BioGraphBench-PPI enables audit-first evaluation of negative-sampling contracts and structural baselines in protein interaction networks.

Introduction: Biomedical graph-learning results are difficult to compare when source versions, identifier rules, data splits, structural features, and sampled-nonedge contracts are incompletely documented. Methods: We developed BioGraphBench-PPI as an audit-first benchmark for protein-protein interaction (PPI) link prediction using STRING v12.0, BioGRID 5.0.260, and two cross-resource no-overlap ablations. For each PPI task, we generated 10 predefined seeds and three balanced sampled-nonedge contracts: random, degree-matched, and two-hop. We compared local heuristics, logistic regression, random forest, histogram gradient boosting (HGB), an auditable node2vec-compatible random-walk baseline, and a two-layer GCN link-prediction baseline, reporting ranking, calibration, runtime, and confidence intervals. Results: Random sampled nonedges yielded high HGB AUPRC values (0.942-0.964), but degree-matched sampled nonedges reduced AUPRC by 0.049-0.259. The GCN baseline showed signal under random negatives in selected graphs (AUPRC 0.526-0.717) but approached chance under degree-matched and two-hop contracts (AUPRC 0.500-0.515). Discussion: BioGraphBench-PPI shows that PPI benchmark conclusions depend strongly on sampled-nonedge design and graph source. The benchmark therefore contributes an auditable evaluation contract rather than a claim that PPI prediction is solved.