Research map / Probabilistic and scientific modeling / Molecules and proteins
Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks
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.
placeholder
Problem
Structure-based molecule optimization requires optimizing both continuous 3D coordinates and discrete atom types against protein targets; existing gradient guidance handles discrete types poorly and often guides only coordinates, causing inconsistencies across modalities.
Approach
Guide the Bayesian flow network's posterior parameters (theta) rather than noisy latents, using a Gaussian latent for discrete types derived from BFN so that joint gradients stay on the probability simplex and remain SE(3)-equivariant; add a backward correction that re-derives the Bayesian update over a sliding window of past steps.
Key results
- MolJO achieves Success Rate 51.3%, Vina Dock -9.05 and SA 0.78 on CrossDocked2020, about 4x the success rate of gradient-based TAGMol (11.1%).
- With top-of-N oracle selection (N=10), MolJO reaches 70.3% Success Rate and Vina Dock -10.50.
- Backward correction with k=130 improves guided Vina Dock from -9.11-level baseline vanilla guided -7.49 to -9.11 (B.C. guided).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply joint gradient guidance to coordinates and atom types rather than single-modality guidance (ablation) | MolJO · CrossDocked2020 molecule optimization | molecular properties and affinity | Helped |
| Guide coordinates alone instead of both coordinates and atom types (ablation) | MolJO · CrossDocked2020 molecule optimization | drug-like properties | Hurt |
| Guide atom types alone instead of both coordinates and atom types (ablation) | MolJO · CrossDocked2020 molecule optimization | affinity | Hurt |
| Add backward correction to guided sampling instead of Vanilla sampling (ablation) | MolJO · CrossDocked2020 guided sampling | Vina Dock Avg.: -7.49 -> -9.11 | Helped |
| Add backward correction to guided sampling instead of SDE sampling (ablation) | MolJO · CrossDocked2020 guided sampling | SA Avg.: 0.68 -> 0.77 | Helped |
| Apply gradient guidance with Vanilla sampling (ablation) | MolJO · CrossDocked2020 guided sampling | QED Avg.: 0.46 -> 0.46 | No clear effect |
| Apply gradient guidance with SDE sampling (ablation) | MolJO · CrossDocked2020 guided sampling | QED Avg.: 0.51 -> 0.50 | Hurt |
| Apply gradient guidance with backward correction (ablation) | MolJO · CrossDocked2020 guided sampling | Vina Dock Avg.: -7.95 -> -9.11 | Helped |
| Apply gradient guidance with backward correction (ablation) | MolJO · CrossDocked2020 guided sampling | SA Avg.: 0.69 -> 0.77 | Helped |
| Replace TAGMol's coordinate-only gradient guidance with MolJO joint guidance | TAGMol · CrossDocked2020 unconstrained | Success Rate: 11.1% -> 51.3% | Helped |
Surprising findings
- Without guidance, the backward-corrected sampler already improves over vanilla sampling (Vina Dock -7.95 vs -7.37)
- Gradient guidance on the vanilla sampler yields little gain (QED unchanged), suggesting suboptimal history limits guidance
Limitations
- Gradient guidance over discrete types relies on an approximated Gaussian latent and bypassed derivative for stability
- Top-of-N variant needs docking oracle calls, so the best numbers are not oracle-free
- Performance on large out-of-distribution molecules is reported only for a subset of baselines
Experimental setup
Datasets: CrossDocked2020
Benchmarks: CrossDocked2020
Metrics: Vina Score Vina Min Vina Dock QED SA Diversity Success Rate Connected ratio
Baselines: TAGMol MolCRAFT DecompOpt RGA AutoGrow4 DecompDiff IPDiff TargetDiff
Builds on: Bayesian Flow Networks MolCRAFT TargetDiff TAGMol classifier guidance
Closest papers by meaning
- A survey and benchmark of high-dimensional Bayesian optimization of discrete sequences (NeurIPS 2024 · similarity 0.88)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.87)
- Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.87)
- Approximation-Aware Bayesian Optimization (NeurIPS 2024 · similarity 0.87)
- Sample Efficient Generative Molecular Optimization with Joint Self-Improvement (NeurIPS 2026 · similarity 0.86)
- Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule (ICML 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.