Research map / Probabilistic and scientific modeling / Molecules and proteins
Physics Aware Neural Networks for Unsupervised Binding Energy 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 CEBind, an unsupervised protein-ligand binding energy predictor trained by applying randomly sampled rigid-body forces to the ligand and enforcing energy conservation plus force matching, reporting higher Pearson correlation and lower memory than DSMBind.
Problem
Predicting protein-ligand binding energy without affinity labels, especially for antibody-antigen complexes where labeled data is scarce.
Approach
Samples forces on ligand atoms, applies them rigidly to perturb the ligand, and trains an energy model so the energy gap between unperturbed and perturbed complexes equals the work of outer forces (energy conservation), with an added force-matching loss; rigid-dynamics perturbations are precomputed in the dataloader without gradients, making training memory-efficient.
Key results
- CEBind reaches Pearson r = 0.652 ± 0.005 on protein-small molecule binding, versus 0.644 for DSMBind (unsupervised SOTA).
- On antibody-antigen binding CEBind achieves r = 0.374 ± 0.009, above DSMBind (0.365) and the best supervised FANN_transfer (0.350).
- CEBind uses 4.89 GB peak GPU memory versus 6.43 GB for DSMBind, with per-batch training time 0.0116 s vs 0.0152 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add energy conservation loss (ablation) | CEBind · PDBbind core set | Pearson correlation: 0.646 ± 0.001 -> 0.652 ± 0.005 | Helped |
| add energy conservation loss (ablation) | CEBind · Antibody-antigen test set | Pearson correlation: 0.362 ± 0.025 -> 0.374 ± 0.009 | Helped |
| add force matching loss (ablation) | CEBind · PDBbind core set | Pearson correlation: 0.646 ± 0.002 -> 0.652 ± 0.005 | Helped |
| add force matching loss (ablation) | CEBind · Antibody-antigen test set | Pearson correlation: 0.361 ± 0.016 -> 0.374 ± 0.009 | Helped |
| remove ESM module (ablation) | CEBind · PDBbind core set | Pearson correlation: 0.652 ± 0.005 -> 0.650 ± 0.004 | No clear effect |
| add ESM module (ablation) | CEBind · Antibody-antigen test set | Pearson correlation: 0.346 ± 0.010 -> 0.374 ± 0.009 | Helped |
| use a general atom-only ligand encoder (ablation) | CEBind · PDBbind core set | Pearson correlation: 0.652±0.005 -> 0.648±0.003 | No clear effect |
| use a general atom-only ligand encoder (ablation) | CEBind · Antibody-antigen test set | Pearson correlation: 0.374±0.009 -> 0.346±0.004 | Hurt |
| apply forces to both ligand and protein instead of ligand only (ablation) | CEBind · PDBbind core set | Pearson correlation: 0.652±0.005 -> 0.665±0.011 | No clear effect |
| apply forces to both ligand and protein instead of ligand only (ablation) | CEBind · Antibody-antigen test set | Pearson correlation: 0.374±0.009 -> 0.362±0.015 | No clear effect |
| train CEBindatom on antibody-antigen data and evaluate on protein-small molecule data (ablation) | CEBindatom · PDBbind core set | Pearson correlation: 0.644±0.003 -> 0.646±0.001 | Helped |
| train CEBindatom on protein-small molecule data and evaluate on antibody-antigen data (ablation) | CEBindatom · Antibody-antigen test set | Pearson correlation: 0.346±0.004 -> 0.320±0.002 | Hurt |
Surprising findings
- The unsupervised antibody-antigen model outperforms supervised models, which are limited by scarce labels.
- Atom-only general encoder without ESM performs comparably on protein-small molecule data but worse on antibody-antigen data.
Limitations
- Forces are applied only to ligands, which may introduce bias; applying forces to both sides gave mixed results.
- Protein-small molecule performance remains below supervised methods such as IGN and PLANET.
- Antibody-antigen binding relies on the ESM module; the atom-only variant degrades on that dataset.
Experimental setup
Datasets: PDBbind V2020 PDBbind core set SAbDab CSM-sb SKEMPI
Benchmarks: PDBbind core set (protein-small molecule) SAbDab test (antibody-antigen)
Metrics: Pearson correlation
Baselines: DSMBind Gauss DSM Contrastive learning ESM-1v ESM-IF MM/GBSA FANN PLANET IGN
Builds on: DSMBind denoising score matching Frame Averaging
Closest papers by meaning
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.85)
- Gradual Optimization Learning for Conformational Energy Minimization (ICLR 2024 · similarity 0.85)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.84)
- Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.84)
- UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules (ICML 2025 · similarity 0.84)
- DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction (NeurIPS 2025 · 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.