Calibrated Variant Effect Prediction at the Residue Level Using Conditional Score Distributions

RaCoon, a residue-aware calibration method on ESM1b, reduces calibration error across variant subgroups and datasets while improving AUROC on multiple VEP benchmarks.

• Why it matters: Clinical VEP needs probability estimates that are reliable not just on average, because subgroup-specific miscalibration can limit trust and usefulness.
• What they did: The work benchmarks 30 VEPs, then proposes residue-level calibration using differential score mapping per variant subgroup, implemented as RaCoon with Gaussian-mixture sampling to cut labeled-data needs.
• The result: With appropriately chosen calibration targets, the approach maintains low calibration error within variant subgroups and can improve discrimination, yielding multicalibrated, interpretable predictions and higher AUROC on several benchmarks.

AI summary of the abstract · may contain errors · not medical advice · read the paper.