Research map / Probabilistic and scientific modeling / Molecules and proteins
PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction
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.
Proposes PSC-CPI, a multi-scale protein sequence-structure contrastive pre-training framework with length-variable augmentation that improves compound-protein interaction prediction and allows single-modality inference.
Problem
Compound-protein interaction prediction suffers when only one protein modality (sequence or structure) is available at inference and when test compounds/proteins are unseen in training, causing domain shift.
Approach
Jointly pre-trains sequence and structure encoders with intra-modality and cross-modality contrastive losses over length-variable protein subsequences/subgraphs, so single-modality encoders retain sequence-structure dependencies at inference.
Key results
- On Karimi with multimodal inference, PSC raises average AUPRC from 16.14 to 18.19 and lowers average RMSE from 1.56 to 1.43 across four splits.
- Under the Unseen-Both setting on Davis, KIBA, and Metz, PSC-CPI reports MSE of 0.455, 0.490, and 0.595, and is best on 5 of 6 metrics versus PerceiverCPI.
- Pre-training with PSC improves unimodal (sequence-only or structure-only) inference, with the best unimodal results comparable to or exceeding multimodal baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intra-modality contrastive learning (ablation) | PSC-CPI · Karimi, four test splits | AUPRC: 25.36 vs 25.42; 24.47 vs 24.67; 10.46 vs 11.03; 11.30 vs 11.65 | Helped |
| add intra-modality contrastive learning (ablation) | PSC-CPI · Karimi, four test splits | RMSE: 1.44 vs 1.42; 1.32 vs 1.31; 1.50 vs 1.47; 1.54 vs 1.52 | Helped |
| add cross-modality contrastive learning (ablation) | PSC-CPI · Karimi, four test splits | AUPRC: 25.06 vs 25.42; 23.95 vs 24.67; 10.88 vs 11.03; 10.34 vs 11.65 | Helped |
| add cross-modality contrastive learning (ablation) | PSC-CPI · Karimi, four test splits | RMSE: 1.47 vs 1.42; 1.36 vs 1.31; 1.48 vs 1.47; 1.58 vs 1.52 | Helped |
| use length-variable rather than length-fixed augmentation (ablation) | PSC-CPI · Karimi, four test splits | AUPRC: 24.69 vs 25.42; 24.00 vs 24.67; 10.69 vs 11.03; 10.41 vs 11.65 | Helped |
| use length-variable rather than length-fixed augmentation (ablation) | PSC-CPI · Karimi, four test splits | RMSE: 1.49 vs 1.42; 1.37 vs 1.31; 1.53 vs 1.47; 1.60 vs 1.52 | Helped |
Surprising findings
- Cross-modality contrasting contributes more to performance than the multi-scale (length-variable) augmentation, suggesting sequence-structure dependency matters more than multi-scale information.
Limitations
- Multi-scale modeling is restricted to the residue level and does not cover atomic scale.
- Efficiency was not explored in depth.
- Pre-training relies on a small number of known sequence-structure pairs; protein graphs for Davis, KIBA, and Metz are predicted with RaptorX-Contact rather than experimental structures.
Experimental setup
Datasets: Karimi Davis KIBA Metz PDBbind BindingDB
Benchmarks: Karimi Davis KIBA Metz
Metrics: AUPRC RMSE MSE CIndex
Baselines: DeepConvDTI GraphDTA DeepAffinity+ HyperattentionDTI TransformerCPI PerceiverCPI Cross-Interaction
Builds on: SimCLR JOAO TransformerCPI PerceiverCPI Cross-Interaction
Closest papers by meaning
- Sequence-Free for Compound Protein Interaction Prediction (AAAI 2026 · similarity 0.92)
- PSBench: a large-scale benchmark for estimating the accuracy of protein complex structural models (NeurIPS 2025 · similarity 0.86)
- Protein Multimer Structure Prediction via Prompt Learning (ICLR 2024 · similarity 0.86)
- ProtCLIP: Function-Informed Protein Multi-Modal Learning (AAAI 2025 · similarity 0.85)
- Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.85)
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs (NeurIPS 2025 · similarity 0.84)
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.