Of the menin co-dependencies, which are actually safe to hit? (run)

This directly addresses the PI's stated requirement: disrupt the target without breaking an essential function, ideally by choosing a protein that is not essential 1.

Nothing came out RED, because the pan-essential genes had already been filtered out by the co-dependency step that produced this list.

This is the finding worth acting on. PSIP1/LEDGF is co-dependent with MEN1 at r = 0.69 in leukemia lines, yet its mean gene effect across 1,200 cell lines is +0.03, meaning cells on average do not need it at all, and its human constraint (LOEUF 0.37) sits above the threshold.

Compare menin itself: mean gene effect -0.36, LOEUF 0.20. Menin is the more constrained protein of the two and the more broadly required.

So on the two axes the PI asked about, the adaptor sitting immediately next to menin in the same tether looks like the better-positioned target, not a fallback. LEDGF is already established as the bridge that links menin to chromatin, so disrupting it should collapse the same tether.

This is a target-selection hypothesis from public data, not a result. The obvious objection is that a protein cells do not need in culture may still matter in a developing human, and LEDGF's role in HIV integration means it has been studied mostly in a different context. But it reframes a question the project had treated as settled: menin is the validated target, not necessarily the best one.

The first run called MEN1 green. The gnomAD lookup for MEN1 had failed transiently, the script recorded the constraint as missing, and the triage rule treated missing data as passing.

That is the wrong default for a safety screen in the most basic way: a tool that upgrades a target because a network call dropped is worse than no tool. MEN1's real LOEUF is 0.204, which puts it firmly in amber, and a second direct query confirmed it.

The rule now has an explicit UNKNOWN state and missing data can never produce a pass. This is the fifth identifier-or-data error caught in this project, and the second one in my own work rather than in the literature.

Three public datasets and a threshold rule cannot establish safety. LOEUF measures selection against heterozygous loss over evolutionary time, which is a poor proxy for tolerating partial pharmacological inhibition for months. DepMap mean gene effect is across adult cancer lines, not normal infant haematopoiesis, which is the tissue that matters here. Human Protein Atlas distribution at the whole-tissue level says nothing about the specific progenitor population at risk.

The thresholds (0.35 for LOEUF, -0.5 for gene effect) are conventional but arbitrary, and several genes sit close to them: BCL2 at 0.333 and MLLT1 at 0.330 are amber only just, and would be green under a 0.30 cut.

The phase-2 readiness review, run independently, reached the same conclusion about LEDGF but made a sharp objection to how this triage ranks: human loss-of-function constraint cannot rank this portfolio, and applied strictly it would have killed menin.

That objection is correct and worth stating plainly. Menin has a LOEUF of 0.204, which this rule flags amber, and yet menin is a clinically validated target with an approved drug and an acceptable short-term safety profile in children. If the metric disqualifies the one target the field has already de-risked, the metric is not measuring the thing we care about.

The reason is that constraint measures selection against germline heterozygous loss across a lifetime. A drug delivers partial, reversible occupancy for months. Those are different exposures, and a gene can be strongly constrained under the first while being perfectly tolerable under the second. Treat the constraint column here as a flag that invites a closer look, never as a filter.

The LEDGF conclusion survives the critique, because it does not rest on constraint. LEDGF's case is that its mean gene effect is +0.03 across 1,200 lines, meaning cells do not need it, while it is co-dependent with menin at 0.69 in leukemia. That is a dependency argument, not a constraint argument, and it is the stronger half of this analysis.

This triage put DPY30 in the green column. The interactome review shows that was wrong on two independent grounds, and both are failures of the rule rather than of the data.

The constraint value is underpowered. DPY30's LOEUF of 0.731 is what put it above the threshold, but DPY30 is a 99-residue protein. Loss-of-function constraint is estimated from the number of truncating variants observed against the number expected, and in a gene this small the expected count is tiny, so the confidence interval is enormous. A LOEUF near 0.73 in a 99-residue protein carries almost no information. My rule read it as though it did.

The biology contradicts it outright. Conditional knockout of Dpy30 in mouse haematopoiesis causes pancytopenia with a block in haematopoietic stem cell differentiation. That is precisely the toxicity this triage exists to detect, and it is exactly the tissue at risk in an infant with leukemia.

DPY30 is reclassified AMBER at best. More usefully, this exposes a structural flaw: the rule should not consume a constraint value without also consuming its uncertainty or the gene's length. A short gene will pass the LOEUF filter almost regardless of its biology.