Abstract
Abstract
Protein language models have become a dominant foundation for sequence-based protein–protein interaction (PPI) prediction, but their generalization remains limited under stringent evaluation, and benchmark accuracy alone cannot reveal whether a PPI predictor learns partner-specific biological signals or exploits contextual shortcuts. Here we introduce AuditPPI, an interpretable framework that transforms sparse-autoencoder (SAE) features from a frozen protein language model into order-invariant pair fingerprints, recasting PPI prediction as an auditable tabular-learning problem. AuditPPI achieves competitive predictive performance across diverse PPI benchmarks while enabling multiscale auditing of the information supporting its predictions. Protein-level audits show that partner-independent participation signals remain substantial in conventional benchmarks, including protein-disjoint benchmark, but are largely insufficient for pair-level discrimination in topology- and degree-controlled benchmark. Pair-level analyses further show that removing protein identity overlap does not eliminate contextual structure: protein-disjoint benchmarking retains strong subcellular co-localization signals and benchmark-dependent feature matching captured by SAE co-activation and absolute-difference. Structural analyses, however, provide little evidence that these predictive signals are specifically grounded in PPI interfaces: feature enrichment at interfaces for protein-disjoint benchmark was no longer detectable after controlling for surface exposure, and contact-specific enrichment was observed for only a small fraction of testable SAE feature pairs. AuditPPI therefore separates predictive success from mechanistic fidelity and provides a practical framework for evaluating whether sequence-based PPI predictors capture partner-specific biological signals instead of contextual shortcuts.