Are the menin partners outliers, or did we pick them? (run)
also called null distribution MEN1 co-dependency · random gene control
Every co-dependency result here so far tested only genes we already believed were partners, which guarantees that partners come out on top. We repeated the test against 377 randomly chosen genes to find out what an unremarkable result looks like, and the four strongest partners beat every single random gene.
The same thing, for a scientist
Builds the missing null for the MEN1 co-dependency work. 377 random protein-coding genes correlated with MEN1 across the same 54 leukemia lines. The null has mean 0.028, sd 0.152 and a heavy right tail reaching 0.641. KMT2A, PSIP1, DOT1L and FLT3 exceed every random gene; the remaining six survivors sit at the 99.4th to 99.7th percentile but inside the range chance reaches.
Why this analysis had to happen1 deeper
Every co-dependency result in this project used a hand-picked panel of about 36 genes drawn from the known interactome. That design cannot answer the question it appears to answer. If you only test genes you already believe are partners, discovering that the top hits are partners is guaranteed by the sampling, not by the biology.
So: 377 random protein-coding genes from the HGNC catalogue, same anchor, same 54 leukemia lines, same method 1.
The null is wider than intuition suggests
| statistic | value |
|---|---|
| mean | 0.028 |
| standard deviation | 0.152 |
| 95th percentile | 0.275 |
| 99th percentile | 0.399 |
| maximum | 0.641 |
| minimum | -0.404 |
That maximum is the important number. With only 54 cell lines, a randomly chosen gene reached a correlation of 0.64 with menin. Anyone reporting a single correlation of 0.5 from a panel this size, without a null, is reporting noise-compatible data.
The top four are genuine outliers. The rest are not.
| Gene | r | percentile | beats every random gene |
|---|---|---|---|
| KMT2A | 0.78 | 100 | yes |
| PSIP1 (LEDGF) | 0.69 | 100 | yes |
| DOT1L | 0.68 | 100 | yes |
| FLT3 | 0.65 | 100 | yes (just) |
| MEF2C | 0.53 | 99.7 | no |
| KAT7 | 0.52 | 99.7 | no |
| MEIS1 | 0.47 | 99.4 | no |
| DPY30 | 0.45 | 99.4 | no |
| BCL2 | 0.43 | 99.4 | no |
| MLLT1 (ENL) | 0.43 | 99.4 | no |
Vindicated: the tether module plus DOT1L and FLT3 sit above all 377 random genes. The hand-picked panel did not manufacture that. Four genes clustering at the very top, when the biology predicts exactly those four, is a real signal.
Deflated: the bottom six sit inside the range chance reaches. Their empirical p-values of 0.003 to 0.006, multiplied across ten comparisons, land at roughly 0.03 to 0.06. That is marginal, not established.
This matters for something I said last cycle. I flagged DPY30 as the intriguing new name in the surviving set. Against a proper null it is at the 99.4th percentile with an empirical p around 0.006 before correction, which is suggestive and no more. It is worth an entry, not a programme.
The control set produced a real lead
The two highest random genes are worth naming, because one of them is probably not noise.
USP22 at r = 0.48, second-highest of 377 random genes. USP22 is the deubiquitinase subunit of the SAGA complex, which is chromatin biology adjacent to everything this project studies. It was picked at random and it scored above six of the ten genes on our curated list. Notably, SGF29 is also a SAGA component and is already on this project's list of documented interactors awaiting an entry. Two SAGA subunits arriving from different directions is worth following.
SIK3 at r = 0.64, the single highest random gene and the one that sets the null's ceiling. A salt-inducible kinase with no obvious connection to the menin tether. Most likely this is the tail of the distribution doing what tails do, and it is a useful reminder of how large that tail is at this sample size. It could also be real; nothing here distinguishes the two.
Limits
377 random genes approximates but does not equal a genome-wide scan of roughly 19,000 protein-coding genes; the true maximum of the null is probably higher than 0.641. The percentiles are therefore optimistic for the borderline genes. A random sample also includes many genes with little variance in gene effect, which biases the null toward zero and makes the partners look better than a variance-matched null would.
A variance-matched or expression-matched null would be the stricter test, and has not been run. The DepMap release version is not stated by the API.
| script | analyses/depmap-null-distribution/run.py |
|---|---|
| status | run |
| run date | 2026-09-18 |
| seed | 20260918 |
| n random with data | 377 |
| n leukemia lines | 54 |
| figure | analyses/depmap-null-distribution/results.md |
null summary — the numbers behind the chart
{
"mean": 0.0282,
"sd": 0.1516,
"p95": 0.2749,
"p99": 0.3987,
"max": 0.6413,
"min": -0.4037
}exceed entire null — the numbers behind the chart
[
{
"gene": "KMT2A",
"r": 0.7831
},
{
"gene": "PSIP1",
"r": 0.686
},
{
"gene": "DOT1L",
"r": 0.6808
},
{
"gene": "FLT3",
"r": 0.6452
}
]inside null range — the numbers behind the chart
[
{
"gene": "MEF2C",
"r": 0.526,
"percentile": 99.72,
"empirical_p": 0.00279
},
{
"gene": "KAT7",
"r": 0.5168,
"percentile": 99.72,
"empirical_p": 0.00279
},
{
"gene": "MEIS1",
"r": 0.468,
"percentile": 99.44,
"empirical_p": 0.00559
},
{
"gene": "DPY30",
"r": 0.4462,
"percentile": 99.44,
"empirical_p": 0.00559
},
{
"gene": "BCL2",
"r": 0.4289,
"percentile": 99.44,
"empirical_p": 0.00559
},
{
"gene": "MLLT1",
"r": 0.4257,
"percentile": 99.44,
"empirical_p": 0.00559
}
]notable random genes
[
{
"gene": "SIK3",
"r": 0.6413,
"note": "sets the null ceiling; salt-inducible kinase with no obvious link to the tether"
},
{
"gene": "USP22",
"r": 0.4774,
"note": "SAGA complex deubiquitinase; scored above six of the ten curated survivors, and SGF29 from the same complex is already on the interactor list"
}
]- The single most important number here is that a randomly chosen gene reached 0.64. Any target-selection decision made from an uncontrolled correlation at this sample size is unsafe.
- Six of the ten genes this project called significant last cycle are, against a proper null, merely suggestive. Treating them as a target list would over-read the data.
- A variance-matched or expression-matched null would be stricter than a plain random sample and has not been run. Would the borderline six survive it?
- USP22 arrived at r = 0.48 from a random draw, and SGF29 from the same SAGA complex is already a documented interactor. Is SAGA a genuine and overlooked part of this biology?
- Is SIK3 at 0.64 simply the tail of the null, or a real dependency nobody has connected to menin? Nothing in this analysis distinguishes them.
- How much higher would the null maximum go with a full 19,000-gene scan rather than 377, and does that move any of the borderline six below significance?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Against a null of 377 random protein-coding genes correlated with MEN1 across 54 leukemia cell lines (mean 0.028, sd 0.152, maximum 0.641), KMT2A (0.78), PSIP1 (0.69), DOT1L (0.68) and FLT3 (0.65) exceed every random gene, while MEF2C, KAT7, MEIS1, DPY30, BCL2 and MLLT1 fall at the 99.4th to 99.7th percentile, inside the range chance reaches.
DepMap CRISPR (Chronos) gene effect via the Open Targets Platform GraphQL API, with a random-gene null drawn from the HGNC protein-coding catalogue; analysis in analyses/depmap-null-distribution/ · Broad Institute DepMap; HGNC; Open Targets Platform; analysis run by this project · Databases and derived analysis 2026What this evidence cannot show
Our own computation, not peer reviewed. 377 random genes is not a genome-wide scan, so the true null maximum is likely higher and the borderline percentiles are optimistic. The null is not matched to the partners on variance or expression, which biases it toward zero. 54 leukemia lines, none of them infant or KMT2A-rearranged-specific. DepMap release version unverified.
Evidence that argues against this (1)
- A randomly chosen gene reached r = 0.64 in this very analysis, which is the clearest possible demonstration that correlations of this magnitude arise by chance at this sample size and should not be interpreted individually. Ben-David U et al. 2018
- ← refutesSix times the null pool, and the criterion I had been using turns out to be the problem (run)
same methodological objection: 'exceeds the null maximum' rises with pool size and is not comparable across analyses
six of the ten 'significant' genes sit inside the range random genes reach
the hand-picked panel did not manufacture the top four; they beat all 377 random genes
LEDGF exceeds every random gene, so its co-dependency is a genuine outlier
strengthens the LEDGF case: outlier status confirmed against a null
FLT3 also exceeds the entire null, deepening the puzzle of its negative targeted trial
uses the random-null result to mark which co-dependencies are genuine outliers
answers the open question it raised, and reverses its verdict on three genes