Oncofusion

start / graph / concept-q146-a-quantile-calling-rule-and-an-absolute-one-answer-different-questions-and-can-order-the-same-pairs-oppositely

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q146-one-universe-jaccard

A quantile calling rule and an absolute one answer different questions, and on the same data they can order the same pairs oppositely

Deciding which genes count as 'targets' by taking the top ten percent is a different decision from using a fixed cutoff, and the two can rank the same pair of samples in opposite orders. So a number like an overlap score means nothing without saying which decision made it.

The same thing, for a scientist

A quantile rule ('the top decile of the N/C ratio among high-N promoters') produces a call set of fixed relative size in every sample - 5% of the universe - and therefore a constant chance floor and a statistic that measures agreement among the STRONGEST calls. An absolute rule ('N/C > 1.75 and binomial p < 1e-5') produces a set whose size tracks both the biology and the library depth - 0 to 44% of the universe across GSE252378 - and therefore measures agreement among ALL calls, marginal ones included. Measured here on one universe with one depth-matching procedure and one null, the two rules place the same same-patient pair at the 37th percentile of comparable pairs (absolute) and the 74th (quantile). Both are correct answers to their own question. The corpus recorded them as one question with two answers.

Go deeper
Why they diverge

The quantile rule cannot express 'this sample has fewer oncoprotein targets', which is the paper's central finding - it calls 5% of promoters in a sample whose oncoprotein is undetectable just as readily as in the deepest library. The absolute rule can, which is why it returns 0 for the revumenib-exposed AML and 4,347 for p279.

But the absolute rule's set sizes are so large (35-44% of the pair universe) that its chance Jaccard floor is already 0.25, leaving almost no dynamic range: observed 0.33 over a floor of 0.25 is an excess of 1.33x. The quantile rule's floor is 0.05 and observed 0.20 is 3.7x. A rule that calls 40% of the universe compresses every overlap statistic toward its floor, and differences between pairs then live in the third decimal place.

So the honest reading is: the absolute rule has the better semantics and the worse dynamic range; the quantile rule the reverse.

What to do
  • Report the rule with the number, always, and the fraction of the universe it calls.
  • When a rule calls more than a quarter of the universe, say so beside the overlap statistic; the statistic is then near its ceiling-of-compression and small differences are not interpretable.
  • Do not compare an overlap computed under one rule to an overlap computed under the other, and do not average them.
  • If the two rules disagree on direction, that is a finding about the data's structure and should be reported, not resolved by preferring one.
How this could bite us back
  • A target list for a menin inhibitor built with a quantile rule will always have the same length whether or not the patient has any oncoprotein left, which is precisely the wrong behaviour for a pharmacodynamic readout.
Still open
  1. Tune an absolute threshold until it calls 5% of the universe. If the ordering then matches the quantile rule, the divergence is about set size and not about threshold type, and the corpus can stop treating the two rules as rival semantics.
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The absolute rule (N/C ratio > 1.75, binomial upper-tail p < 1e-5) is the publication's own site-calling criterion for the oncoprotein score.

    MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024
    PMID 39472576DOI
    What this evidence cannot show

    The published version is applied to the authors' own peak intervals and calibrated on a CD34+ oncoprotein-negative sample not deposited in GSE252378, so its call counts are not the ones reproduced here.

    Evidence that argues against this (1)
    • Because the published rule includes a fusion-negative calibrator and the re-implementations do not, part of the large call-set sizes seen here may be calibration failure rather than a property of absolute rules in general. Janssens DH et al. 2024
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Measured on one universe with one null, the two rules place the p279/p318 pair at the 37th and the 74th percentile of 27 comparable KMT2A::AFF1 pairs respectively, in all three universes tested.

    What this evidence cannot show

    One deposit, one pair of rules, one feature set; the generality of a rule-driven direction flip is not established by a single case.

    Evidence that argues against this (1)
    • The two rules also differ in set size (5% versus 40% of the universe), so the flip could be a set-size effect rather than a threshold-type effect - a confound this beat names and does not resolve. Janssens DH et al. 2024
Connected to