Research map / Probabilistic and scientific modeling / Molecules and proteins
FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
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 FlexSBDD, a flow-matching generative model that jointly generates ligands and updates flexible protein backbone and sidechain conformations, trained with apo-holo pairs plus relaxation/repacking augmentation.
Problem
Structure-based drug design usually treats the target protein as rigid, ignoring induced-fit conformational change, which causes steric clashes and limits generated molecule quality.
Approach
Uses Riemannian flow matching over protein backbone frames (SE(3)), sidechain torsion angles (torus), and ligand atom types, with an E(3)-equivariant scalar-vector network, starting from an apo structure and outputting a generated ligand plus holo protein; augments apo data via OpenMM relaxation and Rosetta repacking.
Key results
- On CrossDocked, FlexSBDD reaches -9.12 average Vina Dock (vs -8.39 for DecompDiff) and 78.5% high affinity.
- Average steric clashes are 1.39 for FlexSBDD vs 6.43 for DecompDiff, and HB acceptors rise to 1.96 vs 1.18.
- Ablations show removing sidechain flexibility hurts more than removing backbone flexibility (Vina Dock -8.86 vs -9.05); removing data augmentation drops Vina Dock to -8.76.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model flexible protein structure during ligand generation | FlexSBDD · CrossDocked | Vina Dock (Avg.): -8.39 -> -9.12 | Helped |
| model flexible protein structure during ligand generation | FlexSBDD · Binding MOAD | Vina Dock (Avg.): -8.46 -> -9.38 | Helped |
| add data augmentation (ablation) | FlexSBDD · CrossDocked | Vina Dock (Avg.): -8.76 -> -9.12 | Helped |
| use scalar-vector dual feature representation instead of EGNN (ablation) | FlexSBDD · CrossDocked | Vina Dock (Avg.): -8.95 -> -9.12 | Helped |
| update the protein backbone structure (ablation) | FlexSBDD · CrossDocked | Vina Dock (Avg.): -9.05 -> -9.12 | Helped |
| update protein sidechain dihedral angles (ablation) | FlexSBDD · CrossDocked | Vina Dock (Avg.): -8.86 -> -9.12 | Helped |
| allow the protein structure to update during ligand generation (ablation) | FlexSBDD · CrossDocked | Vina Dock (Avg.): -8.82 -> -9.12 | Helped |
| model protein flexibility during ligand generation | FlexSBDD · CrossDocked | steric clashes: 6.43 -> 1.39 | Helped |
| model protein flexibility during ligand generation | FlexSBDD · CrossDocked | HB Acceptors: 1.18 -> 1.96 | Helped |
Surprising findings
- Removing backbone flexibility matters less than removing sidechain flexibility, which the authors attribute to backbone rigidity.
- Protein flexibility modeling is claimed to rediscover a KRAS G12C cryptic pocket, with H95 sidechain rotation consistent with prior experimental observation.
Limitations
- Only small molecule design is considered; antibodies, peptides, and nucleic acids are not addressed.
- Limited dataset size restricts scaling of the model.
- Apo-holo pair data is scarce, so augmentation relies on synthetic apo structures from OpenMM and Rosetta.
Experimental setup
Datasets: CrossDocked Binding MOAD ApoBind
Benchmarks: CrossDocked Binding MOAD
Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity JSD bond length/angle steric clashes hydrogen bonds
Baselines: LiGAN AR Pocket2Mol TargetDiff DecompDiff NeuralPlexer
Builds on: flow matching TargetDiff geometric vector perceptron ApoBind
Closest papers by meaning
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.89)
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.88)
- MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space (ICML 2024 · similarity 0.87)
- Binding-Adaptive Diffusion Models for Structure-Based Drug Design (AAAI 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)
- Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule (ICML 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.