The abstract is read at the publisher; the summary is JClub's.
ProtCOVER provides family-aware conformal prediction for missense variant effect predictors on unseen protein families, achieving family-weighted pathogenic coverage of 0.823–0.873 at a 0.90 target across five zero-shot predictors.
- Why it matters: Split conformal coverage can fail when deploying to variants from new, evolutionarily distant protein families, where exchangeability over individual variants no longer holds.
- What they did: ProtCOVER calibrates on homology families from label-free clusters and uses descriptor-shift adaptation via a density-ratio reweighting of variants, with finite-sample family-weighted coverage under an affine ratio condition.
- The result: Under descriptor shifts, ProtCOVER’s mean worse-class shortfall is 0.000 versus 0.010 for group-weighted and 0.029 for hierarchical conformal prediction, and it meets the 0.90 target on new ClinVar variants.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.