The assumption the whole project rests on, tested at last: menin dependency is fusion-specific (run)

The project's therapeutic argument is that menin is druggable because the fusion depends on it. Four analyses here have correlated candidate genes against MEN1 gene effect across leukemia lines. None of them asked whether MEN1 dependency is higher in the lines that actually carry a KMT2A rearrangement — and a correlation across a mixed panel can arise with no fusion-specific effect at all.

The obstacle was annotation, not interest. The Open Targets DepMap API returns a disease label per screen (acute myeloid leukemia, b-cell acute lymphoblastic leukemia) and nothing about rearrangement status, and the DepMap bulk download is captcha-blocked, which this project documented in an earlier dataset audit.

Cellosaurus has it. Cell lines carry sequence-variation records with variation-type: "Gene fusion" and HGNC-linked gene labels, and many are cross-referenced to DepMap ACH ids. Querying var:KMT2A AND dr:DepMap returns 16 DepMap lines with an annotated KMT2A fusion — MV4-11, MOLM-13, MOLM-14, THP-1, RS4;11, SEM, SEMK2, KOPN-8, NOMO-1, Mono-Mac-6, ML-2, SHI-1, OCI-AML-2, HB11;19, ALL-PO and Karpas-45 — of which five carry KMT2A::AFF1, the fusion this project is actually about. Eleven overlap the leukemia panel with MEN1 data 1.

Gene effect is negative when a gene is required, so a more negative difference means more required in the rearranged lines. Each difference is tested both by label permutation and against a null of the same statistic computed on 918 random genes, because plenty of genes differ between any two groups of cell lines.

MEN1, FLT3, PSIP1, MEF2C, KMT2A and BCL2 are more extreme than every one of the 918 random genes. That is the headline, and it is the first direct evidence in this project, from this project's own data, that the menin dependency is a property of the rearrangement rather than of leukemia in general.

It is not the lineage contrast. The rearranged set is part B-ALL and part AML, so a raw split partly recapitulates the myeloid-versus-lymphoid difference that ana-depmap-within-subtype removed. Restricting to AML only (7 rearranged versus 25 other) the MEN1 difference is -0.890, slightly larger, at p = 0.0011.

The section above warned that MV4-11, MOLM-13 and MOLM-14 are FLT3-ITD lines and that FLT3's near-equal difference is therefore probably a passenger of the group rather than of the fusion. That caveat was worth testing rather than merely stating, so it was. Excluding those three leaves 8 rearranged lines — SEMK2, HB11;19, RS4;11, SEM, THP-1, NOMO-1, OCI-AML-2, SHI-1 — against the same 43.

MEN1 survives cleanly. It keeps 87% of its difference on eight lines at p = 0.0001, so the FLT3-ITD lines were not carrying the result.

The caveat about FLT3 was half right and should be corrected rather than left standing. FLT3's difference does fall, from -0.843 to -0.621, so the ITD lines contribute. But it stays highly significant without them, which means FLT3 is genuinely more required in KMT2A-rearranged lines independent of ITD status. Writing it off as an ITD artefact, as the section above did, was too quick.

ACTB is reassuring here. On the permutation test it is not significant either with the ITD lines (p = 0.16) or without (p = 0.19). The global-offset worry rests on its position in the random-gene null, not on the within-panel test, and it remains a reason to treat magnitudes as upper bounds rather than a reason to doubt the ranking.

DOT1L does not survive this. At p = 0.177 it is the one candidate whose group difference is not distinguishable from relabelling, consistent with it sitting lowest of the Holm survivors in the main table. DOT1L's case rests on correlation and on mechanism, not on differential dependency between these groups.

The section above proposed that the ACTB control's position at the 1.4th percentile reflected a mild global offset — rearranged lines being more dependency-sensitive across the board, inflating every difference in the table. That was a mechanism, not an observation, so it was run rather than left as a caveat (analyses/depmap-line-normalised-specificity/).

Each line's own sensitivity was estimated as its mean gene effect across the random panel only — excluding the candidates, so they cannot help set the baseline they are then measured against — and subtracted from every value before the comparison was rerun, with the null rebuilt on the normalised statistic.

There is no offset to remove. Mean per-line sensitivity is -0.1551 in the rearranged lines and -0.1515 in the others: a difference of -0.0036, which is negligible beside differences of 0.3 to 0.9.

Two conclusions, and the second is against me.

First, the main result is not an artefact of differential screen sensitivity. MEN1 keeps 99.6% of its difference. That strengthens the finding and the caveat it was carrying can be discharged: magnitudes no longer need to be read as upper bounds on this account.

Second, my explanation for the ACTB anomaly was wrong. ACTB keeps 98% of its difference and stays at the 1.3rd percentile. Whatever puts it there, it is not a global offset. The better reading is the one the percentile obscured: ACTB's permutation p is 0.169, so it is not significant on its own test, and a single gene's rank in a 1,097-gene null is a noisy statistic — roughly 14 random genes sit below it. Ranking one control at the 1.3rd percentile out of thirteen genes tested is unremarkable.

So the anomaly was largely an artefact of how I read the percentile, not of the data. The percentile column is still the right one for the headline genes, which sit beyond every random gene; it is the wrong tool for adjudicating a single borderline control.

Eleven lines. Every conclusion rests on a group of 11, and 7 in the AML-only test. That is small, and the random-gene null is what makes the comparison interpretable at all.

Cellosaurus annotation is curated from publications and may be incomplete. Any unannotated rearranged line sitting in the 'other' group weakens the contrast, so this bias is conservative — the true difference is if anything larger. But it also means the 'other' group is not cleanly fusion-negative.

Co-occurring lesions are not separated from the fusion, as FLT3 shows plainly.

Cell lines are not infant leukemia. None of these 11 is an infant-derived line, and the population this project exists to serve is infants. This is the same gap every DepMap analysis here has.

Dependency is not druggability. A gene more required in rearranged lines may still have no ligandable pocket and no therapeutic window.