Research map / Probabilistic and scientific modeling / Molecules and proteins
MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space
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 MolCRAFT, an SE(3)-equivariant Bayesian Flow Network that generates ligands in a fully continuous parameter space with a noise-reduced sampler, improving binding affinity and conformational stability for structure-based drug design on CrossDock.
Problem
Structure-based drug design generative models produce molecules with distorted 3D conformations and inferior binding poses ("false positives"), due to mode collapse in autoregressive models and the hybrid continuous-discrete space in diffusion models.
Approach
Apply Bayesian Flow Networks to jointly model continuous atom coordinates and discrete atom types in a continuous parameter space, and use a noise-reduced sampler that updates parameters directly from the predicted clean molecule, avoiding sampling noisy discrete variables at each step.
Key results
- Achieves reference-level Vina Score of -6.59 kcal/mol vs reference -6.36 kcal/mol, beating the best baseline by 0.84 kcal/mol in the size-matched comparison.
- Ranks first on bond length and angle JSD (0.319 and 0.379) among compared methods.
- Sampling is about 30x faster than TargetDiff and DecompDiff (141 s vs 3428 s and 6189 s for 100 samples).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use fully continuous parameter-space generation instead of hybrid continuous-discrete diffusion | MolCRAFT · CrossDocked test, reference-size | Vina Score (Avg.): -5.19 -> -6.59 | Helped |
| Use fully continuous parameter-space generation instead of hybrid continuous-discrete diffusion | MolCRAFT · CrossDocked test, reference-size | Length JSD: 0.348 -> 0.319 | Helped |
| Use noise-reduced parameter-space sampling instead of original sampling (ablation) | MolCRAFT · CrossDocked test, same checkpoint | Vina Score: -5.42 -> -6.51 | Helped |
| Use noise-reduced parameter-space sampling instead of original sampling (ablation) | MolCRAFT · CrossDocked test, same checkpoint | Vina Min: -6.30 -> -7.13 | Helped |
| Use 100 sampling steps instead of original 1000 training steps (ablation) | MolCRAFT · CrossDocked test | Generation time: 10× faster | Helped |
| Use noise-reduced parameter-space sampling (ablation) | MolCRAFT · CrossDocked test, 100 generated samples | Generation time: 141s; 30 × speedup | Helped |
| Use noise-reduced parameter-space sampling (ablation) | MolCRAFT · CrossDocked test, 100 test proteins | Generation success: 96.7% | Helped |
| Generate molecules at larger size | MolCRAFT · CrossDocked test, larger-size | Vina Score (Avg.): -5.67 -> -6.61 | Helped |
Surprising findings
- Sampling discrete atom types at each denoising step adds harmful noise; the noise-reduced sampler yields better molecules with fewer steps.
- The model's RMSD to redocked poses exceeds the reference, which the authors attribute to a train/test distribution shift in the fraction of docked poses.
Limitations
- Vina Dock scores can be inflated by larger molecules, so comparisons require size-controlled settings.
- Evaluated only on CrossDocked with docking-based metrics, which are imperfect proxies for binding.
Experimental setup
Datasets: CrossDocked
Benchmarks: CrossDocked test set (100 proteins)
Metrics: Vina Score Vina Min Vina Dock strain energy clash RMSD JSD SA QED diversity
Baselines: AR Pocket2Mol FLAG TargetDiff Decomp-O Decomp-R
Builds on: Bayesian Flow Networks GeoBFN TargetDiff DecompDiff EDM
Closest papers by meaning
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.87)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.87)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.86)
- DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.86)
- Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule (ICML 2025 · similarity 0.86)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · 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.