The Fanconi core complex is 2nd of 719 complexes for KMT2A-selectivity - but coherence does not explain it, and the explanation this entry was built on is dead

...and the other end of the same list:

Spearman rho = −0.764 between coherence and selectivity rank across 32 genes (negative because a better rank is a smaller number). This is the finding: behavioural membership and disease relevance are the same axis here, and neither was used to compute the other.

In haematological lines the core tightens further. UBE2T rises to +0.406, the highest of any member, with FANCF +0.385 and FANCL +0.368 — the E2, a core subunit and the E3. ATR falls to −0.144 and RPA1 to −0.129.

ATR sits at selectivity rank 17,082 of 17,728 with coherence +0.048, falling to −0.144 in haematological lines. RPA1 is 17,085 with −0.054. RPA2, RPA3, FAN1, POLN and EME2 are all in the same region.

This project has now excluded the replication-stress explanation for the FA signal three times by three different methods: a direct pathway correlation (ATR −0.289, CHEK1 −0.317, both weakly negative), a genome-wide selectivity scan, and now a coherence measure. The arm of the FA pathway that handles replication stress is not merely unsupported here — it is at the bottom of a 17,728-gene ranking.

The honest qualifier is unchanged and stated in every one of those entries: all three analyses read the same DepMap matrix. The instruments differ and the data do not.

ana-usp1-is-in-neither-home tested USP1 alone and found it at roughly half the module's internal coherence. Its own open question objected that USP1 "was tested because it was the druggable one, not because it was the doubtful one" and asked for the rest.

Running all 32: USP1's coherence is +0.072, below the member median of +0.117, while its selectivity rank is 129. It is one of only two members in the top 500 selective with below-median coherence — the other is ERCC1 (+0.099, rank 430).

So the single-gene result was not a post-hoc rescue. USP1 is genuinely the outlier of the set: selective without being coherent. That makes it a lead in its own right and not evidence for the module, which is exactly what the earlier entry concluded before this test existed to check it.

This entry's own open question was the one that could kill it: does coherence predict selectivity in gene sets OTHER than FA? If it does, the rho of −0.764 is a fact about DepMap and not about Fanconi biology. The identical pipeline, on three other sets:

The finding survives. Two real pathways give essentially nothing and NOTCH gives the wrong sign. FA's relationship is roughly six times NOTCH's and thirteen times cholesterol's, in the opposite direction from both.

The random set is the honest complication. At −0.239 it is not zero — there is a baseline tendency for coherence and selectivity to track, almost certainly the variance artefact recorded in this entry's counter-evidence: a gene whose effect barely moves across lines can neither correlate with anything nor show selectivity, so low-variance genes sink on both axes together. FA's −0.764 is about three times that baseline, which is the right way to read it — as an excess over a non-zero floor, not as a relationship where none would otherwise exist.

Worth noting separately: FA's median coherence of 0.1165 is 2–40× the controls' (0.0030, 0.0575, 0.0026). The Fanconi genes are a genuinely coherent unit in this data and the control pathways are not, which is a fact about the pathway independent of any selectivity.

The first round of controls was too easy and this entry said so: NOTCH and cholesterol biosynthesis have median coherence 0.003 and 0.058 — they were never coherent to begin with, so showing they predict nothing proves little. The open question asked for a tightly coherent complex, where coherence is high and the question is genuinely live.

Proteasome assembly is the control that matters. Its median coherence is 0.1029, essentially the same as Fanconi's 0.1165 — a gene set just as tightly co-dependent. And its rho is +0.4280: the opposite sign.

So the relationship is not coherent sets predict KMT2A selectivity. A set matched on coherence gives the reverse. Whatever produces Fanconi's −0.764, it is not a general property of coherent gene sets in DepMap, and the low-variance artefact cannot explain a control that is equally coherent and points the other way.

The splicing control lands between them (coherence 0.0415, rho +0.1818), which is the expected shape for a large heterogeneous pathway rather than a complex.

One control failed to run: the 80S ribosome returned 0 genes carrying a selectivity rank, so ribosomal proteins are either absent from or differently named in the cached scan. That is a gap, not a result, and it is the control most likely to behave like the proteasome.

The previous round rested on one random draw at −0.239, and this entry said so. Five hundred draws of 32 genes, from a pool of 260 random genes fetched once and correlated over 1046 shared lines:

0 of 500 draws reach −0.764. It lies beyond the minimum of the entire null distribution (-0.659). Empirical p < 1/500.

Two things this corrects about the earlier reading.

The baseline is real and larger than I guessed. The null median is -0.140, not zero — an arbitrary 32-gene set shows a mildly negative coherence-selectivity relationship, almost certainly the low-variance artefact. The earlier caveat was right to flag a floor.

And my single draw was unremarkable. −0.239 sat close to the null median of -0.140, well inside the bulk. Reading it as 'a baseline of about a third of FA's' was over-reading one sample from a distribution running -0.659 to +0.628. The null is wide, which is precisely why one draw could not settle it.

The ribosome control failed the first time by returning zero ranked genes. That was diagnosed as a Reactome participant-parsing failure rather than a data gap — 86 RPL/RPS symbols carry a selectivity rank** — and re-run from an explicit symbol list it works.

The ribosome lands on the null median. Its rho of -0.147 is indistinguishable from the random-set median of -0.140, despite a coherence of 0.0881 that sits between the proteasome's and Fanconi's. A large, tightly co-dependent, pan-essential complex behaves on this axis exactly like an arbitrary gene set.

So the three fair controls now bracket the question from three directions — one opposite in sign, one at the null, one mildly positive — and none comes within 0.3 of Fanconi's −0.764, which itself lies beyond the minimum of 500 random draws.

The cost of running it this way is in the caveat: requiring all 81 ribosomal proteins to share cell lines collapses the intersection from ~1,046 lines to 221, so this control is computed on a fifth of the data the others used.

This entry's first open question asked for a sweep over many pathways rather than three chosen after seeing the result. Fourteen Reactome pathways spanning distinct biology, all of comparable size, gene lists deduplicated and fetched once:

Fanconi ranks 1 of 14. That is the good news and it is real.

The bad news is second place. The citric acid cycle comes in at −0.657, within 0.11 of Fanconi, at a median coherence of 0.086 not far below Fanconi's 0.117. A mitochondrial metabolic pathway with no obvious connection to KMT2A biology behaves almost as Fanconi does on this measure.

My three hand-picked controls missed this entirely. Proteasome +0.428, ribosome −0.147, splicing +0.182 — all near or above zero, which made Fanconi look isolated. Sampling fourteen pathways instead of three found one within a tenth of it. That is exactly what a systematic sweep is for, and it cuts against the confidence the earlier controls supported.

So the honest statement is now: Fanconi is the most extreme of fourteen pathways tested, and it is the top of a distribution rather than an isolated outlier. The permutation result stands — 0 of 500 random 32-gene sets reach −0.764 — but a random gene set is not the right comparator if real coherent pathways routinely reach −0.6.

The sweep raised the possibility that the whole relationship is a coherence artefact: the two most coherent sets were also the two most negative, which is the low-variance explanation this entry has carried as counter-evidence from the start. Pooling all 18 gene sets measured across every run and correlating median coherence against rho settles it.

Essentially zero without Fanconi. The apparent coherence-drives-rho pattern was Fanconi itself, and one point cannot establish a trend it is the sole support for.

The table makes it concrete. Cytosolic ribosomal proteins at coherence 0.088 give −0.147. Proteasome assembly at 0.103 gives +0.428. Two of the three most coherent sets in the whole comparison, similar coherence, rho differing by 0.57 and opposite in sign. Coherence is not the driver, and the low-variance explanation can be retired.

What that does and does not restore. It removes the artefact that would have explained Fanconi away, so Fanconi's −0.764 is not a by-product of being tightly co-dependent. It does not restore the earlier 'no control comes within 0.3' framing, because the citric acid cycle at −0.657 is still unexplained — and now it cannot be dismissed as a coherence effect either, since its coherence (0.086) is indistinguishable from the ribosome's (0.088) which gives −0.147.

So after five rounds of controls the position is: Fanconi ranks 1 of 18, its rank is not explained by coherence, and exactly one other pathway comes close for reasons nobody has established.

The remaining puzzle was why a mitochondrial metabolic pathway reaches −0.657. Reading its 34 members' selectivity ranks answers it.

Least selective: SUCLA2 16,524 · LYRM4 16,587 · NFS1 16,602 · SUCLG1 16,927 · IDH3A 17,644 · FXN 17,706.

The pathway as a whole is not selective at all — median rank 8,164 against the 8,864 a uniform set would give, with only 3 of 34 in the top 500. What is selective is a sub-complex: seven of the twelve most selective members are succinate dehydrogenase — SDHA, SDHB, SDHC, SDHD and the assembly factors SDHAF2, SDHAF3, SDHAF4. Complex II is a genuine multi-subunit enzyme, so its internal coherence is high for the same reason the FA core complex's is.

That is structurally the same shape as Fanconi, where the core complex plus ID2 ubiquitination module is coherent and selective while the ATR/RPA arm is neither. So the statistic is not measuring is this pathway selective. It is measuring does this pathway contain a coherent sub-complex that is also selective — and Reactome pathways are mixtures of complexes, so several will.

This makes the Fanconi result more interpretable and less unique at once. Fanconi ranks 1 of 18 because its selective sub-module is the most extreme, not because it is the only pathway with one. The right follow-up is to score complexes rather than pathways, where the question 'is this unit selectively required' is the one the statistic actually answers.

This entry's first open question read: "Score COMPLEXES rather than pathways. Both the Fanconi and TCA results reduce to a coherent sub-complex inside an unselective pathway, so the pathway is the wrong unit."

Run on 2026-09-20 over 719 Complex Portal human complexes, 1,901 genes with DepMap profiles. The analysis is ana-the-complex-level-sweep-kills-the-coherence-mechanism.

Between-complex Spearman of median coherence against median selectivity rank: -0.016. The statistic the pathway analysis is built on does not exist at the unit this entry nominated as the right one.

And SDH is much less special than the previous section made it. It is 38th of 719 - selective, since the median complex sits at 8,175 - but respiratory complexes I, III and IV are also above the median. The electron transport chain trends selective as a whole and SDH is the strongest of four, not a distinct phenomenon.

The finding survives the mechanism. The Fanconi anaemia ubiquitin ligase complex is 2nd of 719, median rank 125 against a uniform 8,864, on nine members. Every mechanistic account this graph has attached to that observation - replication stress, low-variance artefact, coherence, sub-complex structure - has now been tested and failed. What is left is a well-measured dependency with no explanation, which is a more honest place to stand than where this entry started.