Oncofusion

start / graph / ana-pcgf1-loss-lowers-baseline-fusion-occupancy-and-makes-the-dot1l-inhibitor-the-stronger-displacer

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

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

also called PCGF1 loss lowers baseline fusion occupancy · the DOT1L inhibitor displaces more fusion in the knockout · the SGC ruler was not a ruler

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.

The same thing, for a scientist

**Statement.** Cycle 101 scored prediction 5's eviction clause and used SGC0946's genotype difference as a ruler: *"Its genotype difference of 0.2276 is the scale of genotype-level library variation against which VTP's 0.0133 should be read."* Its own counter-evidence flagged that as *"a convenient interpretation of a result that could be real biology"*, and its open question asked whether the 22.8% was a noise floor or a finding. **The ruler was never calibrated. It is now.** **Two independent input pairs give the technical floor**, measured over the same reconstructed `S_FLAG` bins, with no antibody and no drug: | | C1 / NT | deviation from 1 | |---|---:|---:| | input, q125 pair | 1.0967 | **9.7%** | | input, q124 pair | 1.0360 | **3.6%** | **So genotype-level variation in `S_FLAG` share is roughly 4–10%.** Against that: | | value | reading | |---|---:|---| | **VTP** ratio-of-ratios, C1 vs NT | **1.0134** (1.3%) | **inside the floor** | | **SGC** ratio-of-ratios, C1 vs NT | **0.7724** (22.8%) | **2–6× outside it** | | **FLAG DMSO** C1/NT, baseline | **0.8375** (16.3% lower) | **outside it, and against the input bias** | **Three consequences, in order of how well supported they are.** **1. Cycle 101's eviction conclusion is unaffected and better founded.** VTP's 1.3% agreement now sits inside a floor measured from inputs rather than borrowed from a drug arm. *Evicted normally* stands. **2. My use of SGC as a ruler was wrong, and overstated the floor by 2–6×.** That was the specific thing the entry's counter-evidence warned about, and it took one more cycle to test. **3. PCGF1 loss lowers baseline fusion occupancy at the top-FLAG sites by 16.3%** — and the **inputs point the other way** (C1 is 3.6–9.7% *higher*), so a coverage artefact would mask this rather than create it. **The knockout has less fusion at the sites the fusion most occupies, before any drug.** **And the observation that motivated all of it:** in the control, VTP removes more fusion than SGC (24.5% against 11.2%); **in the PCGF1 knockout the order reverses** — SGC removes **31.4%** against VTP's **23.5%**. **A DOT1L inhibitor displaces more of the fusion than a menin inhibitor does, in a documented resistance genotype.** **THE REDISTRIBUTION OBJECTION WAS TESTED BEFORE THIS SHIPPED AND IT FAILS — the effect gets LARGER.** This entry's counter-evidence noted that `S_FLAG` is the *parental* line's top-FLAG set, so a knockout that **moved** the fusion rather than losing it would show less signal there while having the same total. Recomputed with **each genotype scored at its own top 0.1%**: | scored at | NT | C1 | **C1/NT** | |---|---:|---:|---:| | parental top set | 0.008412 | 0.007045 | **0.8374** | | **its own top set** | 0.010289 | 0.008083 | **0.7856** | **At its own best sites the knockout is 21.4% behind, worse than the 16.3% at the parental set.** Redistribution predicts recovery; the opposite happens. **So PCGF1 loss reduces how concentrated the fusion is at its own strongest sites, rather than relocating it.** The top-set turnover is real — NT and C1 top sets overlap 59.2% — but it does not rescue the deficit.

Go deeper
What is solid, what is suggestive, and what has no error bar at all

Solid: the input-derived floor. Two pairs, from two beats, over the same bins, agreeing at 3.6% and 9.7%. That is a genuine measurement of genotype-level variation in this statistic and it did not exist before this cycle.

Solid enough to act on: the 22.8% SGC genotype difference and the 16.3% baseline occupancy deficit both exceed that floor by margins that a 4–10% band does not explain. Both are ratio-of-ratios or directly-opposed-to-bias comparisons, so genotype-level library scale cancels or works against them.

Suggestive and not more: the reversal itself. Comparing SGC to VTP within the knockout is a drug-versus-drug comparison, and the input pairs bound genotype variation, not drug-arm variation. There is no replicate of any FLAG track — n = 1 per genotype per treatment across all six. So the statement "SGC displaces more than VTP in the knockout" rests on two single measurements differing by 10.4%, with no error bar of the right kind. I am reporting the direction and the magnitude and explicitly not defending the reversal as established.

And a structural limit that applies to all of it: GSE260742 has no usable spike-in, which this project established across 419,824 characters of the deposit with zero hits. Every number here is a within-sample share, so a uniform global change in fusion occupancy is invisible, and "less fusion at S_FLAG" strictly means "a smaller share of this sample's own signal at S_FLAG".

Facts and figures
questionis SGC0946's 22.8% genotype difference a noise floor or a finding?
answera finding - it is 2-6x the technical floor measured from two independent input pairs
absence classmeasured-not-analysed
absence class noteAll six FLAG tracks and both input pairs were binned and on disk before this cycle; what had never been analysed is the genotype contrast, because the FLAG series was mislabelled as H2AK119ub until cycle 101.
bite backIf read as 'a DOT1L inhibitor is the better agent in menin-inhibitor-resistant disease', this would be a clinical inference from two unreplicated measurements, in one CRISPR clone, of one murine line carrying KMT2A::MLLT3 rather than the KMT2A::AFF1 this project studies, with no spike-in and no PCGF1 confirmation in the deposit. SGC0946 is a tool compound, not a clinical agent. Nothing here should inform a treatment choice, and the reversal needs a replicated experiment in a human line before it is even a hypothesis worth testing.
technical floor
{
  "q125_input_pair_C1_over_NT": 1.0967,
  "q125_deviation_pct": 9.7,
  "q124_input_pair_C1_over_NT": 1.036,
  "q124_deviation_pct": 3.6,
  "reading": "genotype-level variation in S_FLAG share is roughly 4-10%, measured with no antibody and no drug"
}
against that floor
{
  "VTP_ratio_of_ratios": 1.0134,
  "VTP_pct": 1.3,
  "VTP_verdict": "inside the floor",
  "SGC_ratio_of_ratios": 0.7724,
  "SGC_pct": 22.8,
  "SGC_verdict": "2-6x outside it",
  "FLAG_DMSO_C1_over_NT": 0.8375,
  "baseline_pct": 16.3,
  "baseline_verdict": "outside it, and AGAINST the input bias - inputs put C1 3.6-9.7% HIGHER, so a coverage artefact would mask this rather than create it"
}
three consequences
{
  "1_eviction_conclusion_unaffected": "VTP's 1.3% now sits inside a floor measured from inputs rather than borrowed from a drug arm; 'evicted normally' stands and is better founded",
  "2_my_ruler_was_wrong": "using SGC as the noise scale overstated the floor by 2-6x - the exact thing cycle 101's counter-evidence warned about",
  "3_baseline_occupancy": "PCGF1 loss lowers fusion occupancy at the top-FLAG sites by 16.3% before any drug"
}
the observation that motivated it
{
  "control_NT": {
    "VTP_removes_pct": 24.5,
    "SGC_removes_pct": 11.2,
    "order": "VTP > SGC"
  },
  "knockout_C1": {
    "VTP_removes_pct": 23.5,
    "SGC_removes_pct": 31.4,
    "order": "SGC > VTP - REVERSED"
  },
  "status": "SUGGESTIVE ONLY - a drug-versus-drug comparison within one genotype, and the input pairs bound genotype variation, not drug-arm variation. No FLAG track has a replicate."
}
redistribution test
{
  "ran": "before this entry shipped, because its own counter-evidence named it",
  "question": "is the 16.3% baseline deficit a LOSS or a REDISTRIBUTION?",
  "share_at_parental_top_set": {
    "NT": 0.008412,
    "C1": 0.007045,
    "C1_over_NT": 0.8374
  },
  "share_at_own_top_set": {
    "NT": 0.010289,
    "C1": 0.008083,
    "C1_over_NT": 0.7856
  },
  "top_set_sizes": {
    "parental": 1174,
    "NT": 1224,
    "C1": 1231
  },
  "overlap_with_parental": {
    "NT_pct": 52.4,
    "C1_pct": 66.7
  },
  "overlap_NT_top_with_C1_top_pct": 59.2,
  "verdict": "REDISTRIBUTION REFUTED - scored at its own best sites the knockout is 21.4% behind, worse than 16.3% at the parental set. Relocation predicts recovery; the deficit deepens. PCGF1 loss reduces how CONCENTRATED the fusion is at its strongest sites."
}
Still open
  1. Is an input pair the right floor for a ChIP-derived ratio? Inputs carry no antibody, so they bound library and coverage variation but not ChIP efficiency - plausibly the largest term. If the true floor is wider, SGC's 22.8% may fall inside it and two of this entry's conclusions weaken. A same-antibody same-condition pair would settle it and none exists in this series.
  2. ANSWERED IN THIS ENTRY: a LOSS of concentration, not a redistribution. At each genotype's own top set C1/NT is 0.7856, deeper than the 0.8374 at the parental set. The live question that replaces it: where does the displaced share GO? A within-sample share must sum to 1, so a 21% deficit at the top sites is a 21% surplus somewhere, and nobody has looked at the rest of the genome.
  3. Does the reversal survive stratification? Q125 and Q130 both found aggregates hiding monotone gradients. The same quartiles cannot be used here without circularity - they are defined on FLAG - but a stratifier built from the knockout's own DMSO signal would be independent of the drug contrast.
  4. Why is SGC0946, a DOT1L inhibitor, displacing the fusion at all? It removes 11.2% in the control. Cycle 83 already found a DOT1L inhibitor moving the MLL1 channel as much as a menin inhibitor; this is the same shape in the FLAG channel and nobody has explained the mechanism.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    Over a reconstructed S_FLAG of 1,174 bins (top 0.1% of the q117 FLAG DMSO ranking universe), input share is 0.002442 for the q125 NT input and 0.002678 for the q125 C1 input (ratio 1.0967), and 0.003337 for the q124 NT input and 0.003457 for the q124 C1 input (ratio 1.0360). FLAG share in DMSO is 0.008412 in NT and 0.007045 in the PCGF1 knockout, a ratio of 0.8375. The VTP/DMSO ratio-of-ratios between genotypes is 1.0134 and the SGC0946/DMSO ratio-of-ratios is 0.7724. Within the knockout, SGC0946 leaves a FLAG share of 0.004831 against VTP's 0.005393; within the control, SGC0946 leaves 0.007470 against VTP's 0.006355.

    What this evidence cannot show

    NO FLAG TRACK HAS A REPLICATE - n = 1 per genotype per treatment across all six, so no comparison here has a confidence interval and the input pairs bound GENOTYPE variation only, not drug-arm variation. The SGC-versus-VTP reversal within the knockout is therefore a difference between two single measurements and is reported as a direction, not a result. S_FLAG is reconstructed at 1,174 bins against Q117's stored 1,228, so absolutes must not be quoted against Q117's published figures; every comparison here uses one bin set for all arms so the contrasts survive. WITHIN-SAMPLE SHARE ONLY - GSE260742 has no usable spike-in, so a uniform global change is invisible and 'less fusion' means 'a smaller share of this sample's own signal'. The PCGF1 knockout is one guide, one clone, no rescue, and the deposit contains no PCGF1 measurement to confirm the knockout. The two input pairs are not independent of each other in the way two biological replicates would be: both compare the same two genotypes, so they bound reproducibility of the measurement rather than biological variability between clones.

    Evidence that argues against this (3)
    • The input pairs may be the wrong floor for the comparison they are being used to license. They contain no antibody and no drug, so they capture library and coverage variation but not ChIP-efficiency variation - which is plausibly the largest term in any ChIP ratio and is exactly what differs between an SGC-treated and a VTP-treated immunoprecipitation. A floor of 4-10% derived from inputs could badly understate the real variability of a ChIP-derived ratio-of-ratios, in which case SGC's 22.8% might sit inside the true floor after all and this entry's second and third conclusions both weaken.
    • I raised redistribution as the alternative to loss - S_FLAG is the parental line's top set, so a fusion that MOVED would show less there with the same total - and noted the test was computable from the same file and undone. DONE BEFORE SHIPPING and it FAILS: scored at each genotype's own top 0.1%, C1/NT is 0.7856 against 0.8374 at the parental set, so the deficit DEEPENS from 16.3% to 21.4% where relocation predicts recovery. What survives as a caveat is that top-set membership turns over substantially (NT and C1 top sets overlap 59.2%), so 'its own top sites' is a moving target and the two shares are not measured on the same genomic bins.
    • The headline orders these findings by how well supported they are, which is honest, but the title leads with the two that are outside the floor and the reversal - the most clinically suggestive claim - is the one with no error bar of the right kind. A reader skimming will take away 'a DOT1L inhibitor works better in resistant cells', which is a statement about two unreplicated measurements in one clone of one murine line carrying the wrong fusion for this project. The entry says so in a section; the title does not.
Connected to