The lineage confound, tested: the FA signal is in both halves, the B-lineage half is stronger, and it survives holding lineage constant
also called myeloid-enrichment artefact · lineage split FA · lymphoid-stronger
The cells used to find the Fanconi result were mostly of one blood-cell type, so the finding might have been about that cell type rather than about the KMT2A fusion. Splitting them up shows the opposite of what that worry predicted: the effect is stronger in the minority half, and it is still there when the comparison is made within a single cell type.
The same thing, for a scientist
The index arm is 7 myeloid / 4 lymphoid, and this project has a documented artefact class in which myeloid-enriched index arms manufacture apparent KMT2A selectivity for genes that are really myeloid-identity dependencies. The scoring statistic — `max(z_myeloid, z_lymphoid)` — cannot distinguish that case, because one half satisfies a max on its own. Tested directly: **both halves clear the permutation null at p = 5e-05**, and the **lymphoid half is the stronger one** (FA set mean rank 286 of 1,424 lymphoid versus 400 of 1,433 myeloid). Per gene, seven of the eight headline genes are more depleted in the four B-lineage lines than in the seven myeloid ones — FANCL -0.613 lymphoid against -0.270 myeloid, UBE2T -0.587 against -0.243, FANCF -0.591 against -0.349 — with FANCG the single exception [E1]. **This is the opposite shape from the myeloid-identity artefact class, and it is the right half for infant B-ALL.** A second, coarser signal does exist: FA genes are generally more depleted in B-ALL lines than in AML lines regardless of KMT2A status (non-KMT2A B-ALL headline-eight mean d -0.196; non-KMT2A AML +0.095). But the within-lineage contrast separates the two cleanly: KMT2A-rearranged AML against non-KMT2A AML gives -0.334, and KMT2A-rearranged B-ALL against non-KMT2A B-ALL gives -0.308 [E1]. Lineage moves the baseline; the rearrangement moves the genes on top of it.
Why a max over lineage halves cannot answer this, and what replaced it
The genome-scale sweep scores every gene as max(z_myeloid, z_lymphoid). A gene depleted only in the seven myeloid lines and completely indifferent in the four lymphoid ones receives the same score as one depleted in both. The published results file stores only the max, so the graph carried no record of which case the FA genes were.
The replacement is to score and permute the halves independently — not to look at the components of a max, but to run the whole test twice, once in each half, against its own null 1.
| arm | FA set mean rank | null median | p |
|---|---|---|---|
| both halves | 329.0 | 716.5 | 5e-05 |
| myeloid only (7 lines) | 400.0 | 716.5 | 5e-05 |
| lymphoid only (4 lines) | 286.2 | 712.0 | 5e-05 |
Ranks are out of ~1,430 in a 1,445-gene background panel and are not comparable to the 8,269-gene sweep's ranks.
Per gene, and the one exception
Chronos difference against the 43 other-leukaemia lines, offset-corrected 1.
| gene | d myeloid | d lymphoid | which half is stronger |
|---|---|---|---|
| FANCL | -0.270 | -0.613 | lymphoid, by 2.3x |
| FANCF | -0.349 | -0.591 | lymphoid |
| UBE2T | -0.243 | -0.587 | lymphoid, by 2.4x |
| FANCA | -0.300 | -0.457 | lymphoid |
| FANCI | -0.315 | -0.440 | lymphoid |
| BRIP1 | -0.297 | -0.424 | lymphoid |
| FAAP100 | -0.236 | -0.330 | lymphoid |
| FANCG | -0.343 | -0.265 | myeloid |
Seven of eight point the same way. FANCG is the exception and it is also the gene with the least pan-cancer requirement of the eight (Chronos -0.091 across all lines, 4.6% below -0.5), so it is the member most likely to be noise in either half.
For comparison, the offset control ACTB runs -0.319 myeloid and +0.066 lymphoid — a gene that looks selective in the myeloid half only, which is precisely the artefact shape the FA genes do not have.
The real lineage signal that does exist, and why it is not the explanation
Splitting by disease string, with KMT2A-rearranged lines removed from both arms, shows something genuine: FA genes are more depleted in B-ALL lines than in AML lines. Non-KMT2A B-ALL against the rest of the leukaemia panel gives FA set p = 5e-05 and headline-eight mean d -0.196; non-KMT2A AML gives p = 0.999 and +0.095. These two arms are near-complementary halves of the same 43 lines, so they are one fact stated twice, not two independent findings 1.
Read alone, that could explain the whole result: the index arm contains four B-ALL lines, B-ALL lines are FA-dependent, done. It does not survive the within-lineage test.
| contrast | n index | n comparator | headline-eight mean d | p |
|---|---|---|---|---|
| KMT2A-r AML vs non-KMT2A AML | 7 | 25 | -0.334 | 0.0001 |
| KMT2A-r B-ALL vs non-KMT2A B-ALL | 4 | 9 | -0.308 | 5e-05 |
With lineage held exactly constant, the effect is essentially undiminished from the -0.356 measured against all comers. The rearrangement is doing the work, not the lineage. That is the single most decisive number in this beat.
What this does not rule out
Lineage is assigned from DepMap free-text disease strings, not immunophenotype, and 'b-cell acute lymphoblastic leukemia' covers a wide range of biology. The B-ALL within-lineage contrast is four lines against nine, which is small enough that a single unusual line could move it. And the test controls lineage, not culture history, proliferation rate, TP53 status, or anything else that might differ systematically between KMT2A-rearranged and non-rearranged lines — see risk-fa-signal-cell-line-artefact, which is the objection that remains open.
| kind | compensation |
|---|---|
| severity | tested and not supported — the signal is lineage-wide and stronger in the minority half |
| mitigation | report both halves separately rather than a max, and always run the within-lineage contrast before calling a dependency subtype-selective |
affects
[ "ana-effect-size-vs-z-reranking", "ana-fa-pathway-lineage-and-kmt2a-specificity" ]
index composition
{
"myeloid": 7,
"lymphoid": 4
}fa set mean rank — the numbers behind the chart
{
"both": 329,
"myeloid": 400,
"lymphoid": 286.2,
"null_median": 716.5
}within lineage headline8 mean d
{
"kmt2a_r_aml_vs_other_aml": -0.334,
"kmt2a_r_ball_vs_other_ball": -0.3083
}actb control
{
"d_myeloid": -0.3192,
"d_lymphoid": 0.0656
}- A dependency that is stronger in B-lineage points the therapeutic hypothesis at infant B-ALL, the group least able to tolerate a drug that disables DNA repair and the group with the most remaining lifetime in which a second malignancy could appear.
- Why is the B-lineage half stronger? Is it cell-of-origin, proliferation rate, or an accident of four lines?
- Are FA genes generally more depleted in B-lymphoid than myeloid cells in non-malignant settings too?
- Would an infant-derived KMT2A::AFF1 model strengthen or weaken the B-lineage result?
- Does the same both-halves test rescue or kill any of the other scan hits this project carries?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
The FA/BRCA set beats its permutation null in both lineage halves independently (p = 5e-05 each), the lymphoid half is the stronger one and seven of the eight headline genes are more depleted in the four B-lineage lines, and the within-lineage contrasts give headline-eight mean d of -0.334 (KMT2A-r AML vs non-KMT2A AML) and -0.308 (KMT2A-r B-ALL vs non-KMT2A B-ALL).
Is the Fanconi signal lineage-wide, fusion-specific, and separable from pan-essentiality? Lineage-split, subtype-control and line-label permutation re-test of the FA/BRCA enrichment · Oncofusion knowledge graph, analyses/fa-pathway-lineage-and-fusion-specificity; DepMap Chronos via Open Targets 2026What this evidence cannot show
Four lymphoid index lines and nine non-KMT2A B-ALL comparators. Lineage from DepMap disease strings, not immunophenotype. None of the lines is infant-derived, which is the age group the B-lineage result would matter most for.
Evidence that argues against this (1)
- A lineage-invariant effect is consistent with a KMT2A cause and also with any other property shared by these eleven lines and not by the comparators — aggressive relapsed-disease origin, proliferation rate, TP53 status. Holding lineage constant removes one confound, not all of them. Ben-David U et al. 2018
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
The statistic being tested is max(z_myeloid, z_lymphoid), which by construction can be satisfied by a single lineage half, and the published results file stores only the max — so nothing before this analysis recorded whether the FA signal was lineage-wide.
Effect-size versus z re-ranking of the genome-scale lineage-consistency sweep, with a permutation test of the Fanconi/FA-BRCA pathway · Oncofusion knowledge graph, analyses/effect-size-vs-z-reranking; DepMap Chronos via Open Targets 2026What this evidence cannot show
The per-half values are computed inside that analysis's scoring function; they are simply not serialised to its results file, so this is a reporting gap rather than a methodological error in the sweep.
Evidence that argues against this (1)
- A max is a defensible choice for its original purpose — finding genes that matter in at least one lineage context — and criticising it as a conjunction test is criticising it for something it was not built to do. Oncofusion knowledge graph, analyses/effect-size-vs-z-reranking 2026
tested: both lineage halves, lymphoid stronger, survives holding lineage constant
tested: both lineage halves, lymphoid stronger, survives holding lineage constant