Oncofusion

start / graph / ana-effect-size-vs-z-reranking

Analysisunrated confidence 0 of 5: unrated

Nothing beats menin after all, and the statistic that said otherwise ranks narrowness

also called effect size versus z · d-consistency · Fanconi permutation test

Last cycle we said one gene looked more important than menin for this leukaemia, because we divided each gene's effect by how variable it was elsewhere. Measured in plain units of effect, menin is first of 8,269 and a cluster of DNA-repair genes near the top turns out to be real.

The same thing, for a scientist

Re-scores the genome-scale sweep on the same numerator with and without the comparator-SD denominator. **z-consistency = max(z_my, z_ly); d-consistency = max(d_my, d_ly) in Chronos units.** Spearman between the two rankings is **0.960** over 8,269 genes, and only **14 of the z top thirty appear in the d top thirty** — the agreement is in the bulk and the disagreement is in the tail where every claim lives. Among the 2,199 DEPLETED genes, places gained by the z correlates with comparator SD at **rho −0.923**: in that tail the statistic is close to a pure ranking of narrowness. **On raw effect MEN1 is rank 1 of 8,269 (d −0.855) and FLT3 rank 2 (−0.732)**, reversing last cycle's superlative withdrawal. The FA/BRCA pathway beats a permutation null on both statistics and against both an all-gene and a leukaemia-relevant null.

Go deeper
The same numerator, with and without the denominator

Both statistics use the index-arm mean minus the other-leukaemia mean on line-normalised gene effects. The z divides by the other-arm SD; the d does not.

generank by zrank by dzdcomparator SD
FLT312−4.79−0.7320.153
MEN121−2.12−0.8550.403
PSIP146−1.75−0.4180.239
KMT2A135−1.24−0.4360.351
BRD92213−1.04−0.3290.315
MEIS13031−0.97−0.2490.256
MYB21225−0.54−0.2700.500
DOT1L33561−0.44−0.1780.401
EP30041941−0.40−0.2170.548
ACTB351657220.170.0660.380
CEBPA394966610.220.0900.405
SPI1795782620.990.5870.594

MEN1 is first of 8,269 on raw effect. Last cycle's withdrawal of "nothing beats menin" was a withdrawal of a claim about the z, and the z was the wrong scale for the comparison being made. FLT3 remains an important and underexploited dependency; it is not a larger one.

MYB, DOT1L and EP300 are the genes the z was hiding. Each has a large comparator SD — they matter somewhat in many leukaemia lines — and each moves up by hundreds of places when the denominator is removed. DOT1L at rank 61 of 8,269 by effect size, against 335 by z, is the sharpest case, and it matters because DOT1L is the one target in this list whose clinical failure the project keeps having to explain.

The bias, measured — and my first attempt to measure it had the sign wrong

Places gained by the z is rank_d − rank_z. If the z promotes narrow dependencies, that should be negatively correlated with comparator SD.

populationnSpearman
depleted (d < 0)2,199−0.923
enriched (d ≥ 0)6,070+0.921
pooled8,269+0.564

The pooled figure is an artefact and the first version of this script reported only the pooled figure. Dividing by a small SD pushes a negative d further down and a positive d further up, so the same scaling promotes depleted genes and demotes enriched ones. Pooled over a population that is 73% at or above zero, the sign inverts — and it inverts into a number that looks like a refutation of the bias. Every claim this project makes on the statistic lives in the depleted tail, and there the correlation is −0.92.

The Fanconi cluster is real

Recorded last cycle as an open question: FANCA, FANCG and FAAP100 all in the top ten of a statistic run 8,248 times, so "a coherent cluster, or the expected tail". Twenty-six FA/BRCA pathway genes were fixed as a set before any result was read, and all 26 scored.

nullstatisticmean rankp (one-sided, 20,000 permutations)
all 8,269 scoredz1852.55e-05
all 8,269 scoredd2074.65e-05
leukaemia-relevant (1,579)z1829.85e-05
leukaemia-relevant (1,579)d2106.31e-04

p = 5e-05 is the floor of 20,000 permutations, so the correct reading is "no random gene set of this size did as well". It survives the harder null: 21 of the 26 members are themselves depleted in the comparator arm, and the set still beats 1,579 genes that are.

By raw effect the core complex clusters tightly: FANCF 11, FANCI 15, FANCA 20, BRIP1 22, FANCL 26, FANCG 27, UBE2T 33, FAAP100 36 of 8,269. Eight members of one pathway inside the top forty is not a tail.

What this does and does not license

It licenses saying MEN1 is the most extreme KMT2A-selective dependency in 8,269 genes on effect size, and it requires saying which scale. It does not license a Fanconi therapeutic claim: the permutation treats complex members as independent draws and they are not, physically interacting subunits have correlated gene effects, and the test is therefore anti-conservative for any complex. What it does license is reading the FA result as a hypothesis with a clear next experiment — whether KMT2A-rearranged lines carry replication stress that makes fork protection rate-limiting, which would be measurable directly rather than inferred from a dependency ranking.

Every source behind this entry (4)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    Re-scoring the identical genome-scale sweep with and without the comparator-SD denominator gives Spearman 0.960 between the two rankings over 8,269 genes and only 14/30 overlap in the top thirty. On raw effect (d-consistency, Chronos units) MEN1 is rank 1 at −0.855 and FLT3 rank 2 at −0.732, reversing the z-based ordering in which FLT3 is first at −4.79 against MEN1's −2.12.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20 on the identical sweep to ana-lineage-consistent-genome-scale: same seed 20260921, same panel draw, same line offsets, index arm resolved from Cellosaurus by DepMap accession. 8,269 genes scored (the 8,248 of the original plus the candidate panel and the Fanconi members that the random draw missed). Neither statistic removes the lineage confound: both share the same index arm, the same four-line lymphoid half and the same DepMap disease-string lineage labels, so a disagreement between them is about scaling and nothing else.

    Evidence that argues against this (2)
    • The d statistic has the opposite bias to the z: it ignores how noisy the comparator arm is, so a gene with a large but unstable difference outranks a small reliable one. Neither ordering is correct for every purpose — the z answers "how unusual is this line's dependency" and the d answers "how much more does this line need the gene", and this project has been asking the second question while computing the first. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
    • A Spearman of 0.960 means the two rankings agree almost everywhere, so anyone reading this as "the scan was wrong" is overreading it. The disagreement is confined to the tail, which happens to be the only part of the ranking the project ever quotes. Oncofusion knowledge graph (internal analysis) 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Among the 2,199 genes depleted in the index arm, places gained by the z correlates with comparator SD at Spearman −0.923: in the tail where every claim lives, the z statistic is close to a pure ranking of narrowness. Among the 6,070 enriched genes the correlation is +0.921 and pooled it is +0.564, so the pooled figure has the opposite sign to the effect it is meant to measure.

    What this evidence cannot show

    The first version of this script computed only the pooled correlation and reported +0.564, which reads as evidence AGAINST the bias. The sign inversion is arithmetic, not statistical: scaling by a small SD moves negative and positive values in opposite directions, and 73% of scored genes sit at or above zero.

    Evidence that argues against this (2)
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    The FA/BRCA pathway, 26 genes fixed as a set before any result was read and all 26 scored, beats a 20,000-fold permutation null on both statistics (mean rank 1852.5 by z and 2074.6 by d against a null median of 4135, p = 5e-05 for both) and still beats a harder null restricted to the 1,579 genes depleted in the comparator arm (p = 5e-05 by z, 1e-04 by d). By raw effect FANCF, FANCI, FANCA, BRIP1, FANCL, FANCG, UBE2T and FAAP100 occupy ranks 11 to 36 of 8,269.

    What this evidence cannot show

    p = 5e-05 is the floor of 20,000 permutations and should be read as "no random set of this size did as well", not as a point estimate. FA/BRCA membership is one of several defensible pathway lists.

    Evidence that argues against this (2)
    • The permutation treats gene sets as independent draws and the members of a physical complex are not: subunits of one machine have correlated gene effects across every cell line, so a set drawn from one complex has far less effective independence than its size suggests. The test is anti-conservative for any complex, and eight core-complex members inside the top forty is exactly the configuration where that matters most. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
    • A DNA-repair dependency in KMT2A-rearranged lines could reflect the lines rather than the fusion: immortalised leukaemia lines carry accumulated replication stress from long culture, and the index arm's eleven lines have been passaged for decades. A patient-sample or primary-xenograft test would be needed before this is a statement about the disease. Oncofusion knowledge graph (internal analysis) 2026
  4. E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    MYB, DOT1L and EP300 move up by hundreds of places when the comparator-SD denominator is removed — MYB 212 to 25, EP300 419 to 41, DOT1L 335 to 61 of 8,269 — because each has a large comparator SD (0.500, 0.548, 0.401), meaning each matters somewhat in many other leukaemia lines rather than only in the index arm.

    What this evidence cannot show

    These are graded rather than narrow dependencies by construction of the statistic; the movement is the point and the absolute effect sizes remain modest (d −0.270, −0.217, −0.178).

    Evidence that argues against this (2)
    • A large comparator SD is exactly what a pan-leukaemia dependency with variable penetrance looks like, so promotion on the d scale is not evidence of KMT2A-selectivity — it may be evidence of the opposite. The d statistic cannot distinguish "needed more here" from "needed variably everywhere and a bit more here". Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
    • For DOT1L specifically the clinical answer is already in and it is negative: pinometostat reduced H3K79 methylation as intended in 51 adults and produced 2 complete remissions. A better position in a dependency ranking does not revive a target whose pharmacological test has been run. Blood 2018
Connected to