Oncofusion

start / graph / concept-q125-a-criterion-scored-where-the-control-has-nothing-to-lose-cannot-validate-a-reagent-and-its-arithmetic-will-fire-anyway

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q125-the-h2ak119ub-arm-of-the-pcgf1-knockout

Q124's criterion H2 was written to be able to void Q124's own conclusion, and it cannot - because it is scored on a contrast in which the control genotype shows no effect to lose. Its raw arithmetic FIRES at 0.0813 against a 0.05 threshold, which reads as 'the knockout is a knockout' and is not evidence of anything

also called the uninformative validator · H2 could not void ARM 2 · criterion informativeness, second use

A test meant to check whether a gene really was deleted compared the deleted cells with normal cells under a drug that, it turns out, barely changes the thing being measured in normal cells either. When neither group shows much, a small difference between them proves nothing - but the arithmetic still produces a confident-looking verdict.

The same thing, for a scientist

**Q124 wrote a specification for this beat and listed its criterion H2 above H1 *'deliberately, because H2 can void Q124's own ARM 2 outright'*.** H2 reads: `abs(R_K119ub,KO - R_K119ub,control) >= 0.05` at the FLAG-defined top-0.1%% set under VTP-50469; below 0.05, *'the clone shows the same in-fill as its control and every ARM 2 null from cycle 83 is void'* [E1]. **Measured: |R_C1 - R_NT| = %.4f, so H2 FIRES** and, read literally, the knockout is a knockout and cycle 83's nulls stand [E2]. **It is not evidence of that.** `R_K119ub,NT(VTP)` at the same set is **%.4f** - 0.0383 from unity. Q122 had already measured **0.8473** for the same quantity in a different culture of this line. **The menin inhibitor does not deposit H2AK119ub at this set in aggregate in EITHER genotype, so a small genotype difference is the expected observation whether the knockout worked or not, and H2 cannot distinguish the two hypotheses it was written to distinguish** [E2]. **This was pre-registered, not discovered.** The dispatching brief required, for every criterion, one sentence naming what result would make THAT CRITERION uninformative. For H2 (scored here as V1) the sentence read: *'V1 can only validate a knockout if the control genotype shows a PRC1.1-dependent VTP response to lose... if the control's VTP response is null, a small genotype difference is the expected result under BOTH "the knockout worked" and "the knockout failed"'*, and it was written from Q122's 0.8473 before any file for this beat was downloaded [E3].

Go deeper
The general shape, which is not about PCGF1

A validation criterion is a comparison of two groups, and it is only informative if the reference group exhibits the phenomenon the intervention is supposed to remove. The failure mode is invisible in the criterion's own text, because the text names a difference and a threshold and both are perfectly well-formed. It becomes visible only when the reference group's own effect size is printed beside the difference - which is why this beat's scoring output carries control_own_effect_R_NT_VTP inside every verdict string and labels a passing verdict VACUOUS when the control is within 0.05 of unity.

This is the sibling of the registry's existing entry on a threshold written for a one-sided quantity firing on a negative increment. Both are cases where the arithmetic is right, the verdict string is confident, and the number means the opposite of what the reader will take from it. The defence in both cases is the same: print the quantity that makes the verdict interpretable in the same table as the verdict.

And the criterion was not useless - it was pointed at the wrong contrast

H2's logic is sound. The knockout SHOULD be validated from its expected phenotype rather than from its label. What it got wrong was the choice of contrast: it picked the menin-inhibitor arm, where the control response is null, instead of the DOT1L-inhibitor arm, where Q122 had already measured a control response of 1.4927 at the same set - a 0.49 gain over unity and the largest PRC1-attributable signal in the series. That number was in a committed results.json when H2 was written.

The lesson for writing a validator: before choosing the contrast, look up the largest control-genotype effect already measured in the same channel, and score against that one. The three validations that did work here - the SGC-response blunting at the most-evicted quartile, the drug-free CpG-island localisation test, and the H3K27me3 collapse - all have a control-genotype effect far from unity.

Facts and figures
kindmethodological lesson, measured
criterion text as offeredH2: abs(R_KO - R_control) >= 0.05 at the FLAG-defined top-0.1% set under VTP-50469; below 0.05 every ARM 2 null from cycle 83 is void
the ruleBefore choosing the contrast a validation criterion is scored on, look up the largest control-genotype effect already measured in the same channel and score against that one. Print the control's own effect size inside the verdict string, and label a passing verdict VACUOUS when the control is within the indifference band of no effect.
absence classmeasured-not-analysed
absence class noteThe number that makes this criterion uninformative - Q122's R_K119ub(VTP) = 0.8473 at this exact set - was in a committed results.json in this repository when the criterion was written, and was not read. Nothing new had to be measured to know the criterion would fail; it had to be looked up. Cost: one file read.
absence class assigned2026-09-23 (Q125)
measured
{
  "R_control": 1.0383,
  "abs_difference": 0.0813,
  "threshold": 0.05,
  "raw_arithmetic_verdict": "FIRES",
  "informativeness_verdict": "UNINFORMATIVE",
  "within_genotype_replicate_range_on_the_same_quantity": 0.2623
}
How this could bite us back
  • The same defect would be far more expensive in a wet experiment: a knockout validated by a readout its control does not exhibit produces a clean-looking negative that licenses months of downstream work. Here it cost one re-read of a JSON file.
Still open
  1. How many of this corpus's 600-plus stated falsifiers are scored on a contrast whose control effect has never been measured? A sweep could compute, for every falsifier naming a threshold on a difference, whether the reference arm's own effect size appears anywhere in the entry.
  2. Should the pre-registration template require the control-arm effect size as a named field rather than as prose in an informativeness clause? It was prose here and it worked, but prose is not checkable by script.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    exp-q124-the-h2ak119ub-arm-of-the-pcgf1-knockout-is-twelve-downloads-and-scores-two-falsifiers-at-once states in data.offered_criteria: 'H2 the knockout is a knockout: abs(R_KO - R_control) >= 0.05; below 0.05 voids every ARM 2 null from cycle 83', and in bite_back: 'it is why H2 is listed above H1'. The same criterion is hyp-q124's falsifier F2.

    What this evidence cannot show

    Quoted from an entry in this repository, not from an external source. This entry does not edit that entry and proposes no change to it beyond recording the outcome; the criterion was OFFERED rather than pre-registered, and its author said so.

    Evidence that argues against this (1)
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Measured in GSE260742: R_H2AK119ub(VTP/DMSO) at the FLAG-defined top-0.1% set is 1.0383 in the non-targeting control and 0.9570 in the PCGF1-knockout clone, a difference of 0.0813 which exceeds H2's 0.05 threshold; the control's own distance from unity is 0.0383, and the two replicate conventions of the control alone span 0.2623 on the same quantity.

    What this evidence cannot show

    NO SPIKE-IN NORMALISATION IS POSSIBLE. Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of GSE260742's 158 sample records and no scale factor is deposited, so every cross-condition number here is a WITHIN-SAMPLE SHARE, which cancels the unknown per-sample scale factor exactly and IS BLIND TO A UNIFORM CHANGE BY CONSTRUCTION. Nothing here measures any mark's abundance; everything measures its distribution. EXPOSURE ASYMMETRY: DMSO 72 h and SGC0946 5 uM 72 h are matched, VTP-50469 500 nM is 48 h, so every VTP ratio is a LOWER BOUND on the 72-hour effect - identical in both genotypes, so the genotype DIFFERENCE is unaffected. REAGENTS: H2AK119ub is Ubiquityl-Histone H2A (Lys119) (D27C4) XP Rabbit mAb, NEB 8240S, and it CANNOT distinguish PRC1.1 from canonical PRC1 - RING1A/B is the ligase in both and PCGF1 is one of six PCGF paralogs - nor deposition from loss of a deubiquitinase. The H3K27me3 and H3K4me3 reagents are NOT RESOLVED and are recorded unresolved rather than guessed; `antibod` occurs zero times in the 158 GEO records. S_FLAG is defined by DYKDDDDK Tag (D6W5B) Rabbit mAb, NEB 14793S, on a RETROVIRALLY OVER-EXPRESSED tagged MLLAF9 construct, so it is the over-expressed molecule's sites and not an endogenous fusion's. THE KNOCKOUT is one CRISPR guide, one clone C1, no second clone, no rescue and NO deposited PCGF1 protein or transcript measurement. SYSTEM: murine MLL-AF9 (KMT2A::MLLT3) myeloid blood cells, NOT the infant B-lineage KMT2A::AFF1 disease this graph is about.

    Evidence that argues against this (1)
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    The uninformativeness of this criterion was stated in this beat's PREREGISTRATION.md section 5, written and timestamped before any bin file for the beat existed on disk, in a column headed 'what would make THIS CRITERION uninformative' that the dispatching brief required for every criterion.

    What this evidence cannot show

    The pre-registration is a file in this repository and its timing rests on the recorded check that C:/Temp/q125 contained no bin file when it was written, plus the git history of the commit that adds it. It is not externally timestamped.

    Evidence that argues against this (1)
Connected to