Certified selectivity margins for selection-aware synthetic-lethal target nomination from cancer dependency screens

Genome-wide CRISPR dependency screens are commonly mined by selecting, for each cancer driver, the gene with the largest mutant-versus-wild-type dependency contrast. The selected gene is the maximum over roughly 17,000 candidates, so a confidence bound attached after selection is inflated by the winner's curse. We introduce the Certified Selectivity Margin (CSM), a selection-aware lower-margin certificate for the selected synthetic-lethal (SL) dependency. The deployed CSM uses a Romano–Wolf studentized max-statistic band and validates false-certification by genotype-label permutation. Across DepMap 24Q2/24Q4/25Q2, the CSM false-certifies under a no-SL label-permutation null at mean 0.0173 (worst driver 0.060; 200 label permutations per driver), below the target =0.10, whereas the selection-ignoring bound false-certifies at 0.9992. The primary certified result is ARID1A1B, a SWI/SNF paralog SL that remains CSM-positive after full 24Q2 covariate adjustment for lineage, copy number, expression, screen and sequence QC, inferred efficacy, and growth (n=1150 cell lines, 17,107 genes, B_BOOT=600). EP300 remains positive but boundary-small. The raw recovery view gives CSM 5/7 textbook pairs; the full-covariate stress test preserves 2/7. The same adjustment is hard on the incumbent: official SLIdR raw all-hit recovery is 5/7, but full-covariate recovery drops to 2/7 all-hit and 0/7 WT-filtered exact. The CSM is therefore not a raw-power replacement for SLIdR; it is a complementary certificate layer that reports a selection-corrected effect-size margin on the selected dependency, a quantity permutation p-values and balanced-class calls do not emit. : synthetic lethality; DepMap; cancer dependency screens; post-selection inference; Romano–Wolf correction; certified effect-size margin.