Oncofusion

start / graph / ana-king-verification-of-q154-the-class-survives-every-stratification-and-irx1-individually-does-not

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

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

also called q154 verified · 64 vs 8.05 holds · IRX1 individually is not significant · the class is real, the gene is not

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.

The same thing, for a scientist

Cycle 143 committed Q154's and Q155's output explicitly labelled **NOT KING-VERIFIED**. This entry discharges that for Q154's two load-bearing findings. **Both are verified, and they point in opposite directions.** 🟢 **The AFF1-exclusive class is real.** Using Janssens *et al.*'s own published per-sample target calls (Supplementary Data 2, 1,091 genes x 26 non-drug-exposed samples), **64 genes are called in >=3 of 8 KMT2A::AFF1 samples and 0 of 18 non-AFF1 samples.** Re-run from the agent's script, every figure reproduced to the digit, and **re-run under a different seed (777001) every conclusion held** [E1]: | null | agent (seed 20260924) | king (seed 777001) | verdict | |---|---|---|---| | unstratified | 2.32, p 0.00010 | 2.35, p 0.00012 | holds | | call-rate matched +-15% | 12.12, p 0.00326 | 11.99, p 0.00216 | holds | | **institution-stratified EXACT** | **8.05, p 0.00272** | **8.05, p 0.00272** | **identical - exact enumeration of all 4,410 assignments** | | B-lineage only (16 samples) | 23.45, p 0.00366 | 23.41, p 0.00351 | holds | | patient-level (24 patients) | 12.67, p 0.00029 | 12.68, p 0.00043 | holds | The obvious artefact is controlled: the AFF1 arm has a **21.3%** call rate against **9.6%** in the non-AFF1 arm, which mechanically favours exclusivity, and **restricting to draws whose arm call rate is within 15% of the real one raises the null from 2.3 to 12.1 but leaves p at 0.003.** 🔴 **IRX1 individually does not survive.** The same permutation on IRX1's single row gives **p = 0.0219** unstratified and **0.1190** (institution, exact), **0.1756** (B-lineage), **0.0772** (patient-level). **None clears 0.05** [E1]. IRX1 is called in exactly `p279`, `p318`, `p454` - which I confirmed directly against the authors' table, not via the agent's code [E2]. 🔵 **The primary source agrees on both genes.** Janssens' Supplementary Data 2 lists `IRX1 -> ['p279','p318','p454']` and `IRX2 -> ['p454']`, read straight from the parsed table. Q154's independent gene-body measurement puts IRX2 at **5.145x and 3.575x** in p454's two KMT2A-N libraries - **the same single sample the authors call** [E2]. 🟠 **The tension with cycle 141 is the most interesting thing here and is not resolved.** My own promoter-scale measurement found **332 AFF1-only promoters against a null of 426.4 - BELOW chance.** The authors' own call table gives **64 AFF1-exclusive genes against a null of 2.3-8.05 - far ABOVE chance.** Same deposit, same samples, opposite verdicts, because the instruments differ: a 2x within-sample share gate over 19,404 promoters versus SEACR peaks plus a Bayesian N/C model over a curated 1,091-gene list. **Which instrument is right is now the open question**, and it is not settled by this verification [E3].

Go deeper
What verification did and did not cover

Covered: the permutation arithmetic, its seed-dependence, and the two gene lists against the primary table.

🟠 Not covered: the construction of C:/Temp/q154gb/authors_targets.json from the published .xlsx. I verified that the parsed table contains 1,091 genes and that IRX1 and IRX2 carry the sample sets the entries claim, but I did not re-parse the spreadsheet from the original file. If the parse dropped or mis-assigned samples, every number above would be internally consistent and externally wrong - which is exactly the failure mode this corpus recorded when an off-by-one parser confirmed itself. That re-parse is the single highest value remaining check on Q154.

🔴 Also not covered: Q155's eleven entries. They remain not king-verified, and their hypothesis entry has now been refuted on its central claim by a separate analysis this cycle - which is a reason to verify the rest of them rather than to assume they are wrong.

🔵 One thing the seed test genuinely bought: the institution-stratified null is an exact enumeration of all 4,410 valid assignments, so it returned bit-identical values under a different seed. That is the strongest of the five nulls and the one the headline should quote.

Facts and figures
still unverifiedall eleven Q155 entries; the .xlsx -> authors_targets.json parse
seed original20260924
seed king777001
class observed64
unresolved tensionpromoter-scale 332 vs null 426.4 (below) against authors-table 64 vs null 8.05 (above)
nulls king seed
unstratified2.350unstratified: 2.35institution_exact8.050institution_exact: 8.05call_rate_matched12.0call_rate_matched: 11.99patient_level12.7patient_level: 12.68b_lineage23.4b_lineage: 23.41
p upper king seed
unstratified0.000unstratified: 0.00012patient_level0.000patient_level: 0.00043call_rate_matched0.002call_rate_matched: 0.00216institution_exact0.003institution_exact: 0.00272b_lineage0.004b_lineage: 0.00351
irx1 p king seed
unstratified0.022unstratified: 0.0219patient_level0.075patient_level: 0.0753institution_exact0.119institution_exact: 0.119b_lineage0.176b_lineage: 0.1756
verifies
[
  "finding-q154-in-the-authors-own-call-table-the-aff1-exclusive-class-is-far-above-its-permutation-null",
  "finding-q154-irx1s-own-three-sample-pattern-does-not-clear-0-05-once-institution-or-lineage-is-held-fixed",
  "finding-q154-irx2-is-fusion-occupied-at-gene-body-scale-in-p454-and-the-promoter-scale-zero-was-a-knife-edge"
]
nulls king seed — the numbers behind the chart
{
  "unstratified": 2.35,
  "call_rate_matched": 11.99,
  "institution_exact": 8.05,
  "b_lineage": 23.41,
  "patient_level": 12.68
}
p upper king seed — the numbers behind the chart
{
  "unstratified": 0.00012,
  "call_rate_matched": 0.00216,
  "institution_exact": 0.00272,
  "b_lineage": 0.00351,
  "patient_level": 0.00043
}
irx1 p king seed — the numbers behind the chart
{
  "unstratified": 0.0219,
  "institution_exact": 0.119,
  "b_lineage": 0.1756,
  "patient_level": 0.0753
}
authors table spotcheck
{
  "IRX1": [
    "p279",
    "p318",
    "p454"
  ],
  "IRX2": [
    "p454"
  ],
  "EPHA7": [
    "p279"
  ],
  "AFF1_n": 10,
  "MEIS1_n": 22,
  "HOXA9_n": 19,
  "PROM1_n": 17
}
arm call rates PCT
{
  "AFF1": 21.3,
  "non_AFF1": 9.6
}
measurements
[
  {
    "quantity": "AFF1-exclusive target genes in the authors' own call table, against the strictest null",
    "value": 64,
    "unit": "genes",
    "subject": "called in >=3 of 8 KMT2A::AFF1 samples and 0 of 18 non-AFF1, institution-stratified exact null 8.05, p_upper 0.00272",
    "universe": "1,091 genes x 26 non-drug-exposed GSE252378 samples; verified at two seeds, exact null bit-identical"
  },
  {
    "quantity": "IRX1's individual permutation p once patients, institution or lineage are held fixed",
    "value": 0.119,
    "unit": "p (institution-stratified, exact)",
    "subject": "IRX1 called in exactly p279, p318, p454; 0.0219 unstratified, 0.0753 patient-level, 0.1756 B-lineage",
    "universe": "none of the four clears 0.05; the class IRX1 belongs to does, at p 0.0027"
  }
]
Still open
  1. Re-parse Supplementary Data 2 from the .xlsx independently of the agent's parser. Every number verified here rests on that one file, and a parse error would be invisible to a re-run.
  2. Adjudicate the promoter-scale versus authors-table disagreement. The cleanest test is to restrict this project's promoter rule to the authors' 1,091-gene universe and re-run its null there - if the sign flips, the universe is the explanation.
  3. Verify Q155's eleven entries. One of its hypotheses was refuted this cycle by a separate analysis, which raises rather than lowers the priority of checking the rest.
  4. Decide what the IRX1 result means for the entries that rest on it. Several entries treat IRX1 as a partner-specific address; on this evidence the defensible claim is about the class, and IRX1 is one member of it.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Re-running analyses/q154-gene-scale-occupancy-null/authors_null2.py reproduces every figure in Q154's two findings exactly, and re-running it under seed 777001 instead of 20260924 leaves every conclusion unchanged: the AFF1-exclusive class is 64 observed against nulls of 2.35 (p 0.00012), 11.99 call-rate-matched (p 0.00216), 8.05 institution-stratified exact (p 0.00272, bit-identical because it enumerates all 4,410 assignments), 23.41 B-lineage (p 0.00351) and 12.68 patient-level (p 0.00043); IRX1 individually gives 0.0219, 0.1190, 0.1756 and 0.0753.

    PMID 39472576DOI
    What this evidence cannot show

    This verifies the statistics, not the input. The permutation operates on a parsed copy of the authors' spreadsheet at C:/Temp/q154gb/authors_targets.json which I did not re-derive from the .xlsx; a parsing error would be invisible to this check. The Monte Carlo nulls differ in the third decimal between seeds, as their MCSE (0.016-0.24) predicts, so 'holds' means the conclusion is unchanged rather than the numbers are identical.

    Evidence that argues against this (2)
    • Re-running an agent's own script with a different seed tests seed-dependence and nothing else. It shares every assumption in the code - the arm definitions, the drug-exposed exclusions, the k>=3 threshold - so agreement here is agreement with the agent's framing, not independent confirmation of it. A genuinely independent check would re-implement the statistic from the spreadsheet.
    • The k>=3 threshold was not pre-registered in a form I checked. The same table also reports k>=1 results (416 genes), and if the threshold were chosen after seeing the data the p-values would need a selection correction that none of these nulls applies.
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Read directly from the parsed authors' target table rather than through the agent's analysis code, IRX1's called-sample set is exactly ['p279','p318','p454'] and IRX2's is exactly ['p454'], matching the claims in Q154's entries; for comparison MEIS1 is called in 22 samples, HOXA9 in 19, PROM1 in 17, AFF1 itself in 10 and EPHA7 in 1 (p279).

    PMID 39472576DOI
    What this evidence cannot show

    Read from the parsed JSON, not from the .xlsx itself, so this confirms the agent's analysis is faithful to its input without confirming the input is faithful to the publication. EPHA7 appearing in exactly one sample is noted here because it is a carried open item in this project, not because it was verified further this cycle.

    Evidence that argues against this (1)
    • A gene being called in 3 of 26 samples is a small observation however the null is computed, and IRX2 at 1 of 26 is smaller still. The authors' caller has its own sensitivity threshold, so 'called in p454 only' may reflect detection power rather than biology - the same knife-edge argument Q154 makes against this project's own promoter gate applies to the authors' gate too.
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.journal-article

    Two measurements on the same deposit disagree in direction about AFF1-exclusive targeting: this project's promoter-scale rule found 332 AFF1-only promoters against a label-permutation null of 426.4 (below chance, p_lower 0.1517), while the authors' own call table gives 64 AFF1-exclusive genes against a null of 2.32 unstratified and 8.05 institution-stratified exact (far above chance, p_upper 0.0027).

    PMID 39472576DOI
    What this evidence cannot show

    The two are not measuring the same object: the promoter-scale rule scores a 2 kb window at every one of 19,404 symbols by a within-sample share ratio, while the authors' table is a curated 1,091-gene list produced by SEACR peak calling plus a Bayesian N/C model at p<1e-5 over a master list of 1,692 intervals. Different universes, different callers and different exclusivity definitions, so the disagreement is not yet evidence that either is wrong. No experiment was run this cycle to adjudicate.

    Evidence that argues against this (2)
    • The most likely resolution is unflattering to the loud result rather than to the quiet one: a curated 1,091-gene list is already enriched for genes the authors found interesting, and exclusivity computed inside a pre-filtered universe can be far above a null that permutes labels within that same filtered universe. The permutation holds the matrix fixed, so it controls for the filter's effect on marginal call rates but not for the filter having selected genes on a partner-correlated criterion in the first place.
    • Conversely the promoter-scale null may be the broken one: cycle 144's audit shows 14.69% of the table's symbols have multiple non-overlapping TSSs of which only one is scored, and q153 found 43.9% of occupancy calls sit at windows that cannot be attributed. A rule with that much attribution noise would dilute a real exclusive class toward and past chance.
Connected to