Research map / Probabilistic and scientific modeling / Molecules and proteins
PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces PRING, a graph-level benchmark for PPI prediction across four species that evaluates network reconstruction topology and function-oriented tasks, finding current models over-predict edges and fail to preserve functional structure.
Problem
Existing PPI prediction benchmarks evaluate isolated pairwise classification and ignore whether models can reconstruct biologically meaningful PPI networks with correct topology and function.
Approach
A multi-species, redundancy- and leakage-filtered PPI network dataset with two topology-oriented tasks (intra- and cross-species network construction) and three function-oriented tasks (complex pathway prediction, GO module analysis, essential protein justification), evaluating models by reconstructed graph properties rather than pairwise accuracy.
Key results
- PLM-interact (35M) and (650M) rank first and second on intra-species Human network construction; the highest graph similarity is below 0.5 (PLM-interact 650M BFS GS 0.396).
- Cross-species performance drops by 15.2% (Arath), 25.3% (Yeast), and 35.2% (Ecoli) relative to Human on average.
- In protein complex pathway prediction, PPITrans gets precision 0.863 but recall 0.583, indicating high precision with low recall; best essential protein DO is 0.440 (PLM-interact 650M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PLM-based models rather than naive sequence-based models | PPI prediction models · Cross-species PPI network construction | Graph similarity: 0.17; 0.24 | Helped |
| Train on Human and evaluate on increasingly evolutionarily distant species | PPI prediction models · Cross-species PPI network construction | Graph similarity: 15.2%; 25.3%; 35.2% | Hurt |
| Increase the positive-to-negative training-pair ratio to 1:5 (ablation) | PPITrans and TUnA · Human intra-species PPI network construction, BFS | Network reconstruction performance | Helped |
| Increase the positive-to-negative training-pair ratio from 1:5 to 1:10 (ablation) | PPITrans and TUnA · Human intra-species PPI network construction, BFS | Network reconstruction performance | Hurt |
| Increase the positive-to-negative training-pair ratio (ablation) | D-SCRIPT and Struct2Graph · Human intra-species PPI network construction, BFS | Relative density | Hurt |
| Use pairwise classification metrics in place of graph-level evaluation | PPI prediction evaluation · PPI network reconstruction | Network structure assessment | Hurt |
Surprising findings
- Current models tend to over-predict interactions (relative density above 1), producing overly dense graphs.
- Classification metrics do not fully reflect graph-level structure: higher recall worsens degree distribution mismatch while higher precision improves it.
Limitations
- Only four model organisms (Human, Arath, Ecoli, Yeast); no non-model or pathogen-host PPIs.
- All PPIs treated as binary undirected edges, ignoring interaction types and context-specific dynamics.
- RF2-Lite evaluated on only 10% of test graphs due to computational limits; PLM-interact used pretrained weights for inference only.
Experimental setup
Datasets: UniProt Swiss-Prot Reactome IntAct STRING Gene Ontology CCDS
Benchmarks: PRING
Metrics: Graph Similarity Relative Density Degree Distribution MMD Clustering Coefficient MMD Spectral MMD Pathway Precision Pathway Recall Pathway Connectivity Functional Alignment Consistency Ratio
Baselines: SPRINT PIPR D-SCRIPT Topsy-Turvy PPITrans TUnA PLM-interact (35M) PLM-interact (650M) Struct2Graph TAGPPI RF2-Lite
Builds on: D-SCRIPT PIPR Topsy-Turvy SPRINT PPITrans
Closest papers by meaning
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.87)
- MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding (ICLR 2024 · similarity 0.85)
- Learning to design protein-protein interactions with enhanced generalization (ICLR 2024 · similarity 0.85)
- PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.84)
- PSBench: a large-scale benchmark for estimating the accuracy of protein complex structural models (NeurIPS 2025 · similarity 0.83)
- Rethinking Cancer Gene Identification Through Graph Anomaly Analysis (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.