Research map / 3D vision / 3D generation
Compositional Flows for 3D Molecule and Synthesis Pathway Co-design
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 CGFlow, a flow-matching framework for generating compositional objects with continuous states, and instantiates it as 3DSynthFlow for joint synthesis-pathway and 3D pose co-design in drug discovery.
Problem
Generating compositional objects (e.g., molecules built via synthesis steps) that also have continuous features (3D coordinates) is not handled by standard flow matching or sequential models, which cannot enforce composition validity or give efficient reward credit assignment.
Approach
Extends flow matching with a compositional flow that adds components at discrete time intervals while a state flow interpolates continuous states with component-wise local time; compositional policy is trained with GFlowNet trajectory balance while the state flow model is pre-trained with simulation-free flow matching.
Key results
- On pocket-conditional CrossDocked2020, 3DSynthFlow (high beta) reaches average Vina -9.42 kcal/mol and 36.1% AiZynth success rate, versus RxnFlow -8.85 and MolCRAFT-large -9.25 (3.9% success).
- 3DSynthFlow finds 4.2x more diverse high-scoring modes than RxnFlow after 64,000 samples (4432.3 vs 1062.6) on LIT-PCBA.
- On LIT-PCBA pocket-specific optimization, 3DSynthFlow reports average Vina -10.89 kcal/mol across 5 shown targets, outperforming all baselines, with AiZynth success rate 68.58% across all 15 targets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Co-design 3D molecular structures with synthesis pathways rather than using 2D reaction-based generation | 3DSynthFlow vs. RxnFlow · LIT-PCBA (15-target average) | Vina docking score: -10.46 -> -10.89 | Helped |
| Co-design 3D molecular structures with synthesis pathways rather than using 2D reaction-based generation | 3DSynthFlow vs. RxnFlow · LIT-PCBA (15 targets) | AiZynthFinder success rate: 65.35 -> 68.58 | Helped |
| Co-design 3D molecular structures with synthesis pathways rather than using 2D reaction-based generation | 3DSynthFlow vs. RxnFlow · LIT-PCBA, after sampling 64,000 molecules | Number of diverse high-scoring modes: 4432.3 vs 1062.6 | Helped |
| Co-design 3D molecular structures with synthesis pathways rather than using 2D reaction-based generation | 3DSynthFlow vs. RxnFlow · CrossDocked2020 test pockets | Vina Avg.: -8.85 -> -9.42 | Helped |
| Use high β rather than medium β (ablation) | 3DSynthFlow · CrossDocked2020 test pockets | Vina Avg.: -9.30 -> -9.42 | Helped |
| Use high β rather than low β (ablation) | 3DSynthFlow · CrossDocked2020 test pockets | Diversity Avg.: 0.78 -> 0.69 | Hurt |
| Use high β rather than low β (ablation) | 3DSynthFlow · CrossDocked2020 test pockets | AiZynth success rate Avg.: 36.2% -> 36.1% | No clear effect |
| Use 3D reaction-based generation rather than atom-based generation | 3DSynthFlow (high β) vs. MolCRAFT-large · CrossDocked2020 test pockets | AiZynth success rate Avg.: 3.9% -> 36.1% | Helped |
| Use 3D reaction-based generation rather than atom-based generation | 3DSynthFlow (high β) vs. MolCRAFT-large · CrossDocked2020 test pockets | Vina Avg.: -9.25 -> -9.42 | Helped |
Surprising findings
- Docking-score gains in the pocket-conditional setting are smaller than in pocket-specific optimization, attributed to limited accuracy of the pre-trained pharmacophore proxy.
- RxnFlow (2D) achieves lower synthesis steps (2.17) than 3DSynthFlow (2.39), so 3DSynthFlow does not win on every synthesis metric.
Limitations
- Synthon-based action space excludes nonlinear pathways such as ring-forming reactions.
- Pre-trained pocket-conditional proxy has limited accuracy for docking-score reward, reducing gains in pocket-conditional setting.
- Restricted to two synthesis steps and a limited building block library.
Experimental setup
Datasets: CrossDocked2020 LIT-PCBA Enamine REAL building blocks
Benchmarks: LIT-PCBA CrossDocked2020
Metrics: Vina docking score ligand efficiency AiZynthFinder success rate PoseCheck protein-ligand interactions number of diverse high-scoring modes validity QED diversity time
Baselines: FragGFN SynNet BBAR SynFlowNet RGFN RxnFlow Pocket2Mol TargetDiff DiffSBDD DecompDiff MolCRAFT TacoGFN
Builds on: GFlowNets trajectory balance RxnFlow SemlaFlow conditional flow matching
Closest papers by meaning
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.91)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.91)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.90)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.89)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.89)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · 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.