Research map / Probabilistic and scientific modeling / Molecules and proteins
DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization
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 DecompOpt, a controllable decomposed diffusion model that combines arm-conditioned generation with iterative arm-level optimization for structure-based molecular optimization and controllable tasks like R-group design and scaffold hopping.
Problem
Structure-based drug design generative models only learn the target-ligand distribution, so they cannot reliably produce ligands with desired properties (e.g., binding affinity, synthesizability), and controllable generation tasks (R-group design, scaffold hopping) receive little attention in 3D.
Approach
Decompose ligands into arms and scaffold, condition a decomposed diffusion model on reference arms and their subpockets, and run an evolutionary-style loop that keeps ordered lists of high-scoring arms, so generation and optimization are combined in the decomposed drug space.
Key results
- De novo design: DecompOpt reaches average Vina Dock -8.98 and Success Rate 52.5%, versus -8.39 and 24.5% for DecompDiff and 46.2% for RGA.
- Single-objective optimization: Vina Min average improves from -6.01 (DecompDiff) to -6.72 kcal/mol; QED and SA also improve.
- Scaffold hopping: DecompOpt+Inpainting achieves 93.0% complete rate and 0.35 scaffold similarity versus 89.2% and 0.40 for DecompDiff+Inpainting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine decomposed diffusion with iterative optimization | DecompOpt · CrossDocked2020 test | Vina Dock (Avg.): -8.39 -> -8.98 | Helped |
| combine decomposed diffusion with iterative optimization | DecompOpt · CrossDocked2020 test | Success Rate: 24.5% -> 52.5% | Helped |
| optimize QED as the single objective (ablation) | DecompOpt · Single-objective QED optimization | QED (Avg.): 0.48 -> 0.52 | Helped |
| optimize SA as the single objective (ablation) | DecompOpt · Single-objective SA optimization | SA (Avg.): 0.67 -> 0.74 | Helped |
| optimize Vina Min as the single objective (ablation) | DecompOpt · Single-objective Vina Min optimization | Vina Min (Avg.): -6.01 -> -6.72 | Helped |
| use arm-level rather than whole-molecule optimization (ablation) | DecompOpt · Vina Min optimization | Vina Min (Avg.): -6.62 -> -6.72 | Helped |
| add arm-pocket complexes as conditions instead of using arms alone (ablation) | DecompOpt · CrossDocked2020 optimization | Vina score | Helped |
| use Ref Best instead of Ref Prior checkpoint selection (ablation) | DecompOpt · CrossDocked2020 optimization | Success Rate: 35.4% -> 39.4% | Helped |
| use Pocket Best instead of Pocket Prior checkpoint selection (ablation) | DecompOpt · CrossDocked2020 optimization | Success Rate: 29.2% -> 44.7% | Helped |
| use Opt Best instead of Opt Prior checkpoint selection (ablation) | DecompOpt · CrossDocked2020 optimization | Success Rate: 38.1% -> 52.5% | Helped |
| select the best checkpoint across Ref Prior and Pocket Prior (ablation) | DecompOpt · CrossDocked2020 optimization | Success Rate: 52.5% -> 60.6% | Helped |
| use arm-conditioned scaffold hopping in place of DecompDiff+Inpainting | DecompOpt+Inpainting · Scaffold hopping | Complete Rate: 89.2% -> 93.0% | Helped |
Surprising findings
- Molecule-level optimization is less efficient than arm-level optimization (Vina Min -6.62 vs -6.72 average), supporting decomposition benefits.
- Optimization lowers QED for the main DecompOpt setting (0.48 vs 0.66 for TargetDiff w/ Opt.), indicating trade-offs among objectives.
Limitations
- Multi-objective optimization is not well explored; equal-weight Z-score aggregation is used.
- Diversity is lower than generative baselines (0.60 vs 0.68-0.79 for several), and optimization relies on docking-based oracles.
- Evaluation is on 100 test proteins with docking scores only.
Experimental setup
Datasets: CrossDocked2020
Benchmarks: CrossDocked2020 test set (100 proteins)
Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity Success Rate Validity Uniqueness
Baselines: liGAN GraphBP AR Pocket2Mol TargetDiff DecompDiff RGA TargetDiff w/ Opt.
Builds on: DecompDiff TargetDiff EGNN RGA
Closest papers by meaning
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.89)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.88)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
- Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization (NeurIPS 2024 · similarity 0.88)
- Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design (NeurIPS 2025 · similarity 0.87)
- ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design (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.