Research map / Probabilistic and scientific modeling / Molecules and proteins
Boosting Protein Graph Representations through Static-Dynamic Fusion
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 heterogeneous protein graphs combining distance edges and MD-derived residue/atom motion-correlation edges, and processes them with relational GNNs (RGCN, RGAT and relational variants) for adaptability, binding site, and affinity prediction on MISATO.
Problem
Protein graph representations for machine learning rely only on static structure, ignoring dynamic information available from molecular dynamics trajectories, which limits property prediction for proteins and protein-ligand complexes.
Approach
Augment static distance graphs with correlation edges computed from MD trajectory displacement correlations (after RMSD alignment), forming a heterogeneous graph whose two edge types are handled by relation-specific message passing (RGCN/RGAT) to fuse static and dynamic signals.
Key results
- Combined graph with RGCN raises atomic adaptability Pearson R from 0.5259 (distance) to 0.7326.
- Binding site F1 with R-EGNN rises from 0.3617 (distance) to 0.4142 (combined).
- Binding affinity with R-SS-GNN: combined graph MAE 1.0874 vs 1.2378 for distance graph.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse distance and correlation edges instead of using distance edges alone | Protein graph models · MISATO atomic adaptability | Pearson R: 0.5259 -> 0.7326 | Helped |
| Use correlation edges instead of distance edges | RGCN · MISATO atomic adaptability | MAE: 0.2658 -> 0.2311 | Helped |
| Fuse distance and correlation edges instead of using distance edges alone | Protein graph models · MISATO binding site detection | F1 score: 0.2428 -> 0.2834 | Helped |
| Fuse distance and correlation edges instead of using distance edges alone | R-EGNN · MISATO binding site detection | F1 score: 0.3617 -> 0.4142 | Helped |
| Fuse distance and correlation edges instead of using distance edges alone | R-SS-GNN · PDBbind 2020 binding affinity | Pearson R: 0.7638 -> 0.7873 | Helped |
| Fuse distance and correlation edges instead of using distance edges alone | R-SS-GNN · PDBbind 2020 binding affinity | MAE: 1.2378 -> 1.0874 | Helped |
| Add correlation edges to distance edges | R-EGNN · MISATO atomic adaptability | MAE: 0.2305 -> 0.2321 | Hurt |
| Align trajectory frames before computing correlation edges (ablation) | RGCN with Correlation Graph · Atomic adaptability | Pearson R: 0.5493 -> 0.6426 | Helped |
| Align trajectory frames before computing correlation edges (ablation) | RGCN with Combined Graph · Atomic adaptability | MAE: 0.2192 -> 0.1981 | Helped |
| Align trajectory frames before computing correlation edges (ablation) | RGAT with Correlation Graph · Atomic adaptability | Pearson R: 0.5001 -> 0.6106 | Helped |
| Align trajectory frames before computing correlation edges (ablation) | RGAT with Combined Graph · Atomic adaptability | Pearson R: 0.6633 -> 0.7153 | Helped |
Surprising findings
- Correlation graph alone is worse than distance graph for binding affinity prediction with most architectures.
- R-EGNN with the combined graph performs similarly to the distance graph on adaptability, which the authors attribute to fusion limitations.
Limitations
- Relies on MD trajectories being available, which are not always present for new proteins.
- Fusion mechanism for equivariant architectures (R-EGNN) is preliminary and underperforms on adaptability.
- Only one correlation measure (displacement cosine correlation) is explored.
Experimental setup
Datasets: MISATO PDBbind 2020
Benchmarks: MISATO atomic adaptability MISATO binding site detection PDBbind 2020 binding affinity
Metrics: MAE RMSE Pearson R Spearman R accuracy precision recall F1 score
Baselines: Distance Graph Correlation Graph EGNN-based R-EGNN R-GPS SS-GNN
Builds on: RGCN Relational Graph Attention Network EGNN GPS graph transformer SS-GNN
Closest papers by meaning
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.89)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.88)
- Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure (NeurIPS 2024 · similarity 0.87)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.87)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.87)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
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.