A new calibration layer tells drug-discovery teams which AI binding predictions to trust - and which to ignore - before anyone touches a lab bench.
Researchers built CaliPPer, a post-hoc framework that scores how well a binding prediction model will perform on new antibody or TCR targets without needing fresh labels. It combines a distance metric called Sample-to-Domain Distance with Bayesian recalibration, working at three levels: an overall generalisability score, predicted aggregate metrics like AUROC, and a confidence score for each individual prediction. Tested across ten models, eight architectures and two immune-receptor domains, it correlated with actual performance at 0.80 to 0.92 and improved AUROC by up to 0.20 on epitopes and variants the models had not seen. Applied retrospectively to five published TCR, BCR, MHC-peptide and small-molecule studies, it raised the true discovery rate in every one, including a case where confirmed neoantigen hits went from zero out of five to three out of five.
This matters because binding prediction models are the front door to antibody and TCR discovery, and a model that looks solid in a paper often falls apart on a new dataset. Existing label-free methods like PAPE and M-CBPE only spit out one aggregate number, which is not much use when you need to decide which individual candidates are worth an expensive wet-lab validation run. A per-sample confidence score turns that guesswork into a triage step.
Still, this is a retrospective reanalysis of published data, not a live prediction run on unpublished candidates - the real test is whether it holds up when nobody already knows the answer.