Research map / Probabilistic and scientific modeling / Molecules and proteins
Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number
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 PAFlow, an all-atom flow matching SBDD model with a new categorical conditional flow matching for atom types, protein-ligand affinity guidance, and a pocket-only learnable atom number predictor, reaching -8.31 Avg. Vina Score on CrossDocked2020.
Problem
Structure-based drug design needs 3D molecules with high binding affinity, but existing generators have unstable probability dynamics and sample atom counts from predefined distributions that mismatch pocket geometry.
Approach
Model atom coordinates with a VP flow-matching path and atom types with a newly derived categorical conditional flow matching; guide the coordinate vector field with a protein-ligand affinity predictor's gradient; and predict the ligand atom count from pocket descriptors (N_P, volume, surface area, space size) with noise injection.
Key results
- PAFlow reaches -8.31 Avg. Vina Score on CrossDocked2020 versus -7.07 for the best baseline ALiDiff, with 80.8% High Affinity.
- On Binding MOAD without retraining, PAFlow reaches -9.12 Avg. Vina Score and 83.4% High Affinity.
- With 20 sampling steps, PAFlow reports a 5.5x speedup over TargetDiff-style sampling (717s vs 3968s for 100 molecules) at the 50-step setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add protein-ligand interaction guidance during sampling (ablation) | PAFlow · CrossDocked2020 | Avg. Vina Score: -5.05 -> -8.10 | Helped |
| replace predefined atom-number sampling with the learned atom number predictor (ablation) | PAFlow w/o P · CrossDocked2020 | Avg. Vina Score: -4.49 -> -5.72 | Helped |
| add Gaussian noise with δ = 0.01 to predicted atom numbers (ablation) | PAFlow w/o P · CrossDocked2020 | Avg. Vina Score: -5.72 -> -6.17 | Helped |
| replace diffusion denoising with ODE-based flow matching sampling, remove prior guidance, and use predefined atom-number sampling (ablation) | TargetDiff · CrossDocked2020 | binding affinity | Helped |
| increase the number of sampling steps from 20 toward 200 (ablation) | PAFlow · CrossDocked2020 | Vina Score | Helped |
| increase guidance scale γ (ablation) | PAFlow · CrossDocked2020 | binding affinity | Helped |
| increase guidance scale γ (ablation) | PAFlow · CrossDocked2020 | QED | Hurt |
| add a small amount of Gaussian noise to predicted atom numbers (ablation) | PAFlow · CrossDocked2020 | molecular quality | Mixed |
Surprising findings
- Prior guidance lowers QED (0.53 to 0.49) in exchange for higher binding affinity, a trade-off the authors note.
- Adding small Gaussian noise (delta=0.01) to the predicted atom count improves affinity over the deterministic predictor.
Limitations
- Molecular properties such as QED and SA are not explicitly optimized and are slightly below the best baselines.
- No error bars are reported; results rely on 100 molecules per target.
- Affinity guidance trades off against QED as the guidance scale increases.
Experimental setup
Datasets: CrossDocked2020 Binding MOAD
Benchmarks: CrossDocked2020 Binding MOAD
Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity
Baselines: LiGAN AR Pocket2Mol TargetDiff DecompDiff IPDiff TAGMol ALiDiff MolCRAFT FlowSBDD
Builds on: flow matching TargetDiff conditional flow matching FlowSBDD
Closest papers by meaning
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.89)
- PPFLOW: Target-Aware Peptide Design with Torsional Flow Matching (ICML 2024 · similarity 0.89)
- Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.89)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.89)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.89)
- Generalized Protein Pocket Generation with Prior-Informed Flow Matching (NeurIPS 2024 · 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.