The within-donor null: MEN1 sits at the 99.4th percentile of variation between two sublines of ONE child, and at the 98.9th between donors - menin dependence is not heterogeneous between patients so much as unstable in this background
also called q183 · the within-donor null · menin is unstable not heterogeneous
Two cell lines derived from the same child were reported to need menin very differently, and the question was whether that difference is biology or noise. Measuring 358 genes in the same pair: a typical gene differs by 0.09 between the two lines, and menin differs by 0.71 - the second-largest difference of any gene tested. Menin also differs a lot between children. So menin's requirement is not stable in these cells at all, even within one patient. That is a different and more awkward result than either 'real difference between patients' or 'just noise'.
The same thing, for a scientist
q177 recorded a **2.835×** spread in menin dependence across KMT2A::AFF1 donors and flagged its own weakness: SEM and SEMK2 are two sublines of **one 5-year-old** and differ by **0.7078** on MEN1. Cycle 158 carried the instruction *"never quote 2.835× without 0.71 beside it"* — but 0.71 means nothing without the distribution of within-donor differences, which nobody had computed. Here it is, over a **403-gene random panel** (358 scored in all three lines) [E1]. | | median | p75 | p90 | p95 | max | |---|---:|---:|---:|---:|---:| | **W** — within-donor, \|SEM − SEMK2\| | 0.0880 | 0.1731 | 0.2693 | 0.4057 | **0.7625** | | **B** — between-donor, \|SEM-donor − RS4;11\| | 0.1585 | 0.2715 | 0.4128 | 0.4973 | 1.8321 | 🔴 **MEN1 is an extreme outlier on BOTH.** Its within-donor difference of **0.7078 is the 99.4th percentile of W** — against a median of 0.0880 and a maximum of 0.7625, menin is essentially the most unstable gene in the panel **between two sublines of the same child**. Its between-donor difference of **0.8500 is the 98.9th percentile of B** [E1]. 🟢 **The design does work, which rules out the lazy dismissal.** Median B is **1.80×** median W, so between-donor variation genuinely exceeds subline variation across genes, and **0 of 358 genes** have a within-donor difference as large as MEN1's between-donor 0.85. So the between-donor difference is **not** explainable as subline noise [E1]. 🟠 **But that does not rescue the heterogeneity reading, and this is the point.** Of the three outcomes pre-registered before fetching, the result is the **second**: MEN1's within-donor value is extreme, *which is not the between-donor claim surviving*. **Menin dependence is unstable in KMT2A::AFF1 cells generally — including between two cultures of one patient's leukaemia.** A quantity that moves by 0.71 between sublines is not measuring a fixed property of a patient's fusion. 🔵 **KMT2A behaves the same way and DAXX does not** [E2]. KMT2A itself is at the **90.8th** percentile of W and **96.4th** of B; DAXX is at the **67.9th** and **19.3rd**. So the instability is a property of the **fusion axis**, not of menin alone — and DAXX's evenness, already noted in q177, is confirmed against a proper null. 🟠 **What this means for the menin-interface hypothesis.** If the fusion's requirement for menin were fixed by a retained protein–protein interface, it should be among the *more* reproducible dependencies in cells carrying that fusion. It is among the least. That is not a refutation — an interface can be constant while the consequence of blocking it varies with cell state — but **it does mean menin dependence cannot be used as a readout of the interface**, and several arguments in this corpus have quietly assumed it can.
The control that could have overturned this, run: MEN1 survives, DAXX and DOT1L collapse
The caveat below named the objection: an unselected panel is dominated by genes with mean effect near zero, where a difference of 0.7 is arithmetically impossible, so a strong dependency will rank high for the wrong reason.
The objection is real. Median within-donor difference by mean effect: 0.3374 for genes below −0.80, against 0.0728 for genes above −0.10. Strong dependencies vary 3.63× more than near-zero ones.
MEN1 survives it and gets stronger. Matched against the 40 panel genes nearest it in mean effect (MEN1's is −1.0299), its within-donor difference is the 100.0th percentile — the largest of every comparable-strength gene tested 3.
| gene | mean effect | raw pct | matched pct | scale-free ratio pct | verdict |
|---|---|---|---|---|---|
| MEN1 | −1.0299 | 99.7 | 100.0 | 80.8 | still extreme |
| KMT2A | −0.5325 | 90.8 | 82.5 | 73.7 | weakened |
| DAXX | −0.8556 | 68.2 | 32.5 | 35.4 | collapses |
| DOT1L | −0.5103 | 39.9 | 30.0 | 23.2 | collapses |
🟢 DAXX and DOT1L collapse under matching, and that is a real gain, not a loss. Their middling raw ranks were inflated by the unmatched comparison; corrected, both sit below the median for their strength. DAXX is not merely average in stability — it is more stable than comparable dependencies, which sharpens q177's observation rather than softening it.
🟠 But the two views disagree in degree and both are reported. On a scale-free ratio (within-donor difference divided by mean effect), MEN1 is at the 80.8th percentile, not the 100th — high, but not exceptional. The matched-percentile and ratio views answer slightly different questions — 'largest absolute wobble among equally strong genes' versus 'largest wobble relative to its own size' — and the honest headline is the weaker of the two: menin is in the top fifth of instability by any measure, and the most unstable of its strength class by the stricter one.
(The raw percentile reads 99.4 in the body below and 99.7 here; the difference is one gene that ties with MEN1 once values are rounded to 4dp in the saved JSON.)
Why the pre-registered third option mattered
Three outcomes were written down before the fetch:
- MEN1's 0.71 is ordinary in W → a between-donor 0.85 is ordinary too and the heterogeneity observation dies.
- MEN1's 0.71 is extreme in W → the subline pair is itself anomalous, which is not the between-donor claim surviving.
- W and B are indistinguishable → the design cannot separate donors from sublines for any gene, and the answer is neither 'real' nor 'artefact'.
The result is (2), and without having written (2) down in advance the natural move on seeing 0 of 358 genes reach 0.85 within a donor would have been to declare the between-donor difference real and stop. That reading is available and it is incomplete: it is true and it sits next to MEN1 being the 99.4th percentile of within-donor variation. Both facts are about the same gene in the same cells, and only together do they say what is going on.
Outcome (3) was also genuinely live and was ruled out by measurement rather than assumption: median B / median W = 1.80.
What this does not establish
It is one subline pair. W is built from SEM and SEMK2 alone, so 'within-donor variation' here is really 'variation between these two cultures'. SEMK2 is a distinct subline with its own passage history and possibly divergent genotype; a different pair could give a different W. There is no second within-donor pair among the KMT2A::AFF1 lines to check against.
W conflates three things that this design cannot separate: CRISPR screen technical noise, genuine genetic divergence between the sublines, and Chronos's per-screen normalisation. A gene at the 99.4th percentile of W is unstable for one of those reasons, and knowing which would change the interpretation completely.
The panel is 403 random protein-coding genes, so it is dominated by genes with weak effects near zero, where absolute differences are naturally small. Comparing MEN1 — a strong dependency at roughly −1 — against that panel may overstate its rank. A magnitude-matched null, comparing MEN1 only against genes of similar mean effect, was not run and is the obvious next control. That single omission could move the 99.4th percentile substantially.
| script | analyses/q183-the-within-donor-null-for-chronos/run.py |
|---|---|
| status | executed |
| answers open question from | ana-q177-the-menin-spread-across-aff1-lines-survives-but-rs411-is-also-the-kmt2a-outlier |
| panel size | 403 |
| n scored in all three lines | 358 |
| seed | 20260924 |
| PCT within donor reaching men1 between | 0 |
| median B over median W | 1.8 |
| preregistered outcome realised | 2 |
| verdict | menin dependence is unstable in KMT2A::AFF1 cells including between two sublines of one child; it cannot be used as a readout of the menin interface |
within donor pair
[ "SEM (ACH-000782)", "SEMK2 (ACH-001735)" ]
W — the numbers behind the chart
{
"mean": 0.1291,
"median": 0.088,
"p75": 0.1731,
"p90": 0.2693,
"p95": 0.4057,
"max": 0.7625
}B — the numbers behind the chart
{
"mean": 0.2018,
"median": 0.1585,
"p75": 0.2715,
"p90": 0.4128,
"p95": 0.4973,
"max": 1.8321
}focus — the numbers behind the chart
{
"DAXX": {
"W": 0.1496,
"B": 0.0513,
"W_pct": 67.9,
"B_pct": 19.3
},
"DOT1L": {
"W": 0.0674,
"B": 0.2142,
"W_pct": 39.7,
"B_pct": 65.6
},
"KMT2A": {
"W": 0.2849,
"B": 0.6065,
"W_pct": 90.8,
"B_pct": 96.4
},
"MEN1": {
"W": 0.7078,
"B": 0.85,
"W_pct": 99.4,
"B_pct": 98.9
}
}measurements — the numbers behind the chart
[
{
"quantity": "chronos_delta",
"subject": "median within-donor difference, SEM vs SEMK2",
"universe": "358 random protein-coding genes",
"value": 0.088
},
{
"quantity": "chronos_delta",
"subject": "median between-donor difference, SEM-donor vs RS4;11",
"universe": "358 random protein-coding genes",
"value": 0.1585
},
{
"quantity": "percentile",
"subject": "MEN1 within-donor difference",
"universe": "358-gene within-donor distribution",
"value": 99.4
},
{
"quantity": "percentile",
"subject": "MEN1 between-donor difference",
"universe": "358-gene between-donor distribution",
"value": 98.9
},
{
"quantity": "percentile",
"subject": "KMT2A within-donor difference",
"universe": "358-gene within-donor distribution",
"value": 90.8
},
{
"quantity": "percentile",
"subject": "DAXX between-donor difference",
"universe": "358-gene between-donor distribution",
"value": 19.3
},
{
"quantity": "ratio",
"subject": "median between-donor over median within-donor difference",
"universe": "358 random protein-coding genes",
"value": 1.8
},
{
"quantity": "percentile",
"subject": "MEN1 within-donor difference, MAGNITUDE-MATCHED",
"universe": "40 panel genes nearest MEN1 in mean effect",
"value": 100
},
{
"quantity": "percentile",
"subject": "DAXX within-donor difference, MAGNITUDE-MATCHED",
"universe": "40 panel genes nearest DAXX in mean effect",
"value": 32.5
},
{
"quantity": "fold",
"subject": "within-donor variation of strong dependencies vs near-zero genes",
"universe": "358-gene panel split at mean effect -0.5 and -0.1",
"value": 3.63
}
]magnitude matched — the numbers behind the chart
{
"MEN1": {
"mean_effect": -1.0299,
"W": 0.7078,
"raw_pct": 99.7,
"matched_pct": 100,
"n_matched": 40,
"matched_median_W": 0.2522,
"verdict": "still extreme",
"ratio": 0.6873,
"ratio_pct": 80.8
},
"KMT2A": {
"mean_effect": -0.5325,
"W": 0.2849,
"raw_pct": 90.8,
"matched_pct": 82.5,
"n_matched": 40,
"matched_median_W": 0.1042,
"verdict": "weakened",
"ratio": 0.535,
"ratio_pct": 73.7
},
"DAXX": {
"mean_effect": -0.8556,
"W": 0.1496,
"raw_pct": 68.2,
"matched_pct": 32.5,
"n_matched": 40,
"matched_median_W": 0.2522,
"verdict": "COLLAPSES",
"ratio": 0.1748,
"ratio_pct": 35.4
},
"DOT1L": {
"mean_effect": -0.5103,
"W": 0.0674,
"raw_pct": 39.9,
"matched_pct": 30,
"n_matched": 40,
"matched_median_W": 0.1358,
"verdict": "COLLAPSES",
"ratio": 0.1321,
"ratio_pct": 23.2
}
}median W strong vs near zero — the numbers behind the chart
{
"strong_mean_lt_-0.5": 0.2642,
"near_zero_mean_gte_-0.1": 0.0728,
"fold": 3.63
}- RESOLVED same cycle: the magnitude-matched null was run. The objection was real - strong dependencies vary 3.63x more than near-zero genes - but MEN1 survives at the 100.0th matched percentile, while DAXX and DOT1L collapse to 32.5 and 30.0, i.e. below median for their strength. The scale-free ratio puts MEN1 at 80.8th, and that disagreement is unresolved.
- Is MEN1's within-donor instability specific to this subline pair, or does menin vary this much between any two cultures? DepMap contains other derived-from pairs outside the AFF1 set; testing MEN1's W in ten unrelated subline pairs would separate 'menin is an unstable measurement' from 'these two cultures diverged'.
- Does anything in this corpus infer interface strength from menin dependency magnitude? If menin dependence moves 0.71 between two cultures of one patient, any such inference is unsound and should be found and qualified.
- Check whether KMT2A's own instability reflects guides cutting both the fusion and the wild-type allele. If the measured KMT2A gene effect is a mixture of two requirements, the 'fusion axis is unstable' reading needs rewriting.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Across a 403-gene random panel drawn from the cached HGNC protein-coding set (seed 20260924), 358 genes have Chronos values in SEM, SEMK2 and RS4;11. The within-donor distribution |SEM - SEMK2| has mean 0.1291, median 0.0880, p75 0.1731, p90 0.2693, p95 0.4057 and max 0.7625. The between-donor distribution |mean(SEM,SEMK2) - RS4;11| has mean 0.2018, median 0.1585, p95 0.4973 and max 1.8321. MEN1's within-donor difference is 0.7078, the 99.4th percentile of W; its between-donor difference is 0.8500, the 98.9th percentile of B. No gene of the 358 has a within-donor difference of 0.85 or greater. Median B is 1.80 times median W.
What this evidence cannot show
W rests on a single subline pair, so it conflates screen noise, genuine subline divergence and Chronos per-screen normalisation, and cannot separate them. The panel is unselected protein-coding genes and is therefore dominated by weak effects near zero; a magnitude-matched null was not run. '0 of 358' means below 0.28%, not literally zero.
Evidence that argues against this (2)
- Comparing a strong dependency against an unselected panel is close to comparing an extreme value against a distribution it does not belong to. Genes with mean effect near zero cannot produce a difference of 0.7 in either direction, so MEN1's 99.4th percentile partly measures that MEN1 is a real dependency and most of the panel is not. The rank should be regarded as an upper bound until a magnitude-matched null is run.
- The finding that 0 of 358 genes reach 0.85 within a donor is the strongest quantitative statement here and it supports the OPPOSITE reading from this entry's headline - it says the between-donor difference is not subline noise. The entry resolves the tension by arguing both facts hold at once, but a reader could reasonably take the 0-of-358 result as decisive and treat the within-donor extremeness as a separate curiosity.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
In the same panel and the same lines, KMT2A's within-donor difference is 0.2849 (90.8th percentile of W) and its between-donor difference 0.6065 (96.4th of B); DOT1L's are 0.0674 (39.7th) and 0.2142 (65.6th); DAXX's are 0.1496 (67.9th) and 0.0513 (19.3rd).
What this evidence cannot show
Four genes, one subline pair, one between-donor comparison. The percentile of a single gene against a 358-gene panel is a coarse statistic and small changes in the panel move it.
Evidence that argues against this (2)
- DAXX's stability could be read as evidence it is a more reliable dependency than menin, which would favour the DAXX line - but the same evenness is what q177 argued makes DAXX look like a GENERAL requirement rather than a fusion-specific one. The two readings cannot both be used, and this entry does not settle which is right.
- KMT2A's own instability (90.8th and 96.4th) has an alternative explanation that would undercut the 'fusion axis is unstable' framing: in a KMT2A-rearranged line the CRISPR guides target both the fusion allele and the wild-type allele, and the measured gene effect mixes two different biological requirements whose balance could differ between cultures for reasons unrelated to fusion-axis instability.
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Within-donor difference depends strongly on effect magnitude: median |SEM-SEMK2| is 0.3374 for the 30 panel genes with mean effect below -0.80 and 0.0728 for the 226 with mean effect above -0.10, a 3.63-fold difference between strong dependencies (mean < -0.5, n=42, median 0.2642) and near-zero genes. Matching each focus gene against the 40 panel genes nearest it in mean effect: MEN1 (mean -1.0299) is at the 100.0th matched percentile, KMT2A (-0.5325) at 82.5, DAXX (-0.8556) at 32.5 and DOT1L (-0.5103) at 30.0. On the scale-free ratio |SEM-SEMK2| / |mean effect|, restricted to the 99 genes with |mean effect| > 0.2 (median ratio 0.2947), MEN1 is at the 80.8th percentile, KMT2A 73.7th, DAXX 35.4th, DOT1L 23.2nd.
What this evidence cannot show
Matching uses the 40 nearest genes by mean effect, an arbitrary window that was not swept; MEN1 is the most extreme-effect gene in its neighbourhood, so its matched set is necessarily drawn from one side. The scale-free ratio is unstable for genes with small mean effect, which is why it is restricted to |mean| > 0.2, leaving only 99 genes.
Evidence that argues against this (2)
- The matched and scale-free views give 100.0th and 80.8th for the same gene, a large disagreement, and the entry resolves it by quoting the weaker one. A reader preferring the ratio would conclude menin is merely in the top fifth of unstable genes - notable but not remarkable - and nothing here shows which view is the right one for this question.
- A matched percentile of 100.0 on a window of 40 means 'largest of 40', which is not a strong statistical statement: with 40 draws one gene must be largest, and the probability a given gene tops its own window by chance is 1 in 41. The finding is that menin is at the top rather than that it is significantly so.
answers q177's within-donor caveat: MEN1 is the 99.4th percentile of subline variation, so its dependency is unstable rather than patient-heterogeneous
menin dependence moves 0.71 Chronos units between two cultures of one child's leukaemia - it cannot be used as a readout of interface strength
DAXX sits at the 19.3rd percentile of between-donor variation, confirming against a proper null that it is evenly required
q176 said menin fails on variance not effect; this measures that variance against 358 genes and finds it near the top of the distribution