AI-predicted protein interactomes from AlphaFold-Multimer are now published as catalogues of "high-confidence" complexes and mined for drug targets. But the confidence score they rank on (interface predicted-TM) is overconfident, and the false-discovery rates attached to it are measured on balanced benchmarks with roughly equal real and fake pairs. A real interactome is the opposite: genuine interactions are rare among the thousands of pairs tested, and that rarity is exactly what makes the true error rate dramatically greater than the stated one.The cost of error is real. A target that isn't there sends a drug program down a dead end. I re-audited the CM4AI cell map (Schaffer et al., Nature 2025) with conformal FDR control, which produces honest error rates without assuming the confidence score is calibrated, then refereed the result against DepMap co-essentiality, a functional signal the structural model never saw, behind a strict anti-data-leakage firewall. The headline finding: a benchmark-tuned "10% error" cutoff actually admits 90% false discoveries as interactions grow rare, while conformal control stays bounded. Under it, 35 of 161 (22%) of the paper's own high-confidence edges fail, and the removed edges are significantly depleted of independent co-essentiality support (p=0.016). I then nominated KANSL3 as a missing member of the leukemia-associated MLL1-WDR5 complex, corroborated by co-essentiality and an independent Boltz-2 structure. The result is a reusable trust layer that flags which AI-predicted complexes survive honest error control before anyone acts on them. It works only when a release ships the negative controls to calibrate against, which is why CM4AI could be audited and most current deposits cannot.
Team
