Research map / Probabilistic and scientific modeling / Molecules and proteins
Structure-based drug design by denoising voxel grids
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 VoxBind, a voxel-based score-based model for structure-based drug design that extends neural empirical Bayes to pocket-conditional walk-jump sampling, generating ligands as 3D atomic density grids.
Problem
Generating 3D ligand molecules conditioned on protein binding pockets (structure-based drug design) with high affinity, diversity, and few steric clashes, efficiently.
Approach
Extends neural empirical Bayes and walk-jump sampling to the conditional setting: a single-noise-level 3D U-Net conditional denoiser learns the smoothed conditional score; sampling alternates underdamped Langevin walks on noisy voxel grids with single-step denoising jumps, avoiding point-cloud diffusion and SE(3)-equivariant networks.
Key results
- VoxBind sigma=0.9 achieves mean VinaScore -6.94 and high-affinity rate 71.3% vs. 64.4% for DecompDiff (the best baseline) on CrossDocked2020 test set.
- VoxBind sigma=0.9 has mean steric clashes 5.1 vs 7.1 for DecompDiff and 10.8 for TargetDiff.
- VoxBind takes 492.2 s to generate 100 valid molecules on one A100, vs 3,428 s for TargetDiff and 6,189 s for DecompDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VoxBind σ=0.9 instead of DecompDiff | Pocket-conditional ligand generation · CrossDocked2020 test | VinaScore Avg.: -5.67 -> -6.94 | Helped |
| Use VoxBind σ=0.9 instead of DecompDiff | Pocket-conditional ligand generation · CrossDocked2020 test | VinaDock Avg.: -8.39 -> -8.30 | Hurt |
| Use VoxBind σ=0.9 instead of DecompDiff | Pocket-conditional ligand generation · CrossDocked2020 test | High affinity Avg. (%): 64.4 -> 71.3 | Helped |
| Use VoxBind σ=0.9 instead of Pocket2Mol | Pocket-conditional ligand generation · CrossDocked2020 test | QED Avg.: .56 -> .57 | Helped |
| Use VoxBind σ=0.9 instead of Pocket2Mol | Pocket-conditional ligand generation · CrossDocked2020 test | SA Avg.: .74 -> .70 | Hurt |
| Use VoxBind σ=1.0 instead of DiffSBDD | Pocket-conditional ligand generation · CrossDocked2020 test | Diversity Avg.: .73 -> .75 | Helped |
| Use VoxBind σ=0.9 instead of Pocket2Mol | Pocket-conditional ligand generation · CrossDocked2020 test | Strain energy median (kcal/mol): 205.8 -> 161.9 | Helped |
| Use VoxBind σ=0.9 instead of AR | Pocket-conditional ligand generation · CrossDocked2020 test | Steric clashes per ligand (mean): 4.2 -> 5.1 | Hurt |
| Use VoxBind instead of Pocket2Mol | Ligand sampling · Generate 100 valid molecules | Sampling time (seconds): 2,544 -> 492.2 | Helped |
| Increase noise level from σ=0.9 to σ=1.0 (ablation) | VoxBind · CrossDocked2020 test | VinaScore Avg.: -6.94 -> -6.63 | Hurt |
| Increase noise level from σ=0.9 to σ=1.0 (ablation) | VoxBind · CrossDocked2020 test | Diversity Avg.: .73 -> .75 | Helped |
| Add random translation and rotation data augmentation during training (ablation) | VoxBind σ=0.9 · CrossDocked2020 validation set | Median VinaScore across targets | Helped |
Surprising findings
- Non-equivariant 3D U-Net voxel denoiser competes with or beats SE(3)-equivariant point-cloud diffusion models on SBDD metrics.
- Denoising at a single fixed noise level suffices; the noise process does not displace atoms, which helps avoid clashes.
Limitations
- 3D U-Net memory cost bounds the volume that can be processed, limiting scaling to larger molecules such as nucleic acids and proteins.
- Does not explicitly model synthetic accessibility; SA is lower than Pocket2Mol.
- Relies on OpenBabel for bond assignment and peak detection for atom coordinate recovery.
Experimental setup
Datasets: CrossDocked2020
Benchmarks: CrossDocked2020
Metrics: VinaScore VinaMin VinaDock high affinity QED SA diversity atoms per molecule steric clashes strain energy
Baselines: AR Pocket2Mol DiffSBDD TargetDiff DecompDiff
Builds on: neural empirical Bayes walk-jump sampling VoxMol TargetDiff BAOAB
Closest papers by meaning
- Score-based 3D molecule generation with neural fields (NeurIPS 2024 · similarity 0.89)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.87)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.87)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.87)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.87)
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.