Research map / Probabilistic and scientific modeling / Molecules and proteins
Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows
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 DynamicFlow, a full-atom flow model (ODE and SDE variants) trained on a curated MISATO-derived apo/holo dataset to transform apo pockets and noisy ligands into holo pockets and 3D ligands jointly.
Problem
Structure-based drug design usually treats protein pockets as rigid, ignoring conformational dynamics that affect ligand binding; MD sampling of these transitions is computationally expensive.
Approach
A multiscale full-atom flow matching model (atom-level SE(3)-equivariant GNN plus residue-level IPA Transformer) that jointly generates ligands and transforms apo pockets (backbone frames and side-chain torsions) to holo conformations, with an SDE extension for robustness and an interaction distance loss.
Key results
- DynamicFlow-SDE achieves mean Vina Score -7.65 and 52.5% high-affinity ratio, versus TargetDiff -5.09 and 44.0%.
- DynamicFlow-SDE reaches 73.6% complete rate vs 68.2% for TargetDiff; DynamicFlow-ODE 70.2%.
- Using DynamicFlow-refined pockets as input improves TargetDiff's Vina Score from -5.09 to -9.00 and IPDiff's from -7.55 to -11.04 (per the paper's table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add interaction loss (ablation) | DynamicFlow-ODE · Test set | Vina Score: -6.76±1.39 -> -7.28±1.98 | Helped |
| Add interaction loss (ablation) | DynamicFlow-ODE · Test set | High Affi.: 37.4% -> 51.0% | Helped |
| Add interaction loss (ablation) | DynamicFlow-SDE · Test set | Vina Score: -7.00±1.15 -> -7.65±1.59 | Helped |
| Add interaction loss (ablation) | DynamicFlow-SDE · Test set | High Affi.: 43.4% -> 52.5% | Helped |
| Add residue-level Transformer (ablation) | DynamicFlow-ODE · Test set | Vina Score: -6.23 ± 1.68 -> -7.28 ± 1.98 | Helped |
| Add atom-level EGNN (ablation) | DynamicFlow-ODE · Test set | Vina Score: -6.02 ± 1.63 -> -7.28 ± 1.98 | Helped |
| Add residue-level Transformer (ablation) | DynamicFlow-SDE · Test set | Vina Score: -6.50 ± 1.22 -> -7.65 ± 1.59 | Helped |
| Add atom-level EGNN (ablation) | DynamicFlow-SDE · Test set | Vina Score: -6.13 ± 1.31 -> -7.65 ± 1.59 | Helped |
| Use DynamicFlow-generated pocket | TargetDiff · Test set, rigid pocket input | Vina Score: -5.09±8.81 -> -9.00±3.19 | Helped |
| Use DynamicFlow-generated pocket | TargetDiff* · Test set, rigid pocket input | Vina Score: -5.53±4.09 -> -7.82±3.50 | Helped |
| Use DynamicFlow-generated pocket | IPDiff · Test set, rigid pocket input | Vina Score: -7.55±5.59 -> -11.04±3.76 | Helped |
| Use DynamicFlow-generated pocket | IPDiff* · Test set, rigid pocket input | Vina Score: -6.23±5.26 -> -9.71±3.66 | Helped |
Surprising findings
- Removing the interaction loss sharply lowers the high-affinity ratio for DynamicFlow-ODE (51.0% to 37.4%) while slightly raising Comp. Rate.
- Str2Str performs even worse than the naive noise-perturbation baseline on pocket conformation sampling.
Limitations
- Apo pockets are AlphaFold2-predicted proxies rather than experimentally determined apo structures.
- Evaluation on only 50 test complexes; Vina scores are approximate binding affinity proxies.
Experimental setup
Datasets: MISATO AlphaFold DB predicted structures (apo proxies)
Benchmarks: 50-complex held-out test set from curated MISATO-derived dataset
Metrics: Vina Score QED SA Lipinski logP High Affinity ratio Complete rate minimum RMSD cover ratio pocket volume JSD
Baselines: Pocket2Mol TargetDiff IPDiff Str2Str FrameDiff
Builds on: flow matching discrete flow matching EGNN IPA (AlphaFold2 invariant point attention) TargetDiff
Closest papers by meaning
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.88)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.88)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.88)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.88)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.88)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · 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.