Performance Attribution in pLM-Based Biological Relation Prediction

Protein language models (pLMs) have enabled strong benchmark performance in biological relation-prediction tasks, but aggregate metrics do not identify which sources of information support that performance. We examined three case studies-MetaESI, DeepGNHV, and SAGEPhos-using frozen pLM full-input baselines, endpoint- or site-restricted controls, clean train-derived prior controls, and a restriction-matched selector control. Frozen pLM-derived inputs coupled to generic downstream learners reached AUROC/AUPRC of 0.827/0.703 for the current MetaESI full-input rerun, 0.922/0.708 for a DeepGNHV two-endpoint baseline, and 0.896/0.893 for SAGEPhos. Restricted controls retained task-dependent signal. Most notably, a self-label-excluding MetaESI endpoint-frequency control reached 0.846/0.676 under the row split, numerically close to the frozen ESM2 full-input reference despite using no sequence embeddings. Clean full-catalog DeepGNHV endpoint priors and SAGEPhos kinase/substrate/site-window priors provided supplementary diagnostics rather than direct architecture-contribution estimates. In a separate fixed frozen-pooling LightGBM experiment, GARD-selected pooling did not show higher observed performance than count-matched random token pooling. Endpoint-cold diagnostics showed performance degradation, while train-label shuffling returned discrimination to approximately chance; hard- or matched-negative and family- or homology-aware evaluations were not available across the case studies. These findings do not diagnose leakage, imply memorization, exclude biological learning, or invalidate the evaluated models. Rather, they show that benchmark utility and performance attribution are distinct: architecture-specific, relation-specific, selector-specific, and generalization claims require controls matched to the interpretation being made.