Analysismoderateking

Exactly one gene in 8,248 beats menin, and it is FLT3

also called genome-scale lineage consistency · FLT3 refutation

Last cycle we checked two thousand genes and reported that none scored better than menin for this leukaemia. Checking eight thousand finds one that does: FLT3, a gene that already has approved drugs. The claim was true of the sample and not of the genome.

The same thing, for a scientist

`ana-lineage-consistent-scan-fair-null` reported that nothing in 2,018 genes scored more extreme than MEN1 on max(z_myeloid, z_lymphoid), and recorded as an open question that only a genome-scale run could license the unqualified claim. At **8248 genes** — about 47% of protein-coding genes — **exactly one beats MEN1, and it is FLT3 at -4.79 against MEN1's -2.12.** MEN1 moves from the 0.00th to the 0.01st percentile, PSIP1 to 0.02, and every other conclusion holds: BRD9 0.21, CEBPA 47.7, SPI1 96.2, the pan-essential control ACTB 42.4.

Go deeper

The claim was true of the sample, not of the genome

Line-normalised; consistency is max(z_myeloid, z_lymphoid), so the weaker lineage half caps the score. Percentile against 8248 swept genes.

gene	consistency	z myeloid	z B-lineage	percentile
MEN1	-2.122	-2.122	-2.164	0.012
PSIP1	-1.753	-2.686	-1.753	0.024
KMT2A	-1.242	-1.689	-1.242	0.109
BRD9	-1.044	-1.044	-1.535	0.206
MEIS1	-0.973	-0.973	-1.524	0.291
MYB	-0.54	-1.169	-0.54	2.473
DOT1L	-0.444	-0.444	-0.775	3.952
EP300	-0.397	-0.397	-0.706	4.91
ACTB	0.173	-0.829	0.173	42.422
CEBPA	0.223	-1.807	0.223	47.66
SPI1	0.987	-1.569	0.987	96.217

FLT3 at -4.79 is more than twice MEN1's -2.12. Everything else survives quadrupling the sweep, which is the useful half of the result: the rankings were stable and the superlative was not.

My explanation for it was also wrong

The obvious reading is a co-mutation artefact: MV4-11, MOLM-13 and MOLM-14 carry FLT3-ITD, so of course they need FLT3. The per-line data refutes that.

line	lineage	FLT3 z	MEN1 z
MOLM-13	myeloid, ITD	-13.1	-4.4
SEMK2	B-lineage	-10.2	-3.2
HB11;19	B-lineage	-9.6	-4.0
MV4-11	myeloid, ITD	-8.9	-2.0
SEM	B-lineage	-5.5	-1.5
MOLM-14	myeloid, ITD	-5.4	-2.4
RS4;11	B-lineage	-0.9	-0.2

SEMK2, HB11;19 and SEM are B-lineage KMT2A-rearranged lines without FLT3-ITD and they are among the most FLT3-dependent in the panel. That is consistent with long-established biology — FLT3 is highly expressed in KMT2A-rearranged ALL irrespective of ITD — so the blind sweep rediscovered a known feature of this disease rather than finding an artefact. RS4;11 not being dependent shows it is not universal even within KMT2A::AFF1.

The statistic rewards genes that matter nowhere else

Going genome-scale exposed a limitation of the statistic that 2,018 genes hid. The z-score is scaled by the other-leukaemia arm's standard deviation, and that differs sharply between these two genes:

gene	comparator mean	comparator SD	MOLM-13 raw effect
FLT3	-0.116	0.157	-2.181
MEN1	-0.374	0.404	-2.167

On the raw Chronos scale the two are indistinguishable in MOLM-13 (-2.181 versus -2.167). FLT3 scores three times the z because the comparator distribution is tight: FLT3 matters in almost no other leukaemia line, whereas menin matters somewhat everywhere.

So the statistic systematically favours narrow dependencies over graded ones. That is defensible for finding selectivity and it means cross-gene z comparisons are not effect-size comparisons, which is how the 'nothing beats menin' line was being read.

Facts and figures

script	analyses/lineage-consistent-genome-scale/run.py
status	executed
run date	2026-09-20
n scored	8248
n relevant	1583
n index myeloid	7
n index lymphoid	4
n other leukaemia	43
headline	Exactly one gene in 8,248 beats MEN1 and it is FLT3, whose margin is a scaling effect: raw effects in MOLM-13 are -2.181 and -2.167.
caveat	The relevance filter is a DEPENDENCY proxy for expression, not expression itself: a gene expressed in leukaemia but dispensable is excluded, which makes the relevant null harder than a true expressed-gene null would be and therefore conservative in the direction that matters. The lymphoid half is small and caps the statistic by construction. Lineage comes from the DepMap disease string, a label rather than a signature. The discovery list is unreviewed output of a statistic - high-ranking genes there are candidates for reading, not findings, and pan-essential genes with slight lineage skew are expected to appear.

percentiles

flt3 per line z

discovery top10

percentiles — the numbers behind the chart

{
  "MEN1": {
    "all": 0.012,
    "leukaemia_relevant": 0,
    "consistency": -2.122
  },
  "PSIP1": {
    "all": 0.024,
    "leukaemia_relevant": 0,
    "consistency": -1.753
  },
  "KMT2A": {
    "all": 0.109,
    "leukaemia_relevant": 0.253,
    "consistency": -1.242
  },
  "BRD9": {
    "all": 0.206,
    "leukaemia_relevant": 0.253,
    "consistency": -1.044
  },
  "MEIS1": {
    "all": 0.291,
    "leukaemia_relevant": 0.505,
    "consistency": -0.973
  },
  "MYB": {
    "all": 2.473,
    "leukaemia_relevant": 3.727,
    "consistency": -0.54
  },
  "DOT1L": {
    "all": 3.952,
    "leukaemia_relevant": 5.117,
    "consistency": -0.444
  },
  "EP300": {
    "all": 4.91,
    "leukaemia_relevant": 5.875,
    "consistency": -0.397
  },
  "ACTB": {
    "all": 42.422,
    "leukaemia_relevant": 45.167,
    "consistency": 0.173
  },
  "CEBPA": {
    "all": 47.66,
    "leukaemia_relevant": 50.979,
    "consistency": 0.223
  },
  "SPI1": {
    "all": 96.217,
    "leukaemia_relevant": 95.894,
    "consistency": 0.987
  }
}

flt3 per line z — the numbers behind the chart

{
  "MOLM-13": -13.13,
  "SEMK2": -10.17,
  "HB11;19": -9.55,
  "MV4-11": -8.86,
  "SEM": -5.51,
  "MOLM-14": -5.43,
  "NOMO-1": -4.61,
  "RS4;11": -0.88,
  "SHI-1": -0.85,
  "OCI-AML2": -0.32,
  "THP-1": 0.12
}

comparator variance

{
  "FLT3": {
    "mean": -0.116,
    "sd": 0.157
  },
  "MEN1": {
    "mean": -0.374,
    "sd": 0.404
  }
}

discovery top10 — the numbers behind the chart

[
  {
    "gene": "FLT3",
    "consistency": -4.795,
    "z_myeloid": -4.795,
    "z_lymphoid": -6.708,
    "mean_other_leukaemia": -0.119
  },
  {
    "gene": "JADE2",
    "consistency": -2.055,
    "z_myeloid": -2.055,
    "z_lymphoid": -2.367,
    "mean_other_leukaemia": 0.002
  },
  {
    "gene": "MED13L",
    "consistency": -1.612,
    "z_myeloid": -1.781,
    "z_lymphoid": -1.612,
    "mean_other_leukaemia": -0.364
  },
  {
    "gene": "MPZL3",
    "consistency": -1.515,
    "z_myeloid": -1.536,
    "z_lymphoid": -1.515,
    "mean_other_leukaemia": 0.052
  },
  {
    "gene": "FAAP100",
    "consistency": -1.363,
    "z_myeloid": -1.363,
    "z_lymphoid": -1.94,
    "mean_other_leukaemia": -0.389
  },
  {
    "gene": "CYP2E1",
    "consistency": -1.343,
    "z_myeloid": -1.343,
    "z_lymphoid": -1.444,
    "mean_other_leukaemia": 0.035
  },
  {
    "gene": "FANCA",
    "consistency": -1.329,
    "z_myeloid": -1.329,
    "z_lymphoid": -2.051,
    "mean_other_leukaemia": -0.279
  },
  {
    "gene": "ALAD",
    "consistency": -1.276,
    "z_myeloid": -1.276,
    "z_lymphoid": -1.296,
    "mean_other_leukaemia": -0.254
  },
  {
    "gene": "FANCG",
    "consistency": -1.267,
    "z_myeloid": -1.625,
    "z_lymphoid": -1.267,
    "mean_other_leukaemia": -0.146
  },
  {
    "gene": "NKIRAS2",
    "consistency": -1.196,
    "z_myeloid": -1.196,
    "z_lymphoid": -1.351,
    "mean_other_leukaemia": 0.058
  }
]

How this could bite us back

I published 'nothing in 2,018 genes beat menin' last cycle and flagged the sample size as the open question. It took one run to overturn it. A superlative asserted from a sample is a claim about the sample, and the entry should have said so in its headline rather than its caveats.
The top hit being FLT3 - known for two decades, already drugged, already in this graph - means the sweep validated the method rather than discovering anything. A discovery claim would need the method to surface something nobody had.

Still open

The statistic is scaled by comparator variance, so it ranks narrow dependencies above graded ones. Should the ledger report raw effect difference alongside z, given that FLT3 and MEN1 are indistinguishable on the raw scale in the line where both are most extreme?
Three Fanconi-pathway genes (FANCA, FANCG, FAAP100) appear in the top ten of a blind sweep. Is that a real DNA-repair dependency in KMT2A-rearranged leukaemia, or the expected tail of a statistic run 8,248 times? Nobody has read a paper on it.
RS4;11 is FLT3-independent (z -0.9) while SEM and SEMK2 are strongly dependent, all three KMT2A::AFF1 B-ALL. What distinguishes them, and does it predict anything about which patients a FLT3 inhibitor would help?

Every source behind this entry (1)

E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Sweeping 8248 protein-coding genes (1583 of them depleted in the other-leukaemia arm) on the line-normalised statistic max(z_myeloid, z_lymphoid) finds exactly one gene more extreme than MEN1: FLT3 at -4.79 against MEN1's -2.12. MEN1 sits at the 0.01st percentile, PSIP1 0.02, KMT2A 0.11, BRD9 0.21, MEIS1 0.29; CEBPA at 47.66 and SPI1 at 96.22 remain rejected and the pan-essential control ACTB at 42.42 remains unremarkable. FLT3 dependency is NOT confined to the FLT3-ITD lines: the B-lineage KMT2A-rearranged lines SEMK2 (z -10.2), HB11;19 (-9.6) and SEM (-5.5) carry no ITD and are among the most dependent, while RS4;11 (-0.9) is not dependent at all.
Genome-scale lineage-consistency sweep of 8,248 genes in KMT2A-rearranged leukaemia · Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
What this evidence cannot show
Run by the coordinator on 2026-09-20 to discharge an open question recorded against ana-lineage-consistent-scan-fair-null. 8248 genes is about 47% of protein-coding genes, so this bounds what half the genome contains, not the genome. The index arm is 7 myeloid and 4 B-lineage lines and the statistic is capped by that four-line half by construction. Lineage comes from the DepMap disease string. Chronos measures knockout, not interface blockade. Most importantly, the z-scaling makes cross-gene comparisons sensitive to comparator variance rather than to effect size.
Evidence that argues against this (3)
- FLT3's apparent superiority is largely a scaling effect and the entry should not be read as ranking FLT3 above menin as a target. In MOLM-13 the raw Chronos effects are -2.181 for FLT3 and -2.167 for MEN1, which is the same number. FLT3 earns three times the z because its comparator standard deviation is 0.157 against menin's 0.404 — it matters in almost no other leukaemia line. A statistic scaled by comparator variance rewards narrow dependencies over graded ones. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- FLT3 is also the least novel possible hit. High FLT3 expression in KMT2A-rearranged leukaemia has been known for two decades, FLT3 inhibitors have been trialled in this disease, and the graph already holds a `gene-flt3` entry. A blind sweep whose top hit is the best-known co-dependency in the disease is evidence the method works and not evidence that it discovers. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- The refutation is narrower than it sounds. Quadrupling the sweep moved MEN1 from 0.00 to 0.01 and overturned one superlative; every substantive conclusion — BRD9 surviving, CEBPA and SPI1 failing, ACTB unremarkable — held. The honest reading is that the previous analysis was right about the ordering and wrong to state a maximum from a sample. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026

Connected to

← testsThe Fanconi signal is in both lineage halves, survives the correlation-proof null, and is still there when lineage is held constant
the max(z_myeloid, z_lymphoid) statistic is unpacked into its two halves
refutes →The easy null was not the problem, and nothing in 2,018 genes beat menin
Quadrupling the sweep overturns its 'nothing beats menin' superlative: FLT3 scores -4.79 against MEN1's -2.12. Every other conclusion holds.
supports →FLT3 / CD135 — the strongest non-tether co-dependency, and a target the field may have written off on a badly executed trial
FLT3 is the single most lineage-consistent dependency in 8,248 genes, and not only in the FLT3-ITD lines.
related to →MEN1 / menin — the adaptor the fusion cannot work without
MEN1 sits at the 0.01st percentile of 8,248 genes, second only to FLT3 on this statistic.
supports →Two of the three top new dependencies are myeloid identity, not KMT2A
CEBPA (47.7th percentile) and SPI1 (96.2nd) remain rejected at genome scale.
related to →BRD9 — dispensable in the average cell, rank 39 of 17,728 in KMT2A-rearranged lines, and with an on-target proof most epigenetic targets never get
BRD9 holds at the 0.21st percentile of 8,248 genes.
← refutesNothing beats menin after all, and the statistic that said otherwise ranks narrowness