Target safety triage v2: constraint with its uncertainty attached (run)

Version 1 read loss-of-function constraint as a point value. It had no way to know when that value was uninformative, so short genes passed almost automatically. DPY30 cleared the filter on a LOEUF of 0.731, and DPY30 is a 99-residue protein for which gnomAD expects only 16.2 loss-of-function variants. Its mouse conditional knockout causes pancytopenia with a stem-cell differentiation block, which is precisely the toxicity the screen exists to find.

v2 reads gnomAD's expected loss-of-function count alongside LOEUF and gates on it 1:

For scale: KMT2A expects 345 such variants, FLT3 127, LEDGF 75, menin 58, DPY30 16, BCL2 14.

Only two genes now earn an unqualified pass, and LEDGF is the one that also exceeds the random-gene null. That is the same conclusion v1 reached about LEDGF, arrived at by a rule that no longer has the defect which produced the DPY30 error. A conclusion that survives the correction of the method that produced it is worth more than one that does not.

BCL2 is the other gene whose constraint is underpowered (14.2 expected), but it lands amber on a low LOEUF, which the asymmetry above says is safe to act on.

On the first v2 run, menin's constraint lookup returned nothing and the gene fell through to UNKNOWN. The same lookup succeeds when run on its own, so this is rate limiting rather than missing data.

This is the second time menin specifically has been lost to a transient failure in this analysis family. In v1 it produced a wrong answer, because missing data was treated as passing. In v2 it produced UNKNOWN, which is correct behaviour but still a hole.

The script now retries anything that comes back empty, with a longer pause, and reports what it recovered. Menin came back at LOEUF 0.204 from 58 expected variants and is correctly AMBER. A hole is never accepted on the first failure, which is now a standing rule for every analysis in this project.

This is an addition, not a revision. The table above was run on eleven genes and is left exactly as it ran. Seven further genes were struck in cycle 21 and audited in cycle 22 (ana-target-list-rewrite-audit), and none of them was ever passed through this triage. What follows records them and what the triage would have said.

Four of the seven are red on essentiality alone, which is reassuring: the triage and the counter-evidence sweep agree where they overlap. The other three are the informative cases, because this triage would have waved all three through.

ELL2 is the one that should worry anyone using this table. It scores -0.064 with 1.0% of lines depleted and LOEUF 0.416 at 72.8 expected variants: not essential, not constrained, adequately powered, green on both axes. And a B-cell-conditional Ell2 knockout mouse has curtailed humoral responses, four-fold fewer plasmablasts, a paucity of secreted immunoglobulin and a broken unfolded-protein response (PMID 25238757). The toxicity is real, it is severe, and it is in exactly the compartment an immunosuppressed child cannot spare.

This is the DPY30 failure again in a different disguise. v1 missed DPY30 because a constraint value was underpowered. v2 would miss ELL2 although both values are well powered, because neither axis can see a phenotype that is not proliferation. A CRISPR screen in culture measures growth; gnomAD measures survival to reproductive age. Antibody secretion is invisible to both. The v2 limits section already says no threshold rule substitutes for the mouse and human genetics of the specific target — ELL2 is the case that shows it is not a caveat but the main event.

Gating constraint on power makes the rule honest about what it does not know; it does not make constraint the right instrument. Loss-of-function constraint measures selection against inheriting one broken copy across a lifetime, and a drug gives partial reversible occupancy for months. Those remain different exposures, as the phase-2 review argued when it pointed out that strict constraint filtering would disqualify menin, the one validated target in this field. Menin is still AMBER here, and that is still the right flag rather than a verdict.

The DepMap release version is not stated by the API. Cell-line gene effect is measured in adult cancers, not infant haematopoiesis. And no threshold rule of this kind can substitute for the mouse and human genetics of the specific target, which is what actually caught DPY30.

The beats random null column above uses the criterion ana-depmap-null-deep-pool has since shown to be unstable: exceeding the maximum of a null is an order statistic that rises as the null grows, so it is not comparable between analyses and should not have been a pass/fail gate.

Rerun against 2,743 usable random genes with Holm correction, the co-dependency set that survives is KMT2A, PSIP1, DOT1L, FLT3, MEF2C and KAT7. Two corrections follow for this table:

• MEF2C and KAT7 are recorded here as no and should be yes.
• MEIS1, DPY30, BCL2 and MLLT1 remain no, so those rows are unaffected.

What this does to the conclusion. The headline above — that LEDGF is the only gene to both pass the safety gate and exceed the null — survives, because MEF2C and KAT7 are both AMBER on constraint (LOEUF 0.23 and 0.21) and so still fail the safety gate rather than the co-dependency gate. LEDGF and FLT3 remain the only unqualified passes. But the reason KAT7 is held back has changed: it is constraint, not the absence of a co-dependency signal. That matters, because KAT7 already has chemical matter and the two objections have very different remedies.

The table is left as it was run rather than silently rewritten, so the correction is visible.