Do the other WRAD subunits behave like DPY30? (run, not planned)
also called WRAD discrimination · DPY30 vs RBBP5 ASH2L WDR5
Four proteins form the shared core of every complex that puts the H3K4 mark on DNA, and only one of them, DPY30, tracks menin in leukemia cells. We tested all four the same way to check that this was real and not an artefact of which genes happened to be in the earlier panel.
The same thing, for a scientist
Leukemia-restricted Pearson correlation of DepMap Chronos gene effect against MEN1 across 54 leukemia cell lines, with 5,000-resample bootstrap confidence intervals, 5,000-shuffle permutation p values and Holm-Bonferroni correction across 16 genes. Run specifically to close a gap in the cycle-9 bootstrap, which tested WDR5 but never tested RBBP5 or ASH2L, leaving the claim that DPY30 is distinctive within WRAD partly unsupported.
Why this was worth running
Cycle 9 reported DPY30 as a MEN1 co-dependency that survives multiple-testing correction, and framed it as notable because DPY30's fellow WRAD components did not survive. But the cycle-9 panel of 26 genes contained WDR5 and not RBBP5 or ASH2L. A gene cannot fail a test it was never given. The comparison had to be redone like for like before the claim could be stated 1.
Result2 deeper
| Gene | r vs MEN1 (54 leukemia lines) | 95% CI | crosses 0 | p (perm) | p (Holm) | survives | mean Chronos (all lines) | % depleted |
|---|---|---|---|---|---|---|---|---|
| KMT2A | 0.7831 | [0.671, 0.874] | no | 0.0002 | 0.0032 | yes | -0.206 | 11.4% |
| DPY30 | 0.4462 | [0.142, 0.658] | no | 0.0010 | 0.0150 | yes | -0.390 | 31.5% |
| PRMT1 | 0.3696 | [0.058, 0.610] | no | 0.0054 | 0.0756 | no | -1.399 | 98.7% |
| SETD1A | 0.3593 | [0.115, 0.586] | no | 0.0068 | 0.0884 | no | -0.911 | 94.7% |
| CCNT1 | 0.3486 | [0.038, 0.585] | no | 0.0122 | 0.1464 | no | -0.206 | 8.4% |
| LEO1 | 0.3016 | [0.049, 0.539] | no | 0.0260 | 0.2859 | no | -0.213 | 7.7% |
| SGF29 | 0.2414 | [-0.005, 0.458] | yes | 0.0800 | 0.7998 | no | -0.567 | 58.5% |
| BRD4 | 0.1768 | [-0.119, 0.424] | yes | 0.1970 | 1.0 | no | -1.015 | 94.7% |
| ELL2 | 0.0964 | [-0.223, 0.391] | yes | 0.4777 | 1.0 | no | -0.064 | 1.0% |
| CTR9 | 0.0062 | [-0.289, 0.280] | yes | 0.9630 | 1.0 | no | -0.873 | 92.6% |
| WDR5 | 0.0042 | [-0.281, 0.304] | yes | 0.9774 | 1.0 | no | -1.159 | 98.3% |
| CDC73 | -0.0762 | [-0.306, 0.150] | yes | 0.5939 | 1.0 | no | -1.122 | 98.7% |
| RBBP5 | -0.0809 | [-0.332, 0.164] | yes | 0.5625 | 1.0 | no | -1.105 | 97.6% |
| PAF1 | -0.0953 | [-0.318, 0.149] | yes | 0.5027 | 1.0 | no | -0.832 | 93.3% |
| ASH2L | -0.1529 | [-0.433, 0.118] | yes | 0.2684 | 1.0 | no | -0.333 | 21.4% |
| KDM6A | -0.1916 | [-0.408, 0.052] | yes | 0.1580 | 1.0 | no | +0.096 | 0.8% |
KMT2A reproduces its cycle-9 value to four decimal places (0.7831), and WDR5 reproduces to 0.0042, which confirms the protocol matches 1.
The finding that changes an argument in this graph
This graph's standing explanation for why the hijacked machinery never appears as a co-dependency is that a pan-essential gene has a flat profile and a flat profile cannot correlate with anything. This run confirms that for WDR5, RBBP5, CTR9, CDC73, PAF1 and BRD4, all of which are depleted in over 92% of lines and all of which land near zero.
But it also produces a case the explanation does not cover. ASH2L is not pan-essential — mean Chronos -0.333, depleted in only 21.4% of lines, a distribution close to DPY30's -0.390 and 31.5%. It has variance. And it comes out at r = -0.15. The flat-profile argument therefore cannot be the whole story: two subunits of the same four-protein module with comparable essentiality give opposite answers, so something other than statistical power separates DPY30 from ASH2L.
Two genes that come close and one that should be watched
SETD1A (r = 0.359) and PRMT1 (r = 0.370) both fail Holm but have confidence intervals clear of zero. Both are pan-essential (94.7% and 98.7% of lines depleted), so even a real signal here would not be a target. It is worth recording that a pan-essential gene can show a non-zero leukemia-restricted correlation; the flat-profile argument is a tendency, not a law.
SGF29 (r = 0.241, CI crosses zero) is the one to watch. It is the only gene tested here with intermediate essentiality — mean Chronos -0.567, 58.5% of lines depleted, leukemia mean -0.741 — that is not flagged common-essential. It did not reach significance in 54 lines, and with a CI of [-0.005, 0.458] it is exactly the kind of result that a larger leukemia panel would settle in either direction.
KDM6A sits at r = -0.19 with a mean gene effect of +0.096 and depletion in 0.8% of lines. Knocking it out makes cells grow slightly better. That is a tumour-suppressor profile, and it is the quantitative counterpart of this graph's position that KDM6A is not a KMT2A-fusion partner.
What this cannot tell you
54 leukemia lines is a small sample and every confidence interval here is wide; DPY30's spans 0.14 to 0.66, which is the difference between a marginal association and a strong one. None of the 54 lines is an infant KMT2A-rearranged sample. Correlation across cell lines reflects shared requirement, which can arise from complex membership, shared pathway, shared lineage restriction or correlated expression, and this run does not distinguish those or adjust for lineage within the leukemia subset. The DepMap release version is not stated by the Open Targets API, so the data vintage is unverified.
One convention should be flagged. 'Leukemia line' is assigned by string matching on the disease field, which pulls in T-lymphoblastic and myeloid lines alongside B-ALL. That is the same convention as the cycle-9 run, chosen for comparability, but it is a coarse definition for a project about infant B-ALL.
Coordinator resolution on the DPY30 statistic
Two tests, two questions, and this graph conflated them.
ana-depmap-bootstrap-ci reports DPY30 at Holm-corrected permutation p = 0.015. ana-depmap-null-deep-pool reports DPY30 failing Holm at 0.086, and ana-depmap-within-subtype agrees at 0.085. These do not contradict each other, because they test different null hypotheses:
- A permutation test shuffles the cell-line labels and asks whether the correlation differs from zero. DPY30 passes: its correlation is real.
- A variance-matched random-gene null asks whether the correlation exceeds what comparable random genes achieve. DPY30 fails: a random gene of the same variance does about as well.
So the honest statement is that DPY30's co-dependency with MEN1 is real but not distinctive. It is not withdrawn, and it is not evidence that DPY30 is special. For target selection the matched null is the relevant test, because a correlation no better than a same-variance random gene is not a reason to pick anything.
This distinction applies to every gene in these analyses, not just DPY30. 'Survives Holm' has been used in this graph as if it named one thing; it names two, and the weaker candidates (DPY30, BCL2, MLLT1) are exactly where the two diverge.
This entry's use of the Holm p = 0.015 figure is therefore not wrong, but incomplete. Read it as DPY30's correlation with MEN1 is reliably non-zero, not as DPY30 stands out among the WRAD subunits. The counter-evidence items added by the interactome2 sweep make the same point from the literature side.
| script | analyses/wrad-discrimination/run.py |
|---|---|
| status | run |
| run date | 2026-09-18 |
| n cell lines | 1258 |
| n leukemia lines | 54 |
| n bootstrap | 5000 |
| n permutation | 5000 |
| seed | 20260918 |
| multiple testing | Holm-Bonferroni across the 16 genes tested |
| result summary | DPY30 is the only WRAD subunit whose leukemia-restricted correlation with MEN1 survives correction (r = 0.446). WDR5, RBBP5 and ASH2L all cross zero. ASH2L is not pan-essential, so the 'flat profile cannot correlate' explanation does not cover it, and the DPY30/ASH2L difference must be biological. KDM6A has a positive mean gene effect (+0.096), a tumour-suppressor profile. |
| reproduction check | KMT2A r = 0.7831 and WDR5 r = 0.0042 both reproduce the cycle-9 bootstrap values exactly, confirming protocol equivalence. |
| figure | analyses/wrad-discrimination/results.md |
inputs
[ "DepMap CRISPR Chronos gene effect per cell line via Open Targets GraphQL", "17 genes: WRAD, SET1/MLL catalytic subunits, PAF1 complex, and previously untested interactors" ]
newly tested genes
[ "RBBP5", "ASH2L", "SETD1A", "SGF29", "KDM6A", "LEO1", "CTR9", "PAF1", "CDC73", "ELL2" ]
- This analysis can only ever find selective partners, never targets inside the pan-essential core, so a negative result here is not evidence that a gene is unimportant to the fusion.
- Promoting DPY30 on the strength of r = 0.45 in 54 lines would be promoting a target on a confidence interval that reaches down to 0.14.
- Does the DPY30-MEN1 correlation survive restriction to B-ALL lines alone, and separately to AML lines alone? If it evaporates in both, it was a lineage effect across leukemia subtypes.
- SGF29 sits at r = 0.24 with a CI that touches zero and intermediate essentiality. How many additional leukemia lines would be needed to resolve it, and does DepMap have them?
- SETD1A and PRMT1 both show non-zero leukemia-restricted correlations despite being pan-essential. Is the flat-profile argument weaker than this graph has been assuming?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Across 54 DepMap leukemia cell lines, DPY30 correlates with MEN1 at r = 0.446 (95% CI 0.142-0.658, permutation p = 0.001, Holm p = 0.015) and survives correction, while WDR5 (0.004), RBBP5 (-0.081) and ASH2L (-0.153) all have confidence intervals crossing zero; ASH2L is not pan-essential (mean Chronos -0.333, 21.4% of lines depleted) so its null result is not explained by lack of variance.
DepMap CRISPR (Chronos) gene effect via the Open Targets Platform GraphQL API; leukemia-restricted correlation with bootstrap and permutation testing, script and results in analyses/wrad-discrimination/ · Broad Institute DepMap; Open Targets Platform; analysis run by this project · Database and derived analysis 2026What this evidence cannot show
Our own computation, not peer reviewed. n = 54 leukemia lines, all adult-derived or paediatric cell lines rather than primary infant samples. Bootstrap intervals are wide. No lineage adjustment within the leukemia subset. DepMap release version unverified.
Evidence that argues against this (2)
- Cell-line panels evolve genetically and transcriptionally in culture, so cross-line dependency correlations can reflect culture-acquired state rather than tumour biology. Ben-David U et al. 2018
- This project's own deeper re-analysis withdraws the DPY30 result that this evidence item rests on. Against variance-matched nulls drawn from 2,743 usable random protein-coding genes across the same 54 leukemia lines, the Holm-corrected survivors are KMT2A (r 0.783), PSIP1 (0.686), DOT1L (0.681), FLT3 (0.645), MEF2C (0.526) and KAT7 (0.517). DPY30 is explicitly among those that do NOT clear correction, alongside MEIS1 (Holm 0.067), MLLT1 and BCL2. The earlier 16-gene panel gave DPY30 Holm p = 0.015 using an 'exceeds the maximum of the null' criterion that the deep-pool run shows to be unstable by construction, because the null maximum rises mechanically as the pool grows. The DPY30-versus-other-WRAD-subunits contrast is therefore a contrast between one uncorrected near-signal and three nulls, not between a survivor and three failures. Broad Institute DepMap; Open Targets Platform; analysis run by this project 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The mechanistic division of labour that best explains the DPY30-versus-ASH2L split is that DPY30 governs genome-wide placement of H3K4 methylation through its ASH2L interface while contributing only modestly to catalytic stimulation.
Structural Analysis of the Ash2L/Dpy-30 Complex Reveals a Heterogeneity in H3K4 Methylation. · Haddad JF, Yang Y, Takahashi YH, Joshi M, Chaudhary N, Woodfin AR, Benyoucef A, Yeung S, Brunzelle JS, Skiniotis G, Brand M, Shilatifard A, Couture JF · Structure 2018What this evidence cannot show
This paper was not performed in leukemia models and does not address MEN1. Using it to explain a DepMap correlation is an inference by us, and it is the weakest link in the argument.
Evidence that argues against this (1)
- This is the decisive objection to the whole WRAD thread and it should be read before any of it. The SET domain of MLL1 -- the catalytic activity that WRAD exists to stimulate -- is dispensable for maintaining haematopoietic stem cells AND for supporting leukemogenesis driven by the MLL-AF9 fusion oncoprotein in mouse. On Mll1 deletion, what is lost at MLL1 target genes is H4K16 acetylation, and inhibiting SIRT1 is sufficient to restore both the acetylation and target-gene expression. The conclusion the authors draw is that recruited MOF activity, not MLL1's intrinsic methyltransferase activity, is what maintains these genes. If H3K4 methylation by MLL1 is not required for MLL-fusion leukemia, then the subunit whose job is to enable that methylation is not a target in this disease, whatever its biochemistry. Mishra BP et al. 2014
- related to →Six times the null pool, and the criterion I had been using turns out to be the problem (run)
Superseded on the key number: the deep-pool re-run places DPY30 among the genes that fail Holm correction
- supports →DPY30 — the smallest COMPASS subunit, and the only one that behaves like a selective dependency
DPY30 is the only WRAD subunit surviving correction
closes the RBBP5/ASH2L gap left by the cycle-9 panel
DPY30's GREEN call rests on a LOEUF that is underpowered at 99 residues
nominal leukemia correlation r=0.30, fails Holm
- ← refutesASH2L — the WRAD subunit that has DPY30's essentiality profile and none of its co-dependency signal
ASH2L has DPY30-like variance and no MEN1 correlation, breaking the flat-profile explanation
the data pattern the hypothesis was built from
- ← supportsLost from the KMT2A half: the WIN motif and the whole WRAD interface (aa 3762-3767, Arg3765)
supplies the assembly requirement (R3765) that subunit-level discrimination has to work around
the WRAD discrimination analysis is unaffected: its own counter-evidence already withdrew the DPY30 result it rested on
The WRAD discrimination analysis is a dependency-correlation argument and is untouched by this adjudication, which concerns physical association; the two have been read as one.
The DPY30-versus-MEN1 co-dependency was never an argument about physical association with the fusion and cannot substitute for one.