Evidence-resolved benchmarks separate biological context recognition from direct-partner specificity in protein interaction models

Sequence-based models enable scalable protein-protein interaction (PPI) prediction, but whether high scores identify direct physical partners or broader biological compatibility remains unclear. Here we show that biological-context recognition and direct-partner specificity are separable prediction capabilities. We introduce ProtNegDB (http://protnegdb.bjmu.edu.cn/), an evidence-resolved resource distinguishing experimentally supported non-binding from coordinate-verified noncontact within PDB biological assemblies. Using context-matched benchmarks, we find that five sequence-based predictors show limited ability to distinguish direct partners from noncontacting members of the same assembly, whereas PLM-Interact readily ranks both above unrelated proteins. Representation analyses reveal that adaptation to PPI prediction can strengthen biological-context encoding without yielding transferable direct-contact specificity. AlphaFold3 shows greater, though still incomplete, direct-partner resolution, suggesting that explicit multimer modeling can help refine partner specificity. In PLM-Interact, negatives defined by reconstituted direct-binding, cellular binary/proximity and co-complex assays occupy distinct score regimes. Across independent datasets and multiple models, experimental direct non-binding and assembly noncontact further show reproducible, model-dependent scoring patterns. High sequence-based PPI scores should not be interpreted as direct-interface probabilities. Mechanistically interpretable evaluation instead requires explicit consideration of interaction-evidence semantics and partner specificity.