Research map / Probabilistic and scientific modeling / Molecules and proteins
Generalized Protein Pocket Generation with Prior-Informed Flow Matching
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 PocketFlow, a flow-matching model for protein pocket co-design that learns protein-ligand interaction types and uses affinity and interaction-geometry guidance at sampling, evaluated on CrossDocked, Binding MOAD, PPDBench and PDBBind RNA.
Problem
Designing protein pockets that bind ligands with high affinity and structural validity; existing physics-based and template methods are slow, while deep generative methods ignore protein-ligand interaction priors and mostly handle only small molecules.
Approach
Multimodal Riemannian flow matching over residue frames, sidechain torsions, residue types and interaction types, with a lightweight affinity predictor and distance/angle interaction-geometry guidance applied during ODE sampling, using a sidechain ensemble to make residue-type sampling differentiable.
Key results
- On CrossDocked, PocketFlow reaches 52.19% AAR, 0.67 scRMSD and -8.236 Vina versus -7.012 Vina for RFDiffusionAA, a average improvement of 1.29 Vina over RFDiffusionAA across datasets.
- Average steric clashes in generated pockets drop to 1.21 versus 4.59 in the test set and 3.58 for RFDiffusionAA, with 4.12 hydrogen bonds on average.
- On peptide and RNA benchmarks PocketFlow is comparable to RFDiffusionAA (e.g., 48.19% vs 46.85% AAR on PPDBench) with ddG improvements, using a model pretrained only on small-molecule data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PocketFlow instead of RFDiffusionAA | small-molecule pocket design · CrossDocked | Vina: -7.012±0.09 -> -8.236±0.16 | Helped |
| use PocketFlow instead of RFDiffusionAA | small-molecule pocket design · Binding MOAD | Vina: -8.020±0.11 -> -9.370±0.24 | Helped |
| use PocketFlow instead of RFDiffusionAA | peptide-binding pocket design · PPDBench | ∆∆G: -0.62±0.05 -> -1.06±0.04 | Helped |
| use PocketFlow instead of RFDiffusionAA | RNA-binding pocket design · PDBBind RNA | ∆∆G: -0.45±0.07 -> -0.78±0.07 | Helped |
| add affinity guidance (ablation) | PocketFlow · CrossDocked | Vina: -7.375±0.10 -> -8.236±0.16 | Helped |
| add affinity guidance (ablation) | PocketFlow · Binding MOAD | Vina: -8.380±0.19 -> -9.370±0.24 | Helped |
| add interaction geometry guidance (ablation) | PocketFlow · CrossDocked | Vina: -8.120±0.14 -> -8.236±0.16 | Helped |
| add interaction geometry guidance (ablation) | PocketFlow · Binding MOAD | Vina: -9.197±0.22 -> -9.370±0.24 | Helped |
| add affinity and interaction geometry guidance (ablation) | PocketFlow · CrossDocked | Vina: -7.135±0.13 -> -8.236±0.16 | Helped |
| add affinity and interaction geometry guidance (ablation) | PocketFlow · Binding MOAD | Vina: -8.241±0.18 -> -9.370±0.24 | Helped |
| add interaction-type learning and interaction geometry guidance (ablation) | PocketFlow · CrossDocked | Vina: -7.968±0.15 -> -8.236±0.16 | Helped |
| add interaction-type learning and interaction geometry guidance (ablation) | PocketFlow · Binding MOAD | Vina: -9.031±0.17 -> -9.370±0.24 | Helped |
Surprising findings
- Affinity guidance slightly worsens scRMSD, and larger guidance strength gamma increases scRMSD while improving Vina score.
- Without learning interaction types, peptide and RNA generalization drops markedly (e.g., PPDBench AAR 36.41% vs 48.19%).
Limitations
- Trained only on protein-small molecule datasets; peptide and RNA generalization is zero-shot with fixed ligand structures.
- Affinity guidance can hurt structural validity (scRMSD) and high guidance strength reduces sequence/structure quality.
- No wet-lab validation of generated pockets.
Experimental setup
Datasets: CrossDocked Binding MOAD PPDBench PDBBind RNA
Benchmarks: CrossDocked Binding MOAD PPDBench PDBBind RNA
Metrics: AAR scRMSD Vina Score Rosetta ddG Rosetta-Vienna RNP-ddG steric clashes hydrogen bonds salt bridges hydrophobic interactions pi-pi stacking
Baselines: DEPACT dyMEAN FAIR RFDiffusionAA
Builds on: FrameDiff Riemannian flow matching RoseTTAFold All-Atom FAIR PLIP
Closest papers by meaning
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.90)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.89)
- Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment (ICLR 2024 · similarity 0.88)
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.87)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.86)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 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.