Auditing Protein-Protein Interaction Signals with Sparse Autoencoder Fingerprints

Protein language models have proven effective for predicting protein-protein interactions (PPIs), yet their generalization abilities are still questionable. A new framework called AuditPPI has been developed to address this issue. AuditPPI utilizes sparse-autoencoder (SAE) features extracted from a frozen protein language model to create order-invariant pair fingerprints. This approach transforms PPI prediction into an auditable tabular-learning problem.

AuditPPI demonstrates competitive predictive performance across various PPI benchmarks. It also enables comprehensive auditing of the information supporting predictions at different scales. At the protein level, the presence of partner-independent participation signals is consistent across conventional benchmarks, including protein-disjoint datasets. However, these signals are insufficient for accurately distinguishing pairs within the topology- and degree-controlled benchmark.

Further pair-level analysis reveals that removing protein identity overlap does not eliminate contextual structures. The protein-disjoint benchmark still exhibits strong subcellular co-localization signals and feature matching captured by SAE co-activation and absolute difference. Despite this, structural analyses provide limited evidence that these predictive signals are specifically associated with PPI interfaces.

Feature enrichment at interfaces for protein-disjoint benchmark data is no longer detectable when controlling for surface exposure, and contact-specific enrichment is observed for only a small fraction of testable SAE feature pairs.

AuditPPI effectively separates predictive success from mechanistic fidelity. By converting sparse-autoencoder features into order-invariant pair fingerprints, it offers a practical framework for evaluating whether sequence-based PPI predictors capture partner-specific biological signals or rely on contextual shortcuts.