Oncofusion

start / graph / ana-the-disjointness-is-the-base-rate-a-matched-permutation-refuted-this-beats-own-hypothesis-the-hour-it-was-written

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q138-kmt2a-aff1-surfaceome

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

also called the matched falsifier · the base rate explains the zero

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.

The same thing, for a scientist

The surfaceome screen found 0 of 15 fusion-enriched surface antigens to be parenchymally clean, and this beat wrote a hypothesis proposing that fusion-driven antigens are *inherently* normal-tissue-exposed because the fusion's output is a re-used developmental programme. The hypothesis carried a pre-written matched-comparison falsifier, and it was run the same hour. For each of the 15 fusion-enriched genes, all non-fusion-enriched genes in the same 169-gene index-expressed surface pool within +/-20% of its B-arm nTPM were taken as expression-matched controls. **14 of 15 had at least one match; IFITM2 at 1,603 nTPM had none and is reported as dropped, not silently excluded** [E1]. | quantity | value | |---|---:| | median parenchymal maximum, fusion-enriched | **97.4 nTPM** | | median parenchymal maximum, expression-matched controls | **106.9 nTPM** | | ratio | **0.911** | | permutation p, 20,000 draws, seed 20260923 | **0.789** | **The hypothesis is refuted.** Fusion-enriched antigens are, if anything, marginally *cleaner* than expression-matched controls, and the difference is indistinguishable from noise. **What actually explains the screen's zero** is the base rate: only **3 of 169** index-expressed surface proteins are parenchymally clean - **1.8%** - which by itself predicts 0.27 clean genes among 15. Observing zero is the expected result [E1, E2].

Go deeper
What this does and does not change about the screen

It does not change the screen's result. 0 of 169 surface, index-expressed genes are both fusion-enriched and normal-tissue clean; that is a measurement and it stands. CSPG4 at 124.5 nTPM in blood vessel and LAMP5 at 297.8 in basal ganglia are unchanged, and they remain concrete, specific objections to two specific published antigens.

It changes the explanation, and the explanation is what a reader would act on. The finding is not 'fusion-driven antigens are a bad class'. It is that the normal-tissue axis is brutal for almost every surface protein a leukaemia cell expresses - 98.2% of them fail it - and that fusion-driven antigens are ordinary members of that population. A group choosing between a fusion-driven antigen and a lineage antigen gets no normal-tissue advantage either way, and should choose on escape resistance instead, where fusion-driven antigens do have a genuine argument.

Stated as a rule this project can reuse: an empty intersection of two rare properties is not evidence that they are anticorrelated.

Why this is recorded as its own entry rather than a caveat

This project has repeatedly found that a proportion reported without its null becomes a claim, and that the null is supplied - if at all - by somebody other than the beat that reported the proportion. Here the beat supplied its own null, the null killed the beat's own hypothesis within the hour, and the hypothesis entry carries the refutation inside its falsifier field rather than being deleted.

The deletion would have been tidier and worse: a reader would not learn that the tempting mechanistic story was tested and failed, and the next beat would propose it again 2.

Facts and figures
scriptanalyses/kmt2a-aff1-surface-target-screen/falsifier_matched.py
statusrun
falsifier scoredhyp-no-surface-antigen-is-both-fusion-driven-and-normal-tissue-clean-in-kmt2a-aff1
verdictREFUTED
observed median parenchymal max97.4
matched median parenchymal max106.9
ratio0.911
permutation p0.78851
n permutations20000
seed20260923
n fusion enriched15
n with a match14
base rate parenchymally clean3/169 = 1.8%
result summaryThe screen's empty intersection is the base rate; fusion-driven antigens are ordinary members of a population in which 98.2% of surface proteins fail a normal-tissue filter.
dropped
[
  {
    "gene": "IFITM2",
    "b_arm_mean_nTPM": 1603,
    "reason": "no expression-matched control in the pool"
  }
]
How this could bite us back
  • This entry refutes a hypothesis written by the same beat an hour earlier, which is the right outcome but also means the hypothesis was written without checking a null that took four minutes to compute. The lesson is to run the null before writing the hypothesis, not after.
  • A refutation at n=14 could itself be wrong, and a future beat quoting 'fusion-driven antigens are not dirtier' as established would be over-reading a null from a small sample.
  • If the normal-tissue filter rejects 98.2% of everything, it may be the filter and not the biology that is being measured - in which case both this entry and the screen it scores are measuring a threshold.
Still open
  1. At what normal-tissue threshold does the clean fraction become large enough for the screen to be informative? A sweep from 1 to 50 nTPM would show whether 5 nTPM is a reasonable operating point or an arbitrary cliff, and it has not been run.
  2. Is the 1.8% clean rate specific to leukaemia-expressed surface proteins, or would any 169 surface proteins expressed in any cell type give the same? A second lineage would answer it.
  3. With n=14 this test excludes a 2-fold effect. What sample size would be needed to exclude a 1.3-fold one, and is the fusion-enriched set ever going to be large enough?
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Among 169 index-expressed surface genes, the 15 fusion-enriched ones have a median parenchymal maximum of 97.4 nTPM against 106.9 for expression-matched non-fusion-enriched controls (within +/-20% nTPM), ratio 0.911, permutation p = 0.789 over 20,000 draws with seed 20260923; 14 of 15 genes had a match and IFITM2 (1,603 nTPM) had none.

    What this evidence cannot show

    The matching is on expression level only, not on gene family, chromosomal context or transcript length. The permutation resamples within each gene's own matched set, which is the correct null for the base-rate objection but does not control for anything else. n = 14 in the observed group is small, so the test has limited power to detect a modest true difference; it excludes a 2-fold effect, not a 1.3-fold one.

    Evidence that argues against this (1)
    • A null result at n=14 is not strong evidence of no effect. The hypothesis could be true with an effect size the test cannot resolve, and the verdict is properly 'this design cannot support the hypothesis' rather than 'the hypothesis is false'. Oncofusion Knowledge Graph 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Only 3 of 169 index-expressed surface genes are parenchymally clean at the pre-registered 5 nTPM threshold - a base rate of 1.8% - which predicts 0.27 clean genes among any 15 drawn from the pool.

    What this evidence cannot show

    The 1.8% base rate is a property of the 5 nTPM threshold, which was pre-registered but is a judgement. At a 25 nTPM threshold the clean fraction would be far larger and the base-rate argument correspondingly weaker.

    Evidence that argues against this (1)
    • If the threshold is too strict - which the CD19 control failure suggests is possible - then both the zero and the base rate are artefacts of the threshold, and the whole screen measures the threshold rather than the biology. Oncofusion Knowledge Graph 2026
Connected to