The comparator arm was never screened for pathway-equivalent lesions — and it made us understate our own result
also called control arm contamination · comparator audit
Everything this project says about what makes KMT2A-rearranged leukaemia different is a comparison against other leukaemia cell lines. Nobody had checked whether those comparison lines carry the same biology under a different name — and two of them do. Removing them makes the differences slightly bigger, not smaller, so the findings survive; but the check should have been run before the findings, not after.
The same thing, for a scientist
Every DepMap number in this project compares 11 KMT2A-rearranged lines against 43 other leukaemia lines. A line-by-line Cellosaurus audit found the comparator is not a clean negative control: U-937 (ACH-000406) carries PICALM-MLLT10, a HOXA/MEIS1 driver whose transformation is Dot1l-dependent exactly as MLL-AF10's is, and OCI-AML3 (ACH-000336) carries NPM1c with DNMT3A R882C, a menin-inhibitor-sensitive genotype. Recomputing the difference statistic without them across a 12-gene panel: **9 of 12 genes become MORE selectively required**, shifts of 0.007–0.035, and no reported conclusion changes.
What was wrong, and why it was predictable
The comparator was built by one filter: is this leukaemia line KMT2A-rearranged? Everything that failed that test became a negative control. Nobody asked the second question — does this line run the same programme by another route?
Two do. U-937 carries PICALM-MLLT10 (CALM-AF10), which over-expresses HOXA5/9/10 and MEIS1 at levels similar to MLL-translocated leukaemia and whose transformation is abolished by Dot1l inactivation. OCI-AML3 carries NPM1c, a genotype menin inhibitors are approved against. Both were sitting in the negative arm of every selectivity measurement this project has published.
The direction of the resulting bias is not ambiguous. Putting menin- and HOXA-axis-dependent lines into the comparator makes the comparator resemble the index group, which shrinks the measured difference. The error pushes toward the null, so the reported numbers should be underestimates. That is a prediction, and the point of this entry is that it was then measured rather than asserted.
Measured: the findings get stronger, slightly
Difference = mean(KMT2A-rearranged) − mean(other leukaemia). More negative = more selectively required.
| gene | with contaminants | cleaned | shift | direction |
|---|---|---|---|---|
| MEN1 | -0.8885 | -0.9101 | -0.0217 | more |
| KMT2A | -0.5571 | -0.5770 | -0.0200 | more |
| CDK6 | -0.5145 | -0.5498 | -0.0354 | more |
| MEIS1 | -0.3108 | -0.3210 | -0.0102 | more |
| DOT1L | -0.2615 | -0.2819 | -0.0204 | more |
| AFF1 | -0.2462 | -0.2533 | -0.0071 | more |
| USP22 | -0.2206 | -0.2332 | -0.0126 | more |
| MCL1 | -0.4333 | -0.4509 | -0.0176 | more |
| BCL2 | -0.3906 | -0.3816 | +0.0091 | less |
| RNF20 | +0.0864 | +0.0802 | -0.0062 | more |
| AQR | +0.2541 | +0.2605 | +0.0065 | less |
| FLT3 | -0.8462 | -0.8434 | +0.0028 | less |
9 of 12 move toward more selective, exactly as the bias direction predicts. The three that do not — BCL2, AQR, FLT3 — move by +0.003 to +0.009, which is noise at this sample size.
No conclusion in this project changes. MEN1 stays the most selectively required gene and becomes slightly more so. CDK6 shows the largest shift (−0.035) and its gap from CDK4 widens. The two negatives the project has reported both survive: AQR stays on the wrong side (+0.254 → +0.261) and RNF20 stays positive (+0.086 → +0.080), so neither the AQR nor the H2Bub-writer negative was an artefact of a dirty comparator.
The rule this should have followed
A negative control group must be screened for pathway-equivalent lesions, not only for the index lesion. For this disease that means excluding, or at minimum declaring, PICALM-MLLT10, NUP98 fusions, NPM1c and any other HOXA/MEIS1 driver before a line is used as a comparator.
The same audit found three further problems in the panel that are not contamination but do affect what can be claimed: MOLM-13 and MOLM-14 are the same patient — one peripheral-blood draw at relapse — so the index arm is 11 lines from 10 patients and any monocytic count is 6 lines from 5 patients. OCI-AML3 and MUTZ-8 are FAB M4 myelomonocytic rather than M5. And MOLM-14's Cellosaurus record carries no monocytic disease term at all, so the '6 of 7 monocytic' figure this project has quoted is not reproducible from the database alone.
| script | analyses/control-arm-contamination/run.py |
|---|---|
| status | executed |
| run date | 2026-09-19 |
| n genes | 12 |
| n more selective after cleaning | 9 |
| headline | 9 of 12 genes become more selective after cleaning; shifts 0.007-0.035; no conclusion changes; both reported negatives (AQR, RNF20) survive. |
| standing rule | A non-KMT2A-rearranged control group must be screened for pathway-equivalent lesions - PICALM-MLLT10, NUP98 fusions, NPM1c, other HOXA/MEIS1 drivers - not only for KMT2A rearrangement. |
| caveat | This checks the DIRECTION and SIZE of the bias on a panel, not the full 17,728-gene ranking - ranks are not recomputed and could move by more than the difference statistic suggests for genes clustered near a threshold. Removing two lines from 43 is also a small perturbation; the deeper problem the audit raises is that nobody screened the comparator for pathway-equivalent lesions at all, and these two are the ones that were found, not necessarily the only ones. |
contaminants
{
"ACH-000406": "U-937 — carries PICALM-MLLT10 (CALM-AF10), a HOXA/MEIS1 driver",
"ACH-000336": "OCI-AML3 — carries NPM1c + DNMT3A R882C, menin-inhibitor-sensitive"
}panel — the numbers behind the chart
{
"MEN1": {
"with": -0.8885,
"cleaned": -0.9101,
"shift": -0.0217
},
"KMT2A": {
"with": -0.5571,
"cleaned": -0.577,
"shift": -0.02
},
"CDK6": {
"with": -0.5145,
"cleaned": -0.5498,
"shift": -0.0354
},
"MEIS1": {
"with": -0.3108,
"cleaned": -0.321,
"shift": -0.0102
},
"DOT1L": {
"with": -0.2615,
"cleaned": -0.2819,
"shift": -0.0204
},
"AFF1": {
"with": -0.2462,
"cleaned": -0.2533,
"shift": -0.0071
},
"USP22": {
"with": -0.2206,
"cleaned": -0.2332,
"shift": -0.0126
},
"BCL2": {
"with": -0.3906,
"cleaned": -0.3816,
"shift": 0.0091
},
"MCL1": {
"with": -0.4333,
"cleaned": -0.4509,
"shift": -0.0176
},
"AQR": {
"with": 0.2541,
"cleaned": 0.2605,
"shift": 0.0065
},
"RNF20": {
"with": 0.0864,
"cleaned": 0.0802,
"shift": -0.0062
},
"FLT3": {
"with": -0.8462,
"cleaned": -0.8434,
"shift": 0.0028
}
}panel facts
{
"MOLM-13 and MOLM-14": "same patient, one draw at relapse (PMID 9305600)",
"index_arm": "11 lines from 10 patients",
"monocytic_count": "6 lines from 5 patients; MOLM-14 carries no monocytic Cellosaurus term"
}- An error that happens to point the safe way is the easiest kind to under-react to. The finding here is not 'our numbers were fine' but 'we never checked, and the check took twenty minutes' — and the next contaminant may not be conveniently conservative.
- Quantifying the bias on twelve genes chosen because they are the project's headline results is a biased sample by construction. A gene whose apparent selectivity came entirely from the HOXA-axis contamination would not be in this panel, because it would never have looked interesting enough to include.
- Are U-937 and OCI-AML3 the only pathway-equivalent lines in the comparator? The audit found these two; nobody has screened all 43 for NUP98 fusions or other HOXA/MEIS1 drivers.
- Does the full 17,728-gene ranking move for genes clustered near a rank threshold? This panel measures the difference statistic, not the ranks.
- Should the scan be re-run with a declared, screened comparator as standard, given the whole project rests on it?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
Removing U-937 (PICALM-MLLT10) and OCI-AML3 (NPM1c/DNMT3A) from the 43-line comparator makes 9 of 12 panel genes MORE selectively required in KMT2A-rearranged lines. MEN1 moves -0.8885 to -0.9101, CDK6 -0.5145 to -0.5498. AQR remains positive (+0.2605) and RNF20 remains positive (+0.0802), so both reported negatives survive the correction. Shifts are 0.007–0.035 and no conclusion changes.
Control-arm contamination check: DepMap selectivity with and without pathway-equivalent comparator lines · Oncofusion knowledge graph; analyses/control-arm-contamination 2026What this evidence cannot show
Computed by the coordinator on 2026-09-19 against live per-line DepMap data. This checks the DIRECTION and SIZE of the bias on a 12-gene panel; it does NOT recompute the 17,728-gene ranking, and a gene clustered near a rank threshold could move further than its difference statistic suggests. Removing two lines from 43 is a small perturbation, and the contaminants were identified by one audit - they are the ones that were found, not provably the only ones. Line membership was reconstructed by name matching against DepMap's own labels, which is how the original scan grouped them.
Evidence that argues against this (1)
- Reporting that the error was in the conservative direction is comfortable and should be read carefully. It is true for these twelve genes and it does not license the panel: a gene whose selectivity is genuinely driven by the HOXA axis would be the case where contamination bites hardest, and such a gene would look unremarkable in both computations. The deeper problem is also not fixed by this measurement - nobody screened the comparator at all, and the fix is a screening rule applied before the next analysis, not a reassuring recomputation after this one. Oncofusion knowledge graph; analyses/control-arm-contamination 2026
Closes its open question: screening all 46 comparator lines finds the same two contaminants and no third.
Its comparator contained a HOXA/MEIS1 driver and a menin-sensitive genotype; the verdict stays inconclusive but the reason changes from 'too few controls' to 'the controls were not controls'.
- related to →The whole protein-coding genome: menin is still first, and a metabolic pathway appears (run)
Quantifies the bias in this scan's comparator: the error runs toward the null, so its ranks are conservative.
- related to →Six countermeasures, no head-to-head — so rank their targets: CDK6 is 13th and CDK4 is 16,052nd
Built on the same comparator; CDK6 shows the largest shift and its gap from CDK4 widens.
- related to →Six of 85 proteins are both selectively required and drugged — and MEN1 is not close to the others
Same comparator; the selectivity axis is conservative.
Same comparator arm.
- supports →AQR (Aquarius) - a spliceosomal RNA helicase whose loss impairs MLL-rearranged leukaemia stem cells
AQR stays positive after cleaning (+0.261), so the negative reported against it is not an artefact of the contaminated comparator.
The line-by-line audit that found the contamination.
Shows the 'pathway-equivalent contaminant' label was too coarse: U-937 is menin-independent, so it was never contamination for a menin contrast.
extends the preclinical control-arm problem to the t(11;19) panel