Oncofusion

start / graph / ana-the-arginine-axis-is-mechanistically-real-and-genetically-non-selective-in-kmt2a-rearranged-lines

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

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

also called arginine axis dependency · TDRD3 DepMap · PRMT6 PRMT7 non-dependency · WRAD selectivity

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.

The same thing, for a scientist

Cycle 56 added five entries to the arginine axis - four mark nodes and `gene-prmt7` (Q103), plus `gene-tdrd3` (coordinator) - and every one argued mechanism without touching genetic dependency. `gene-tdrd3`'s own open question asked for this. Answered. **The instrument validates itself before anything else is read.** DepMap Chronos gene effect from the Open Targets Platform, 1,258 screens for the well-covered genes, leukaemia lines split into 16 Cellosaurus-annotated KMT2A-rearranged versus other leukaemia, two-sided permutation test at 20,000 permutations: | control | KMT2A-r | other leuk | diff | perm p | |---|---:|---:|---:|---:| | **MEN1** | **-1.2629** | -0.3745 | **-0.8885** | **5e-05** | | **KMT2A** | **-0.8113** | -0.2542 | **-0.5571** | **5e-05** | | DOT1L | -0.6042 | -0.3427 | -0.2615 | 0.055 | | GAPDH (negative) | -1.7299 | -1.7504 | +0.0205 | 0.89 | **In that instrument, the arginine axis is flat.** Nothing reaches p<0.05, and the direction is often wrong: | gene | mean all lines | diff KMT2A-r minus other | perm p | |---|---:|---:|---:| | **TDRD3** | **+0.0671** | +0.0622 | 0.31 | | TOP3B | -0.1507 | **+0.0841** | 0.090 | | PRMT1 | **-1.3984** | -0.0878 | 0.39 | | PRMT5 | **-1.1947** | +0.0786 | 0.62 | | CARM1 | -0.3222 | +0.0536 | 0.61 | | PRMT3 | +0.0419 | -0.0383 | 0.37 | | PRMT6 | -0.0666 | -0.0425 | 0.23 | | PRMT7 | -0.1513 | -0.0241 | 0.67 | **Three readings, in order of how much they should change behaviour.** First, **TDRD3 is not a dependency in any line** (+0.067 across 1,258): the reader this project wrote up today is dispensable everywhere, which is consistent with H4R3me2a having at least two structurally unrelated readers. Second, **PRMT1 and PRMT5 are strongly essential and equally so in both groups** - they are pan-essential, not fusion-selective, and this graph treats PRMT1 as an attractive target. Third, **PRMT3, PRMT6 and PRMT7 are non-dependencies outright**, which reproduces Q103's figures to the digit (it reported PRMT6 -0.0666 and PRMT7 -0.1513). **The WRAD surprise, stated with its weakness.** WDR5 (+0.2075, p=0.075) and ASH2L (+0.1104, p=0.094) trend toward being LESS essential in KMT2A-rearranged lines, which is the opposite of what this graph's WRAD reasoning would predict. Neither survives any correction and both are nominal, but the direction is worth recording rather than discarding. **So the axis is mechanistically real and, on this evidence, not actionable as a fusion-selective vulnerability.** Q103 anticipated exactly this: it called the blocking marks' writers 'brakes' in a HOXA-driven leukaemia rather than dependencies.

Facts and figures
scriptanalyses/arginine-axis-dependency-kmt2a-versus-other-leukaemia/run.py
resultsanalyses/arginine-axis-dependency-kmt2a-versus-other-leukaemia/results.json
testsmechanism-arginine-methylation-gates-kmt2-access-independently-of-menin-and-the-win-motif
depmap releaseNOT STATED - the Open Targets GraphQL depMapEssentiality payload carries no release identifier (rule 83). Absolute Chronos values move between releases; the KMT2A-versus-other CONTRAST is the durable part.
n screens for well covered genes1258
n kmt2a rearranged depmap lines16
methodPer-line Chronos gene effect from the Open Targets Platform GraphQL API; KMT2A-rearranged lines identified from Cellosaurus by an annotated Gene fusion whose HGNC xref is KMT2A, intersected with DepMap accessions; leukaemia lines selected by substring match on diseaseFromSource; two-sided permutation test, 20,000 permutations, seed 20260922, minimum 5 lines per side.
low coverage genesRPL13A returns only 394 screens in this slice and POLR2A returns 0 - measured, not inferred. Both were chosen as pan-essential controls and were the wrong choice; coverage is a property of the gene in this dataset, not of the response. Their means are flagged weakly estimated in results.json and POLR2A yields no contrast at all.
the instrument bug this analysis shipped with firstThe Open Targets `depMapEssentiality` field is a LIST OF ONE BLOCK PER TISSUE - 29 blocks for TDRD3. The first version of this script read `[0]["screens"]` with a fallback that only fired when that was empty, which it never is. It returned 25 of 1,257 screens for PRMT6 and reported a full table of plausible-looking means off 2% of the data. What caught it was the pan-essential controls looking impossible (RPL13A at 8 lines), not the code. A payload guard now refuses to report anything unless the best-covered panel gene exceeds 1,000 screens.
shown for fusiontrue
absence classmeasured-not-analysed
instrument validation
{
  "MEN1": "-1.2629 vs -0.3745, diff -0.8885, p 5e-05",
  "KMT2A": "-0.8113 vs -0.2542, diff -0.5571, p 5e-05",
  "DOT1L": "-0.6042 vs -0.3427, diff -0.2615, p 0.055",
  "GAPDH negative control": "diff +0.0205, p 0.89",
  "reading": "The two dependencies this field is built on are recovered at 5e-05 and the negative control shows nothing, so a null on the arginine panel is a null and not a dead instrument."
}
arginine panel results
{
  "TDRD3": "mean_all +0.0671 (1,258 screens); diff +0.0622, p 0.31 - not a dependency anywhere",
  "TOP3B": "mean_all -0.1507; diff +0.0841, p 0.090 - trend is LESS essential in KMT2A-r",
  "PRMT1": "mean_all -1.3984 - strongly essential; diff -0.0878, p 0.39 - pan-essential, not selective",
  "PRMT5": "mean_all -1.1947 - strongly essential; diff +0.0786, p 0.62 - pan-essential, not selective",
  "CARM1": "mean_all -0.3222; diff +0.0536, p 0.61",
  "PRMT3": "mean_all +0.0419; diff -0.0383, p 0.37 - non-dependency",
  "PRMT6": "mean_all -0.0666; diff -0.0425, p 0.23 - non-dependency; reproduces Q103's figure exactly",
  "PRMT7": "mean_all -0.1513; diff -0.0241, p 0.67 - non-dependency; reproduces Q103's figure exactly"
}
the gated complex
{
  "WDR5": "mean_all -1.1607 (pan-essential); diff +0.2075, p 0.075 - trend toward LESS essential in KMT2A-r",
  "ASH2L": "mean_all -0.3316; diff +0.1104, p 0.094 - same direction",
  "RBBP5": "mean_all -1.1065; diff +0.0690, p 0.51",
  "DPY30": "mean_all -0.3898; diff -0.0680, p 0.36",
  "KMT2D": "mean_all -0.3892; diff +0.0751, p 0.42",
  "reading": "WDR5 and RBBP5 are strongly essential everywhere. The two nominal trends both point the wrong way for this graph's WRAD reasoning and neither survives correction."
}
the audit this entry called for found nothing
{
  "what_was_feared": "This entry's open question said ten DepMap analyses might have read one tissue block and inherited the same 2%-of-the-data bug this script shipped.",
  "audited_cycle_57": "41 python scripts in analyses/ and scripts/ touch depMapEssentiality.",
  "scripts_with_the_bug": "0. Every one of them iterates the tissue blocks.",
  "and_the_shape_was_already_documented_here": "analyses/unwritten-interactors-dependency-triage/run.py carries the comment 'depMapEssentiality is a LIST of tissue blocks, each with its own screens.' The project knew. I was the only one who got it wrong.",
  "the_tg0_pattern_is_not_the_bug": "Several scripts read tg[0].get('depMapEssentiality') where tg = data.targets - that indexes the queried TARGET, not the first tissue block, and then iterates the blocks correctly. Some of those scripts additionally read the returned approvedSymbol back and record an error if it differs from the symbol asked for, which is the same readback guard I had to learn from CTH resolving to VSIG2.",
  "so": "The concern is closed as a non-issue, and the honest version of the lesson is narrower than the entry first implied: this was one script's bug, not a systemic one, and the codebase's existing convention was already correct."
}
the coverage question also closes
{
  "tested_cycle_57": "15 genes: RPL13A, RPL5, RPS11, POLR2A, POLR2B, SF3B1, ACTB, GAPDH, TUBB, MEN1, KMT2A, TDRD3, PRMT5, EEF2, PSMB2",
  "full_coverage_1257_to_1258_screens_29_blocks": 13,
  "thin_or_absent": "RPL13A 394 screens / 22 blocks; POLR2A 0 / 0",
  "the_tell_that_it_is_not_systemic": "RPS11 (1,258) and POLR2B (1,258) - the paralogues of the two thin genes - are fully covered, as are ribosomal RPL5 and proteasomal PSMB2.",
  "consequence_for_this_analysis": "Only the two pan-essential controls were affected, and they were the wrong controls. No panel gene of interest was thin: every arginine-axis gene returned 1,257-1,258."
}
How this could bite us back
  • This is a negative result about genetic dependency and must not be read as saying the arginine axis is unimportant or that PRMT inhibitors do not work. A knockout screen measures proliferation after removing one gene; it cannot see a redundant reader, a brake, a differentiation effect, or a combination effect, and all four are live readings of this axis.
  • The opposite misreading is the more dangerous one. PRMT1 and PRMT5 ARE strong dependencies here - they are simply not selective - so an inhibitor is expected to hit normal proliferating tissue at a comparable dose. Nothing here supports dosing either in a child.
  • Q103's brake model predicts that inhibiting PRMT6 would HELP the leukaemia, not hurt it, because PRMT6 gain lowers HOX transcription. Any therapeutic use of this axis has to resolve that sign before anything else.
  • The two nominal WRAD trends are almost certainly chance at 17 tests with no correction. They are recorded as a direction and should not be cited as a finding.
Still open
  1. Does a drug-sensitivity readout show what the knockout screen cannot? PRMT5 inhibitors are in trials and PRMT6/PRMT7 tool compounds exist; a viability screen under a menin inhibitor is the combination form that a single-gene knockout cannot see, and Q103's brake model predicts ANTAGONISM rather than synergy for PRMT6 inhibition.
  2. Is TOP3B's wrong-direction trend real? It is the enzyme TDRD3 delivers and the only gene on the reader arm with any nominal signal, at p 0.090 in the direction of LESS essential in KMT2A-rearranged lines.
  3. Would a co-dependency analysis be more informative than a selectivity one? The axis may act through the wild-type KMT2A allele rather than the fusion, in which case the right question is whether TDRD3 or PRMT6 loss co-varies with KMT2A or WDR5 dependency across lines, not whether it is selective for the fusion.
  4. ANSWERED, cycle 57, in the reassuring direction on both halves. (1) The slice is NOT systematically incomplete: of 15 genes spanning ribosomal, polymerase, spliceosome, housekeeping and project genes, THIRTEEN return 1,257-1,258 screens across 29 tissue blocks - including RPS11 and POLR2B, the paralogues of the two thin ones. RPL13A (394 screens, 22 blocks) and POLR2A (0 screens, 0 blocks) are gene-specific gaps, not a slice-wide truncation, and this analysis had the bad luck to pick both as its pan-essential controls. (2) The tissue-block bug is not systemic either: 41 python scripts in this repo touch depMapEssentiality and ZERO read [0] - every one iterates the blocks, and analyses/unwritten-interactors-dependency-triage/run.py already carries the comment 'depMapEssentiality is a LIST of tissue blocks, each with its own screens.' The convention was correct and documented; this script was the sole offender. What remains genuinely open is narrow: WHY POLR2A has no Chronos data at all, which is a DepMap question rather than a question about this project.
  5. No infant-derived KMT2A::AFF1 line is in DepMap with usable screens, which is now the third distinct beat this cycle to hit that wall. The seven lines this graph names are the standing answer and none is screened.
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    In DepMap Chronos gene-effect data served by the Open Targets Platform, with leukaemia lines split into 16 Cellosaurus-annotated KMT2A-rearranged versus other leukaemia and tested by two-sided permutation at 20,000 permutations, the instrument recovers both known selective dependencies and shows nothing for the negative control: MEN1 -1.2629 in KMT2A-rearranged against -0.3745 in other leukaemia (diff -0.8885, p 5e-05); KMT2A itself -0.8113 against -0.2542 (diff -0.5571, p 5e-05); DOT1L -0.6042 against -0.3427 (diff -0.2615, p 0.055); GAPDH diff +0.0205 (p 0.89).

    What this evidence cannot show

    NO DEPMAP RELEASE IS STATED because the API does not expose one (rule 83), so absolute Chronos values are not reproducible across releases even though the contrast should be. The KMT2A-rearranged group is 16 lines, dominated by myeloid lines, and contains NO infant-derived KMT2A::AFF1 line - this project has documented that seven such lines exist and none is in DepMap with usable screens. Leukaemia membership is a substring match on diseaseFromSource and will miss oddly-named entries. Permutation p-values on groups this size are unstable and the 5e-05 figures are at the resolution floor of 20,000 permutations rather than exact.

    Evidence that argues against this (1)
    • A validated instrument on the controls does not guarantee power for the panel. MEN1's effect is enormous (-0.89 difference) and DOT1L's, at -0.26, only reaches p 0.055 - so this design detects large selective effects and is genuinely underpowered for modest ones. A null on a gene with a true difference of, say, -0.15 would be expected here, and several arginine-panel differences are in that range.
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    No gene on the arginine axis is a selective dependency in KMT2A-rearranged lines. Difference in mean gene effect (KMT2A-rearranged minus other leukaemia) with permutation p: TDRD3 +0.0622 (p 0.31), TOP3B +0.0841 (0.090), PRMT1 -0.0878 (0.39), PRMT5 +0.0786 (0.62), CARM1 +0.0536 (0.61), PRMT3 -0.0383 (0.37), PRMT6 -0.0425 (0.23), PRMT7 -0.0241 (0.67). None reaches p<0.05 and several point the wrong way. Separately, the genes split by absolute essentiality: PRMT1 (-1.3984) and PRMT5 (-1.1947) are strongly essential in all lines, while TDRD3 (+0.0671), PRMT3 (+0.0419), PRMT6 (-0.0666) and PRMT7 (-0.1513) are not dependencies at all. The PRMT6 and PRMT7 figures reproduce Q103's independently reported values to four decimal places.

    What this evidence cannot show

    A CRISPR knockout screen cannot see a factor whose function is redundant, and H4R3me2a has at least two structurally unrelated readers - TDRD3's Tudor domain and the tandem PHD(4-6) inside KMT2D - so a null on TDRD3 is uninformative about the MARK. The same logic protects the marks generally: a gate that can be bypassed by a second reader, or that acts on the wild-type allele rather than the fusion, is invisible to a single-gene knockout. No multiple-testing correction is applied across 17 genes, which matters mainly for the two nominal trends and not for the nulls. Gene effect is a proliferation readout; a factor that changes differentiation or drug sensitivity without changing growth does not appear.

    Evidence that argues against this (2)
    • The strongest objection to reading this as a negative for the axis is that the axis was never proposed as a dependency. Q103's own report called the writers of the blocking marks 'brakes' in a HOXA-driven leukaemia and predicted that the therapy would invert rather than the mechanism fail - PRMT6 gain LOWERS HOX transcription, PRMT7 knockdown RAISES KMT2D-catalysed H3K4me3. A brake is expected to be a non-dependency in a knockout screen, and its removal would be predicted to help the leukaemia, not hurt it. So these nulls are consistent with the mechanism rather than against it.
    • PRMT1 and PRMT5 being pan-essential is a real problem for the therapeutic reading in the opposite direction: they ARE strong dependencies, just not selective ones, so inhibiting them is expected to harm normal proliferating tissue at a similar dose. This graph treats PRMT1 as an attractive target and that attractiveness rests on fusion-cofactor biology rather than on a selectivity window, which this analysis cannot supply.
  3. E3weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.database

    Two members of the complex the marks gate trend toward being LESS essential in KMT2A-rearranged lines than in other leukaemia: WDR5 diff +0.2075 (p 0.075) and ASH2L diff +0.1104 (p 0.094). Both are nominal and neither survives correction. In absolute terms WDR5 (-1.1607) and RBBP5 (-1.1065) are strongly essential across all lines, while ASH2L (-0.3316), DPY30 (-0.3898) and KMT2D (-0.3892) are moderate.

    What this evidence cannot show

    Two nominal p-values out of 17 tests is approximately what chance produces, so this is recorded as a direction rather than a finding, and the confidence is weak deliberately. Both are also confounded in the same way as everything else here: the KMT2A-rearranged group is myeloid-dominated and contains no infant AFF1 line, and a lineage difference between the groups would move WRAD subunits for reasons unrelated to the fusion.

    Evidence that argues against this (1)
    • If this direction were real it would be awkward for the project rather than for the axis: this graph reasons extensively about WRAD as the complex the fusion exploits, and a fusion line LESS dependent on WDR5 than other leukaemia lines is not what that predicts. The most likely explanation remains chance or lineage, and the honest position is that the analysis is underpowered to tell those apart - but it should not be quietly dropped because it is inconvenient.
Connected to