The easy null was not the problem, and nothing in 2,018 genes beat menin
also called fair null sweep · lineage consistency discovery scan
Last cycle we checked some genes against seven hundred random ones and worried the comparison was too easy, because most random genes do nothing in leukaemia. We redid it against two thousand genes and then against only the four hundred that do something here. The answer barely moved, and nothing we had not already picked scored better than menin.
The same thing, for a scientist
Two open questions recorded against `ana-lineage-consistent-dependency-null` answered on one dataset. **The flattering-null worry was largely wrong.** Against {} swept genes and then against the {} that are actually depleted in the other-leukaemia arm (mean effect <= {}), percentiles barely move: MEN1 0.00 -> 0.00, PSIP1 0.00 -> 0.00, KMT2A 0.10 -> 0.25, **BRD9 0.25 -> 0.50**, MEIS1 0.30 -> 0.50. The rejected candidates stay rejected (CEBPA 48.76 -> 47.88, SPI1 95.79 -> 96.26) and the pan-essential control stays unremarkable (ACTB 43.2 -> 43.1). **And the unbiased sweep found nothing better than the hand-picked panel's top:** the best of 2,018 random genes scores -1.36 against MEN1's -2.14 and PSIP1's -1.74 — though it does outscore BRD9 (-1.07) and KMT2A (-1.23).
Both nulls, side by side
Statistic = max(z_myeloid, z_lymphoid); the weaker lineage half caps the score. Percentile: lower is more extreme.
| gene | consistency | pct vs all 2018 | pct vs 401 relevant | verdict |
|---|---|---|---|---|
| MEN1 | -2.138 | 0.0 | 0.0 | survives |
| PSIP1 | -1.738 | 0.0 | 0.0 | survives |
| KMT2A | -1.232 | 0.099 | 0.249 | survives |
| BRD9 | -1.066 | 0.248 | 0.499 | survives |
| MEIS1 | -0.996 | 0.297 | 0.499 | survives |
| MYB | -0.533 | 3.221 | 4.239 | median |
| DOT1L | -0.461 | 4.163 | 4.988 | median |
| EP300 | -0.409 | 5.154 | 6.234 | median |
| ACTB | 0.181 | 43.211 | 43.142 | median |
| CEBPA | 0.23 | 48.761 | 47.88 | median |
| SPI1 | 0.991 | 95.788 | 96.259 | opposite tail |
The correction here is to my own caution, not to the result. The earlier entry's counter-evidence said the random-gene null flatters any gene with real biology and that a fair comparator "could move BRD9 materially". It moves it from 0.25 to 0.50, which is not material. The worry was reasonable and it was wrong, and that is worth recording as plainly as a confirmed suspicion would be.
What the unbiased sweep found, and what it did not
Top of 2,018 genes scored blind on the same statistic:
| gene | consistency | z myeloid | z B-lineage |
|---|---|---|---|
| PLIN3 | -1.362 | -1.454 | -1.362 |
| FANCF | -1.34 | -1.34 | -2.233 |
| NCALD | -1.203 | -1.203 | -1.551 |
| IREB2 | -1.168 | -1.887 | -1.168 |
| EIF4E1B | -1.147 | -1.147 | -1.235 |
| BBC3 | -1.0 | -1.0 | -1.605 |
| DDHD2 | -0.978 | -1.217 | -0.978 |
| SDHAF4 | -0.971 | -1.37 | -0.971 |
| AKR1B15 | -0.929 | -0.929 | -0.98 |
| PSMG2 | -0.896 | -0.896 | -0.902 |
Nothing beat menin. MEN1 at -2.14 and PSIP1 at -1.74 are both more extreme than the best gene in a 2,018-gene sweep (-1.36). For a project whose central claim is that the menin axis is the specific thing, an unbiased sweep failing to find anything better is the strongest form that claim has taken here — because the sweep could have refuted it and did not.
But two random genes outscore BRD9, which tempers it. BRD9 at -1.07 sits below PLIN3 (-1.36) and FANCF (-1.34), consistent with its 0.25th percentile: roughly five of 2,018 genes beat it. BRD9 is top-five, not unique.
The listed genes are unreviewed output of a statistic, not findings. Nobody has read a paper about PLIN3 or FANCF in this disease, several are the kind of broadly-required gene expected to drift into such a list, and the right next step is reading rather than writing entries.
| script | analyses/lineage-consistent-scan-fair-null/run.py |
|---|---|
| status | executed |
| run date | 2026-09-20 |
| n scored | 2018 |
| n relevant | 401 |
| relevance cut | -0.2 |
| n index myeloid | 7 |
| n index lymphoid | 4 |
| n other leukaemia | 43 |
| headline | Restricting the null to leukaemia-relevant genes does not move any candidate materially (BRD9 0.25 -> 0.50). No gene in 2,018 scored more extreme than MEN1 or PSIP1; two outscored BRD9. |
| 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 — the numbers behind the chart
{
"MEN1": {
"all": 0,
"leukaemia_relevant": 0,
"consistency": -2.138
},
"PSIP1": {
"all": 0,
"leukaemia_relevant": 0,
"consistency": -1.738
},
"KMT2A": {
"all": 0.099,
"leukaemia_relevant": 0.249,
"consistency": -1.232
},
"BRD9": {
"all": 0.248,
"leukaemia_relevant": 0.499,
"consistency": -1.066
},
"MEIS1": {
"all": 0.297,
"leukaemia_relevant": 0.499,
"consistency": -0.996
},
"MYB": {
"all": 3.221,
"leukaemia_relevant": 4.239,
"consistency": -0.533
},
"DOT1L": {
"all": 4.163,
"leukaemia_relevant": 4.988,
"consistency": -0.461
},
"EP300": {
"all": 5.154,
"leukaemia_relevant": 6.234,
"consistency": -0.409
},
"ACTB": {
"all": 43.211,
"leukaemia_relevant": 43.142,
"consistency": 0.181
},
"CEBPA": {
"all": 48.761,
"leukaemia_relevant": 47.88,
"consistency": 0.23
},
"SPI1": {
"all": 95.788,
"leukaemia_relevant": 96.259,
"consistency": 0.991
}
}discovery top10 — the numbers behind the chart
[
{
"gene": "PLIN3",
"consistency": -1.362,
"z_myeloid": -1.454,
"z_lymphoid": -1.362,
"mean_other_leukaemia": -0.041
},
{
"gene": "FANCF",
"consistency": -1.34,
"z_myeloid": -1.34,
"z_lymphoid": -2.233,
"mean_other_leukaemia": -0.272
},
{
"gene": "NCALD",
"consistency": -1.203,
"z_myeloid": -1.203,
"z_lymphoid": -1.551,
"mean_other_leukaemia": -0.057
},
{
"gene": "IREB2",
"consistency": -1.168,
"z_myeloid": -1.887,
"z_lymphoid": -1.168,
"mean_other_leukaemia": -0.301
},
{
"gene": "EIF4E1B",
"consistency": -1.147,
"z_myeloid": -1.147,
"z_lymphoid": -1.235,
"mean_other_leukaemia": 0.02
},
{
"gene": "BBC3",
"consistency": -1,
"z_myeloid": -1,
"z_lymphoid": -1.605,
"mean_other_leukaemia": -0.098
},
{
"gene": "DDHD2",
"consistency": -0.978,
"z_myeloid": -1.217,
"z_lymphoid": -0.978,
"mean_other_leukaemia": 0.033
},
{
"gene": "SDHAF4",
"consistency": -0.971,
"z_myeloid": -1.37,
"z_lymphoid": -0.971,
"mean_other_leukaemia": 0.089
},
{
"gene": "AKR1B15",
"consistency": -0.929,
"z_myeloid": -0.929,
"z_lymphoid": -0.98,
"mean_other_leukaemia": -0.128
},
{
"gene": "PSMG2",
"consistency": -0.896,
"z_myeloid": -0.896,
"z_lymphoid": -0.902,
"mean_other_leukaemia": -0.71
}
]- This analysis was run to check a worry the coordinator recorded against its own previous analysis, and it found the worry unfounded. That is a better outcome than confirmation would have been, but it is the fourth consecutive coordinator analysis whose conclusion favours the menin axis, and each was designed after seeing the result it was checking.
- The discovery list is unreviewed statistical output and should not be mined for entries. Writing up PLIN3 because it topped a blind sweep would be exactly the error this project made with CEBPA and SPI1, in a new form.
- The sweep is 2,018 of ~17,700 genes. A genome-wide run on this statistic is the only thing that licenses 'nothing beats menin' without qualification, and the full scan's stored results do not keep per-line values so it cannot be recomputed from them.
- PLIN3 and FANCF outscore BRD9 and have no KMT2A literature. Is either real, or are they the expected tail of a statistic run 2,018 times? Reading two papers settles it.
- The relevance filter and the statistic share a term, since both are defined against the other-leukaemia arm. Would an expression-based filter, from an independent dataset, give the same agreement between nulls?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Scoring genes on max(z_myeloid, z_lymphoid) against the other-leukaemia arm, line-normalised with offsets from the sweep panel only, over 2018 swept protein-coding genes and then against the 401 of them depleted in the other-leukaemia arm (mean effect <= -0.2): MEN1 and PSIP1 fall at the 0.00th percentile under both nulls, KMT2A 0.10 then 0.25, BRD9 0.25 then 0.50, MEIS1 0.30 then 0.50. CEBPA sits at 48.76 then 47.88 and SPI1 at 95.79 then 96.26; the pan-essential control ACTB at 43.21 then 43.14. Restricting the null to leukaemia-relevant genes therefore does NOT materially change any candidate's standing. Separately, no gene in the 2,018-gene sweep scored more extreme than MEN1 (-2.14) or PSIP1 (-1.74); the best swept gene scored -1.36, which does exceed BRD9 (-1.07) and KMT2A (-1.23).
Lineage-consistency sweep of 2,018 genes with a leukaemia-relevant null · Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026What this evidence cannot show
Run by the coordinator on 2026-09-20 to discharge two open questions recorded against ana-lineage-consistent-dependency-null. The relevance filter is a DEPENDENCY proxy for expression, not expression: a gene expressed in leukaemia but dispensable is excluded, making the relevant null harder than a true expressed-gene null and therefore conservative. The index arm is 7 myeloid and only 4 B-lineage lines, and the statistic is capped by that four-line half by construction, so every number is a four-line measurement carrying a genome-scale percentile. Lineage comes from the DepMap disease string. 2,018 genes is a sample of ~17,700, so the sweep bounds what a random ninth of the genome contains, not the genome.
Evidence that argues against this (3)
- The sweep covers about one ninth of protein-coding genes, so 'nothing beat menin' means nothing in a 2,018-gene sample did. A genome-wide run on this statistic has not been done and is the only thing that would license the unqualified claim. At this sample size roughly nine genes genome-wide could be expected to match the best swept gene. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- Two arbitrary genes outscoring BRD9 is the part of this result that argues against the project's own framing. BRD9 was presented last cycle as the one output-layer nomination that held up; at the 0.25th percentile it is roughly fifth of 2,018, which is a good gene rather than a distinctive one, and PLIN3 and FANCF have no KMT2A literature behind them at all. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- The two nulls agreeing is partly an artefact of how the relevance filter was built. Genes were selected for being DEPLETED in the comparator arm, and the statistic measures depletion in the index arms relative to that same comparator, so the filter and the statistic share a term. A filter built on expression data rather than on dependency would not share it and might not agree. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
Discharges both open questions it recorded: the leukaemia-relevant null does not move any candidate materially, and an unbiased sweep found nothing beating MEN1.
BRD9 holds at the 0.50th percentile under the stricter null, but two arbitrary swept genes outscore it, so it is top-five rather than distinctive.
No gene in a 2,018-gene blind sweep scored more extreme than MEN1 on the lineage-consistency statistic.
PSIP1 holds at the 0.00th percentile under both nulls alongside MEN1.
CEBPA stays at the median of the null under the stricter comparator (48.76 -> 47.88), confirming the myeloid-identity reading.
SPI1 stays in the opposite tail under the stricter comparator (95.79 -> 96.26).
Quadrupling the sweep overturns its 'nothing beats menin' superlative: FLT3 scores -4.79 against MEN1's -2.12. Every other conclusion holds.