PPI-Caliper puts a false-discovery-rate guarantee on protein language model interaction predictions, until the data shift
- PPI-Caliper is a model-agnostic wrapper that applies conformal calibration to pick protein interactions at a user-set false discovery rate, using held-out labels and no retraining.
- Across five species and two protein-language-model predictors, each with two versions of human training data, the empirical error rate stayed at or below target when calibration data matched the test data.
- Under distribution shift, thresholds carried over from another setting turned conservative when predictions improved but could fail to control errors when performance dropped.
💡 Points Based on the abstract
For target discovery the benchmark AUC was never the real question; it is how many of the top-ranked interactions you can afford to chase before the follow-up assays dry up. An explicit FDR dial answers that, and the honest shift result matters most: calibrating on a well-studied organism then applying to a poorly annotated one is exactly where it can break. Try it on an in-house interactome with orthogonal validation before trusting it.
Error-Controlled Discovery of Protein Interactions via Conformal Calibration ↗