Error-Controlled Discovery of Protein Interactions via Conformal Calibration

Protein language models (PLMs) excel at predicting protein interactions; however, achieving high benchmark performance does not guarantee accurate discoveries. Researchers have developed PPI-Caliper, a model-agnostic framework that utilizes conformal calibration to identify interactions at a user-defined false discovery rate (FDR) while utilizing held-out labels and without retraining the predictor.

The framework was tested across five species, two PLM-based predictors, and two versions of human training data for each predictor. PPI-Caliper maintains the empirical FDR at or below the target level when the calibration process is conducted under in-distribution. When faced with distribution shift, the transferred thresholds tend to become overly conservative if the predictive performance improves; however, the framework may fail to control FDR if the performance deteriorates.

PPI-Caliper links pre-trained interaction scores to explicit error targets and highlights the importance of evaluating calibration in addition to predictive accuracy for reliable biological discoveries.