start / graph / ana-the-eviction-clause-was-reachable-all-along-and-the-fusion-is-evicted-normally-in-the-pcgf1-knockout

Analysismoderate

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

also called the eviction clause was reachable all along · the fusion is evicted normally in the PCGF1 knockout · GSM8123407-412 are FLAG not H2AK119ub

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.

The same thing, for a scientist

**Statement.** `hyp-q117` prediction 5: *"In a PCGF1 or BCOR knockout, the fusion is evicted normally but the vacated sites are not silenced, and the programme returns on drug withdrawal."* Q125 scored the **silencing** clause in cycle 88 and found it **attenuated, not abolished**. The **eviction** clause was declared **UNREACHABLE** by Q124 and left unscored by Q125, both for the same stated reason: *no FLAG, MLL1 or MENIN track exists in the NTvKO arm of GSE260742 by title.* **By title.** Q125 saw the discrepancy and filed it `UNRESOLVED`: six files, `GSM8123407–412`, are titled *"H2AK119ub rep1"* while carrying `ChIP_FLAG` in their GEO filename and in the TDF-internal `trackNames` their depositors' igvtools wrote. It proposed `elink gds→sra` as the resolution and did not relabel them. **Resolved this cycle from three independent sources, no download:** | accession | GEO title | depositor's filename | |---|---|---| | `GSM8123407` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_CLONE1_DMSO.tdf` | | `GSM8123408` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_CLONE1_SGC.tdf` | | `GSM8123409` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_CLONE1_VTP.tdf` | | `GSM8123410` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_NT_DMSO.tdf` | | `GSM8123411` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_NT_SGC.tdf` | | `GSM8123412` | H2AK119ub **rep1** | `ChIP_`**`FLAG`**`_NT_VTP.tdf` | against `GSM8123417`/`GSM8123420`, titled *rep2*, which are `ChIP_`**`K119UB`**`_*.tdf`. The **third** source is the binned files already on disk, whose headers read `trackNames=[ChIP_FLAG__NT_DMSO.filt.bam]`. **And the fourth is an anomaly nobody had explained:** Q131 found `REP1/REP2` share ratios of **3.1716** and **2.8412** inside `S_FLAG` and excluded them as inadmissible. That is precisely what a FLAG ChIP over an H2AK119ub ChIP predicts in bins **defined** as top-FLAG. **So this is a complete FLAG occupancy series** — two genotypes × three treatments — **and all six were already binned on disk** under Q125's misleading `q125_bin_K119_REP1_*` name. **Zero downloads.** **The result.** FLAG share within reconstructed `S_FLAG` (top 0.1%, 1,174 bins against Q117's stored 1,228): | | DMSO | VTP | **VTP/DMSO** | |---|---:|---:|---:| | **control (NT)** | 0.008412 | 0.006355 | **0.7554** | | **PCGF1-KO (C1)** | 0.007045 | 0.005393 | **0.7655** | **`E1` — EVICTED NORMALLY**, a **1.3%** relative difference inside the pre-registered 20% band. `U1` passes: VTP genuinely evicts in the control, removing ~24% of fusion share. **And the `U2` control makes the agreement meaningful rather than assumed.** SGC0946, a DOT1L inhibitor not expected to evict the fusion, gives **0.8879** in control against **0.6858** in the knockout — a **22.8% genotype difference, 17× larger than VTP's 1.3%.** The two genotypes agree on VTP eviction far more closely than they agree on an unrelated drug. **So prediction 5 is now scored in both halves: eviction CONFIRMED, silencing ATTENUATED rather than abolished.** **THE BATCH OBJECTION WAS CHECKED BEFORE THIS SHIPPED AND IT FAILS.** This entry's own counter-evidence asked whether the two VTP tracks were processed together in a way the SGC tracks were not, which would make the 1.3% agreement a batch property. **All four are one batch:** submitted Mar 03 2024, **Illumina NovaSeq 6000**, with **consecutive SRA accessions** — `SRX23818684` (KO SGC), `685` (KO VTP), `687` (NT SGC), `688` (NT VTP) — and consecutive BioSamples `SAMN40239754/755/757/758`. **The accessions are grouped by genotype, not by drug**, which is if anything the opposite of a VTP-specific batch. The contrast between 1.3% and 22.8% therefore cannot be attributed to batch structure.

Go deeper

What this does to the resistance story, which is the reason it matters

Q125's resistance-q125-a-polycomb-deficient-cell-is-evicted-normally-and-silenced-less argued that a PCGF1-deficient cell would look fully drug-engaged on an occupancy readout while being functionally resistant. Both halves of that are now measured rather than one measured and one assumed. The knockout is evicted to within 1.3% of the control — a pharmacodynamic assay reading target engagement would score it as a complete responder — while Q125 showed its silencing is attenuated (H2AK119ub Q1 1.4450 → 1.1325 (paired) and 1.0151 → 0.8842 (rep2), both conventions falling, H3K27me3 2.0007 → 1.0134).

That is the sharpest form this project has of its own standing threat. Occupancy-based PD is the readout a trial would reach for, and here it is blind by construction: the drug does exactly what it is supposed to do at the target, and the consequence downstream does not follow.

Q130 sharpened it further in the other direction — H3K79me2, stratified by eviction depth, also fails to see the genotype, with the difference sitting at the quartile where the drug works least. So two independent readouts score a resistant genotype as engaged, and only the repressive marks distinguish it.

What is reconstructed rather than reproduced, and why the contrast survives it

S_FLAG here is my reconstruction, not Q117's stored set. Taking the top 0.1% of the 1,174,321-bin ranking universe in q117_bin_FLAG_DMSO.tsv gives 1,174 bins against Q117's published 1,228 — 4.4% short, inside the 15% tolerance declared before running but not the same bin set.

That voids comparison to Q117's published numbers and does not touch this result. The NT-versus-C1 contrast is computed over one set of bins for both genotypes, so any difference between my reconstruction and Q117's affects both arms identically and cancels in the ratio of ratios. What it does mean is that the absolute 0.7554 should not be quoted against Q117's eviction figures, and any future beat wanting that comparison must use Q117's stored cuts.

Every number is a within-sample share. GSE260742 has no usable spike-in — this project established that across 419,824 characters of the deposit with zero hits — so a uniform global change in FLAG occupancy is invisible here and all ratios are relative.

Facts and figures

question	is the fusion evicted normally in a PCGF1 knockout? - prediction 5's first clause
verdict	E1 - EVICTED NORMALLY. Prediction 5's eviction clause CONFIRMED.
downloads required	0
U2 reading	SGC0946's 22.8% genotype difference is 17x VTP's 1.3%, so the two genotypes agree on VTP eviction far more closely than on an unrelated drug
why it matters	a PD assay reading target engagement would score the PCGF1-deficient cell as a complete responder - evicted to within 1.3% of control - while its silencing is attenuated. Q130 showed H3K79me2 is blind to the same genotype. Two readouts score a resistant genotype as engaged.
absence class	measured-not-curated
absence class note	The FLAG series was deposited, public and already binned on this machine; what was absent was a correct label. Q124 declared the clause unreachable from the titles and Q125 flagged the discrepancy without resolving it.

share S FLAG

R VTP over DMSO

R SGC over DMSO

the relabelling

{
  "flag_series_mislabelled_as_H2AK119ub_rep1": {
    "GSM8123407": "ChIP_FLAG_CLONE1_DMSO.tdf",
    "GSM8123408": "ChIP_FLAG_CLONE1_SGC.tdf",
    "GSM8123409": "ChIP_FLAG_CLONE1_VTP.tdf",
    "GSM8123410": "ChIP_FLAG_NT_DMSO.tdf",
    "GSM8123411": "ChIP_FLAG_NT_SGC.tdf",
    "GSM8123412": "ChIP_FLAG_NT_VTP.tdf"
  },
  "genuine_H2AK119ub_rep2": {
    "GSM8123417": "ChIP_K119UB_CLONE1_DMSO.tdf",
    "GSM8123420": "ChIP_K119UB_NT_DMSO.tdf"
  },
  "sources": [
    "GEO supplementary filenames",
    "TDF-internal trackNames visible in the binned headers on disk",
    "the 3.1716 and 2.8412 REP1/REP2 share ratios Q131 found and excluded, which are what a FLAG ChIP over an H2AK119ub ChIP predicts inside top-FLAG bins"
  ],
  "still_unrun": "elink gds->sra, the orthogonal check Q125 proposed"
}

sflag

{
  "universe_bins": 1174321,
  "reconstructed_bins": 1174,
  "q117_stored_bins": 1228,
  "note": "reconstruction voids comparison to Q117's published absolutes; the NT-vs-C1 contrast is unaffected because both arms use the same set"
}

share S FLAG — the numbers behind the chart

{
  "NT_DMSO": 0.008412,
  "NT_VTP": 0.006355,
  "NT_SGC": 0.00747,
  "C1_DMSO": 0.007045,
  "C1_VTP": 0.005393,
  "C1_SGC": 0.004831
}

R VTP over DMSO — the numbers behind the chart

{
  "NT": 0.7554,
  "C1_PCGF1_KO": 0.7655,
  "relative_difference": 0.0133
}

R SGC over DMSO — the numbers behind the chart

{
  "NT": 0.8879,
  "C1_PCGF1_KO": 0.6858,
  "relative_difference": 0.2276
}

prediction 5 now scored in both halves

{
  "eviction_clause": "CONFIRMED - evicted normally, 1.3% difference",
  "silencing_clause": "ATTENUATED not abolished (Q125, cycle 88) - H2AK119ub Q1 1.4450 -> 1.1325, H3K27me3 2.0007 -> 1.0134",
  "withdrawal_clause": "still UNREACHABLE - no washout arm exists in any of the 158 samples",
  "BCOR_arm": "still never tested - the series' only genetic perturbation is the PCGF1 knockout"
}

batch check

{
  "ran": "before this entry shipped, because its own counter-evidence named it",
  "all_four_samples": "submitted Mar 03 2024, Illumina NovaSeq 6000, ChIP-Seq",
  "sra": {
    "GSM8123408_KO_SGC": "SRX23818684",
    "GSM8123409_KO_VTP": "SRX23818685",
    "GSM8123411_NT_SGC": "SRX23818687",
    "GSM8123412_NT_VTP": "SRX23818688"
  },
  "biosample": {
    "GSM8123408": "SAMN40239758",
    "GSM8123409": "SAMN40239757",
    "GSM8123411": "SAMN40239755",
    "GSM8123412": "SAMN40239754"
  },
  "verdict": "the batch objection FAILS - all four are one submission on one instrument with consecutive accessions grouped by GENOTYPE rather than by drug, so the VTP agreement is not a batch the SGC pair lacks"
}

Still open

Run elink gds->sra on GSM8123407-412. It is the one genuinely orthogonal check - the GEO filename and the TDF trackName may both derive from the same internal naming, so the three sources agreeing is weaker than it looks. Q125 proposed this one request and it still has not been run.
ANSWERED: no. All four are one submission, one instrument, consecutive SRA and BioSample accessions grouped by genotype. The batch objection fails. The surviving version is that n = 1 per cell gives no confidence interval at all.
Is SGC0946's 22.8% genotype difference a noise floor or a finding? I used it as a ruler. Cycle 83 showed a DOT1L inhibitor moves the MLL1 channel as much as a menin inhibitor, so PCGF1 loss changing the fusion's response to DOT1L inhibition would be a result rather than variation.
Re-run this on Q117's STORED S_FLAG cuts rather than a reconstruction, so the absolute ratios can be compared to Q117's published eviction figures. My 1,174 bins against 1,228 is close enough for an internal contrast and not for an external one.

Every source behind this entry (1)

E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
GSM8123407 through GSM8123412 are titled 'H2AK119ub rep1' in GEO and their supplementary files are named ChIP_FLAG_{CLONE1,NT}_{DMSO,SGC,VTP}.tdf; GSM8123417 and GSM8123420, titled rep2, are ChIP_K119UB files. The binned copies on disk carry TDF-internal trackNames reading ChIP_FLAG__NT_DMSO.filt.bam. Over a reconstructed S_FLAG of 1,174 bins (top 0.1% of a 1,174,321-bin universe from q117_bin_FLAG_DMSO.tsv), FLAG share is 0.008412 in NT DMSO, 0.006355 in NT VTP, 0.007045 in PCGF1-KO DMSO and 0.005393 in PCGF1-KO VTP, giving VTP/DMSO ratios of 0.7554 and 0.7655, a relative difference of 1.3%. The same contrast for SGC0946 gives 0.8879 and 0.6858, a relative difference of 22.8%.
analyses/the-eviction-clause-was-reachable-all-along/run.py and results.json, this repository; GEO sample records for GSM8123407-412, GSM8123417 and GSM8123420 fetched 2026-09-23; bins from C:/Temp/q117 and C:/Temp/q125, computed with zero downloads · Oncofusion knowledge graph, coordinator · internal analysis, reproducible from the repository 2026
What this evidence cannot show
S_FLAG IS RECONSTRUCTED, NOT Q117's STORED SET - 1,174 bins against 1,228, so the absolute ratios must not be quoted against Q117's published eviction figures; the NT-versus-C1 contrast survives because both arms use the same reconstructed set. WITHIN-SAMPLE SHARE ONLY: GSE260742 has no usable spike-in, so a uniform global change is invisible. n = 1 per genotype per treatment - there is no replicate of the FLAG series, so no confidence interval is available and the 1.3% agreement could be coincidence at this sample size. The 20% band was set before looking but is arbitrary, which is why the SGC arm is reported as the scale of genotype-level variation. The PCGF1 knockout is one guide and one clone with no rescue and no deposited PCGF1 measurement, a limitation this project recorded when Q125 first used it. The mislabelling is established from filenames and TDF headers, not from the depositor's methods or from SRA metadata - elink gds->sra, which Q125 proposed, still has not been run.
Evidence that argues against this (3)
- I flagged that a 1.3% agreement against 22.8% on the control drug is almost too good and could be a batch property if the VTP samples were processed together in a way the SGC samples were not. CHECKED BEFORE SHIPPING and it fails: all four samples were submitted Mar 03 2024 on one NovaSeq 6000 with consecutive SRA accessions SRX23818684/685/687/688 and BioSamples SAMN40239754/755/757/758, grouped by GENOTYPE rather than by drug. What remains unaddressed is the deeper version: n = 1 per cell means no confidence interval exists, so 1.3% could still be coincidence even without a batch explanation.
- Reading the SGC arm as 'genotype-level library variation' is a convenient interpretation of a result that could be real biology. SGC0946 inhibits DOT1L, and this project has already shown (cycle 83) that a DOT1L inhibitor moves the MLL1 channel as much as a menin inhibitor does. A 22.8% genotype-dependent difference in FLAG occupancy under DOT1L inhibition might mean PCGF1 loss changes how the fusion responds to DOT1L inhibition - which would be a finding, not a noise floor - and I have used it as a ruler without testing it as a result.
- This entry converts a clause from UNREACHABLE to CONFIRMED on the strength of a relabelling, and the relabelling rests on filenames rather than on the depositor's stated methods. If the GEO titles are right and the filenames are legacy artefacts of the depositor's pipeline, then six files have been silently reinterpreted and a prediction scored on the wrong channel. The three sources agreeing are not fully independent - the TDF trackName and the supplementary filename may both derive from the same internal file naming - and elink gds->sra, the one genuinely orthogonal check, remains unrun.

Connected to

tests →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
scores prediction 5s EVICTION clause, previously declared unreachable; it confirms, and Q125 had already scored the silencing clause as attenuated
supports →A cell that cannot deposit H2AK119ub is still evicted normally by a menin inhibitor and silences its vacated sites less - so a fusion-occupancy pharmacodynamic marker reads FULL ENGAGEMENT in exactly the genotype that responds worst, and this is now the fourth independent reason this corpus distrusts its own candidate markers
measures the half that entry had assumed - the knockout is evicted to within 1.3% of control, so an occupancy PD assay scores it a complete responder
supports →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
explains the 3.17 and 2.84 aggregate anomaly that entry excluded as inadmissible: those pairs are a FLAG ChIP over an H2AK119ub ChIP inside top-FLAG bins
related to →PCGF1 — the PRC1.1 scaffold whose LOSS causes menin-inhibitor resistance, so it is a restraint on the leukaemia rather than a dependency of it
the PCGF1 knockout is evicted normally and silenced less
related to →KMT2A / MLL1 — the gene that breaks
FLAG occupancy of the KMT2A fusion is the measured channel
← refutesMeasured 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
corrects that entry's use of SGC0946 as a noise ruler: the real floor from two input pairs is 4-10%, so SGC's 22.8% is a finding and not variation. Its eviction conclusion is unaffected and better founded
← refutesAuditing 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
that entry quoted the paired convention's 1.4450 without naming it; amended to carry both, conclusion unaffected because both conventions fall
← supportsEnumerating 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
that beat established GSM8123407-412 are FLAG ChIPs mislabelled H2AK119ub rep1, which is why REP1 is excluded here rather than scored as a third convention
← related toAnswering 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
its finding that six H2AK119ub-titled files are FLAG is a fact about those files; applied blanket to the string rep1 it hid H3K79me2's second channel
← supportsThe 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
an independent title-versus-filename sweep recovers exactly the same six mislabelled FLAG files, GSM8123407-412, no more and no fewer
← supportsIn 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