, and only the shape says which drug was given.
siMA6 confirmed junction-spanning by reconstruction - 8 KMT2A bases plus 13 AFF1 bases - and it targets the 1362/1363 breakpoint, not the one this graph modelled as canonical
A short RNA used to switch off the fusion gene was said to sit exactly across the join between the two halves. Rebuilding the join from scratch shows it does - and shows which of two possible joins the cells in question actually carry.
Six countermeasures, no head-to-head — so rank their targets: CDK6 is 13th and CDK4 is 16,052nd
Six different research groups have each proposed a different drug to add when menin inhibitors stop working, and nobody has compared them. We could not run that comparison, but we could ask a cheaper question of each one: is the protein it targets actually more needed by this leukaemia than by others? Four of the six clear that bar and four do not — and one widely used drug class turns out to contain one target near the top of the list and one near the bottom.
Six files titled H2AK119ub rep1 are FLAG ChIPs, so prediction 5's eviction clause - declared UNREACHABLE by Q124 - was computable from bins already on disk. It is now scored and it CONFIRMS: VTP-50469 removes the same share of fusion occupancy in the PCGF1 knockout as in the control, 0.7655 against 0.7554, a 1.3% difference against 22.8% genotype variation on a control drug
A prediction this project has carried for months says that in cells lacking a particular repressive protein, the drug should still strip the cancer-driving fusion off DNA but the freed sites should fail to switch off. The second half was measured last month; the first half was declared impossible because the necessary files did not exist. They did exist - six of them are labelled as one experiment in the public database while the depositor's own filenames say they are another - and they were already sitting processed on this machine. Measured, the first half of the prediction holds: the drug removes just as much fusion in the mutant cells as in normal ones.
Six gene bodies are AFF1-exclusive and recurrent at gene scale; the authors' own pipeline confirms one of them, refutes three and has never measured two - and IRX1 is the one it confirms
Exactly six genes in the whole genome are hit by the fusion protein in three or more of the eight AFF1 leukaemias and in none of the other eighteen patients. Five of those six have a non-AFF1 patient sitting just under the cut-off, and the scientists who made the data call three of them in non-AFF1 patients too. IRX1 is the one where the nearest non-AFF1 patient is eight times below the bar and the original pipeline agrees it is AFF1-only.
Six GSE260742 samples titled 'NTvKO ... H2AK119ub rep1' carry ChIP_FLAG filenames AND a ChIP_FLAG string in the TDF header their own depositors' igvtools wrote - two of three independent sources say FLAG, the shape test is near-indecisive, and the channel of half the replicate arm cannot be settled from the deposit
Six public files are described in one place as measuring a silencing mark and in two other places as measuring the fusion protein itself. Which is right decides whether half of a replicate set is usable, and whether a completely different question becomes answerable. The data itself does not clearly say.
Six of 85 proteins are both selectively required and drugged — and MEN1 is not close to the others
We already knew, separately, which proteins this leukaemia needs more than other leukaemias do, and which proteins anyone has ever made a molecule for. Putting the two lists side by side sorts every protein in the graph into four groups. Only six are in the group you could start work on tomorrow, and fifteen are things the disease clearly needs that nobody has a molecule for at all.
Six times the null pool, and the criterion I had been using turns out to be the problem (run)
Two earlier versions of the same test disagreed about three genes. Running it again against six times as many random comparison genes settles the disagreement, and shows the disagreement was caused by how we were scoring the test rather than by the genes themselves.
Stratifying by eviction depth does NOT rescue H3K79me2 as a resistance-detecting readout: at the most-evicted quartile the PCGF1-knockout and the control are 0.6123 against 0.5860, a difference of 0.0263, while the whole genotype effect sits at the LEAST-evicted quartile (-0.1401) - the opposite of eviction-graded
A cell engineered to carry a known resistance change loses the activating tag just as completely as a normal cell does at the places the drug clears the fusion from. Splitting the measurement up did not reveal a hidden difference there - the only difference is at the places the drug barely touches.
Stratifying the MOLM13 co-occupied bins by signal amplitude qualifies this beat's own headline: at the strongest tenth the MLL1-N channel falls to 0.5304 and H3K79me2 to 0.7614, so 'MLL1-N does not move' and 'the mark does not move' are both properties of the aggregate - while the ordinal swap that refuted the four-way ordering survives in every stratum
Splitting the sites by how much signal they carry shows that at the very strongest ones everything moves, including the two things reported as not moving. The surprising ordering result survives the split; the reassuring flatness does not.
Summit containment passes the positive control that the bin-overlap measure failed, in all three HA batches
An earlier check of this dataset used a method that claimed the destroyed cancer protein was still sitting on most of its sites - which cannot be true. Switching to a stricter method gives 2-9% instead of 40-93%, which is what a destroyed protein should look like.
SUPERSEDED IN PART, CYCLE 56 - the MI-2-2 response is ~9x fusion-selective (that stands), but 'does not touch HOXA or MEIS1' was a fact about a weak compound: VTP-50469 collapses MEIS1 -5.25 log2 in the same line
A public dataset tested a menin-blocking drug on two leukaemia cell lines carrying the fusion this project studies plus one line without it, and the effect was about nine times larger in the fusion-carrying lines. We also reported that the disease key genes did not budge, and cycle 56 showed that was only true of this weak drug: a stronger one from the same class switches them off hard in the same cells.
SWI/SNF enumeration for KMT2A-rearranged leukaemia: every subunit searched, every count dated, and what was read and rejected
This is the bookkeeping entry for the SWI/SNF beat: which genes were searched, with which alternative names, what each search returned on which date, which papers looked relevant and were not, and which subunits were deliberately left without an entry.
SWI/SNF split by sub-complex and by lineage: every KMT2A-selective subunit is in the ncBAF module and none is in canonical BAF
We scored 28 SWI/SNF genes for how much KMT2A-rearranged leukaemia cell lines need them, splitting the cell lines into myeloid and B-lineage so that a myeloid gene could not masquerade as a KMT2A gene. Only four genes passed, and three of them belong to the same minority flavour of the machine.
Systematic enumeration of menin-inhibitor resistance mechanisms: seven entries written, thirteen candidates rejected with reasons
This is the working record of a systematic sweep of the published literature for every way a leukaemia can escape a menin inhibitor. It lists the mechanisms that turned out to be real and got their own entry, and — just as importantly — the ones that turned out not to be documented, so nobody has to search for them again.
Target safety triage v2: constraint with its uncertainty attached (run)
Our first safety screen passed a gene because a genetic statistic looked reassuring, when in fact that statistic is meaningless for a protein that small. This version refuses to use the statistic unless there is enough data behind it to mean anything.
TB051939 is rank 12 of 26 and owns fewer singletons than its even share; the library that owns 22.4% of singletons destroys no exclusivity at all; and dropping any single comparator moves the exclusive set by a median of 2 genes
Three of the four genes that looked specific to this fusion were seen in only one patient sample, so the obvious worry was that that sample is simply noisy and finds things everywhere. It is not - it is middle of the pack, and produces fewer one-off findings than an average sample. A different sample does produce a fifth of all the one-offs, but removing it changes nothing, because its extra findings are in genes that were never candidates. Removing any single comparison sample shifts the answer by about two genes out of 63.
TB920488, the other HOXA-negative IRX1-negative blast, has the highest GMP-over-B-lineage ratio of any B-lineage sample in the deposit - and it fails both the matched null and the N-over-C test
The second of the two patients whose fusion protein is on neither usual gene set does look, on one measure, as though it has moved to myeloid genes. Two stricter checks say no.
Ten affinity measurements for this contact and not one uses a fusion protein, full-length KMT2A or a nucleosome
Every published number for how tightly the fusion grips DNA was measured on a small piece of the protein, alone, against a short piece of naked DNA. Nobody has ever measured the actual fusion protein binding actual chromatin.
Ten of menin's 33 IntAct partners come from one BRET screen for MUTATION-created interactions — and five are gains caused by a named allele
A database says menin binds BRAF, PTEN, TP53, EGFR and AKT1, and it does not: those rows all come from one screen that deliberately looked for interactions created by cancer mutations. Writing them in as menin partners would have put five false entries into the graph at once.
Tested in an independent deposit with an epitope-tagged fusion: of the pivotal paper's three named losers under a menin inhibitor, MEIS1 and JMJD1C reproduce and MEF2C does not; of its three named retained loci, the HOXA cluster reproduces and MYB does not - and GAPDH, the negative control, loses half its share too
The paper that launched menin inhibitors listed which genes lose the fusion protein under the drug and which keep it. Re-measuring that list in a different laboratory's public data, two of the three named losers and the famous retained cluster come out the same, two do not, and even the housekeeping control loses half its signal - which is the scale everything else has to be read against.
Testing every pattern against the sentence it must not match
Three of our own quality measures turned out wrong in one cycle, each because a text-matching rule caught something it was supposed to ignore. This checks every such rule against an example it must reject.
Testing my own diagnosis from last cycle: 809 of 1,052 entries that state a quantity in prose already carry it as typed data - 76.9% - so the corpus was machine-readable, the detector was simply weak, and the reporting field I imposed on every future beat rests on a different and narrower justification than the one I gave
Last cycle a tool I built to spot when this project measures the same thing twice did not work, and I concluded the cause was that the numbers are written into sentences instead of into proper fields — then changed the instructions for every future analysis on that basis. Checking the claim shows it is wrong: three quarters of those numbers are already in proper fields. The tool was just weak. The instruction change is still worth keeping, but for a different reason than I gave.
Testing this corpus's own recorded refusals instead of its gap list overturned the first refusal tried, and the refusal had named the falsifier itself
Last cycle showed this project's list of missing gene write-ups was empty - everything on it had already been decided against. So this cycle tested the decisions themselves. The very first one checked turned out to be wrong, and the evidence that refutes it had been published five years before the refusal was written.
Testing whether every citation points at the paper it claims: 314 scoreable blocks, 242 scoring below 0.3 on title overlap, and every low scorer read is CORRECT - because this project's citation title is a descriptive sentence by house style, which means a PMID pointing at the wrong paper is undetectable by any automated check over the corpus alone
Last cycle this project confirmed every one of its 1,908 cited paper identifiers points at a real paper. That is not the same as pointing at the RIGHT paper, and this project has twice cited an identifier that led somewhere real and entirely unrelated. Testing for that turns out to be impossible with what is stored: the project's notes record a description of why a paper matters, never the paper's own title, so there is nothing to compare against. The fix is cheap and is now specified.
The '0.400, keeps 1,894 of 2,045 calls' in the dispatch brief is two independent per-sample call counts read as a retention ratio; the intersection is 1,116 of a 2,812 union
A summary of an earlier result said one sample kept 1,894 of another's 2,045 target genes. Those are just the two samples' separate totals sitting next to each other in a table; the number they actually share is 1,116.
The 'named but unwritten' queue flags ALL EIGHT of this project's negative-control loci plus one alias-filed entry - 20.5% of it, after I corrected my own reimplementation which had double-counted collisions the tool already blocks
**Corrected: about a fifth, not well over a third.** Part of what I counted as noise was already being filtered by the tool itself — I rebuilt the tool's logic instead of running it, and left out one of its steps. --- One of this project's tools lists genes that get talked about a lot but never written up, as a queue of work to do. Auditing it: every one of the eight genes the project uses as negative controls is in that queue - they appear in dozens of analyses precisely because they are the controls, and the tool cannot tell that apart from a real gap. Adding the already-known name clashes and one gene whose entry exists under a different name, well over a third of the queue is noise the project generated itself.
The 'never duplicate an entry' rule has held across 1,761 entries - and the audit that checked it would not have caught the near-duplicate that prompted it
This project insists on one entry per thing and has never had a way to check. Comparing all 1,761 entries against each other finds no real duplicates, and the handful of similar-looking pairs turn out to be exactly the distinctions the rule exists to preserve, like a drug and the mechanism it works through.
The 'raw library promoter mean' that two cycles treated as library coverage tracks the in-promoter fragment FRACTION at Spearman +0.993 and true sequencing depth at only +0.292, and the oncoprotein share correlates with true depth at -0.255
A recent conclusion said one patient sample looked low on fusion protein only because it had been sequenced shallowly. It was not sequenced shallowly - it sits in the middle of the pack for read depth. The number that was called 'coverage' is really a measure of how concentrated a sample's signal is at genes, which is a different thing.
The 'sixteen percent of GEO sample labels contradict their own metadata' finding is a one-line off-by-one: 2,280 of 2,280 titles were attributed to the wrong sample, and with the record boundary fixed the rate falls from 96 of 586 comparable pairs (16.4%) to 7 of 656 (1.1%) - and all seven, read by hand and named here, are tokeniser artefacts rather than contradictions
A previous analysis in this project concluded that one archived experiment in six is labelled inconsistently with its own description, and the worst dataset two in five; the code that found this was reading each experiment's name off the next experiment's record. With the bug fixed the rate is 7 in 656 (1.1%), and reading all seven by hand shows each is a wording mismatch in our own matching code rather than a real contradiction.
The 0.069 chance floor quoted beside the 0.400 same-patient Jaccard is the cohort MEDIAN's null, and this pair's own floor is 0.267
An earlier analysis said one patient's two leukaemias shared six times more target genes than chance. The 'chance' number it used belonged to a different, much smaller comparison; using the right one, the excess is about one and a half times, not six.
The 1 kb bin-overlap retention measure this project uses for eviction FAILS a degron positive control: it reports the fusion as still occupying 40-93% of its own sites in cells where the fusion has been destroyed, because a collapsing ChIP signal makes MACS call FEWER but 1.9x WIDER peaks - so it fails hardest exactly where the protein is most completely gone
This project measures whether a drug knocks a protein off DNA by asking how many one-kilobase chunks of the genome still overlap a detected signal, so we tested that yardstick on a case where the answer is certain - the protein had been deliberately destroyed - and the yardstick said it was still there. The reason is that when a signal collapses the peak-detection software returns a few very broad, blurry regions that blanket the chunks without actually sitting on the original spots.
The 110 MLL-AF4-specific preys are 14.5% mitochondrial and 34% nuclear; the 42 pan-fusion preys are 2.4% mitochondrial and 88% nuclear — Fisher p = 0.042 on a nuclear-extract pulldown
The pulldowns were done on nuclear extracts, so anything that lives in mitochondria should not be there. One in seven of the proteins said to be unique to MLL-AF4 is a mitochondrial enzyme, against one in forty of the proteins shared by all seven fusions. That pattern is what mass-spectrometry background looks like.
The 17-promoter AFF1-only recurrent core is not more recurrence than a label permutation of the same matrix gives
Seventeen gene switches look like they are turned on only in the KMT2A::AFF1 leukaemias and in three or more of the eight samples, which feels like a shared programme. When we shuffle which eight of the twenty-seven patients we call 'the AFF1 group' and re-count, a random group of eight gives nine such genes on average and twenty-one at the 95th percentile - so seventeen is unremarkable.
The 2015 paper did not discover the KMT2A acetylation site - it inherited it from the 2009 Choudhary acetylome, which was run in a cell panel that includes MV4-11, a KMT2A::AFF1 line
The famous clock paper did not find the chemical tag on MLL1 itself - it looked the positions up in a 2009 survey that catalogued acetylation across thousands of proteins. That 2009 survey was done partly in MV4-11, a leukaemia cell line that carries the exact fusion this project studies. So the tag may already have been seen in the right kind of cell, six years before anyone connected it to the clock, and sixteen years before anyone asked about the fusion.
The 24 genes that respond to BOTH menin inhibitors in RS4;11 are led by the ATF4 stress programme - CHAC1 -5.11, STC2 -4.40, TRIB3 -3.10 - and CHAC1 and TRIB3 are named zero times in this graph
Comparing a weak and a strong menin-blocking drug in the same leukaemia cells, only twenty-four genes respond to both, and most of them belong to one stress-response programme the cell runs when it is short of nutrients. Two of the strongest are genes this project has never once mentioned, and the single existing paper linking menin to that programme is about a different fusion cancer entirely.
The 297 absence-bearing entries that are not named absence-* split cleanly by entry type: the 109 gene and protein entries are bounded by title/abstract search (7 of 7 hand-read), the analyses are bounded by deposited data (0 of 6), and the weighted abstract-level share is 60% - indistinguishable from the 59% measured on the 39 named ones
Last cycle sorted thirty-nine write-ups of things nobody has measured and found that most rested on a search of titles and summaries only. Reading a random sample of the several hundred similar claims that were never sorted shows the same picture, so the problem is about five times bigger than the first count suggested.
The 3.7% was mostly the query: DOT1L recall is 19.6% once you ask about leukaemia
Last cycle a check said we had read only 3.7% of recent DOT1L papers, the worst of any topic we measured. Re-running it with the word "leukaemia" added to the search puts the number at 16-20%, roughly in line with the rest of the project, because most DOT1L papers are about other diseases entirely.
The 332 AFF1-only promoters are BELOW chance: a label-permutation null over the same occupancy matrix expects 426, so the figure the IRX1 partner-specificity line rests on is not an enrichment - and my own brief told the agent the opposite
A count of genes that look switched on only in the samples carrying one fusion partner was treated as evidence that partner is special. Shuffling which samples count as that partner produces MORE such genes, not fewer, so the count is not evidence of anything.
The 332-promoter AFF1-only set is BELOW the count a random eight-sample arm produces (null 425)
Three hundred and thirty-two gene switches are switched on somewhere in the eight KMT2A::AFF1 leukaemias and nowhere else. Shuffling which eight patients we call the group gives an average of four hundred and twenty-five such switches. The AFF1 group is, if anything, slightly LESS distinctive than a random group of eight.
The 50-RBP BioID2 proximity map in K562 - the first modern proximity dataset this graph can reach, and AFF1 is in it as prey
A 2024 study attached a biotin-tagging enzyme to 50 RNA-handling proteins inside human leukaemia cells and recorded everything each one brushed past. AFF1 - half of the fusion this project studies - shows up in that record, and nobody here had ever looked.
The 82.9% population scored twice as many defect-shaped notes as the same-source one, and reading six of them says the comparison is confounded by our own DepMap numbers
We asked whether counter-arguments that cite a second paper are better supported than ones reusing the first. The scores say they are worse - but reading the flagged cases shows most are quoting our own calculations alongside a paper, which the test cannot tell from a misattributed fact.
The absence-class checker globbed a filename instead of the field, so it walked 40 of 243 entries and reported them as the whole population - and the 40 it called invalid turn out to be two independently re-invented vocabularies
A tool that checks how we label the gaps in our own knowledge was only looking at the files whose names began with the word 'absence', which was one in six of them. Fixing it showed that where authors had written something the tool called invalid, they were usually describing a distinction the official label list had no word for.
The ABT1/KRR1/UTP3 nucleolar triple: pre-registered, then adjudicated - it is not three proteins on two baits but a seven-member ribosomal block in one screen plus one hit in a neighbourhood that shows no nucleolar enrichment, and two of the four pre-registered discriminators were void exactly as the pre-registration predicted
Three proteins that build ribosomes in the nucleolus showed up as partners of the two proteins this project studies, which could have meant those proteins do something in the nucleolus. Writing down the test before looking, then looking, the pattern turns out to be what the experiments themselves would produce by accident.
The active chromatin landscape does not lose a single peak call within 24 hours of the fusion being destroyed: H3K79me2 holds 99.7% at both 3 h and 24 h, H3K9ac 100.0% at both, H3K27ac 83.3%/67.2%, H3K4me3 99.6%/78.9% - while the authors, measuring signal at fusion target genes, report a marked H3K79me2 reduction by 24 h
We looked at four chemical tags on DNA packaging one day after the cancer protein was destroyed and found that essentially none of their detectable locations had disappeared, while the original authors - looking only at the specific genes the protein controls, and measuring signal strength rather than presence - did see a clear drop. Both can be true, and the gap between them is the most instructive number in this analysis.
The AFF1 promoter window holds three genes' TSSs, not two, and the third is protein-coding - Q153's 92.5% is confirmed and its two-gene framing is not
This project measures AFF1's on-switch with a 2,000-letter window, and three different genes begin inside it, not two, one of them protein-coding. One of the three is missing from the project's reference table entirely and another is in it under a different one of its own three start sites 42,000 letters away, so nothing the project could run would have noticed either.
The AHD inhibitor's deposited arm is MOLM-13 and its in vivo arm is nowhere: a half-resolution that names the limit exactly
Two entries note that a small-molecule AHD inhibitor paper never says which fusion its animal model carried. The deposited sequencing is six samples of one human leukaemia line, MOLM-13, which carries MLL-AF9 - and the animal experiment left no deposited data at all, so that half of the question stays open.
The AML-over-ALL retention ratio is 93.5% denominator: the GMP-like numerator barely moves across 23 samples while the B-lineage denominator varies fourteenfold, so the ratio measures loss of the B-lineage class and not gain of the GMP-like one
A popular way to say 'the fusion moved to myeloid genes' is to divide myeloid signal by lymphoid signal. Across 23 patients the myeloid number hardly changes at all and the lymphoid number changes enormously, so the division is really reporting the lymphoid number upside down.
The anchor-matched ranking, run: the KMT2A exon-11 CTCF anchor is rank 58 of 1,625 anchors matched on baseline strength - the 3.6th percentile - and the matched MEDIAN anchor also falls to 0.63x, so 'CTCF falls only at the KMT2A anchor' is refuted while 'the KMT2A anchor loses more than most' survives weakly
Last cycle this project found that a loop-anchor protein falls off the KMT2A gene when one enzyme is removed and does not fall off at eleven other places it checked - but none of those eleven was a loop anchor. Compared against twelve thousand real loop anchors of its own strength the KMT2A anchor loses more protein than about 96 of every 100, while the typical matched anchor still loses a third and three hundred of them lose more than half, so the effect is stronger here than average rather than unique to this gene.
The assumption the whole project rests on, tested at last: menin dependency is fusion-specific (run)
Everything this project argues assumes leukemia cells need menin because they carry the fusion, and nobody here had ever checked it against our own data. Splitting the cell lines by whether they carry the fusion, the ones that do need menin far more — beyond what any of 918 randomly chosen genes would predict.
The atlas heuristic is scoped to six baits instead of to publication scale, so it misses HuRI, OpenCell and five other proteome-wide screens - 49.5% of rows credited with a 'focused' publication have only atlas support, and all five of q178's credible interactors fail
To avoid treating a giant automated screen as if it were a dedicated experiment, this project labels publications as 'atlas' or 'focused'. The rule counts how many partners a paper contributes to six specific proteins - so a screen of fifty thousand interactions that happens to touch only a few of those six is filed as a focused study. Checking every borderline paper against PubMed: seven large screens were mislabelled, including two of the largest human interaction maps ever published. Half of all 'focused' support in the list is actually atlas support, and every one of the five partners that looked worth writing up rests on one.
The authors' own deposited table puts menin at 0.87 of its diagnostic-sample signal over their oncoprotein-target peaks, in the sample their text describes as having menin reduced to undetectable levels - the two statements are about different denominators, and only one of them travels
The paper says menin became undetectable in the drug-treated sample. The authors' own published spreadsheet, covering the same genes in the same sample, says menin kept 87% of its signal. Both are honest: the first describes what a genome browser picture looks like, the second is a number from a within-library calculation that cannot see a whole-cell loss at all.
The authors' signal floor removes 92% of non-occupied genes from the menin table, and all of the occupied ones survive
The published table of menin measurements does not contain every gene - the authors kept only genes with enough signal. All the fusion's target genes make the cut; nine out of ten of the others do not. So 'menin did not change elsewhere' means 'at the other places where menin was detectable', not 'everywhere in the genome'.
The B-ALL and the AML of one infant share no more fusion-occupied promoters than two unrelated infants do - lineage, not patient identity, sets where KMT2A::AFF1 sits
One child's leukaemia was profiled twice, before and after it changed from a lymphoid to a myeloid cancer. The genes the fusion protein sat on in the two states overlapped no more than the genes it sat on in two completely different children.
The bigWig refusal of cycle 57 is overturned: pybigtools installs on this box under Python 3.12, reads GEO tracks by HTTP range request without downloading them, and passes the control loci the refusal itself demanded
Last cycle this project correctly established that the standard Python library for reading these genome signal files would not build on this Windows machine, and correctly refused to guess. A different library, built in Rust with ready-made Windows packages, installs in seconds and reads the same files straight off the internet without downloading them.
The blind spot is closed: all four K1130-and-leukaemia full-text hits are reviews restating the source paper, and the citation-network bound survives 26 nomenclatures rather than five
Only four papers in the open literature mention both this exact protein spot and the word leukaemia, and all four are review articles that simply repeat the original finding rather than testing anything. Re-running the search with 26 different ways of writing 'leukaemia fusion' instead of five still finds nobody who has done the experiment.
The bounded-absence population went 39 to 73 in two days, and the abstract-level defect class went from 26 of 39 to 2 of 34 (Fisher p = 1.3e-5)
This project writes down things nobody has measured. A survey two days ago found thirty-nine such write-ups and warned that most rested on nothing more than a keyword search of paper summaries. There are now seventy-three, and of the thirty-four written since that warning only two rest on a keyword search.
The cannot-be-occupied floor never binds at gene-body scale in any of the 27 samples, so q147's gene-scale rule is a bare 2.0x gate - and the max-versus-95th-percentile difference between q147 and q152 is immaterial there
This project's occupancy test has two parts: the signal must be at least twice an average gene switch, and it must beat a floor built from eight genes that cannot possibly be switched on in a leukaemia. At the gene-body scale, that floor is always tiny - never above 1.65, when the bar is 2.0 - so the second part of the test never does anything. The test is really just the first part.
The CD34+ oncoprotein-negative sample that calibrates the published N/C threshold is in GSE159608, not GSE252378 - so the 1.75 threshold cannot be re-derived from the deposit that carries the 27 patients it is applied to
The published method sets its cutoff by checking how many false hits it makes in a healthy control sample. That control lives in a different public dataset from the patient samples, so anyone re-analysing the patient data alone has no way to check the cutoff.
The chain's biggest risk is generalising a myeloid mouse result to infant B-ALL, and the graph cannot measure how much of it does
This project's most-documented mistake is taking a result from one leukaemia and one mouse model and quietly applying it to the infant leukaemia it actually studies. Trying to count how often the central mechanism story does that: it cannot be counted, because nothing in the graph records which system each piece of evidence came from.
The cheapest unrun test in this beat's queue: GSE260742 deposits an H3K79me2 ChIP under the DOT1L inhibitor SGC0946 in the same murine MLLAF9 cells as its FLAG and MLL1 tracks, which removes the catalysis directly instead of indirectly and would separate 'the mark is stable' from 'residual DOT1L is sufficient' - one download
There is a second drug in the same public dataset that blocks the tag-writing enzyme directly rather than through menin. Comparing the two would say whether the tag survives because it is stable or because enough enzyme is left, and it is a single file.
The check I called decisive was not: the haem signal is still open (run)
Yesterday's genome-wide scan found four genes from the haem-building pathway among the top hits in fusion-carrying leukemia cells, and the obvious worry was that this reflects what kind of blood cell they are rather than the fusion. One version of that worry is now ruled out. The other could not be tested properly, because only four suitable comparison cell lines exist.
The check this project named nine cycles ago and never ran: the 34 least-verified papers in the corpus are attached to the right claims, and every low-scoring pair is support living in a supplement or a structure rather than an abstract
Every earlier check in this project asked whether a citation's ID points at a real paper, never whether that paper actually backs up the sentence it is attached to. This run read the one subset small enough to read - the 34 papers nobody transcribed - and found all of them correctly attached.
The chromatin compartment that the field treats as 'the fusion's sites' is smaller and less shared in MOLM13 than in RS4;11: 467 of 1,389 MLL1-N-top bins also carry top menin (33.6%) against 692 of 1,390 (49.8%), and the MLL1-N and H3K79me2 compartments barely overlap in either line
The places where the MLL1 antibody and the menin antibody both peak - the best available guess at where the fusion sits - are only a third of the MLL1 peaks in one cell line and half of them in the other. The compartments are not the same size and the histone tag sits somewhere else again.
The citation index built to stop false gap claims made one: its regex matched prose but not the schema's own pmid field, missing 73.7% of the corpus's citations, and the single-cell gap it selected was refuted by a paper already cited in fourteen entries
Two cycles ago this project built a search tool over every paper its notes cite, specifically to stop it claiming that something had never been done when it had. The tool reported 1,908 papers and that every one was real, which sounded like a clean bill of health. It was not: the tool could only see citations written one of the two ways this project writes them, and it missed nearly three quarters of them. The gap it then pointed at — nobody has studied these patients one cell at a time — was refuted by a paper the project had already cited fourteen times.
The clearest clinical rationale for a LEDGF agent - second line after menin-inhibitor failure - has never been tested, and one 2025 measurement now argues against it
The obvious use for a LEDGF drug is in patients whose leukaemia has stopped responding to the menin drug. Nobody has ever tested a LEDGF-directed agent in a menin-resistant model, and a new measurement suggests the LEDGF grip may already be loosened by the menin drug, which would weaken the idea.
The clock/SIRT1-regulated MLL1 acetylation site maps to KMT2A K1130 and K1133 in human numbering, just N-terminal to the CXXC and 208 residues before the breakpoint cluster - so it is retained in every KMT2A fusion
A 2015 paper found that the MLL1 protein is switched on and off by adding and removing acetyl groups at two specific spots, under the control of the body clock and cellular energy levels. Those two spots turn out to sit inside the exact piece of the protein that every leukaemia fusion keeps.
The cohesin-degron break map was run in the one human cell line where the KMT2A element does not break - so premise (i) is not underpowered, it is unaddressed
To ask whether cohesin causes a DNA break, you need a cell where that break happens. The one experiment that removed cohesin used a bowel-cancer cell line in which the KMT2A spot shows no drug-induced break at all - so the experiment could not have detected an effect either way.
The cohort paper this graph said had certainly measured IDH status was opened: 96,537 characters of full text contain IDH1 zero times, IDH2 zero times, TET2 zero times and DNMT3A zero times, so the measured-not-reported class is confirmed by reading rather than by inference
Whole-genome sequencing of thirty-six infants and children with relapsed KMT2A leukaemia would have detected mutations in the IDH genes as a matter of course, and this project assumed the result was simply never written down. Opening the paper confirms it: the two gene names appear nowhere in it.
The comparator arm was never screened for pathway-equivalent lesions — and it made us understate our own result
Everything this project says about what makes KMT2A-rearranged leukaemia different is a comparison against other leukaemia cell lines. Nobody had checked whether those comparison lines carry the same biology under a different name — and two of them do. Removing them makes the differences slightly bigger, not smaller, so the findings survive; but the check should have been run before the findings, not after.
The comparison that founded 'the CXXC contact is too weak' puts a 300 mM NaCl protein-DNA measurement against a 50 mM NaCl protein-peptide measurement
This project says the DNA grip is a thousand times weaker than the menin grip. The two numbers were measured in salt concentrations six times apart, and salt hurts a DNA grip far more than it hurts a protein-protein grip, so the comparison is biased in exactly the direction the argument needed.
The compound in the only infant menin-inhibitor occupancy sample is revumenib, named in the abstract - and neither the dose nor the duration of exposure is recorded in the paper, the GEO record or the authors' repository
The drug the child was on is revumenib. How much of it, and for how long before the cells were taken, is written down nowhere we can find - and this project has already established that dose is exactly where menin-inhibitor chromatin results stop being comparable to one another.
The computed-but-unread scan ran, passed its own declared informativeness control at 60.3% cited, and produced a headline nobody can act on: 36,110 uncited values across 182 results files, dominated by ranks, taxids and chromosome extents. The useful population is not 'uncited numbers' but 'PRE-REGISTERED quantities that went uncited', which is what the one real instance actually was
Twice recently this project has found that a number it had already calculated and saved was sitting unread in its own files and, once read, changed an interpretation - so this cycle built a tool to find all such numbers systematically. The tool works correctly and its internal check confirms it is matching properly, but its answer is thirty-six thousand uncited numbers, almost all of which are internal bookkeeping like list positions and chromosome lengths that nobody would ever cite. The genuinely useful question is narrower: which quantities did a study formally commit in advance to measuring, and then never report? That is what both real instances were, and it is a far smaller list that the next cycle should build instead.
The confounder every previous analysis flagged and none removed, removed (run)
Every earlier version of this test compared leukemia cell lines as if they were one group, but they are not: most are myeloid and the rest are lymphoid. Two genes that simply matter more in myeloid leukemia would look linked even if they never touch. Comparing each line only against others of its own subtype removes that, and the link between menin and the fusion machinery survives almost untouched.
The convenient deposited promoter table finds 3 gained windows where the genome-wide re-binning finds 53, because 83% of the gained set is not at a RefGene TSS - so the processed file that makes this dataset cheap to re-use is structurally blind to the answer
This dataset comes with a ready-made table of signal at gene starts, which is what anybody reusing it would open first. Running the same question against the raw files instead finds seventeen times more changed places, because most of them are not at gene starts at all.
The corpus citation index reads `PMID nnnnnnnn` in prose and never the `pmid` citation field — it sees 1,908 of 3,991 papers, undercounts 872 of the rest, and its headline single-cell absence was an artefact
A tool built to list every paper this project cites was actually listing only the papers someone happened to mention by number in a sentence, not the papers recorded in the citation slot where citations belong. Because of that, it reported that the project cites only eight single-cell papers and none about this leukaemia — when in fact it cites eighteen, and the single most important single-cell paper on KMT2A-rearranged infant leukaemia is cited in thirteen separate entries.
The corpus holds 42 of 42 pan-fusion partners and 1 of 110 MLL-AF4-specific ones — the inverse of the standing priority, and every mechanism that produced it was a reasonable local decision
The lab's standing instruction is to write up every partner of our own fusion. The graph has written all 42 partners that every MLL fusion shares and one of the 110 that only ours has. Every step that led there was sensible on its own — but the result is the opposite of the instruction, and this entry sets out why, and whether it should be fixed.
The counter-note defect detector mistakes cross-referencing for unsupported assertion, which is why the different-source rate looked twice as bad
A tool here tries to spot counter-arguments that state specific facts the paper they cite does not actually contain, and it reported that counter-arguments citing a different paper were about twice as likely to be defective as ones citing the same paper - but nobody had read the thirty-one it flagged. Reading them, almost none is defective: the 'unsupported facts' are mostly gene names, absent from the cited abstract because the whole purpose of those notes is to bring in knowledge from elsewhere, so the tool was penalising exactly the behaviour that makes a counter-argument worth having.
The coverage checker the standing brief mandates overwrites its own HGNC index mid-loop, so every symbol after the first no-match is answered wrongly - and the wrong answer is 'treat any answer as unsafe'
Before writing about any gene, every agent here is told to run one script and quote its verdict. That script quietly breaks itself partway through a list: after the first gene it cannot find, it forgets how to translate old gene names, and it then tells you a perfectly resolvable old name does not exist. Run the same name on its own and it answers correctly.
The coverage guard withheld 30 of 351 pairs and named every one: 26 contain 152985, which calls zero promoters, and a Jaccard with a zero-call sample is a division, not a measurement
Some comparisons were refused rather than reported, because one of the two samples had no detectable signal at all. Refusing to print a number there, and saying which comparisons were refused and why, is the point.
The coverage tool's id arm cannot reach five of its nine prefixes - 225 entries - and the prose arm covers for it in 338 of 338 cases, so the reported false clear happens only under a flag that disables the prose arm
An agent reported that the project's safeguard against re-writing things it had already decided against was silently failing. The mechanism it described is real, but the safeguard has a second check that catches every affected case - unless you run it with a speed flag that switches that check off, which is what the agent had done.
The CRISPR-ChIP H3K79 partition is a mouse MLL-AF9 result: resolved from GEO after the full text proved unreachable
The screen this project treats as its strongest independent support - the one splitting the H3K79 chromatin tag into a fusion-driven half and a normal half - was run in mouse leukaemia cells carrying MLL-AF9, not the infant fusion MLL-AF4. We could not read the paper, so we read the raw data records it deposited, and every single H3K79 experiment in them is mouse MLL-AF9.
The CRM1 address is not a menin-free address: NPM1c reaches HOXA the same way and NPM1c AML is menin-inhibitor-sensitive in the clinic
One candidate fusion reaches the HOXA genes by hitching a ride on an export protein, which looked like a way to switch those genes on without needing menin. But the same export protein carries a very common leukaemia mutation to the same genes and those patients do respond to menin drugs, so the route was the wrong thing to look at.
The cross-entry contradiction screen does not work, and the reason is that this graph writes carefully
The existing tool for finding places where this graph disagrees with itself can only look inside a single entry or at a single paper, so it cannot see two entries citing two different papers that say opposite things - which is exactly the disagreement found this cycle. I tried to build a tool for that and it does not work. The reason is almost funny: the graph's own house style is to say precisely what it does and does not know, so words like 'only', 'exactly' and 'no' are everywhere as marks of care, and a machine looking for disagreement reads all of them as disagreement.
The curated interaction database knows less about AFF1 than this graph does
We checked our list of AFF1's binding partners against the big public database of protein interactions, expecting to find things we had missed. The database knows of 25; we already have 126 entries.
The CXXC-DNA affinity is a ten-point ladder from 44 uM to 33 nM, not two numbers 1,300-fold apart, and every step but the last is under 5.5-fold
Two published numbers for how tightly the fusion's CpG-gripping module holds DNA differ 1,300-fold, and this project built an argument on the weaker one. Reading the six primary papers shows they are not in conflict: there are ten measurements forming a smooth ladder, and only the final step - the one this beat could not resolve - is large.
The deferred KMT2A partners: what the counts actually are, what was written, what was rejected, and three numbers the previous enumeration got wrong
Before writing anything, this beat counted how much real evidence exists for each KMT2A fusion partner that an earlier round had put off. Nine entries were written, several partners were rejected with the numbers behind each rejection, and three counts in the earlier list turned out to be wrong.
THE DEFLATING NULL, RUN: matching on gene length AND MCF-7 expression roughly HALVES the H3K9me3 depletion over the KMT2A gene body - from the 7.6th-15.0th percentile of class-matched CTCF anchors to the 18.2nd-23.4th of 385 matched transcribed gene bodies - so the methyl-poor breakpoint domain is partly a fact about transcription and does not dissolve into one
This project has said four times that the region where KMT2A breaks is unusually free of a silencing tag. Compared against other genes of the same size that are switched on just as much, about half of that unusualness disappears - but not all of it.
The deposit names its first replicate "paired" and its later ones rep1/rep2/rep3, so GSM8123403 "H2AK119ub paired" and GSM8123420 "H2AK119ub rep2" are one condition in two replicates - which this project has called "the two processing conventions" since cycle 88. GSE260742 holds 17 genuine same-condition replicate groups, 13 of them histone marks, refuting the claim repeated since cycle 100 that it has none and never will
A public dataset labels repeated runs of the same experiment in an inconsistent way: the first repeat is called "paired" and later ones "rep2" or "rep3". Because of that, this project spent nineteen cycles treating two repeats of the same experiment as two different ways of processing one experiment, and separately concluded that the dataset contained no repeats at all. Reading the dataset's own list of 158 samples shows seventeen sets of repeats, thirteen of them for the chemical marks the project measures.
The deposit paper's own supplementary table calls IRX1 an oncoprotein target in exactly p279, p318 and p454 - the same three samples this project reached by a different instrument
This project decided which patients had the fusion protein sitting on the IRX1 gene using a simple rule of its own: twice the average gene switch, in two antibodies. The scientists who generated the data used a completely different method - peak calling with a statistical model - and published their answer in a supplementary spreadsheet. Both come out with the same three patients.
The deposit's two lineage-switching patients behave in opposite directions: the untreated pair keeps 1126 of its 2045 oncoprotein-target promoters across the switch, while the menin-inhibited pair falls from 345 to zero - so lineage switching does not by itself clear the oncoprotein
Two children in this dataset switched their leukaemia from one type to another. The one who was not on the drug kept most of its fusion-bound genes; the one on the drug lost all of them. The switch alone does not remove the fusion protein.
The deposited dm6 spike-in table covers the eighteen H3K79me2 samples and nothing else - the ten MEN1 and DOT1L ChIPs in the same series have no fly chromatin, so every occupancy number from them is a within-sample contrast and no gain-or-loss statement on an absolute scale is available
This study included a clever internal ruler that lets you say how much of a signal was really lost. The ruler was only put into one half of the experiments, and not into the half this analysis needed it for.
The deposited GSE127507 processed signal table is an Excel export in which every value at or above 1,000 is a quoted comma-separated string, so a plain numeric parse silently zeroes 34,955 of 386,298 cells - and those cells are the entire top of all eighteen distributions
The ready-made results table in this public dataset came out of a spreadsheet, so every number over 999 is stored with a comma inside quotation marks. A normal number-reading program turns exactly those into zero - and those are all the strongest signals, which is the only part anyone wants.
The deposited RS4;11 H3K79me2 table in GSE149183 is two-thirds MOLM13: 38,580 of its 57,794 rows are byte-for-byte identical to the MOLM13 file, so the index fusion is measurable at only 19,214 TSS entries
A public data file named for the leukaemia cell line this project cares about most turns out to contain another cell line's numbers in two-thirds of its rows - the same values, digit for digit. Anyone who used the whole file, as this project did in an earlier cycle, was mostly measuring the wrong cells.
The depth-matched discriminator I called the only available test: thinning 44 of 54 libraries down to A73051's in-promoter fragment count moves their oncoprotein share by a median of +0.0003, 39 of 39 hold and 0 collapse - so depth could never have explained A73051, and my cycle-135 worry was misconceived rather than merely wrong
One patient's sample has the fusion protein sitting almost nowhere, and the worry was that it only looked that way because the sample was sequenced less deeply. Artificially cutting the deep samples down to its size leaves their measurements essentially unchanged, so depth was never the explanation.
The dispatch brief scoped this deposit's oncoprotein-score arm at two patients and a cell line; it is 27 patient samples across six fusion partners, and the scoping error was inherited from a beat that only needed three of them
I was told the relevant part of this dataset covered two children and one lab cell line. Counting every record showed 27 patient samples with the same complete set of antibodies, which makes the planned comparison an order of magnitude stronger than expected.
The DOT1L inhibitor removes menin from fusion promoters too, which a pure interface model does not predict
A drug aimed at a completely different enzyme - DOT1L, which writes a chemical mark on histones - also strips menin off the fusion's target genes. Menin's position on chromatin is therefore not controlled only by its handshake with the fusion.
The DOT1L partner IntAct calls TCF4 is TCF7L2 - the second UTP18/WDR5-class symbol mis-assignment found in this interactome, and gene-tcf4 is declined
A database lists TCF4 as a partner of DOT1L. In the paper it comes from, TCF4 is the field's nickname for a different gene, TCF7L2. The two are unrelated proteins, so the entry this project was about to write would have been about the wrong protein.
The easy null was not the problem, and nothing in 2,018 genes beat menin
Last cycle we checked some genes against seven hundred random ones and worried the comparison was too easy, because most random genes do nothing in leukaemia. We redid it against two thousand genes and then against only the four hundred that do something here. The answer barely moved, and nothing we had not already picked scored better than menin.
The epigenetics coverage test inspects data{} KEY NAMES and never a field's value, so 71 entries that state a histone mark INSIDE a data field score zero - and four more are missed by one character
The tool that scores epigenetics coverage looks only at the NAMES of the fields in an entry, never at what those fields say. So an entry with a field called 'mechanism' whose text reads 'deposits H3K4me3 at target promoters' counts as having no epigenetics at all.
The epitope that the whole shared-epitope argument turns on is not documented anywhere this beat could reach: three vendor routes for Bethyl A300-086A returned HTTP 404, so 'MLL1 N-terminus, retained in the fusion' rests on the depositors' own target label and on general fusion architecture, not on an immunogen range
Everything this project says about one antibody hiding the drug's effect depends on which part of the protein that antibody recognises. That part has never been written down in anything we could open.
The excess overlap survives a structure-respecting null in both beats - permutation p = 0.0005 in all six cells - but every published excess figure shrinks 2.4- to 4-fold
The finding that two samples share more target genes than chance is real and holds up under a much tougher test. What does not hold up is how big the effect was said to be: it is two to four times smaller than reported.
The experiment that refutes cycle 55's headline was already public: VTP-50469 collapses MEIS1 32-fold in the same line where MI-2-2 left it untouched - and HOXA9 was never actually measured by either dataset
Last cycle we found that a menin-blocking drug left the leukaemia's key genes untouched, and said plainly that we could not rule out the drug simply being too weak. A stronger drug from the same class was already tested on the same cell line in public data, and it switches those genes off hard, so the weak-drug explanation was the right one.
The falsifier that decides the beat, run: 65.6% of the fusion's own top sites overlap a CpG island and 1 of the 52 sites it formally 'gains' under a menin inhibitor does - 1.35-fold over a class-matched null, Fisher p = 0.53 - so the gained set does not carry the signature of CXXC-driven binding and the redistribution model is refuted on its own predicted substrate
The fusion protein grips DNA at unmethylated CpG islands, and two thirds of its normal sites are islands. Of the fifty-two places it appears to gain under the drug, one is - which is what you would get by chance, so those places are not new fusion binding sites.
The Fanconi core complex is 2nd of 719 complexes for KMT2A-selectivity - but coherence does not explain it, and the explanation this entry was built on is dead
A gene being listed in a pathway does not mean it acts with the rest of it. We measured which Fanconi genes actually behave together across a thousand cell lines, and found that those are exactly the ones that matter in this leukaemia.
The Fanconi signal is in both lineage halves, survives the correlation-proof null, and is still there when lineage is held constant
Last cycle this project found that the Fanconi DNA-repair genes are needed more by KMT2A-fusion leukaemia cells than by other leukaemia cells, but could not say whether that was a real property of the fusion or an accident of which cell lines were used. This analysis pulls the result apart four ways, and it holds up in every one of them.
The Fanconi signal is real, and it sits in the ubiquitination module
The Fanconi repair pathway has kept turning up in this project without anyone testing whether that was a coincidence. It is not. And the part of the pathway responsible is a specific chemical-tagging step, not the DNA-repair machinery people usually think of.
The federated sweep this graph prescribes for the missing index-fusion interactor cannot be run: PSICQUIC free text returns ZERO for `mll_af9`, an entity that provably exists and returns 32 records under its accession - so every provider's zero in that sweep is uninterpretable
One entry here says the next thing to do is search every interaction database for the fusion protein this project is about. Running that search returns nothing everywhere - but it also returns nothing for a fusion protein we know is in there, so the search cannot tell the two apart and the recommendation does not work as written.
The first measurement of any histone mark outside the top 0.1% of fusion sites: by ratio the drug's mark response is confined to the peaks as pre-registered, while in ABSOLUTE share the movement at the decile below is 5 to 33 times larger - though that is partly guaranteed by the stratum being 100x bigger. H3K79me2 falls at the peaks (0.8065) and is not measurably changed at D1, and H3K4me3 and H3K79me2 both break the floor at the lowest decile
Two earlier studies here showed a drug changing four chemical marks on DNA packaging, but both only looked at the strongest one-in-a-thousand sites, and this is the first time anyone has looked anywhere else. Measured as percentages the changes really are concentrated at those strongest sites, which supports what was done before, while in sheer quantity far more movement happens in the surrounding region - though that is partly just because the surrounding region is a hundred times bigger.
The first replicated share ratio this corpus has ever had: two title-declared replicates of ONE murine culture span 0.2623 on the same quantity - 4.5x the 0.0584 between-culture spread this corpus uses as its indifference threshold - and they disagree in SIGN about whether the PCGF1 knockout blunts the DOT1L-inhibitor response
Almost every chromatin number in this project comes from a single sample, so nobody knew how much such a number wobbles. Here the same measurement exists twice, and the two copies disagree by more than four times the amount this project treats as meaningful - and on one question they point in opposite directions.
The first same-condition H3K79me2 noise floor measured anywhere in this thread: 11.17% at D1 and 26.44% at the TOP 0.1% stratum, against 45.72% and 62.63% for H2AK119ub in the same arm, the same condition and the same pairing shape - 4.1x tighter. A mark floor transferred from H2AK119ub does not describe H3K79me2.
Until now the repeat-to-repeat wobble for one chemical tag on DNA was assumed to apply to the others. Measuring it directly for a second tag gives a number four times smaller, so the assumption was making that tag's results look far less reliable than they are.
The five verdicts in one place, plus what this beat read and did not use
A single table saying which kinds of claim about the KMT2A::AFF1 protein hold up once you ask what reagent was actually used, and which do not. It also records the papers that looked relevant and were not used, so a future reader does not repeat the search.
The four AFF4 two-hybrid partners are three rows of one screen, not three experiments - and in that same screen GOLGA2 has 249 partners while AFF4 has five
Four proteins were listed as AFF4 partners with three database rows each, which reads like three separate experiments. All twelve rows come from one yeast two-hybrid screen, and three of the four proteins stick to hundreds of other things in that very screen.
The four eviction quartiles are not equivalent stretches of genome: all three GSE260742 input tracks give a monotone increasing share across Q1 to Q4 at Spearman -1.000, 1.15 / 1.17 / 1.13-fold - small, real, in the same direction as the H3K79me2 gradient, and it cancels exactly in a within-genotype drug/vehicle ratio
Even before any antibody is used, the raw DNA from these cells is not evenly represented across the four groups of stretches - there is about fifteen percent more of it in one group than another, in a steady trend. It is small, and it drops out of the comparison being made, but it means the groups are not interchangeable.
The four-way ordering given a second cell line: menin, DOT1L and the MLL1-N channel reproduce between KMT2A::AFF1 and KMT2A::MLLT3 to within 0.04 of a share ratio, and H3K79me2 - the one channel the ordering was built around - diverges by 0.21 to 0.30 and changes places with MLL1-N
Repeating last cycle's four-antibody measurement in a second leukaemia cell line shows that three of the four answers come out almost identically, which is about as good as this kind of experiment gets. The fourth - the chemical tag on the histone - behaves completely differently, and it is the one the whole ordering was built around.
The fusion-biased peak at the AFF1 locus fires in 23 of 27 samples across all six partners, so it is not the fusion visiting its own partner
The hottest spot next to AFF1 lights up in almost every patient, whichever partner gene the fusion happens to carry. If the fusion were drawn there because AFF1 is its own partner, it should light up in the AFF1 patients and not the others. It does not.
The fusion-oncoprotein TurboID experiment this graph proposes has already been published - for PML::RARA, RUNX1::RUNX1T1, CBFB::MYH11, NUP98::HOXA9 and SET::NUP214, and not for any KMT2A fusion
Two groups have already attached a biotin-tagging enzyme directly to leukaemia fusion proteins and listed everything each one touches. They did it for five fusions. None of the five was a KMT2A fusion.
The fusion-specific score's per-promoter reliability inside one patient is r = 0.183, and the median correlation between two DIFFERENT patients is 0.196 - so the ordering question cannot be resolved at promoter resolution, and the corpus's 0.82-0.97 figures cannot be carried to this readout
Measured twice in the same patient's cells with two sets of antibodies, this readout barely agrees with itself. Two different patients agree about as well, which means the method cannot tell whether different patients really rank their genes the same way.
The gain question is answerable today for free: GSE127508 contains MLL1n ChIP-seq in a fusion-only line, plus and minus VTP-50469, and nobody has called peaks on the treated sample
The data needed to ask whether the fusion moves instead of leaving has been sitting in a public archive since 2019. It includes a cell line whose only copy of the protein is the fusion, treated and untreated. Nobody has run the analysis that looks for new sites.
The gap instrument's 'not mentioned anywhere in the corpus' reads only ids, titles and aliases - it was wrong about 9 of 9 this cycle
A script tells us which of menin's partners the project has never discussed, and it was wrong about all nine it named. It only looks at entry names, so anything discussed inside an entry is invisible to it.
The genome build of GSE127507 is hg19, verified from the chromosome extents written inside each of the nine downloaded .tdf files - 9 bp from hg19 and 2,818,723 bp from hg38 on chr1, in every file - and independently from the deposited TSS table's own coordinates
Before measuring anything, the reference genome each data file was aligned to was read out of the files themselves rather than taken from the deposit's description. All nine agree it is the 2009 human reference, and so does the separate processed table.
The GEO sweep that q110 named as free and did not run: SUV39H1 chromatin occupancy HAS been measured in two human cell types, six samples, hg38 bigWigs already deposited - so the writer half of the H3K9me3 absence is measured-not-analysed, not never-measured
A published entry here said nobody could even ask an archive where the main H3K9 silencing enzyme sits, because the big public atlas holds no experiment for it. The atlas really does hold none, but a different archive holds six samples from two human cell types, already processed into files anyone can read.
The GMP-like peak class is fusion-enriched in exactly one of GSE252378's 27 KMT2A-rearranged samples - p318, the sample it was derived from - once the null is matched on how coverable those promoters generically are
The published idea is that when a leukaemia changes identity, the KMT2A fusion protein moves onto the genes of a myeloid progenitor. Scored at the gene switches of all 27 patients in the deposit, that myeloid gene set looks busy everywhere - but so does any equally readable set of genes, and once that is corrected for, only the single patient the gene set was built from still stands out.
The GMP-like promoter class carries no N-over-C excess in 25 of GSE252378's 27 samples: whatever occupies it is seen equally by the fusion-side and the wild-type-side antibody
Two antibodies are used: one sees both the fusion protein and the normal protein, the other sees only the normal one. At the myeloid gene set they give the same answer in almost every patient - which means the thing sitting there is not the fusion.
The GMP-over-B-lineage occupancy contrast tracks the sample's own lineage - and institution, which is the identical split in this deposit - and not its HOXA status
The myeloid-versus-lymphoid signal ratio separates the myeloid patients from the lymphoid ones, which is what you would expect and tells you nothing new. It does not separate the patients whose fusion has left the HOXA genes from the ones whose fusion is still there.
The grading rule applied to twenty citations, and ORCID tested as a clustering key and rejected: zero coverage before 2010
We wrote a rule for how confident to be in each piece of evidence, then actually applied it. We also tested whether a standard author identifier could tell laboratories apart, and it cannot - the older papers do not carry one.
The GSE149183 per-gene columns are RAW, not spike-in scaled: every one of 1,040,292 values in both files sits exactly on a 0.04 grid, which no per-library scale factor survives - so two entries in this graph called an unscaled number absolute, and the true collapse is 37-91x, not 5.8x
A public data file's name promised that a fruit-fly yardstick had already been applied to its numbers. A one-line arithmetic check shows it had not: every number is an exact multiple of 0.04, which is what raw counts look like and what scaled counts never look like. Applying the yardstick properly makes the measured drug effect about ten times larger than this project had recorded.
The GSE260742 H2AK119ub pair that shares a processing state - rep2 against rep3, both bare `.tdf` - is not an unconfounded replicate pair: rep3 was submitted 2024-10-01 and rep2 2024-03-03, through a different file-naming pipeline. No pairing in this deposit isolates replicate variance alone.
A dataset's repeated samples looked like a fair like-for-like comparison because their filenames matched. Their submission dates do not: they were deposited seven months apart through different pipelines, so comparing them still mixes two things together, just different ones.
The H2BK34ub HOXA9/MEIS1 result is HeLa: the full text contains the word leukaemia zero times
An entry on the histone mark H2BK34ub already suspected that its HOXA9/MEIS1 result came from a non-leukaemic setting. Reading the full text confirms it and names the cells: HeLa, a cervical cancer line. The word leukaemia does not appear in the paper at all.
The H3K27me3 absence this graph keeps asserting is narrower than stated: the measurement exists, in the genetic arm of the paper the graph cites for the phenocopy
This project has been saying that nobody has ever measured the silencing mark H3K27me3 in a cell treated with a menin drug. That is very nearly right and not exactly right: the measurement exists, in the same paper the project already cites, but in cells where the menin gene was deleted rather than drugged.
The H3K27me3 and H3K4me3 NTvKO arms - twelve files nobody had ever binned, listed as second priority - carry the largest effect in this cycle (H3K27me3 2.0007 at the most-evicted quartile, collapsing to 1.0134 in the PCGF1 knockout) and the free internal control that rules out a library artefact (H3K4me3 runs the opposite way, Spearman -1.000)
Two sets of files that everyone had ranked lower turned out to matter most. One shows the strongest switching-off signal anywhere in this work; the other, by moving in the opposite direction in exactly the same places, is the reason we can be confident the first is real and not a quirk of the measurement.
The H3K79me2 eviction gradient against a DMSO-decile-matched null drawn from outside the fusion's sites at the same quartile sizes: the null is flat at 1.030-1.057, none of 20 draws reached the observed span of 0.6005, and the residual retains Spearman -1.000 - but the restricted-population trap is real and the baseline IS graded 2.2-fold
The four groups of DNA stretches being compared are not alike to begin with - the group that loses the most fusion also starts with the most of the tag. To check that the pattern is not just that difference reappearing, the same comparison was run on matched stretches elsewhere in the genome, twenty times. None of the twenty produced anything like the observed pattern.
The H3K79me2 eviction gradient is not a ratio of sums dominated by a few bins: the MEDIAN per-bin log2 share ratio is itself monotone at Spearman -1.000, the fraction of bins individually below 1 falls 89.3% to 24.7%, the bootstrap span interval is [-0.651, -0.545], and no single bin moves the span by more than 0.0057
The headline number adds up signal across three hundred DNA stretches at a time, so it could in principle be set by one or two unusual ones. Checking the typical stretch instead of the total gives the same answer, and removing the single most influential stretch barely changes it.
The HBO1/MOZ modules are not selective as modules: one paralogue per family carries the signal, and no member clears both a dependency and a pan-essentiality bar
A sweep of 719 protein machines put a family of histone-modifying machines at the top for this leukaemia. Tested gene by gene, the signal turns out to come from one specific member of each family while its near-identical siblings do nothing - and none of them is both important and specific enough to be a target.
The histone/ribosomal shape in the read queue is three classes, not one - H2BC21 is distinguishable AND essential at -0.70, so neither existing rejection ground reaches it, and HMGA1 is a harder dependency than MEN1 or KMT2A
Three more histone-like proteins turned up in the list of things we have not written about, and it looked like they could all be dismissed the same way the last one was. Measuring them showed the easy dismissal only works for one of them.
The HOXA/IRX1 anti-correlation does not reproduce in KMT2A-ENL, the only other partner group that can carry one
In KMT2A::AFF1 leukaemia, the more the fusion sits on HOXA the less it sits on IRX1. In KMT2A-ENL - the only other group with enough patients to test this - there is no such relationship, and if anything it points the other way.
The HOXA/IRX1 mutual exclusivity is reproduced at the transcript level in 48 KMT2A::AFF1 infant ALL cases (Fisher two-sided p < 1e-6) and is ABSENT in KMT2A::MLLT1 and in KMT2A-germline controls
Infant leukaemias with the AFF1 partner come in two flavours: HOXA-on/IRX1-off or IRX1-on/HOXA-off, almost never both. We reproduced that split in 48 patients. In leukaemias with a different partner, the split does not exist - they are simply HOXA-on.
The human CXXC/CpG-reading family is exactly twelve proteins, and TET2 is not one of them
We counted every human protein that carries the little zinc module KMT2A uses to read unmethylated CG letters in DNA. There are exactly twelve, and TET2 - the gene most often broken in leukaemia in this pathway - is not one of them, because its CXXC module was split off onto a separate gene early in evolution.
The Human Protein Atlas lists AFF1 as mitochondrial and nucleoplasmic at only 'Approved' reliability, with mitochondria as the single-cell-variable compartment, against nucleoplasm for KMT2A, menin and most of the pathway - a localisation lead that needs an orthogonal check
A public protein atlas places the leukaemia fusion's partner protein partly in the mitochondria, unlike its nuclear partners. The evidence is a single lower-confidence antibody, so it may well be an artefact.
The immunophenotype beat's own ledger: what was enumerated, what was written, what was dropped and why, and what could not be read
This is the working record of the beat that wrote the surface-marker entries: which markers were considered, which got an entry, which were deliberately left out and the reason, and which papers could be found but not read. It exists so that someone repeating this work next year repeats the decisions and not just the searches.
The imports that were never necessary: H3K79 methylation in KMT2A::AFF1 has its own literature, and this graph barely cites it
Not every borrowed result needed borrowing. There is a body of work done directly in MLL-AF4 leukaemia showing the H3K79 tag matters there, and this project has been leaning on mouse MLL-AF9 papers instead of citing it.
The index fusion DISCARDS the SIAH degron: the junction is at AFF1 residue 347, the degron is at human 282-304, and three entries had it the other way because they measured a murine motif against an exon-numbering junction
The cancer protein is built by joining half of one gene to part of another. A short tag that normally tells the cell to destroy the second protein turns out to sit in the piece that gets thrown away, so the cancer protein has no such tag. Three entries in this project said the opposite.
The index fusion's 3' partner has ONE observed binary interaction in IntAct and forty-two residues of solved structure - every SEC contact attributed to AFF1 is an AFF4 result wearing AFF1's name
AF4 is the partner gene in the commonest infant leukaemia fusion, and everything this field says about how it assembles the elongation machine comes from experiments where a whole complex was pulled out at once. Of its forty-three recorded partners, exactly one has ever been shown to touch it one-on-one, and the only piece of AF4 anyone has ever solved a structure of is forty-two amino acids long.
The infant menin-inhibitor occupancy experiment has been done and deposited - menin, KMT2A and ENL CUT&RUN under revumenib in ONE infant patient sampled before and after a lineage switch, not two patients - but it is KMT2A-ENL, so the index-fusion absence survives, and the same deposit holds untreated KMT2A::AFF1 libraries this corpus cites once
This project has repeatedly recorded that nobody has measured where menin sits on the DNA of an infant's leukaemia cells while a menin-blocking drug is present. That experiment turns out to have been done and made public — in two infant patients — but their leukaemias carry a different version of the same broken gene than the one this project studies. The gap is real, and it is now much narrower and much better described.
The IntAct interactome gap is closed - all 135 'unwritten' interactors are mentioned and 60% carry an explicit decision - while 95 of the 110 partners unique to the INDEX fusion have never been mentioned at all
**Corrected: the number of leftover partners worth writing is zero, not one.** The single candidate turned out to rest entirely on a yeast assay from a genome-wide screen, which my filter had misread as a focused study. --- The project's standing instruction says roughly a third of known interaction partners have been written up, implying two thirds are waiting. Checking each one: every single supposedly missing partner is already discussed somewhere, and most were examined and explicitly turned down for good reasons. So that gap is essentially closed. But a different list - the partners found only with this project's own fusion, in the one experiment that looked - is almost entirely untouched: 95 of 110 are never mentioned anywhere. The work that remains is the opposite of where it was thought to be.
The interactome coverage figure is computed on 651 of the 1,015 IntAct rows available - 64% - and the gap is worst for menin and LEDGF, the two baits the menin-interface hypothesis rests on
The number this project uses to say how much of the leukaemia machinery's interaction map it has written is calculated from a database connection that returns about two thirds of the available records. The missing third is worst for the two proteins the project's main hypothesis is about.
The interactome instrument is migrated to IntAct REST: 277 partners, 118 written, 42.6% - the denominator grew by 60 and coverage fell 3.5 points, which is the instrument getting honest
The tool this project uses to ask how much of the fusion's protein network has been written down was only seeing about two thirds of the database it was asking. Now that it sees all of it, the job looks bigger and less finished than we have been saying.
The interactome tail is orderable after all, just not by evidence: 62 of the 96 unwritten partners are already DISCUSSED somewhere in this corpus and need only promoting to entries, while 34 have never been read at all. Two cycles looked for a priority order in evidence strength and found none, because the ordering that exists is over how much work is left, not over how good the evidence is
This project wants its own page for every documented protein partner of the fusion it studies, and ninety-six partners still lack one - but two attempts to work out which to do first both failed, because the published evidence behind them is identical and their functional signatures are indistinguishable from random. Splitting them a different way solves it: sixty-two are already discussed somewhere in the project's own notes and only need that discussion turned into a page, while thirty-four have never been read at all. The first group is much cheaper to finish and should go first.
The IRX2 / IRX2-DT geometry re-derived from three annotations: 215 bp and 89.2% under RefSeq, 437 bp and 78.2% under Ensembl, and the verdict does not move
Two genes sit back to back on chromosome 5 with their on-switches only about 200 base pairs apart, so the 2,000-base window this project measures at IRX2 is mostly the other gene's window as well. Three separate gene catalogues were checked and they disagree about the exact distance, but all of them say the window is at least three-quarters shared.
The junction-peptide absence was a claim about mass-spectrometry datasets bounded by a PubMed abstract search: PRIDE holds three human MLL-AF4 proteomics projects, so measured-not-searched is now named rather than asserted
This project says nobody has ever reported the short piece of protein that straddles the join in a fusion protein. That statement was made by searching paper summaries, but the claim is about raw data files - and three sets of the right kind of raw data are sitting in the public proteomics archive.
The K1130/K1133 numbering is HUMAN, stated in the paper's own words, and the immunising peptide proves it a second time without relying on that sentence
A 2015 paper said a protein called MLL1 is switched on and off by a chemical tag at two spots, numbered 1130 and 1133 - but proteins are numbered differently in mice and humans, and most of that paper's experiments were in mice. Reading the full text settles it: the numbers are the human ones, and the exact stretch of protein the authors used to make their detection antibody matches the human sequence letter for letter.
The KDM6 contradiction resolved one step: GSK-J4 inhibits UTX too, so it should ANTAGONISE a menin inhibitor - and nobody has combined them
This graph held two published claims pointing opposite ways: one says the cell's ability to strip a repressive chemical tag is what makes menin-inhibitor drugs work, while the other says a drug that BLOCKS that stripping kills leukaemia cells. Chasing the reagent rather than the biology moved it forward - the blocking drug turns out not to be selective for one of the two enzymes that do the stripping but to block both, including the one the first story depends on, which turns a vague disagreement into a sharp prediction that the two drugs should fight each other, and nobody has yet put them in the same dish.
The KMT2-family beat's own ledger: every query run, every gene dropped, every paper read and not used
This is the working record of one research beat: what was searched, what was found, what was deliberately left out and why. It exists so that someone repeating the work next year repeats the decisions and not just the searches.
The KMT2A C-terminal site class gives VTP-50469/DMSO = 1.13 - an apparent GAIN of H3K4me3 - and its IgG negative control sits at 2.4-2.6, so the class is background and the number must not be reported as a result
One of our three site definitions produced an eye-catching answer in the opposite direction. The no-antibody control showed the same pattern, which means that definition was picking out places where any experiment accumulates data rather than places the protein actually is. We are recording the number together with the reason it cannot be used.
The KMT2A exon-11 CTCF anchor carries a CBX3 peak and a BRCA1 peak in three of the very ENCODE experiments Q35 called empty - the peaks are in ENCODE's own pre-IDR files, present in both replicates, and deleted by the IDR threshold at global IDR 0.227 and 0.059
Last cycle this graph concluded that two DNA-repair proteins are simply not present on the KMT2A gene, because a strict peak-calling filter found nothing in thirteen public datasets. Re-reading the same datasets with the filter turned down shows the proteins are there, sitting on one small 147-base-pair element, and one of them missed the cut-off by nine thousandths.
The KMT2A-N signal at the AFF1 locus is bimodal, neither mode is on AFF1's TSS, and the flank test I ran first said the opposite
Looking at the DNA fifty letters at a time instead of two thousand at a time shows two separate hotspots around AFF1, and AFF1's own starting point sits in the dip between them. The bigger hotspot is just outside the window this project measures, sitting on a different gene's start.
The KMT2A-r AML patient table numbers MLLT10 on isoform P55197-1, not the canonical sequence: S674 is S658 and S705 is S689
A published patient dataset lists two phosphate spots on the AF10 protein, and against the standard version of that protein one is not even the right kind of amino acid while the other lands on a matching letter in the wrong place. Both are explained by the study using a shorter version of the protein: once you allow for that, both spots line up exactly.
The KMT2A-rearranged DepMap arm is not MTAP-deleted, and the gene that looked like the test cannot answer it
Last cycle we worried that a few of this project's best-looking drug-target numbers might be an accident of which cell lines we happened to compare, because cells missing a gene called MTAP are unusually sensitive to that whole chemical pathway. The answer is no: the most reliable warning signs point the other way, so the numbers are not an artefact of the panel.
The KMT2A::MLLT1 mechanism sweep: what was searched, what was read and not used, and the three retrieval traps that would have produced confident wrong answers
This is the audit trail for the mechanism half of this beat: every search that was run with its result, every paper that looked relevant and was set aside with the reason, and the three searches that returned a clean, believable zero for the wrong reason. It exists so that someone re-running this beat next year repeats the decisions rather than only the searches.
The LEDGF hole, measured: every published LEDGF-removal experiment removes it BEFORE or AT transformation, and none treats an established leukaemia in an animal
Scientists have shown many times that leukaemia cells cannot get started if you take LEDGF away first. Nobody has ever taken LEDGF away from a leukaemia that already exists in a living animal, which is the only version of the experiment that tells you whether a LEDGF drug would help a patient.
The length-matched control I demanded of myself: 42% of the abstract-to-full-text rise is reproduced by an unrelated same-field paper, so the specificity I claimed for it was roughly half what I reported
Yesterday I showed that reading whole papers instead of their summaries roughly tripled how well citations matched their claims, and then wrote down the obvious objection: a longer document shares more words with anything. Scoring each claim against unrelated papers of similar length shows about two-fifths of that improvement was the document's length rather than its contents.
The leukaemia term hid the answer for DPY30 — a rule-77 demonstration with clean rule-67 controls
Searching for DPY30 and any cell-state word inside the leukaemia literature returns nothing. Dropping the word leukaemia from the same search returns the paper that answers the question. The search was working perfectly and the conclusion would have been wrong.
The lncRNA-DOT1L-LAMP5 axis was measured on H3K79 in MOLM-13 alone: eight human samples, one fusion, no index-fusion arm
An entry on the lncRNA LAMP5-AS1 notes that the MLL fusions used are not named. The deposited chromatin sequencing is eight human samples from a single cell line, MOLM-13, which carries MLL-AF9 - so the H3K79 part of the story was never measured in the infant fusion.
The lymphoid look of the core does not survive an activity-matched null inside the same 332, and the pre-registered power check says 17 genes cannot carry the statistic anyway
Three of the seventeen genes are recognisable immune-cell genes, which makes the set look like a B-cell programme. But a random seventeen genes from the same pool already contain about one, and once the comparison is matched for how active the genes are, three is not more than chance. The apparent programme is a low-number coincidence.
The lymphoid-restricted DAXX dependency survives three nulls and two comparator restrictions - and ATRX, MEN1 and DOT1L do not survive the same test in the same lines
Three entries in this graph asked the same unanswered question: is it real that cells with this fusion need DAXX, and only the B-lineage ones? We ran the test, hard, four different ways - and DAXX passed every one, while the two genes this project treats as its flagship dependencies did not pass in the same cells.
The MCF-7 arm, and the clause that saved it: SUV39H1 at the exon-11 anchor reaches the 96.3rd percentile of 56,171 same-cell-type CTCF anchors in one replicate - but all six non-anchor KMT2A intervals are elevated too (medians 73-87), so the pre-registered FIRES clause does not fire, the verdict is NEITHER, and four replicates of one antibody in one cell line span 51.8 to 96.3 at the same 147 bp
In breast cancer cells the silencing enzyme does look present at the spot this project cares about - but it looks just as present everywhere else across the whole gene, so nothing about that spot is special. And four repeats of the same experiment in the same cells disagree wildly with each other, which is the most useful thing this arm produced.
The measurement everybody said was missing exists: in SEM cells KMT2A::AFF1 occupies 20% of unmethylated CpG sites, CXXC1 50% and KDM2B 89%, and only the fusion tracks menin (r2 = 0.96)
A 2017 paper about a different drug quietly measured, in human infant-leukaemia cells, where the KMT2A::AFF1 fusion sits on DNA and where two rival proteins that read the same DNA signal sit. The fusion uses only a fifth of the available sites, its rivals use half and nearly all of them, and what marks out the fusion's fifth is the presence of menin.
The measurement q148 called the one thing that could overturn its headline: scored as ~20 kb intervals instead of 2 kb promoters, the authors' GMP-like class is enriched in 0 of 24 untreated non-circular samples, while the two control classes reach 19 and 10 - so the promoter window was not the limitation
A previous beat found no sign of a progenitor gene programme, and its own best objection was that it had only looked at narrow windows around gene starts. Looking at the full wide regions the original authors defined gives the same answer, while two comparison sets behave as expected.
The measurement that would separate a fusion-partner effect from a lineage effect does not exist: GSE127507 deposits H3K79me2 for RS4;11 and MOLM13 and for ML-2 it deposits DOT1L, MENIN, MLL1n and an input and no H3K79me2 at all
Two of the three cell lines in this dataset were tested for the histone tag and the third was not. The third is the one that would tell us whether the difference we found is about the fusion partner or about the type of leukaemia.
The Mediator/BET deposition is 48 of 48 mouse MLL-AF9/NrasG12D samples: resolving the KMT2A status behind the MED23 verdict
This graph's verdict on the Mediator subunit MED23 rests on a study whose abstract never says whether its cells were KMT2A-rearranged. All forty-eight deposited samples are mouse leukaemia cells labelled MLL-AF9/NrasG12D, so they are KMT2A-rearranged - but mouse, myeloid, and the wrong partner.
The menin gap survives matching on menin's own signal, but only two deciles of the two strata overlap
Genes the fusion sits on are also the busiest genes, so a difference between them and the rest can be an artefact of how busy they are rather than of the fusion. Correcting for that shrinks the menin effect by about two-fifths but does not remove it - and the honest caveat is that the two groups barely overlap, so the correction rests on a narrow slice of genes.
The menin inhibitor in this head-to-head is at 50 nM and all three DOT1L inhibitors are at 100 nM - the comparison is dose-biased AGAINST the menin inhibitor - and compound 10's target engagement at exactly 100 nM is demonstrated inside the same series by a spike-in-normalised H3K79me2 collapse of 5.6-fold by day 3
A sceptic would ask whether the menin drug only looked better because it was given at a higher dose. It was given at half the dose of the others, and the same study separately proves the other drugs were working hard at the dose they got.
The menin inhibitor removes three-quarters of the fusion from its own promoters, replicated in two batches
The clinical strategy assumes that blocking the menin-MLL handshake knocks the cancer fusion protein off the genes it controls. In this dataset it does: four days of a menin inhibitor leaves about a quarter of the fusion at its target genes, and the same result appears in two separately sequenced experiments.
The menin-dependence test I proposed in the cycle-146 log cannot fire by construction - and what the one pair CAN answer is that 85% of gene-body occupancy is menin-dependent while histone clusters preferentially survive
Last cycle I proposed testing whether the fusion's exclusive target genes need menin, using the one patient in the dataset sampled before and during menin-inhibitor treatment. That test is impossible for a reason built into how the gene list was defined. What the pair can show is that the drug removes most binding - and that what it fails to remove is mostly histone genes, which look like background.
The menin-inhibitor plus PARP-inhibitor experiment is genuinely unrun, but its rationale should be DNA repair, not HOXA9 - and the evidence comes from a disease where menin is a tumour suppressor
This project recommended an untried drug combination on the reasoning that menin inhibitors should mimic switching off HOXA9, and a properly controlled search confirms nobody has run it. The check also turned up a better reason to try it from a different cancer, where losing menin directly cripples DNA repair - but there menin is protective, so the logic runs the opposite way round.
The menin, ENL and DOT1L replicates of the drug-exposed relapse sample agree with each other at Spearman 0.11 to 0.25, below the floor this project set in the previous cycle, so no quantitative claim from that arm survives
In the sample taken while the child was on the drug, running the same experiment twice gave almost unrelated answers for menin, ENL and DOT1L. Two runs of the same measurement that disagree that badly cannot settle anything - including the published claim that ENL stayed put.
The menin/H3K79 partition this graph leans on was measured in MURINE MLL-AF9 - resolved from GEO, because the abstract does not say
A key screen split the H3K79 chemical mark into a menin-dependent half and a menin-independent half. The paper is paywalled and its summary never says which cells were used - so we looked up the data it deposited, which names them: mouse MLL-AF9 leukaemia cells.
The menin/Polycomb split between NLRC5 and CIITA does not extend to their pathways, and the MHC locus is structurally invisible to the method that found it
Two genes switching on the two halves of the immune-visibility system sit in opposite categories in a published dataset, which is a tidy story that does not hold for the rest of the machinery - the other genes are distributed exactly as chance would put them. The genes that matter most cannot be examined by this method at all, because that part of the genome exists in many alternative versions.
The methylome co-mutation table, recomputed: IDH2 up at p = 0.0004 and not p < 0.0001, IDH1 and DNMT3A flat, and the abstract disagrees with its own Table 2
The claim that IDH2 mutations are commoner in KMT2A-rearranged leukaemia was checked by recalculating it from the paper's own numbers. The enrichment is real. The p-value quoted everywhere - including in this project's own notes - came from the paper's abstract and is ten times stronger than the paper's own table says.
The MHC was never invisible - the resolver was wrong - and once the classical chains are joined the class I machinery is bivalent, not unmarked
I reported that a family of immune genes could not be examined because their names match several database entries, and drew a rule from it - the rule was half wrong, because a standard registry names one official entry per gene and brings them all back. Redoing the test shows the fragment-display machinery carries the activating and the silencing mark at once, which is a weak lead rather than a result.
The MLL-fusion interactome recovers corepressor SCAFFOLDS and not their HDAC1/2 catalytic core: NuRD splits 5/7 remodelling arm versus 0/7 deacetylase arm (Fisher p = 0.021), and HDAC1, HDAC2, RBBP4 and RBBP7 are absent from all seven baits
NuRD is a two-part machine: a motor that slides nucleosomes and a separate deacetylase that strips acetyl marks off histones. Sorting the published MLL-fusion pulldown by which part each protein belongs to shows that the motor half is well represented and the deacetylase half is completely missing - and the same is true of the rival Sin3 complex, whose scaffold shows up while its deacetylase does not.
The monocytic confound is worse than recorded: two of the six rearranged lines are one patient, and the four-line control group contains a HOXA-driven fusion and two NPM1c/myelomonocytic lines
This project's iron-and-haem finding depends on comparing leukaemia cell lines that carry the fusion against ones that do not, and the earlier check said the comparison simply had too few cells lines to settle. Looking up every line's record one by one shows the problem is not only the count: two of the fusion lines came from the same patient, and one of the four comparison lines carries a different fusion that drives the same genes.
The mouse fusion fraction is 0.73 and the human proxy fusion fraction is 0.21-0.24 - a three-fold gap in the direction a retroviral over-expression model predicts, so the one place the masking factor can be calibrated against a tag is the one place its value does not transfer
The mouse model that carries a clean tag also carries far more of the cancer protein than a human patient's cells do, because it was put there artificially. So the number you can measure cleanly is not the number you need.
The MRN asymmetry adjudicated: RAD50 -0.114 sits with NBN -0.047 and not with MRE11 -0.542, so it is not one subunit against two - and yet MRE11 and RAD50 covary across 1,208 cell lines at r = 0.554, the same as BRCA1 against BARD1, against a background of zero
Two entries in this graph noticed that one member of a three-protein machine looks eleven times more important than another in the big knockout screens, and neither could explain it. Fetching the third member shows the machine's parts rise and fall together across a thousand cell lines exactly as a true partnership should, while their average scores are simply on different scales - so the puzzle was about the yardstick, not about the biology.
The murine MLL1 ChIP that sits beside a FLAG-tagged fusion in GSE260742 uses Bethyl A300-086A - the identical antibody to the human MLL1n track in GSE127507 - which no GEO record in either series would tell you, and which is what makes a masking calibration against a tag possible at all
The mouse experiment that has a clean tag on the cancer protein also ran the same antibody the human experiments use. That was not written in the public database record for either dataset; it took opening a supplementary spreadsheet attached to the paper.
The murine-coordinate sweep is a NEGATIVE result - zero new defects in 3,086 entries - and its actual yield was a fourth entry that had the degron right before Q191 did
Last cycle three entries were wrong because they used a mouse measurement as if it were a human one. A search of all 3,086 entries for the same mistake found none left. It did find a fourth entry that had already got the answer right.
The named-but-unwritten queue is keyed on gene symbols, so CD15 is invisible to it at 23 entries
We built a tool that finds molecules this graph talks about a lot but has never written an entry for, and it works by looking up gene names. Some of the most useful markers in this disease are not made by a single gene - CD15 and CD65s are sugar structures assembled by enzymes - so the tool can never see them no matter how often we mention them, and CD15 is mentioned in 23 entries. There is a second and sneakier version of the same problem: TdT is a real gene, but we call it by its enzyme nickname in 13 entries and use its actual symbol in only 3, so the tool sees a quarter of the discussion and decides it is not important enough to report.
The named-but-unwritten queue's top two entries were a web API and a survival endpoint, and MYC was the real answer all along
A tool in this project lists molecules the graph mentions often but has never written up, ranked by how often they appear, and its top two answers were REST and EFS - but in every case checked the graph was saying 'the REST web interface' and 'event-free survival', the commonest survival statistic in childhood cancer, and both happen to be real gene names too so the tool could not tell the difference. With those and two others blocked, the honest top of the list is MYC, a gene mentioned in a hundred entries and never written up.
The null the claim needed: the KMT2A BCR ranks 72 of 18,082 chromosome-11 bins for its H3K9me1 response - top 0.40% - while the same locus ranks 3,877th for H3K9me3 and 10,276th for H3K9me2
Saying a change at one gene is 'one of the strongest in the genome' only means something if you rank it against everywhere else. Ranked against every comparable stretch of chromosome 11, the KMT2A gene is in the top half of one percent for one tag and unremarkable for the other two.
The obvious result was that VTP-50469 cuts H3K4me3 at HOXA9 three-fold in SEM. The pre-registered housekeeping control says GAPDH, RPL13A and TUBB fell just as far, so the obvious result was the normalisation
Our first look said the menin inhibitor slashed the chemical mark at the classic leukaemia genes by about three-fold - exactly the expected answer. But genes that have nothing to do with leukaemia fell by the same amount, which means the drop was an artefact of how the data are scaled. We had written down in advance that this exact pattern would mean 'not biology', so we threw the result away.
The occupancy rule this project has used since q145 calls a tenth of ALL promoters occupied, and no entry using it records that
The test this project uses to decide whether a gene is 'switched on' by the fusion protein says yes for about one gene in ten, chosen at random. That does not make the test wrong, but it means a single gene passing it is much less surprising than it looks, and nobody had written that number down.
The oncoprotein-target site count for sample 148752 is 46, 345 or 745 depending on the interval set and the universe, and the corpus has been quoting one of the three with no label
How many genes a cancer protein sits on depends entirely on which stretches of DNA you decided to count. The same sample gives 46, 345 or 745 depending on that choice, so a bare count means nothing without saying which regions were used.
The one experiment that would separate the H3K9me1/2 rearrangement brake from H3K9me3 was run, deposited as eight public samples, and reported nowhere in the paper
The paper that found a silencing tag controls whether the KMT2A gene breaks also measured a closely related tag that should have behaved differently, and never showed the result. The data are sitting in a public archive, eight samples, and nobody has looked.
The only absolute-scale H3K79me2 deposit this project has contains no IgG or no-primary library, so the 1-31% residual measured after DOT1L inhibition cannot be separated from antibody background - the single cheapest missing control in the dataset
After nine days of the drug, genes still show one to thirty per cent of their original chromatin signal. Nobody knows whether that is real leftover mark or just the antibody sticking where it should not, because the experiment includes no control for that - and neither does any similar public experiment we could find.
The only deposited H3K79me2 kinetic series for a menin inhibitor is non-monotone - a minimum at 16 h and a recovery at 24 h that survives common support and two rank statistics - and its 24 h point shows MORE redistribution than the 48 h main-arm sample, so the 48 h/72 h asymmetry in cycle 82's F stands and F keeps the words lower bound
A public dataset follows the H3K79 tag for 4, 8, 16 and 24 hours after the menin drug goes on, which could have told us whether a number this project committed yesterday is near-final or only a floor. It does not: the tag is lowest at 16 hours and back up at 24, five ways of measuring it disagree about which way that last step goes, and the number keeps its floor.
The only second mark with a drug arm in SEM is H3K27ac, and its replicates agree at r = 0.21-0.29 against H3K4me3's 0.78-0.83 - so GSE159608's SEM EPZ-5676 experiment has one usable readout, not two
There is a second chemical mark measured under one of the two drugs, which would have been a useful cross-check. Its repeat measurements barely agree with each other, so it cannot check anything. Recording this matters because the dataset looks from its labels as though it has two readouts.
The only two trials stopped on a stated risk-benefit judgement are GSK2879552 (LSD1) studies in AML, while iadademstat, another LSD1 inhibitor, ran a KMT2A-rearranged-enriched expansion with an acceptable safety profile - so the stop looks molecule-specific, and its reason is unread
Two early trials of one leukaemia drug that blocks the enzyme LSD1 were closed because the balance of benefit and risk did not favour continuing. A different LSD1 drug went on to treat people with KMT2A-rearranged leukaemia, so the problem was probably that molecule, but the reason was not written down in the registry.
The order in which the complex leaves fusion-occupied promoters, all fifteen channels on one scale
Three hours after the fusion is destroyed, its partners come off its target genes in a definite order: DOT1L first, then menin, then the elongation machinery, then the histone marks - and one mark does not move at all. The fusion itself is essentially all gone, so everything else is only partly following it.
The PAF complex species is mouse - and its only deposited series is Hoxa9-ER, not an MLL fusion at all
An entry inferred, rather than knew, that a PAF complex result was done in mouse cells. The deposited data confirms mouse - twelve samples from female C57BL/6 mice - but also shows that the deposited experiment is not an MLL fusion experiment at all; it is a HOXA9 time course.
The panobinostat H2Bub result is SEM and KOPN8 against a REH comparator: three human lines, two KMT2A partners, one non-KMT2A control
An entry on the drug panobinostat notes that the paper does not name which KMT2A fusions or cell lines it used. The deposited data names them: SEM and KOPN8, two human infant-leukaemia lines with different KMT2A partners, compared against REH, a leukaemia line with no KMT2A rearrangement at all.
The paper's four comparison loci, marks read against FISH calls: the H3K9 mark response agrees with the cytogenetic response at two loci and contradicts it at two - the mark is neither necessary nor sufficient
The paper tested four other genes to show its mechanism is specific: two broke and two did not. Reading its own raw data shows the chemical tag it blames moved at one of the genes that did not break, and did not move at one of the genes that did.
The paralogue split survives the expression control and the ALAS2 control does not: ING4 is transcribed MORE than ING5 and is not required at all
A gene that a cell never switches on will look unnecessary in any experiment that switches genes off, so we checked whether the genes that looked unnecessary were switched on. Most were. One was not, and that one was being used as evidence.
The partner-matched control arm is not 241 versus 219 - it is 153 versus 101, and in that subset the two groups make an identical fusion protein
Published breakpoint data on 302 patients splits this leukaemia into groups finer than anyone has used. Two of those groups - 153 patients and 101 patients - build exactly the same fusion protein and differ only in whether a mirror-image gene also exists, which is as close to a controlled human experiment as this field is ever likely to get.
The per-library 'cannot be occupied' floor spans 0.00x to 4.01x an average promoter, and one core gene's AFF1-exclusivity rests entirely on it
The occupancy test uses eight genes that should never be switched on in leukaemia as a per-sample noise floor. In one sample that floor reads four times an average gene, so in that sample almost nothing can pass. One of the seventeen 'AFF1-only' genes is only AFF1-only because of that one sample's high floor.
The permutation cycle 155 asked for: the two-instrument exclusive intersection is four genes not one, 34 of the 42 could never have been in it, and the surviving IRX1 claim gives p=0.092 against a patient-level null
We had one encouraging result left about IRX1: two different ways of reading the same experiment both said it was switched on only in this fusion. We finally ran the fairness test we had been promising - shuffle which patients count as having the fusion, and see how often the same pattern shows up by luck. It shows up about one time in eleven, which is not rare enough to mean anything. Two other things also went wrong: there were really four such genes, not one, and three quarters of the genes we were comparing against could never have qualified in the first place. This closes the line of work.
The positive control this corpus has never had: degrading MLL::AF9 clears its own tag from 93-97% of its strongest chromatin sites within 3 hours, and two INDEPENDENT sequencing batches agree to within 4.6 points - the first technical noise floor ever measured for a ChIP eviction comparison in this project
Every claim in this project about a drug knocking a cancer protein off DNA has been measured against another drug, never against simply deleting the protein - so we checked a public dataset that did the deletion, and three hours after the protein is destroyed its own molecular label is gone from about 95% of the places it used to sit. The same experiment was run twice independently and the two runs agree closely, which is the first time anyone here has been able to say how much of such a number is real and how much is noise.
THE POSITIVE RESULT: the writer G9a, the reader HP1a and the scaffold KAP1 all sit on the KMT2A PROMOTER in MCF-7 at the 87.6th-98.9th percentile of 2,000 ACTIVE promoters - surviving the control that says an active promoter is ChIP-hot for everything - while none of the three is above the 75th percentile at the exon-11 anchor 52 kb away
Three proteins that work together to silence DNA all sit on the switch that turns KMT2A on, more strongly than they sit on almost any other switched-on gene. None of them is at the spot 52 kilobases away where the gene breaks in this leukaemia.
The PRC2 dependency profile, lineage-split and expression-controlled - and the first time this project's paralogue rule came back positive
We scored every part of the Polycomb machine against a large public database of which genes cancer cells need, split the leukaemia lines by lineage, and then checked whether the parts that looked unnecessary were actually being made at all. The pattern exactly matches what the mouse genetics predicted: the interchangeable parts look unnecessary, the irreplaceable ones do not.
The pre-registered magnitude test failed because the statistic was bounded by construction - top-decile share spreads only 1.1-1.4x between samples and cannot express what the published oncoprotein score says varies
I planned to test whether the strength of the fusion's binding differs between patients, using a measure that turned out to be incapable of showing a big difference even if one existed. The test is reported as void rather than as a negative result.
The pre-registered three-branch test of 'the HOXA-negative blast's oncoprotein is on the GMP programme' returns NEITHER/UNINFORMATIVE, on four independent gates
We wrote down in advance what would count as proof that the fusion protein moves to myeloid genes, what would count as proof it does not, and what would make the whole test say nothing. The third one happened, four times over.
The premise that GSE260742 'carries a ChIP-Rx spike-in' does not survive reading the deposit: zero of 158 sample records mention a spike-in, the only ChIP-Rx experiment in the paper is an H2AK119ub panel that is not identifiable among the deposited samples, and 152 of 158 samples ship only an unnormalised igvtools .tdf
Before comparing a protein's position on DNA between a drug and a control, you have to know how the two experiments were put on the same scale, and here they were not. The public files are raw coverage with no scaling information at all, so any fold-change between conditions would be measuring the scaling rather than the biology.
The prescribed GEO lookup returns zero for every paper this project has ever resolved, including the one it was written from
The recipe this project wrote down for finding a paper's deposited data does not work. Asking the data repository for a paper by its PubMed number returns nothing - for all twenty-six papers tested, including the one the recipe was originally written from. A different call returns the answer.
The prey-only interactome is 36.3% written, not 0% - the census, by identity, of all 212 human prey-only partners of the six core proteins
The project has been saying that the 'leftover' half of the interaction list - proteins seen only as passengers in somebody else's experiment - is completely unwritten. Counting them one by one against what is actually on disk, 77 of 212 already have an entry.
The promoter table's second blind spot is not annotation drift at 1.8% - it is one row per symbol at 14.7%, and the case that motivated the question is not displaced at all
This project scores gene activity in small windows built from a 2020 gene annotation. The worry was that newer annotations have moved those windows. They mostly have not - 82% sit on exactly the same spot. The real problem is different: a gene can have several start sites, and the table keeps only one.
The PTM-database route runs: K1130 and K1133 acetylation are independently re-observed in deposited mass spectrometry, K1133 in the CBP/p300 acetylome itself - and K1130 is also a ubiquitination site nobody has mentioned
Rather than trusting the literature, we asked the raw protein-measurement databases directly: has anyone's machine ever seen this chemical tag on MLL1? Yes - in four separate published experiments, none of which set out to study MLL1 at all. One of them also shows a completely different tag on the same spot, which no paper has ever mentioned.
The published peak classes do transfer from 20 kb peaks to 2 kb promoters - the GMP-over-B-lineage ratio swings 8.56-fold across the pair that defined them - but only in the between-sample direction the classes were built in
Before using someone else's gene lists you have to show your own version of their measurement still sees what they saw. It does: in the one patient whose before-and-after samples produced the lists, the myeloid-to-lymphoid balance flips nearly ninefold.
The rarest thing in the deposit, measured: across 59,821 CTCF anchors in the SAME nuclei, SUV39H1 occupancy and H3K9me3 correlate at Spearman 0.148 and 0.078 - and a quarter to two-fifths of the enzyme's own strongest sites carry no more of its mark than the median anchor
This dataset is unusual because it measured the enzyme and the chemical tag it makes in the very same cells. They barely track each other: where the enzyme is strongest, the tag is often no higher than average, which means the enzyme is doing something at many sites other than leaving its mark - or the antibody is not finding only the enzyme.
The reciprocal-antibody ratio splits the six-gene AFF1-exclusive class in two: H3C15 and SPAG7 sit at or below background, the three most enriched are the three the authors' caller says are not exclusive, and IRX1 is the only member that is both
The dataset has two antibodies against the same protein: one recognises the part kept in the fusion, the other the part lost from it. Comparing them tells you whether a site looks like the fusion or like the normal protein. Of the six genes this project calls fusion-specific, two show no fusion signal at all, and only one is both fusion-enriched and confirmed unique to this partner.
The referral-denominator audit: 97 frequency claims, 63 qualified, and one real defect where the ledger predicted seven
Most of what this graph knows about which gene partners KMT2A is broken with comes from one reference laboratory that receives samples sent to it from around the world, so its numbers describe who gets referred rather than who gets ill. The project's own to-do list said a warning about this had been copy-pasted into at least seven entries and that later corrections had reached only one of them. Checking it properly, that is not what happened: the warning sits verbatim in two entries, and of the ninety-seven places where the graph quotes a number from that series, sixty-three already carry the warning and exactly one was missing it in a way that mattered.
The reopening watchlist read for the first time: 0 of 10 candidate papers meet their condition, 8 of 10 watched decisions returned nothing, and not one of the ten titles names the gene it was retrieved for
A script watches for new papers that would overturn decisions this project has made, and nobody had ever read what it found. Reading it: none of the ten papers it flagged actually bears on the decision it was flagged for, and none of them even mentions the gene in its title.
The residual fraction this graph says is absent from the entire literature, measured: FLAG-MLLAF9 enrichment at its own top 1,228 sites falls from 15.9x the genome median to 3.0x under VTP50469 - about 18% residual - and 12.8% of those sites stay in the top 0.1%
When the menin drug is applied, the leukaemia fusion protein keeps only about a fifth of the grip it had on its favourite stretches of DNA. That fifth is the number every occupancy-based drug-response test would have to be built on, and until now nobody had published it.
The residue-type numbering check passed a site that is in the wrong place
We checked whether a list of measured protein sites used the same numbering we do, by asking whether each position held the right kind of amino acid. It turned out that test is too easy to pass by luck.
The resolution floor: every deposited H3K9 track in GSE210480 is 1,000 bp binned, so the 147 bp exon-11 question cannot be answered from the processed files - and the planned analysis that asked for 500 bp windows was asking for something that is not there
The public files for this experiment average the signal over 1,000-base blocks. The piece of DNA the whole argument is about is 147 bases long. So the files can tell you about the neighbourhood but not the address, and the finer version of the question needs the raw sequencing reads instead.
The REST-wide direction audit Q104 asked for: scaffold-as-bait rows DO exist in the 36% of the record PSICQUIC hides - KMT2A is a bait in 96 of 376 rows across 18 publications and PSIP1 in 11 of 114 across 4, but the most recent is 2015 and 2012, and none of the thirteen never-discussed partners comes from one
Somebody checked earlier whether anyone had ever tagged the proteins this project studies and asked what sticks to them, and found nobody had - but only in a slice of the database. Re-running it over the whole database shows they have been tagged, just not since 2015 for one and 2012 for the other.
The resting-state H3K9me2 depletion over KMT2A DOES replicate in a human CD34+ haematopoietic progenitor - measured in 3 of 3 progenitor ChIPs, where the locus sits in the 7th, 11th and 37th longest H3K9me2-free gap on chr11 - but it is not a progenitor property: granulocytes, K562 and 6 of 10 primary AML samples show the same gap
The first step of the published KMT2A-breakage brake is that a silencing tag is unusually scarce across the stretch of DNA where the gene breaks, and until now that had only been shown in an eye-derived cell line. It is now measured in real human blood stem cells and it holds - but it also holds in mature white cells, in a leukaemia line and in most leukaemia patients' cells, so it is a property of the place rather than of the cell.
The RNA layer scored for KMT2A-selectivity across 79 genes, lineage-split — PTBP1 is the only menin-shaped hit, and the IGF2BP family shows no co-dependency
We asked whether the cell's RNA-handling machinery is needed more by KMT2A-rearranged leukaemia cells than by other leukaemia cells, testing 79 genes at once. Only one, PTBP1, stands out — and the three IGF2BP siblings do not back each other up in any way this kind of experiment can detect.
The rule-64 complex-form query that the RPA2 decline flagged as unrun has now been run, and it finds a published damage-state RPA occupancy measurement AT the KMT2A breakpoint cluster region - so two of this corpus's absence entries carry ids that are broader than the claims inside them
Last cycle an entry in this graph said nobody had checked whether the whole RPA complex - rather than its individual subunits - had ever been linked to the leukaemia gene KMT2A, and recorded that as an open question. We ran that check. It found a 2010 paper that measured RPA sitting on the exact stretch of the KMT2A gene where the cancer-causing breaks happen, after chemotherapy damage.
The same baseline-matched test that kills the target effect under EPZ5676 fires hard under compound 10 - at d6 and d9, in both cell lines, 4 of 4 contrasts (p = 0.0012 to <0.0001) - and in MOLM13 HOXA9 holds 24.5% of its H3K79me2 where the genome median holds 4.5%
Two drugs that block the same enzyme do not strip the chromatin mark the same way. The clinical-stage one removes it evenly from every gene. A newer one removes it everywhere too, but leaves several times more of it behind at exactly the genes a KMT2A fusion depends on - and this shows up in both cell lines and at both later timepoints.
The same occupancy rule passes 0.61% of gene bodies and 10.75% of promoters, so IRX1 at 3 of 27 is 18x its gene-scale expectation where at promoter scale it was below it
The test this project uses to decide whether the fusion protein is sitting on a gene passes about one promoter in ten - so any promoter looks occupied in two or three of the twenty-seven patients just by chance. Applied to whole gene bodies instead, the same test passes only about one gene in a hundred and sixty, so chance gives you a sixth of one patient. That is why the IRX1 result looks weak at one scale and strong at the other: the two numbers were never comparable.
The same-condition H2AK119ub noise floor in GSE260742 spans twentyfold across the deposit's three experimental arms - D1 |R-1| of 1.89% in the PCGF1-KO clone, 10.49% in the non-targeting control and 37.04% in the MLLAF9 Control arm - at one mark, one lab, one pairing shape and one external stratifier. A single same-condition floor for this deposit does not exist.
When the same experiment is repeated, the two runs never match exactly, and how badly they disagree is the yardstick for deciding whether a real effect is real. In one public dataset that yardstick turns out to be twenty times longer in one part of the experiment than in another, so borrowing one number for the whole dataset silently changes what counts as a finding.
The score's amplitude correlates with sequencing depth at Spearman -0.474 (p = 0.012) - a real and significant depth dependence that cleared the pre-registered voiding threshold of 0.6 only because the threshold was set loosely
Samples sequenced more deeply give a smaller spread on this measure, which means part of the spread is measurement noise rather than biology. The pre-registered check for this passed, but only because the bar was set low.
The screen this beat was dispatched to run 'because it has never been run' was published in Haematologica in 2022, used the same instrument, and is cited in this graph already - in an entry filed under innate immunity
This project asked for a search of every surface protein in KMT2A-rearranged leukaemia, on the grounds that nobody had done one. Somebody had: a 2022 paper filtered leukaemia expression data through exactly the same published surface-protein list and found LAMP5 - and this project already quotes that paper, but filed under inflammation rather than under targets, so nobody looking for antigens could find it.
The SEC-family target sets, reconstructed as a full three-way Venn: 27.5% of the family's transcriptional output is invisible to AFF4 depletion — and the two 'distinct' sibling sets overlap 8-20x more than chance
The three versions of the transcription-speeding machine were each switched off in turn, and the genes that changed were counted. Most of the action belongs to the main machine, but about a quarter of the affected genes respond only to the two side machines — and the two side machines overlap with each other far more than you would get by chance.
The selection control, run at last: two biologically identical vehicle-treated input libraries produce a MONOTONE decile gradient from 0.9040 to 1.2349 - wider than the real FLAG gradient cycle 103 called a finding - when the strata are cut on the pair itself, and are flat within 0.73% and non-monotone when the strata are cut externally. Monotonicity across deciles is therefore NOT evidence against a coverage artefact. Cycle 103 is safe because it cut externally
A common way to analyse this kind of data is to sort the genome into bins by how much signal they carry, then look for a trend across the sorted bins. This shows that if you do the sorting using the very samples you are comparing, you get a clean, steady trend out of two samples that are biologically identical - bigger than the trend a real biological effect produced. Sorting with a third, independent sample removes it completely.
The serine arm replicates in a third KMT2A::AFF1 line from an independent laboratory - PHGDH -2.77, PSAT1 -2.17, GPT2 -2.34, ASNS -1.16 - but the ISR output genes do not, and it survives PCGF1 and BCOR knockout
A finding from one leukaemia cell line was retested in a third line, measured by a different laboratory asking a different question, and the amino-acid-making machinery switched off there too. The stress-response genes that had looked like part of the same story did not repeat, so that half of the original reading was probably wrong.
The serine/one-carbon arm is NOT a pure ATF4 output: it is menin-selective where the ISR arm is compound-independent (median margin -0.55 vs -0.02, p=0.011), and it is ABSENT in KMT2A::MLLT3 AML where the ISR arm is present
This graph has argued for four cycles about whether a menin inhibitor switches off serine manufacture because it switches off the stress factor ATF4, or by some separate route. Two measurements say separate route: a DOT1L inhibitor copies the stress-gene shutdown almost exactly but only half-copies the serine shutdown, and in a different kind of KMT2A leukaemia the stress genes still fall while the serine genes do not move at all.
The shared-epitope masking calibration, run against a FLAG tag and then refuted by its own control: the MLL1 antibody reports 2.23-fold more retained signal than the tag at the fusion's own sites, but MLL1-dominated non-fusion sites fall HARDER than the fusion does (0.1069 against 0.1793), so no mixture of fusion and wild-type MLL1 can produce the observed 0.3995
The obvious way to work out how much of an antibody's signal comes from the cancer protein is to compare it with a clean tag on that protein in the same cells. Doing it gives a tidy-looking answer of about three quarters - and the control built in to check the assumption shows the arithmetic cannot be right.
The six open read-queue members split three rejected and three earning entries - and the flattest dependency in the set, ASB7, turns out to be a negative regulator of H3K9me3 via SUV39H1 degradation
Six candidate partners were left undecided because the quick numbers could not settle them. Reading what each protein actually does settled all six, and the one with the least impressive numbers turned out to be the most interesting.
The six unwritten DOT1L partners are not six findings - they are nine spokes of ONE GFP-DOT1L pulldown, and the source paper already graded them
All six DOT1L partners this project was missing come from a single experiment in a single paper - one pulldown of tagged DOT1L out of HeLa cells - and that paper's own authors already marked them as its weakest hits. Counting them as six separate discoveries overstates the evidence sixfold.
The specificity control that should have clinched this beat inverts instead: the C-terminal MLL1 epitope the fusion does NOT contain loses MORE than the N-terminal epitope it does contain (68.8% vs 78.4% retained), and because the deposit records different cell-line names for these four samples than for every other one, the arm is reported as uninterpretable rather than interpreted
We had what looked like a perfect internal check - one antibody sees both the cancer fusion and the normal version of the protein, another sees only the normal version, and destroying the fusion should have moved the first while leaving the second alone. It did the opposite, and rather than invent a story for that we are recording that the check failed and why we cannot trust it.
The spike-in ruler this thread has been blocked on for eight cycles is deposited as a 2 MB text file, and it includes RS4;11 - the index fusion: under DOT1L inhibition the fly fraction rises 4.9x to 23.4x against matched DMSO, so H3K79me2 collapses GLOBALLY by roughly that factor, and a within-sample share would have shown almost none of it
All of this project's chromatin measurements are proportions - what share of a signal sits in one place versus another. Proportions cannot detect a change that affects everything at once, because the shares still add to one. Fixing that needs a known amount of fruit-fly material added to each sample as an external ruler, and the project had concluded no such ruler was available. One is, in a public two-megabyte text file, for a cell line carrying the exact fusion this project studies. It shows that a drug of the kind used here can wipe out the signal across the whole genome by five- to twenty-fold - a change the proportion-based method would barely register.
The standing duplicate-check exits 126 and returns nothing at 3,000 files; it caused no duplicate ids, but the audit it should have been doing found two superseded entries that never name their successor - one of them a live safety verdict
Every cycle the coordinator runs a command to list everything already written, so as not to write the same thing twice. That command silently stopped working once the collection passed a few thousand files - it returns nothing, and returning nothing looks exactly like 'nothing matches, go ahead'. Nothing was actually duplicated, because a different check catches that. But looking properly turned up two entries that were replaced by better versions and never mention it, and one of them still tells readers a drug target passed a safety screen that its replacement says it failed.
The standing priority asks for localization on every interactor; the graph has it on 14% of genes and 24% of its most central ones, and DOT1L has none at all
The lab asks that every protein partner record where in the cell it sits and which cells it matters in, and about one gene entry in seven has any location at all - counting generously. The protein targeted by this field's second drug class has no location recorded anywhere in its entry.
The standing worry about same-source counter-evidence, read at n=50: 84% are the author's own caveat or a point of logic, 16% could have cited independent literature, and none of fifty is unsupported
When this project records an objection to a finding, it sometimes cites the same paper the finding came from, and we have worried for many cycles that this is a weakness. Reading fifty of them at random shows it usually is not: 84% are the study's own stated limitation or a plain point of logic, which the paper itself is the right source for.
The stratification passes its own positive control: the fusion tag separates the two strata 62-fold
Before believing any result that splits genes into 'where the fusion was' and 'where it was not', you have to check that the split actually finds the fusion. It does: the fusion's own tag drops to 1% of its starting level in the first group and stays at 61% in the second.
The stricter null, and it partly reverses the last one (run)
Our previous control compared each candidate gene against random genes, but many random genes barely vary at all and so can never look interesting. Comparing each candidate only against random genes that vary as much as it does changes the answer for three genes we had written off.
The super elongation complex, sampled from the RNA side: ELL, ENL, CDK9 and cyclin T1 are all in the proximity record - ELL2, EAF1/2, AF9 and AF10 are at zero
Using a public database we counted, for every part of the elongation machine, how many tagging experiments have ever found it sitting nearby. Some parts appear repeatedly; several appear not once.
The taspase-1 N/C antibody control separates fusion-occupied from wild-type KMT2A promoters in every primary infant sample, and it is the C-terminal arm that does the work while being the less reproducible arm
Because the normal KMT2A protein is cut in half and the cancer fusion keeps only the front half, an antibody to the back half sees only the normal protein. That difference cleanly separates the genes the fusion sits on from the genes normal KMT2A sits on, in all three patient samples.
The ten reported menin-inhibitor escape substitutions are each one nucleotide change from the MEN1 sequence - which the sources could not have shown otherwise - so the useful outputs are a 25-variant validation panel and a burden threshold, not the headline
Every drug-resistance change reported so far in the drug target, menin, needs only one letter of DNA to change. That sounds alarming, but it is partly built in: the lab method used to find these changes can only make one-letter changes, and patients' cells change one letter at a time. So it does not show that two-letter escapes do not exist. What it does give is a list of 25 other one-letter changes at the same spots that nobody has tested, which any new drug should be tried against, and a rough rule for when a tumour is large enough that some resistant cells are likely to be there before treatment starts.
The third silent-failure class, audited after being named in four king logs: exactly one of seven bigWig analyses is raw-coverage, the single incommensurable pair is already labelled three times, and the cross-pipeline ratio comparison was bounded at the point of measurement
Two ways of measuring a protein on DNA give numbers that cannot be compared directly, and this project worried for four cycles that it might have mixed them. It mixed them once, and had already written the warning next to it.
The twelve partners with no entry are three times LESS spoke-expanded than the interactome they came from - 23.1% spoke-only against a 70.6% null, Fisher p = 0.00068
Everyone assumed the leftover partners nobody had written up were the weakest ones. Measured against the whole database, they are the opposite: three quarters of them rest on experiments somebody actually did, where across the interactome as a whole only three in ten do.
The two beats' opposite answers are produced by the calling rule, not by the promoter universe: three universes give one answer, two rules give two
Two earlier analyses of the same data reached opposite conclusions. Running every combination shows the difference came from how they decided what counts as a target gene, not from which genes they looked at.
The two instruments' opposite verdicts decompose: the CALLER explains the 16-fold magnitude gap and the UNIVERSE explains the patient-level survival - and inside the shared universe the only gene both call AFF1-exclusive is IRX1
Two ways of measuring where the fusion protein sits gave opposite answers on the same patients. Testing them against each other shows the disagreement has two separate causes, not one - and that the single gene both methods agree on is IRX1.
The two silent-failure classes found in cycle 59 did not propagate through the corpus - 13 candidates, all legitimate on reading, and the audit corrected itself twice before it found anything
Two ways of getting a wrong number without any error message turned up last cycle. This checked whether the same mistake had been made anywhere else in the project, read every candidate, and found none - though the checking tool had to be fixed twice first, in exactly the way it was built to detect.
The two TARGET B-ALL studies in cBioPortal are 868 of the same patients assayed twice, so joining them DOES answer the question I had just written off as unanswerable: all four ARID1A-mutant B-ALL patients are definitively KMT2A-negative, three of them with truncating calls. The answer is zero of three - and at a 2.7% KMT2A-positive rate an empty overlap has probability 0.92, so it is an upper bound and not a finding
The previous cycle found two children whose leukaemia carried both genetic changes this project studies together, but they had the wrong fusion and the wrong lineage, so the question stayed open for the disease that matters here. Checking the public databases suggested at first that no study could answer it, because the one study that sequences the first gene seemed to record nothing about the second - but the two studies turned out to be the same children assayed twice, so the records could simply be joined. Joined, they give an answer: all four patients with a mutation in the first gene were tested for the second and none had it. That sounds decisive and is not, because with so few patients an empty result is what you would expect about nine times in ten even if the two changes were completely unrelated.
The two VTP-50469 H3K4me3 libraries hold 3.3% and 3.8% of their reads in the DMSO-defined peak set against 8.9% for DMSO and 9.4% for EPZ-5676 - one global fact that explains every raw locus difference in the SEM menin-inhibitor arm, and that cannot be resolved into drug effect or library quality
One number explains almost everything that looked like a drug effect here: the two menin-inhibitor samples have about two and a half times less of their data sitting on the sharp peaks the mark normally makes. That could be the drug genuinely spreading the mark out, or it could be that those two samples simply worked less well. The deposited data cannot tell the difference, and saying which it is would be guessing.
The unbiased half of the token audit: 0 of 15 never-measured tokens are wrong, the entries my filters could NOT see are the best-scoped ones, and two scoping fields already existed that I had claimed to add
Last cycle I checked the gap labels most likely to be wrong and found them mostly right, so this time I checked a random set of the ones my checks could not see at all. None of the fifteen was wrong, and the entries my tools were blind to turned out to be the most carefully written ones in the corpus.
The unbounded oncoprotein score spans 1.73-12.32 log2 units within a sample - it is not the compressed statistic q142 was stuck with - and the pre-registered magnitude verdict is STILL void, because the score's amplitude is no larger than the disagreement between two antibodies against the same end of KMT2A
I rebuilt the measurement with a scale that can show big differences, and it does show them. But a control built from the same data shows the scale moves just as much when nothing biological changes, so the headline question still cannot be answered.
The unnamed AML lines are NB4 and THP-1 - and NB4 is not KMT2A-rearranged at all
An entry on the lncRNA HOXA-AS2 notes that the paper's AML cell lines are not named. The full text names two: NB4 and THP-1. Only THP-1 is KMT2A-rearranged; NB4 carries a completely different fusion, and it is the line the paper uses most.
The unnamed Cre is Mb1-Cre, which acts at the pro-B stage: the one resolution in this beat that lands on the right cell
An entry on DOT1L in normal B-cell development notes that the paper never says which genetic switch was used or when it acts. The deposited data names it: Mb1-Cre, which turns on right at the pro-B stage - the same developmental point where infant KMT2A::AFF1 leukaemia arises.
The unnamed KMT2A-rearranged B-ALL line is SEM, and the deposition says KMT2A::AFF1 in a genotype field
An entry on USP22 says the paper names a KMT2A-rearranged B-ALL line without specifying it, and does not say which fusion that line carries. The deposited sequencing names both: the line is SEM and the fusion is written out in full as KMT2A::AFF1 - the exact fusion this project is built around.
The unnamed osteolineage Cre is Prx1-Cre, which is not osteoblast-restricted - and that changes what the negative result means
An entry on osteoblasts rests on a negative result whose validity depends on which genetic tool the authors used, and the abstract never names it. The deposited data and full text name it: Prx1-Cre, which switches genes off in a broad mesenchymal population rather than in mature bone-forming cells - so the negative result is about the wrong cells to be read as an osteoblast result.
The unstated-system worklist: 26 sources carry a caveat of this class, 18 now resolve, and 13 of 16 new resolutions are imports
We searched the whole graph for entries that give up on a question by saying the source does not state what system it used, found twenty-six such sources, and went and asked the question. Two-thirds answered - and four out of five of the answers turned out to be a different fusion, species, lineage or tissue from the one this project studies.
The washout data cannot say whether the serine/ASNS collapse is durable - the washout control never recovers, ASNS is undetectable in the line, and the housekeeping drifts in opposite directions in the two matrices
We asked whether the drug's effect on the cell's amino-acid machinery lasts after the drug is removed, using public data that removes it. The data cannot answer: the genes that should bounce back as a check never bounce back, and the one gene the question turns on is not even switched on in that cell line.
The whole deposit is hg38 even where the filename does not say so, and the guard that checked it was worth running
Only some of the files in this dataset are labelled with the version of the human genome they were mapped to. Mixing versions would have silently produced a wrong answer, so the version was checked directly from the coordinates - and all of them turn out to be the same, newer version.
The whole protein-coding genome: menin is still first, and a metabolic pathway appears (run)
The earlier scan covered a quarter of our genes and put menin first; this one covers all of them and menin is still first, so nothing hidden in the rest outranks it. The new result is that four genes from a single chemical pathway — the one cells use to build haem — all rank very high, a pattern that was not visible before.
The whole-peak-set overlap metric cannot tell a MEN1 ChIP from a DOT1L ChIP in the same untreated cells (Jaccard 0.357, inside the 0.310-0.487 range of all 45 sample pairs) - so on the same data it scores cycle 83's threat as SURVIVING while the pre-specified locus set scores it as FALLING
One popular way of comparing these experiments turns out to be useless here: it cannot even tell apart two experiments that measured two different proteins in untreated cells. Using it gives the opposite answer to the careful method, on exactly the same data.
The within-donor null: MEN1 sits at the 99.4th percentile of variation between two sublines of ONE child, and at the 98.9th between donors - menin dependence is not heterogeneous between patients so much as unstable in this background
Two cell lines derived from the same child were reported to need menin very differently, and the question was whether that difference is biology or noise. Measuring 358 genes in the same pair: a typical gene differs by 0.09 between the two lines, and menin differs by 0.71 - the second-largest difference of any gene tested. Menin also differs a lot between children. So menin's requirement is not stable in these cells at all, even within one patient. That is a different and more awkward result than either 'real difference between patients' or 'just noise'.
The WRAD degron absence survives, and the reason is worth more than the verdict: a 3.7 nM chemical degron of WDR5 with 2-hour kinetics already exists IN MV4;11, and it still cannot see the fusion
This project records that nobody has used rapid protein-removal methods to ask whether a particular protein machine touches the leukemia fusion. Searching by drug name rather than by disease name turned up a compound that removes one of those proteins from the exact leukemia cell line in question within two hours - and it still does not answer the question, because the readout it uses cannot tell the fusion protein apart from the normal one.
The ZNF384 family is not the inverse of KMT2A::AFF1 - it is the SAME grammar written in the opposite gene order, and three of the four claims in the inverse framing are false
The idea being tested was that ZNF384 leukaemia is KMT2A leukaemia turned back to front: a DNA-binding protein with a chromatin regulator stuck on, rather than a chromatin protein with an elongation factor stuck on. Three of the four things that idea assumed turn out not to be true. What is true is less tidy and more interesting: both families keep the half of the protein that decides WHERE to go, and throw away everything else.
Thirteen of q152's seventeen AFF1-only core promoters are absent from the deposit paper's whole target-gene table, and the four that are present are AFF1-exclusive there too
An earlier round of this project picked out seventeen genes that seemed to be switched on by the fusion only in the AFF1 leukaemias. When we look those seventeen up in the list the original scientists published, thirteen are not in it at all - their method never counted them as targets anywhere, in any patient. The four that are there are indeed AFF1-only.
This corpus argued from the literature that intersecting fusion peaks with partner peaks cannot attribute occupancy, and eighty cycles later a beat measured that partner antibody sitting at the negative control in primary patient samples - two independent derivations from the same paper, with no edge between them and one wrong premise
Years of this project's notes argued, from a 2010 paper, that you cannot prove where a fusion protein sits by checking where its partner sits — because the partner is normally at those same places anyway. A recent analysis measured exactly that in real patients' cells and found the partner antibody performing no better than a control that was meant to mean nothing. The prediction and the measurement were made independently and neither knew about the other.
This cycle's word 'silenced' means two repressive marks up and one active mark down at 2 kb resolution - GSE260742 deposits no RNA-seq, no nascent transcription and no reporter for either NTvKO genotype, so whether the vacated sites stop being transcribed is unmeasured in the experiment that established they are marked for silencing
We can see the chemical 'off' labels being added to the DNA the drug clears, but nobody measured whether those stretches actually stopped being read. The word 'silenced' in this work means labelled for silencing, not observed silent.
This graph recorded 'nobody has looked at AKAP95 in any KMT2A fusion' one cycle before reading the 2026 paper that did - KMT2A returns 0, MLL returns 2, and every cell-line token returns 0 because cell lines live in figures
We wrote down that no one had studied AKAP95 in MLL-rearranged leukaemia. A paper published in 2026 does exactly that. Searching for the modern gene name found nothing; searching for the old name found it immediately.
This graph states 624 falsifiers and records the outcome of 17 - a figure this entry first published as 8, because the detector did not know the corpus's own ALL-CAPS convention for recording one
This project is good at writing down how each of its claims could be proved wrong, and almost never comes back to record what happened when anyone checked. Of 624 such stated tests, 17 have a recorded outcome - a count this analysis first got wrong, and undercounted by half.
This repo's "the pan-fusion core is 7.4-fold cleaner per unit signal than the singleton tail" divides one control column by a denominator summed over seven baits — on a per-bait denominator the ratio is 1.06 and p goes from 0.074 to 0.91
An earlier analysis here concluded that the proteins shared by all seven fusions are far cleaner than the fusion-specific ones, by dividing how much each protein appears in the negative control by how much it appears overall. But "overall" added up all seven pulldowns for a shared protein and only one for a specific protein, so the shared proteins were handed a denominator seven times too big. Put on the same footing, the two groups are indistinguishable.
Three archives, one answer: SUV39H1 has no ENCODE target record under any label (three routes, all empty), ChIP-Atlas's entire hg38 SUV39H1 holding IS the four GSE139196 MCF-7 samples re-indexed, and KMT2A is not among the 26 genes ChIP-Atlas calls SUV39H1 targets
We checked every public archive that could hold a map of where this enzyme sits on human chromosomes. The big reference atlas holds nothing at all, the second archive holds only a re-labelled copy of one breast-cancer experiment, and in that experiment the KMT2A gene is not on the list of genes the enzyme is called a regulator of.
Three CXXC-pathway activities converge on rDNA - KDM2B represses Pol I, KMT2A and SETD1A bind the whole locus, MLL-AF4 is predominantly nucleolar - and no paper contains two of them
The nucleolus is the factory that builds ribosomes, and its DNA is full of unmethylated CG letters. Three separate groups have shown that KDM2B shuts that factory down, that KMT2A sits all over its DNA, and that the leukaemia fusion piles into the nucleolus and disrupts it. No single study has ever looked at more than one of those at a time.
Three hours after the fusion is degraded to 2.5-7.1% of its own sites, NOT ONE of the eight complex components profiled beside it falls below 68% retention genome-wide - menin 80.0%, DOT1L 68.7%, CDK9 83.6%, AFF4 92.7%, RNA Pol II 99.6%, pS2 98.3%, pS5 86.8% - so nothing in the complex is held on chromatin by the fusion at genome scale
When the leukaemia-driving fusion protein is destroyed you might expect its partners to fall off DNA with it, but across eight partners measured in the same experiment none does - they stay at two-thirds to all of their strongest spots while the fusion itself drops below a tenth. Whatever holds these proteins on DNA across the genome, it is not the fusion.
Three measurements the export axis needs and nobody has published: XPO1 occupancy in KMT2A-r, H3K79 under an XPO1 inhibitor, and CRM1 occupancy under a menin inhibitor
Three specific experiments would settle most of the open arguments about XPO1 in leukaemia. We searched for each one properly, checked that the search could find things when they exist, and found none of them published.
Three of seventeen KMT2A-GERMLINE infant ALL cases are IRX1-high, one at 10.37 - so a high IRX1 in an infant leukaemia is not by itself evidence of anything fusion-related
Seventeen babies in the same study had leukaemia without any KMT2A fusion. Three of them still had IRX1 switched on, one of them strongly. So IRX1 being on does not mean the fusion put it there.
Three of the four AFF scaffolds have a named human Mendelian syndrome and the index fusion's partner has none — a 0/4 that is a fact about AFF1 and not about how hard anyone has looked
Of the four related scaffold proteins, three cause a named inherited syndrome when they go wrong — FRAXE, KINSSHIP and CHOPS. The fourth is AFF1, the one fused to KMT2A in infant leukaemia, and it causes no named inherited disease at all.
Three quarters of KMT2A::AFF1's top partners are shared with other MLL fusions — and the two analyses that said otherwise were both artifacts of how the source table was built
We asked whether the fusion we study grabs a distinctive set of proteins compared with six related fusions. The first two ways of measuring it gave opposite answers, and both were wrong for reasons hidden in how the published table was assembled. Measured properly, it is unremarkable: most of what it binds, its relatives bind too.
Three recurrent breakpoints give three junction peptides - and the one nobody uses needs no missed cleavage at all
To find the fusion protein by mass spectrometry you need to know the mass of the join. There are three common joins, so there are three masses - and a test built on one of them would come back empty in patients carrying either of the others.
Three unnamed things at once: the DOT1L-PRC1.1 deposition names the fusion, the cell types and both inhibitors with concentrations
An entry on DOT1L and Polycomb says the paper names neither its cell systems, nor its MLL fusion, nor the drugs it used. Three deposited datasets name all three: mouse MLL-AF9 cells plus human MOLM13, MV4;11, K562 and erythroid and stem-cell lines, treated with the DOT1L inhibitor SGC0946 and the menin inhibitor VTP-50469 at stated doses and times.
Twenty-five of thirty-six analyses that measure this graph were computed against a graph that no longer exists, and seven stamp a date that never moves
This project keeps small programs that measure itself, and their saved answers are quoted later as if still true. Most of them were computed when the graph was a third of its current size, and one such answer had quietly reversed while it was still being cited as a live problem.
Two entries record zero chemical matter against the CXXC-DNA interface. There is one: dimethyl fumarate, IC50 about 1 uM, and it is already an approved drug
This project's drug dossiers say nobody has ever made a molecule that blocks the fusion's DNA grip. One has been published: dimethyl fumarate, a drug already given to people with multiple sclerosis, blocks the grip at about one micromolar in a single-molecule experiment. It is a sticky, unselective molecule rather than a designed drug, but it is not nothing.
Two hypotheses in this graph make the same prediction on every lesion that exists, and the experiment queued to test them picks the one arm where neither speaks
This project wrote two separate explanations for why some cancers respond to menin drugs and others do not, one from leukaemia and one from sarcoma, and never checked whether they actually disagree. They do not disagree about any real tumour, and the experiment already queued to settle the question was aimed at the single tumour where one of them makes no prediction at all.
Two instruments on the same cell line give opposite answers about DOT1L under a menin inhibitor: the depositors' own TSS-window table puts DOT1L at 0.6638 beside menin's 0.2513, and 2 kb genome-wide bins put it at 0.2219 beside menin's 0.1657 - and only the genome-wide instrument can see the 86% of the co-occupied set that is not at a TSS
Using the summary table the original authors deposited, the DOT1L protein looks like it stays put when the drug is added. Going back to the raw coverage files and looking at the whole genome instead of just gene starts, it leaves alongside menin. The table can only see a small slice of the places that matter.
Two KMT2A-rearranged patients share oncoprotein-target promoters at a median Jaccard of 0.132 against an independence expectation of 0.069 - an enrichment of only 1.63x, with 5 of 351 pairs at or below chance
Two children's leukaemias share about an eighth of their fusion-bound genes. That sounds low, but two random lists of the same length would already share a fifteenth, so the real overlap is only about 1.6 times better than chance.
Two measurement defects confirmed: a denominator summed over seven baits made the pan-fusion core look 7.5x cleaner than it is (corrected fold 1.06, p 0.91), and my own gap report counts as coverage in the join that measures the gap
Two ways of measuring turned out to be broken. The first divided a one-column number by a seven-column number, which made one group of proteins look seven times cleaner than another purely by arithmetic; corrected, the two groups are identical. The second is subtler and is mine: I measured how many proteins the project had never discussed, then wrote a report discussing them - and the next run of the same measurement counted that report as coverage. Writing about a gap closes it, by that measure.
Two new contaminated tokens for the registry - p75 (12 of 27 hits are TNF receptor) and AF-10 (an IL-6-responsive B cell line) - and SEM firing a second time, in a different query, one cycle after being written down
Short protein nicknames are dangerous search terms because other fields use the same letters. Searching for a gene nicknamed p75 returned mostly papers about an unrelated nerve-growth receptor, and searching for AF10 returned heart-rhythm papers and a cell line that happens to share the name. One trap that was written down a single cycle ago caught this beat again in a completely different search.
Two of seven interactome read-queue partners should not get a gene entry, and the reasons are a pseudokinase with no nuclear pool and a membrane-skeleton protein whose two leukaemia hits are a gene-name artefact
Seven proteins were nominated for their own entries because databases list them as touching the leukaemia machinery. Two of them should not get one, and this records exactly why, so nobody has to repeat the work.
Two of ten hand-typed genomic intervals in this beat were on the wrong chromosome or a megabase off, and neither produced an error - one would have published the exact opposite of the truth about PAX5
Two of the ten gene locations I typed from memory were wrong. Neither caused an error message; they just returned believable numbers, one of which said the opposite of what the data actually shows.
Two of the nineteen non-AFF1 samples are menin-inhibitor treated, so the specificity arm is effectively seventeen - and removing them does not rescue the core
The claim 'switched on in KMT2A::AFF1 and in none of the other nineteen patients' counts two patients who were on a menin inhibitor, a drug that strips the oncoprotein off DNA. Finding nothing in a treated patient is almost free, so the comparison is really against seventeen patients. Redoing it that way does not change the answer.
Two of the seventeen core promoter windows are shared with a neighbouring TSS - PBLD overlaps HNRNPH3 by 38.2% and CCP110 overlaps GDE1 by 7.8%
Each gene's switch is measured as a two-kilobase stretch centred on where transcription starts. If a second gene starts nearby, the same stretch covers both and you cannot tell which gene the signal belongs to. Two of the seventeen have that problem.
Two of the three top new dependencies are myeloid identity, not KMT2A
The cell lines used to find genes this leukaemia needs are mostly one blood-cell type. So genes that any cell of that type needs look like they are specific to our disease. Splitting the lines by cell type separates the two, and two of the three new top hits turn out to belong to the cell type.
Two of three KMT2A::EPS15 infant ALL cases carry IRX1 at KMT2A::AFF1 magnitude, while the only KMT2A::EPS15 patient with occupancy data has no oncoprotein at IRX1 at all
Two of the three babies in a public dataset whose KMT2A fusion joined the EPS15 gene had IRX1 switched on as strongly as any t(4;11) baby. The one EPS15 patient whose chromatin was profiled - a different child - had no fusion protein sitting on IRX1 at all.
Two processing batches inside one GEO series: the FLAG and KDM2B tracks put 1.7-1.9% of their coverage in ENCODE blacklist regions and the BCOR, RING1B and input tracks put 14-46%, with a single 2 kb bin at chr2:98,666,000 carrying 18.1% of the BCOR SGC0946 track on its own
Half the public files in this dataset were cleaned of artefacts before deposit and half were not, and nothing in the record says so. In the uncleaned ones, a single two-kilobase stretch of repetitive DNA holds up to a fifth of the entire experiment's signal, which quietly distorts every comparison that divides by the total.
Two quality checkers audited: one was skipping a quarter of its input, the other did not exist and failed its own defining case twice before it worked
We checked our own checking tools. One was quietly ignoring a quarter of what it was supposed to look at, and the other had to be rebuilt twice before it could catch the mistake it was written to catch.
Two thirds of the curated interactome of this project's six core proteins is spoke-expanded - and the pairs this graph chose to write are significantly better than the ones it did not
When a scientist pulls one protein out of a cell, dozens of others come with it, and the database turns that one experiment into dozens of separate 'A binds B' rows that nobody ever tested one-on-one. Across the six proteins at the centre of this project, 64% of database rows and 71% of partner pairs are of that inferred kind - but the pairs this project actually wrote up are meaningfully better evidenced than the ones it skipped.
Two-thirds of the load-bearing evidence names no lineage, which is why this graph cannot audit its own commonest failure
The mistake this project makes most often is treating a result from a mouse, or from a different leukaemia, as though it were about the infant disease this graph is about. Two-thirds of evidence items name no lineage and nearly two-fifths name no experimental system, so the graph cannot check itself for that mistake.
Two-thirds of this graph's core-complex interaction evidence comes from experiments where our protein was not the bait, and LEDGF has one bait-supported partner out of thirty-six
The database this project draws fusion partners from records whose experiment each interaction came from, and nobody had looked. Two-thirds of the partners turned up incidentally in somebody else's net rather than being deliberately fished for - and for LEDGF, central to this project's main hypothesis, it is thirty-five out of thirty-six.
Two-window deconvolution: at a typical shared promoter only 1.8% of the signal can be proved to belong to the gene it is filed under, and at IRX2's one near-positive it is 19.7%
When two overlapping windows both report a number, the difference between them can only come from the small stretches they do not share. That difference sets a floor on how much of a gene's signal is genuinely its own. Across this dataset the floor is usually under 2%, meaning almost all of a shared-window measurement is unattributable arithmetic.
U4 was assumed, so I tested it: full text raises the low claim-to-paper overlap scores with a +0.108 median margin over same-field decoys (42% of the raw rise is document length), and every one of the seven pairs that stays low is a blind spot in my own instrument rather than a misattached citation
Yesterday I explained away the worst citation scores by saying the evidence must be in the body of the paper rather than its summary, and wrote down that I had assumed this rather than checked it. Checking it showed the scores do improve for the right paper, though about two-fifths of the improvement is just that whole papers are longer, and the handful that stayed low turned out to be faults in my measuring tool rather than bad citations.
Under EPZ5676 the canonical KMT2A-fusion targets look like they lose H3K79me2 far more than the genome (p=0.017 and p=0.0004), and against a baseline-matched null the entire effect vanishes (p=0.34 and p=0.22) - eight of the ten are simply top-decile genes
The genes a KMT2A fusion is famous for switching on appear to lose their chromatin mark much faster than average under a DOT1L inhibitor. They do not. They start with far more of the mark than average, and once you compare them with other genes that also start high, they behave like any of them.
Under revumenib the oncoprotein leaves chromatin globally, so 'does IRX1 become a menin-inhibitor escape route' cannot be answered from this deposit
In the one child in this dataset sampled while on a menin-blocking drug, the fusion protein has come off the HOXA genes - and off everything else too. So you cannot use that sample to ask whether it moved to IRX1 instead.
Under VTP-50469 in RS4;11, the 209 bins that rise to the top of the MENIN distribution are 3.7-fold CpG-island-enriched over a class-matched null at p = 5.2e-07, while the MLL1-N gained set is not enriched at all - so menin's residue is island-biased and an MLL1-N gain is not
When the drug strips menin off its usual places, the places where menin is still most concentrated afterwards are unusually rich in CpG islands, unlike the equivalent set for the MLL1 antibody. The simplest explanation is not that menin moves, but that the menin that survives the drug was never the kind that needed MLL1.
UniProt's menin 214-390 FANCD2 region is not a surface: it spans three spatial lobes and 29.9% of menin's solvent-accessible area, so its twenty-residue overlap with the gamma-H2A.X motif is what chance gives you
A database says menin's middle third binds the repair protein FANCD2, and a 2026 paper says a thirty-residue stretch inside that middle third binds a different repair protein, so it looked as if menin might have one DNA-repair docking site with several tenants. Measuring the actual structure says no: the FANCD2 region is almost a third of menin's whole outer surface spread across three separate lumps of the protein, and a thirty-residue motif dropped anywhere at random would land inside it this deeply about three times in ten.
UROS: searched and empty — the scan's rank-15 hit has no leukemia literature, and no entry was written
A scan run by this project said that leukemia cells with the KMT2A fusion depend on UROS, an enzyme that makes the red pigment in blood. A search of the medical literature found nothing at all connecting this enzyme to leukemia, so no entry was written about it — and that empty result is recorded here on purpose.
USP1 does not behave like part of the Fanconi module, and does not track ID1 either
We thought USP1 might belong to the repair module our scan implicated, or else to a rival pathway. Across 1,208 cell lines it behaves like neither, which means we should stop counting it as evidence for the first one.
UTP18 is not a KMT2A partner - it is WDR5, mis-curated into IntAct from the founding paper, and it manufactured the only KMT2A-bait gap in the interactome recount
A public interaction database lists a nucleolus protein called UTP18 as part of the MLL machine, but the 2004 paper it says this came from never mentions UTP18 at all. Line by line, the database's list of that paper's proteins is the paper's own list with WDR5 - a protein this project already has an entry for - replaced by UTP18.
Validating the wrong-paper year check against this project's own recorded identifier errors - 16 of 23 would have been flagged, 7 would have passed silently on a shared year, and the figure is biased optimistic because every pair in the set is an error somebody caught
A check built last cycle compares the year each note records against the year of the paper its link actually opens, to catch a working link pointing at the wrong paper. This asks whether that check would have caught the mistakes this project has already made and written down. It would have caught about seven in ten — and the three it would have missed are the dangerous kind, where the wrong paper happens to share a year with the right one.
Verdict: 'eleven cell lines, none infant-derived' is right about the screen and wrong about the material — at least ten infant KMT2A::AFF1 lines exist
This project has been saying that no cell line of this leukaemia comes from a baby. Checking every line by its database accession shows at least ten do, including one from a nine-day-old and one that is already sitting in the collection the project's dependency data comes from.
Verdict: q147's IRX1 headline survives as a measurement and not as an inference - the pre-registered branch is NEITHER, on U5 and on a FALLS condition that fired
The question was whether the fusion protein sitting on IRX1 in three AFF1 patients and none of eighteen others is more than chance. Measured against a proper null at the right scale, it is: only six genes in the whole genome do that, where chance gives about one. But those three samples are two children from one hospital, and once you count them as two patients the result stops being significant - so the finding is real as a measurement and cannot yet carry the conclusion that IRX1 needs the AFF1 partner.
Verified against the live records rather than inherited: HUDEP-2 and mESC each carry a DMSO and an SGC0946 H3K79me2 sample and no VTP-50469 sample, so K562 remains the only fusion-free line in GSE260742 able to score a menin-inhibitor effect - and cycle 82's request for a second and third fusion-free line cannot be met from this series
The strongest challenge to this project's account of how the menin drug works rests on one cell line without the fusion. Two other fusion-free cell lines are in the same dataset, but each was only treated with the other drug, so neither can be used to check the result. That check needs data nobody has deposited here.
We have read a tenth of the recent field, and the gap found a real interactor
Every quality check this project had built looked inside the graph. This one looks outward and asks how much of what the field has published since 2020 we have actually read. The first run surfaced a laboratory we had never cited whose work is squarely on our own fusion.
What 160 infants a year actually permits: the sample-size arithmetic
A proper two-arm trial big enough to prove a new drug improves survival in infant leukemia would need about 300 babies, which even with every country in the world contributing takes six years to recruit and nine or ten to report. A single-arm study answering a narrower question needs about forty babies and one year.
What fills the space: RING1B and BCOR arrive at the fusion's sites under BOTH drugs, but the DOT1L inhibitor puts them everywhere on the fusion's territory while the menin inhibitor puts them only in the quartile where the fusion is most completely evicted - and at Meis1 RING1B rises 16-fold
When the fusion protein is knocked off the DNA, a repressive complex moves in behind it. Which drug is used changes the pattern: the DOT1L drug brings the complex in across all the fusion's sites equally, while the menin drug brings it in only at the sites the fusion actually vacates.
What is left after the missing arm was read: there is no H3K9me1, me2 or me3 measurement at the KMT2A breakpoint region in ANY haematopoietic cell, and the founding paper's own archive contains none because all 60 of its samples are epithelial
The experiment nobody had looked at has now been read, and it answered its question. But every sample in that archive came from eye cells. Nobody has ever measured these chemical tags on the KMT2A gene in a blood cell, which is the only place the disease happens.
What is left of the resection absence after the compute was done: the resting state is now MEASURED and negative in 13 ENCODE datasets, so the whole question collapses onto the DAMAGE state, which is never-measured for all six proteins in every cell type on Earth
Last cycle this graph said nobody had looked for repair proteins on the KMT2A gene, and that for two of them public data could answer it. We ran that analysis and the answer is no. What remains cannot be answered from any existing dataset, because nobody has ever measured these proteins at this gene in a cell that has actually been damaged.
What maps a DNA break, and the fact that none of the modern instruments has ever been pointed at KMT2A in a blood progenitor
Three good methods now exist for finding every DNA break in a genome at single-base resolution. None of them has been published on the KMT2A gene in a blood stem cell, which is why the field still argues about where the break is and what makes it.
What the CTCF-protects paper actually cites: it names Gothe and Atkin in ONE direction-free sentence and declares the direction unresolved, and the two papers that ran the reciprocal experiment are absent from all 81 of its references
This graph said the CTCF-protects paper cites the CTCF-breaks papers and ignores the disagreement. Read directly, the paper cites them in a single sentence that deliberately leaves the direction open and then says the direction has never been settled - and the two papers that did settle it, in the opposite direction, it never cites at all.
What the Polycomb beat searched, what it counted as absent, what it read and rejected, and the two absences it got wrong before hand-back
A record of every search this beat ran, every gap it found, and - importantly - the papers it read and decided not to use. Someone re-running this work next year should be able to repeat the searches and the judgements, not just the searches.
What this beat looked at and did not write: rejected candidates, with counts for both halves of every search
A record of every enzyme and every histone mark this beat considered and chose not to write an entry for, with the reason and the search counts that justify it. Knowing what was checked and found empty is as useful as knowing what was found.
What this beat read and did not use: bounded absences, bounded rejections and one bounded unreadable
A record of the searches we ran, the papers we looked at and decided not to use, and the one paper we wanted and could not read. Without this, a future reader cannot tell what was considered from what nobody ever looked for.
What this beat searched, read and set aside - including one cell line the accession database has not caught up with, and one query that returns a clean zero
A record of the searches run, the papers read and deliberately not used, and the things that could not be read at all. Its purpose is so that someone re-running this next year repeats the decision, not just the search.
What this beat searched, what it rejected, and the two papers it could not read - after a fifth route rescued the third
A record of exactly which searches were run to survey how this leukaemia is measured, which papers were found and deliberately not used, and which papers could be found but not read. It exists so that someone repeating this survey next year repeats the decisions as well as the searches.
What was searched for a structure and what came back empty: the KMT2A interface layer enumerated by accession pair, with the graph's own unmapped interfaces listed
This is the bookkeeping behind the structural interface entries: every protein pair that was searched in the world's structure database, what came back, and what came back with nothing. The short version is that the parts of KMT2A the leukaemia fusion throws away are extremely well solved, and the parts it keeps are not.
What was searched, what came back empty, and what was read and not used - the P-TEFb reservoir beat's audit trail
This is the working record behind the other entries in this beat: every literature search that was run, every one that returned nothing, and every paper that looked relevant and was then set aside with a reason. It exists so that someone repeating this work next year repeats the decisions, not only the searches.
When a 2 kb promoter call fires at a window shared with another gene, the neighbour fires too 91.3% of the time - against 4.3% for random promoter pairs
If two genes' measurement windows overlap, and the test says one of them is switched on, it says the other one is switched on as well nine times out of ten. For two unrelated genes picked at random the same test almost never agrees. So the failure is caused by the overlap, not by the test being vague.
When one infant's KMT2A::AFF1 B-ALL became an AML, EBF1 lost fusion occupancy and PAX5 kept it - the B-lineage transcription-factor programme is not vacated as a block
In the same child, before and after the leukaemia switched from lymphoid to myeloid, the fusion came off one master B-cell gene but stayed on another. The B-cell programme is not simply switched off wholesale.
Where else in a child is the target switched on? (run)
A drug reaches the whole body, not just the cancer. This asks, for every protein on our target list, how much of it normal organs make compared with bone marrow, which is the tissue we are actually trying to treat. Two candidates turn out to be made far more in the brain and the pancreas than in marrow.
Where else in the body each target is expressed, and the two controls that say the method works
Every drug target here was picked for what it does in leukaemia cells, but none of them exists only there. This checks all 54 normal human tissues for each one, so a child cured at one is not injured at another.
Which immune-evasion routes are actually documented in KMT2A-rearranged disease, and which are borrowed from AML and assumed - a ranked ledger with the bounded absences that define it
People talk about leukaemia hiding from the immune system as if the same tricks were known for every kind. For this one, only a handful of tricks have actually been measured, and the most famous ones - losing the identity card, switching on the PD-L1 brake - have never been looked for at all.
Which KMT2A exon carries which interface: exon 1 is the menin-binding motif, exons 5-6 are the CXXC zinc finger (computed, not recalled)
Every paper on KMT2A-rearranged sarcoma describes its fusions by exon number, while this project describes KMT2A by protein domain, so the two cannot be compared until the exons are translated into residues. This entry does that translation once, from live Ensembl and UniProt records, so other entries cite it instead of guessing.
Which leukemia co-dependencies survive uncertainty? (run, with p-values)
Our leukemia results came from only 54 cell lines, few enough that chance can invent a pattern. We resampled and reshuffled the data five thousand times each to see which correlations survive: ten did, and one we had previously highlighted did not.
Which of our phospho-sites are actually seen in patients
Most of what this project knows about phosphorylation of these proteins comes from cell lines, but a study of 74 patients measured it directly and left the table in a supplementary file. Some of our sites are there, one we had ruled out is still absent, and two are numbered on the wrong version of the protein.
Wild-type KMT2A is one of the most promiscuously labelled proteins in the proximity record - 25 of 50 baits - and menin is one of the least, at 3
Counting how many tagging experiments find each protein nearby, KMT2A turns up in half of them and menin in almost none. That gap is either a real difference in how the two proteins sit in the nucleus, or just a difference in how much of each there is.
With H3K79me2 added, the eviction-depth panel over one set of 1,228 bins under one drug is H3K27me3 +1.000, H2AK119ub +1.000, H3K4me3 -1.000 and H3K79me2 -1.000 - a clean repressive-up / active-down split that no shared library, blacklist or regression artefact can produce
Four different chemical tags were measured on the same stretches of DNA in the same cells under the same drug. The two 'off' tags go up where the fusion is most removed and the two 'on' tags go down there. A technical glitch in the measurement could not push two tags one way and two the other.
WITHDRAWN - "thirteen genuinely absent" was wrong: 14 of the 15 are adjudicated in corpus prose and the fifteenth is not an interactor. Superseded by ana-zero-of-fifteen-again-the-same-worklist-the-same-thirteen-and-the-third-recurrence
An old audit listed fifteen genes that were promised as entries and never written. Thirteen really are missing. The other two were deliberately handled a different way, and the entry that handled them explains the choice in its own title - the audit just could not see it.
WITHDRAWN: I identified the promoter-matrix column mean as library coverage and called A73051 depth-confounded; the column mean is the enrichment fraction, A73051 is at the 37th-39th percentile of true depth, and the share-versus-depth correlation is negative
One unusual patient sample has the fusion protein sitting almost nowhere, and the obvious question is whether the protein is scarce everywhere or just missing from one gene. This deposit cannot tell the two apart, because the only way to measure it also measures how deeply the sample was sequenced, and nothing in the experiment separates those.
Within one arm, swapping the processing-marker confound for a submission-batch confound takes the same-condition H2AK119ub floor from 37.04% to 16.84% at D1 (tighter, 0.45x) and from 37.37% to 61.89% at the TOP 0.1% stratum (worse, 1.66x). The brief's question - tighter, wider, or unchanged - has no single answer, and it is worse exactly where this thread scores its claims.
Two repeats of an experiment were compared in two ways, each removing a different bookkeeping difference between them. Removing one made the repeats agree better across most of the genome and noticeably worse at the small number of hottest spots - which are the spots every conclusion in this project depends on.
Working the queue deferred for four cycles: the 40 are three different things, and one absence claim the graph had already refuted ten entries away
We keep a list of places where the graph says nobody has done an experiment, and we had not re-checked them for months. Checking found that most were fine, that many were not really claims at all, and that one had been answered by a paper the graph already knew about elsewhere.
XPO1 mutation in acute leukaemia, counted with denominators and a positive control: 1 of 739 B-ALL, 0 of 584 paediatric AML, 0 of 1,103 adult AML
We asked a public cancer-genomics database how often the XPO1 hot spot is mutated in childhood and adult acute leukaemia, and checked first that the same query finds it where it is known to be common. It is essentially absent from acute leukaemia and abundant in CLL, exactly as the literature implies.
Zero of eight unwritten AFF4/MEN1 partners has an experimentally coded nuclear localisation, and the two that are called nuclear are called it by inference
The proteins at the centre of this disease live in the nucleus, and every one of them has an experiment saying so. Of the eight database-listed partners this beat examined, not one has an experiment placing it in the nucleus - two are called nuclear, but only by guesswork from related proteins.
Zero of fifteen, again: the same worklist, the same wrong intermediate answer of thirteen, and the third recurrence of a gap list that cannot see a decision
An old audit file listed fifteen genes as promised-but-never-written. None of them is actually a gap: the project had already written, or explicitly decided against, every one. This is the third time that same list has been mistaken for real work.
Patterns from finance, law, engineering and ecology, used to generate testable biology rather than decoration.
A genuine small-molecule disruptor of an oncogenic oligomer DOES exist - 7.44 against the RUNX1::RUNX1T1 NHR2 tetramer - which corrects this graph's record; and its eighteen-year potency curve says the class is a chemistry problem, which is what this graph concluded for the wrong reason
This project has been asking whether anyone has ever made a small molecule that pulls apart the glue holding a fusion protein's copies together. The answer is yes - a compound called 7.44 does it to the RUNX1-ETO fusion, and newer versions of it are about ten times better. But even the best of them only works at concentrations tens of thousands of times higher than a real drug, which tells you the idea is sound and the chemistry is not there yet.
A pocketless fusion already had its protein-protein-interface drug taken to phase I/II - and TK216 turned out to be a microtubule destabiliser, not an EWS::FLI1 antagonist. This is the closest thing the field has to a controlled failure of the interface hypothesis.
Ewing sarcoma is driven by a fusion protein with no drug-shaped pocket, exactly like the leukaemia fusion this project studies, and the field's answer was the same: find a small molecule that breaks the fusion's grip on a partner protein. That molecule was taken all the way into patients over fifteen years - and then a genetic screen showed it actually kills cells by poisoning microtubules, and the trial showed almost no activity.
Agency law — actual authority, apparent authority, and what revocation cannot undo
Menin acts on behalf of the fusion protein, the way an agent acts on behalf of a principal. Law has spent centuries working out exactly what happens when you revoke an agent's authority, and almost none of it is instantaneous — which is a strange and useful thing to notice about a drug we dose every twelve hours.
An H3K4me3-reading PHD finger fused to NUP98 makes a potent leukaemia oncoprotein; the H3K4me3-reading PHD finger of KMT2A is deleted from every KMT2A fusion and reinstating it abolishes transformation. Same module class, opposite sign - and the one NUP98 fusion that carries KMT2A's PHD cassette is the weakest member of its family.
A small protein module called a PHD finger reads a chemical tag on chromatin that marks genes as switched on. Bolt that module onto the NUP98 partner and you get a fast, aggressive leukaemia, and breaking its ability to read the tag stops the leukaemia entirely. But the very similar module belonging to KMT2A is thrown away by every KMT2A fusion, and putting it back destroys the fusion's ability to transform cells - so the same kind of reader is required in one family and forbidden in the other.
Another fusion family did the experiment this beat wants: in FUS::ERG and CBFA2T3::GLIS2 AML, MHC class I and II are absent at diagnosis, EZH2 drives it, and an EZH2 inhibitor restores presentation
A different, rarer fusion leukaemia switches off both kinds of immune identity card before treatment even starts, and the reason is a polycomb enzyme. A licensed drug against that enzyme turns the cards back on and lets immune cells kill the leukaemia - which is the experiment nobody has done for KMT2A.
BCAT1 in core-binding-factor AML - the closest analogue in another fusion family, where TWO different fusions converge on one metabolic enzyme by rewiring the chromatin at its locus
A different family of leukaemia fusions - the ones that break the core-binding-factor genes - switches on an enzyme that handles branched-chain amino acids, and the leukaemia then depends on it. Two unrelated fusions do this the same way, by changing how the DNA is folded at that gene, and an old epilepsy drug that blocks the enzyme helped mice live longer.
Buying an already-impaired asset — what diligence would you do, and what is the biological equivalent?
A buyer never assumes an asset is in the condition the seller describes; they inspect it on the day they buy. We give menin inhibitors to children whose leukemia has already been changed by earlier treatment, and we almost never inspect the leukemia first.
CBFA2T3::GLIS2 supplies its own CAR-T antigen - FOLR1 came out of a list of fusion-specific genes in an isogenic human model. KMT2A-rearranged leukaemia is treated with a LINEAGE antigen it can abandon, or with one a drug has to create; nobody has built the fusion-specific surface list for KMT2A::AFF1.
In one infant leukaemia, researchers built a human model of the fusion, compared it with normal cells, and pulled out a list of surface proteins that only the fusion's cells make - then built a CAR-T cell against one of them. The leukaemia this project studies is treated with markers that belong to the cell type rather than to the fusion, which is why the disease can escape by changing cell type, and nobody has made the equivalent fusion-specific list for it.
Children with NPM1::ALK lymphoma immunise themselves against their own fusion protein, and the antibody titre predicts relapse - while no autoantibody against any KMT2A fusion has ever been reported. The difference is not the junction; it is that ALK is an ectopic self-antigen and KMT2A and AFF1 are not.
In one childhood cancer, patients' own immune systems spontaneously attack the fusion protein driving the disease, and the strength of that response predicts whether the cancer comes back. Nobody has ever found the equivalent response in children with KMT2A-rearranged leukaemia - and the likely reason is that the two halves of the KMT2A fusion are proteins the body already knows, so there is almost nothing foreign for the immune system to see.
Competitive release and adaptive therapy — what ecology says happens when you remove the dominant species
In an ecosystem, removing the dominant species does not leave an empty space; the species it was holding down expand into it. Wiping out the B-cell leukemia with a CD19 drug and watching a myeloid leukemia appear has exactly that shape.
Differentiation syndrome is the one human pharmacodynamic readout KMT2A-rearranged leukaemia already shares with APL - and in APL it is a complication of a cure, while here it is a complication of a 30% response rate
In a different leukaemia (APL), a drug that switches the fusion protein off makes the cancer cells grow up into normal blood cells instead of dying, and so many do it at once that patients get a dangerous inflammatory reaction called differentiation syndrome. Patients on menin inhibitors get the same reaction, which is the strongest human evidence that disrupting the menin interface really does release the differentiation block - but in APL that reaction comes alongside a cure and here it does not.
Duration and reinvestment risk — a 4.6-month response is a short-duration instrument, and short duration changes everything about how you value it
A remission that lasts under five months is worth very little on its own. In finance, a short-lived investment is valued almost entirely by what you can move the money into next, not by how good the investment itself was.
EWS::FLI1 builds its own enhancers at GGAA microsatellites, and a single inherited base change that lengthens one repeat raises the fusion's binding and the patient's risk. KMT2A fusions occupy pre-existing, invariant CpG islands - which is why this disease has no germline modifier of its target set, and why its mouse models work at all.
In Ewing sarcoma the fusion protein makes new gene switches out of repetitive DNA that varies between people, and one inherited single-letter change that makes a repeat longer both raises how tightly the fusion binds and raises the person's risk of the disease. The leukaemia fusion this project studies lands on ordinary, fixed features of the genome instead, so there is no equivalent inherited dial - which is a loss for finding risk genes and a gain for making the biology reproducible.
Fault trees, minimal cut sets and common-cause failure — reliability engineering on a buffered oncogenic program
Engineers who design systems that must not fail have exact tools for finding the smallest combination of parts whose simultaneous failure brings the system down, and for the uncomfortable fact that backup systems tend to fail together for shared reasons. The leukemia program is a fault-tolerant system, and neither tool has been pointed at it.
Fiduciary conflict of interest — law does not remove a conflicted agent, it separates the mandates
Menin works for two bosses with opposite interests, using the same office for both. When a professional is caught in that position, the law almost never bars them from working; it separates the two jobs. Current drugs do the opposite and shut the whole office down.
In inv(16), disrupting the fusion interface swaps BRG1 for RING1B at three MYC enhancers - the cleanest version anywhere of the experiment nobody has run on menin
In a different leukaemia, a drug that breaks apart a fusion protein frees a normal transcription factor to sit on three switches that control MYC, and when it sits there it kicks off the machine that turns genes on and replaces it with a machine that turns them off - and the cells die. That is exactly the kind of measurement nobody has made for menin inhibitors in the leukaemia this project studies.
In pancreatic neuroendocrine tumours MEN1 and DAXX/ATRX are the two commonest mutations of the same tumour - and both are LOST there while leukaemia needs both intact
There is a rare pancreatic cancer in which the two most frequently broken genes are menin - the protein this whole project is about - and the pair of chaperones this beat is about. In that cancer both are destroyed; in this leukaemia the cells need both of them working. The same two things, wired in opposite directions.
MBTD1 is the effector-recruiting half of a leukaemia fusion and the addressing half of a sarcoma fusion while donating nearly all of itself to both - so the junction, not the parts, assigns the role
The same protein appears in two different cancers' fusion genes - at the back of one and at the front of the other - and in both cases almost the whole protein is present. That means the two fusions do not differ because different pieces of it were kept; they differ because of which side of the join it sits on, and this is the second time this graph has found that the join, not the ingredients, decides what a fusion does.
Pentimento and restoration — the fusion overpaints the cell, and cleaning reveals whatever was underneath
Painters change their minds, and the earlier version slowly shows through the newer paint. A fusion protein paints a leukemia over a normal blood cell, and when a drug strips that paint away, what appears is whatever the painter started with.
Portfolio theory — a hedge only works against uncorrelated risk, and nobody has estimated the correlation matrix of resistance routes
Combining two drugs only helps if the ways the leukemia escapes them are genuinely different. Investors have a precise way to check that, and it needs a table of how each escape route relates to every other one. That table has never been built for this disease.
Severability (contract law) — what survives when a clause is struck, and what survives a KMT2A break
When a court strikes one clause out of a contract, settled rules decide whether the rest of the agreement still stands. A chromosome break does something structurally identical to a protein, and those rules make a prediction about where breaks are allowed to fall.
Term structure of risk — a steeply falling hazard curve is a statement about heterogeneity, not about time
Most relapses in this disease happen very early and then the risk falls away. In credit markets, a risk curve that falls like that is read as a sign that you are looking at two different kinds of borrower mixed together, not one borrower whose risk declines.
The same reader domain, mutated instead of fused, causes a childhood kidney tumour
The part of ENL that this project wants to drug is also broken in a kind of childhood kidney cancer — not by being fused to another gene, but by small mutations that change how it reads chromatin. The children affected are younger than average, and their tumours switch on the same genes.
ZMYND11::MBTD1 is the KMT2A-fusion architecture assembled from different parts - and the part it throws away is a direct physical brake on KMT2A itself
A rare leukaemia protein is built on the same plan as the fusion this project studies - a piece that finds a place on packaged DNA welded to a piece that summons a gene-activating machine - but every single component is a different molecule, and it switches on the same handful of genes. The twist nobody has written down is that the half it keeps comes from a protein whose normal job is to hold back the very enzyme this project is about, and the translocation cuts off exactly the bit that does the holding back.