PRING — From Protein Pairs to Interaction Networks

Evaluation Tasks

• Topology · Intra-species: Reconstruct Human subgraphs and assess similarity in density, degree distribution, clustering coefficient, and spectrum.
• Topology · Cross-species: Train on Human and transfer to Arath, Yeast, and E. coli to test structural recovery across species.
• Function · Complexes/Pathways: Predict subgraphs and evaluate pathway precision, recall, and connectivity.
• Function · GO Modules: Detect enriched communities and assess functional alignment and consistency.
• Function · Essential Proteins: Identify key proteins using centrality measures and evaluate precision@100 and distribution overlap.

Dataset at a Glance

• Sources: UniProt, Reactome, IntAct, and STRING (confidence > 0.7).
• Quality Control: Only SwissProt-annotated proteins, mapped to four species using NCBI Taxonomy.
• De-redundancy & Leakage Prevention: MMseqs2 with ≤40% sequence identity, functional ID filtering, and non-overlapping protein splits.

Why It Matters

Traditional binary classification metrics cannot fully capture the structural consistency of networks. PRING introduces graph-aware evaluations that better reflect the requirements of real biological analyses.

Reference

Zheng X.*, Du H.*, Xu F.*, Li J.*, Liu Z.†, Wang W., Chen T., Ouyang W., Stan Z. Li, Lu Y.†, Dong N.†, Zhang Y.†.
PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs. NeurIPS 2025 Datasets & Benchmarks.
[Paper] [Code] [Data]