Error-Controlled Discovery of Protein Interactions via Conformal Calibration

PPI-Caliper uses conformal calibration to select protein interactions at user-set FDR levels, keeping empirical FDR at or below target under in-distribution calibration across five species.

• Why it matters: Strong benchmark accuracy for protein language model predictors may not translate into trustworthy interaction discovery, so error rates must be controlled explicitly.
• What they did: The model-agnostic framework calibrates held-out labels to set interaction thresholds at specified FDR levels without retraining, evaluated with two PLM-based predictors and two human training-data versions across five species.
• The result: Under distribution shift, transferred thresholds often become conservative when predictive performance improves, but may fail to control FDR when performance deteriorates, highlighting the need to assess calibration alongside accuracy.

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