Research map / Label-efficient and robust learning / Recognition and long-tail
Learning to design protein-protein interactions with enhanced generalization
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.
Builds PPIRef, a non-redundant 50K-interface PPI dataset deduplicated with the new iDist algorithm, pre-trains an SE(3)-equivariant Equiformer-based PPIformer via masked residue modeling, and fine-tunes it for ΔΔG prediction with a log-odds-ratio formulation on leak-free SKEMPI v2.0 splits.
Problem
Predicting how mutations affect protein-protein binding affinity (ΔΔG) generalizes poorly because PPI datasets are redundant and standard train-test splits (e.g., on SKEMPI v2.0 and DIPS) leak near-duplicate interfaces, so reported performance is overestimated.
Approach
Combine a fast structural-deduplication embedding (iDist) to build a non-redundant large PPI pre-training corpus, pre-train a coarse residue-level SE(3)-equivariant transformer with masked amino acid prediction, and reuse the pre-training log-likelihood as a thermodynamically motivated log-odds ratio for zero-shot and fine-tuned ΔΔG prediction.
Key results
- On 5 non-leaking SKEMPI v2.0 test PPIs, PPIformer reaches Spearman 0.44±0.03 vs 0.38 for GEMME and 0.24 for RDE-Network, a claimed 183% relative improvement over RDE-Network in ranking.
- On the SARS-CoV-2 antibody case study, PPIformer achieves P@1 = 100%, while the best baselines do not place a favorable mutation at rank 1.
- iDist is about 480 times faster than iAlign on 1,646 PPIs (15 s vs 2 h on 128 CPUs) with 99% precision and 97% recall for near-duplicate detection; PPIRef50K is deduplicated from 322K to 45K interfaces.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PPIformer instead of supervised RDE-Network | SKEMPI v2.0 held-out PPIs · SKEMPI v2.0 held-out PPIs | Spearman: 0.24 -> 0.44 ± 0.03 | Helped |
| Use PPIformer instead of Flex ddG | SKEMPI v2.0 held-out PPIs · SKEMPI v2.0 held-out PPIs | Spearman: $0.55$ -> 0.44 ± 0.03 | Hurt |
| Use PPIformer to rank SARS-CoV-2 antibody mutations | Human antibody mutation retrieval · SARS-CoV-2 antibody mutation retrieval | P@1: 0.00 -> 100 | Helped |
| Use PPIformer instead of RDE-Network for SARS-CoV-2 antibody mutation retrieval | Human antibody mutation retrieval · SARS-CoV-2 antibody mutation retrieval | P@10%: 6.12 -> 4.08 | Hurt |
| Use PPIformer instead of RDE-Network for staphylokinase mutation retrieval | Staphylokinase mutations · Staphylokinase mutation retrieval | P@5%: 50.00 -> 75.00 | Helped |
| Use PPIformer instead of RDE-Network for staphylokinase mutation retrieval | Staphylokinase mutations · Staphylokinase mutation retrieval | P@10%: 62.50 -> 87.50 | Helped |
| Use PPIformer instead of MSA Transformer for staphylokinase mutation retrieval | Staphylokinase mutations · Staphylokinase mutation retrieval | P@1: 100 -> 100 | No clear effect |
Surprising findings
- Flex ddG force-field simulation still outperforms machine-learning methods on the non-leaking SKEMPI split (Spearman 0.55), contrary to earlier claims that deep learning surpasses force fields.
- Standard PDB-disjoint SKEMPI splits leak heavily: 57% of test PPIs have a homologous PPI in training, so reported ML performance is overestimated.
Limitations
- Non-leaking evaluation uses only 5 held-out test PPIs, so variance is high.
- On the SARS-CoV-2 case study PPIformer does not rank the 5 annotated mutations as highly as DDGPred and RDE-Network in top-10% precision.
- Coarse residue-level representation ignores side-chain rotamer details.
Experimental setup
Datasets: PPIRef50K PPIRef300K PPIRef800K DIPS SKEMPI v2.0
Benchmarks: SKEMPI v2.0 non-leaking 5-PPI test folds SARS-CoV-2 antibody optimization case study (P36-5D2) staphylokinase thrombolytic activity case study
Metrics: Spearman Pearson precision recall ROC AUC MAE RMSE P@1 P@5% P@10%
Baselines: Flex ddG GEMME MSA Transformer ESM-IF RDE-Network Rosetta FoldX DDGPred
Builds on: Equiformer iAlign RDE-Network ESM-IF
Closest papers by meaning
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.86)
- MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding (ICLR 2024 · similarity 0.85)
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs (NeurIPS 2025 · similarity 0.85)
- Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning (ICML 2024 · similarity 0.84)
- A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer (ICLR 2025 · similarity 0.83)
- PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · 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.