Research map / Probabilistic and scientific modeling / Molecules and proteins
Towards Stable Representations for Protein Interface 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 ATProt, an adversarial-training framework with a stability regularizer on spectral (Bernstein) graph encoders that makes protein representations robust to bound-unbound flexibility for protein interface prediction.
Problem
Protein interface prediction degrades when models trained on bound structures are tested on unbound or predicted structures, because conformational flexibility perturbs representations.
Approach
Treat protein flexibility as an adversarial perturbation and enforce Lipschitz stability of spectral graph encoders via a differentiable regularizer; for BernNet, bound the maximum absolute slope of the filter by max |theta_i - theta_{i+1}| times K.
Key results
- On DB5.5 with ESMFold structures, ATProt-Bern reaches 0.929 MedAUC versus 0.906 for the best baseline dMaSIF.
- After DIPS pre-training and DB5.5 fine-tuning, ATProt-Bern achieves 0.937 MedAUC with ESMFold and 0.928 with AlphaFold2 inputs.
- Without the stability regularization, ATProt-Bern drops from 0.928 to 0.901 on Native-Unbound (DB5.5 training).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply ATProt-Bern instead of BIPSPI | Protein interface prediction · DB5.5 Native-Bound | MedAUC: 0.937 (0.008) -> 0.932 (0.017) | Hurt |
| apply ATProt-Bern instead of BIPSPI | Protein interface prediction · DB5.5 Native-Unbound | MedAUC: 0.911 (0.017) -> 0.928 (0.014) | Helped |
| apply ATProt-Bern instead of BIPSPI | Protein interface prediction · DB5.5 ESMFold | MedAUC: 0.896 (0.008) -> 0.929 (0.014) | Helped |
| apply ATProt-Bern instead of BIPSPI | Protein interface prediction · DB5.5 AlphaFold2 | MedAUC: 0.887 (0.013) -> 0.925 (0.011) | Helped |
| add stable regularization to ATProt-Bern (ablation) | ATProt-Bern · DB5.5 Native-Bound | MedAUC: 0.934 (0.009) -> 0.932 (0.017) | No clear effect |
| add stable regularization to ATProt-Bern (ablation) | ATProt-Bern · DB5.5 Native-Unbound | MedAUC: 0.901 (0.011) -> 0.928 (0.014) | Helped |
| add stable regularization to ATProt-Bern (ablation) | ATProt-Bern · DB5.5 ESMFold | MedAUC: 0.897 (0.010) -> 0.929 (0.014) | Helped |
| add stable regularization to ATProt-Bern (ablation) | ATProt-Bern · DB5.5 AlphaFold2 | MedAUC: 0.901 (0.012) -> 0.925 (0.011) | Helped |
| add stable regularization to ATProt (ablation) | ATProt · DIPS Native-Bound | MedAUC: 0.909 (0.013) -> 0.909 (0.007) | No clear effect |
| add stable regularization to ATProt (ablation) | ATProt · DIPS ESMFold | MedAUC: 0.882 (0.013) -> 0.906 (0.007) | Helped |
Surprising findings
- Native-Bound performance of ATProt is slightly below BIPSPI (0.932 vs 0.937 for the best baseline, BIPSPI), despite the method being designed for flexibility robustness.
- Baseline methods, especially NEA and SASNet, show clear drops when moving from bound to unbound or predicted structures.
Limitations
- Does not incorporate adversarial-training-based classifiers into the framework.
- Native-bound performance is slightly below BIPSPI.
- DB5.5 contains only 253 complexes, limiting evaluation of unbound structures.
Experimental setup
Datasets: DB5.5 DIPS
Benchmarks: DB5.5 DIPS
Metrics: MedAUC
Baselines: BIPSPI SASNet dMaSIF DTNN NEA
Builds on: BernNet NEA SGC ChebNet induced fit flexibility
Closest papers by meaning
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.86)
- Flexibility-conditioned protein structure design with flow matching (ICML 2025 · similarity 0.86)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.85)
- SaProt: Protein Language Modeling with Structure-aware Vocabulary (ICLR 2024 · similarity 0.84)
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.83)
- Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property 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.