Research map / Probabilistic and scientific modeling / Molecules and proteins
Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration
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 CByG, a gradient-guided conditional Bayesian Flow Network for 3D structure-based drug design, plus a benchmark suite extending evaluation to multiple docking tools, PoseBusters, AiZynthFinder synthesizability and kinase selectivity.
Problem
Existing SBDD generative models are judged mainly by docking affinity and SA score, and diffusion-based gradient guidance struggles with hybrid continuous/categorical 3D molecular data; selectivity and realistic synthesizability are largely unevaluated.
Approach
Reformulates BFN parameter updates as gradient-based Bayesian updates (score-like terms for continuous coordinates and softmax-based terms for categorical atom types), enabling property guidance in parameter space with an uncertainty-aware BNN property predictor, and pairs it with a broader evaluation scheme.
Key results
- CByG achieves 93.6% average high-affinity rate versus 76.6% for the best baseline (TargetOpt), with Vina dock score -9.16.
- CByG reaches SA 0.84 average and 94.9% PB-valid, outperforming baselines in those columns.
- On the kinase selectivity test, CByG reaches 78.3% success rate versus 68.6% for TargetOpt and 58.2% for TargetDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CByG gradient-based Bayesian guidance instead of diffusion-based guidance | 3D molecular generation for structure-based drug design · CrossDocked test proteins | Guidance score stability | Helped |
| Add guidance injection to CByG (ablation) | CByG · Kinase selectivity test set | Success rate: 62.6 % -> 78.3 % | Helped |
| Add guidance injection to CByG (ablation) | CByG · Kinase selectivity test set | Delta score: -1.39 | Helped |
| Use CByG rather than the strongest listed baseline | 3D molecular generation for structure-based drug design · CrossDocked test proteins | Diversity: 0.73 -> 0.61 | Hurt |
| Use CByG rather than the strongest listed baseline | 3D molecular generation for structure-based drug design · CrossDocked test proteins | SMINA score: -6.03 -> -7.74 | Helped |
| Use CByG rather than the strongest listed baseline | 3D molecular generation for structure-based drug design · CrossDocked test proteins | High affinity average: 76.6% -> 93.6% | Helped |
Surprising findings
- High SA scores do not reliably indicate retrosynthetic feasibility; models with high SA scores had retrosynthesis success below 50%.
- Roughly half of generated molecules, even from CByG, remain synthetically infeasible under AiZynthFinder.
Limitations
- Generated molecules remain largely synthetically infeasible by retrosynthesis evaluation
- Lower molecular diversity than some baselines, attributed to conditional generation narrowing distributions
- Selectivity evaluation relies on a small set of 38 kinase inhibitors
Experimental setup
Datasets: CrossDocked2020
Benchmarks: CrossDocked2020 test (100 proteins) PoseBusters AiZynthFinder PoseCheck kinase selectivity set
Metrics: SMINA score GNINA score Vina score high affinity rate SA score PB-valid rate diversity retrosynthesis solved rate selectivity success rate delta score
Baselines: AR Pocket2Mol TargetDiff DecompDiff MolCRAFT RGA DecompOpt TacoGFN ALIDiff TargetOpt
Builds on: Bayesian Flow Networks TargetDiff classifier guidance Tweedie's formula
Closest papers by meaning
- Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.90)
- Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.90)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.90)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.89)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.89)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.89)
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.