Research map / Graph learning / Graph self-supervised learning
REBIND: Enhancing Ground-state Molecular Conformation Prediction via Force-Based Graph Rewiring
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 REBIND, an encoder-decoder graph transformer that adds non-bonded edges chosen by Lennard-Jones force magnitude (degree-compensated) to improve ground-state 3D conformation prediction from 2D graphs.
Problem
Predicting ground-state 3D molecular conformations from 2D molecular graphs; existing GNN and graph transformer methods model inter-atomic forces poorly for non-bonded pairs, causing large errors on low-degree atoms.
Approach
Rewire the molecular graph by adding edges between non-bonded atom pairs with the largest Lennard-Jones forces (computed from the encoder's initial conformation), in a degree-compensating manner, split into attractive and repulsive adjacency matrices that are added as residual attention biases in the decoder.
Key results
- On QM9 test, REBIND reduces C-RMSD by 22.46% and E-RMSD by 23.75% versus the best baseline (GTMGC), with D-MAE 0.254 vs 0.281.
- On GEOM-DRUGS, REBIND achieves up to 4.25% C-RMSD and 8.06% E-RMSD reduction over the best baseline.
- Adding force-aware edges to GraphGPS (LP) on QM9 reduces C-RMSD by 5.11% and E-RMSD by 8.74%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use force-aware graph rewiring with Lennard-Jones interactions | REBIND · QM9 test | C-RMSD: 0.414 -> 0.321 | Helped |
| use force-aware graph rewiring with Lennard-Jones interactions | REBIND · Molecule3D random test | C-RMSD: 0.709 -> 0.699 | Helped |
| use force-aware graph rewiring with Lennard-Jones interactions | REBIND · Molecule3D scaffold test | E-RMSD: 1.275 -> 1.182 | Helped |
| use force-aware graph rewiring with Lennard-Jones interactions | REBIND · GEOM-DRUGS test | E-RMSD: 2.830 -> 2.602 | Helped |
| add bond adjacency to Laplacian positional embedding (ablation) | REBIND · QM9 | C-RMSD: 0.5294 -> 0.3524 | Helped |
| add force-aware edges instead of distance-based proximity edges (ablation) | REBIND · QM9 | C-RMSD: 0.3522 -> 0.3209 | Helped |
| add force-aware edges instead of distance-based proximity edges (ablation) | REBIND · QM9 | E-RMSD: 0.6807 -> 0.6103 | Helped |
| integrate force-based graph rewiring | GINE · QM9 test | C-RMSD: 0.867 -> 0.801 | Helped |
| integrate force-based graph rewiring | GATv2 · QM9 test | E-RMSD: 1.358 -> 1.198 | Helped |
| integrate force-based graph rewiring | GraphGPS (LP) · QM9 test | E-RMSD: 1.064 -> 0.971 | Helped |
Surprising findings
- Distance-based proximity edges (A^near) are outperformed by force-based edges, with up to 9% C-RMSD improvement.
- Existing GNNs show large error gaps between low-degree and high-degree atoms (up to 0.488 RMSD deviation), suggesting bond-only and distance-based modeling fails for non-bonded interactions.
Limitations
- Relies on UFF parameters and Lorentz-Berthelot mixing rules for LJ force computation, a fixed physical approximation.
- Evaluation focused on single ground-state conformation prediction; not validated for downstream property prediction or drug discovery applications, which are left as future work.
Experimental setup
Datasets: QM9 Molecule3D GEOM-DRUGS
Benchmarks: QM9 Molecule3D (random split) Molecule3D (scaffold split) GEOM-DRUGS
Metrics: D-MAE D-RMSE C-RMSD E-RMSD
Baselines: RDKit-DG RDKit-ETKDG GINE GATv2 DeeperGCN-DAGNN GraphGPS (RW) GraphGPS (LP) GTMGC Torsional Diffusion
Builds on: GTMGC GraphGPS Molecule3D benchmark pipeline Lennard-Jones potential Universal Force Field (UFF)
Closest papers by meaning
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.87)
- $\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials (NeurIPS 2024 · similarity 0.86)
- Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks (ICML 2024 · similarity 0.86)
- WGFormer: An SE(3)-Transformer Driven by Wasserstein Gradient Flows for Molecular Ground-State Conformation Prediction (ICML 2025 · similarity 0.86)
- Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.85)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.85)
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.