Research map / Probabilistic and scientific modeling / Molecules and proteins

Generative Flows on Synthetic Pathway for Drug Design

ICLR 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 5 citations (OpenAlex)

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 RxnFlow, a GFlowNet over synthetic pathways built from 1.2M Enamine building blocks and 71 reaction templates, using action space subsampling to train over huge action spaces and a non-hierarchical MDP for adding objectives or blocks without retraining.

Problem

Generative models for drug design often ignore synthesizability, and existing synthesis-pathway GFlowNets restrict action spaces (small block libraries, more reaction steps), which reduces diversity and synthetic feasibility.

Approach

Train a trajectory-balance GFlowNet over reaction-based synthetic pathways with a random action-space subsampling scheme (importance-weighted forward policy estimation) so that millions of building blocks can be used at near-constant cost, plus a non-hierarchical joint (template, block) MDP and action embeddings so the policy can adapt to modified or expanded block libraries without retraining.

Key results

• On pocket-specific optimization across five LIT-PCBA targets, RxnFlow achieves the highest hit ratios (e.g., 71.25% on ESR_ago vs 30.75% for SynFlowNet).
• On CrossDocked2020 pocket-conditional generation, RxnFlow reaches an average Vina score of -8.85 kcal/mol and 34.8% synthesizability, with 100% generation success and 0.81 diversity.
• Generation time only roughly doubles when the action space grows 100-fold, per the building block library scaling study.

Measured effects reported in the paper

Surprising findings

• FragGFN+SA (adding a cheap SA-score objective) does not meaningfully improve synthesizability, suggesting optimizing cheap synthesizability proxies is suboptimal.

Limitations

• Uniform subsampling forces exploration and may limit exploitation; prioritized subsampling is left to future work.
• Synthesizability is assessed with AiZynthFinder restricted to USPTO reactions and the Enamine block library, which is computationally intensive and a proxy for real synthesis.
• Docking-based rewards can be hacked by increasing molecule size, requiring property constraints such as QED.

Experimental setup

Datasets: Enamine REAL building blocks CrossDocked2020 LIT-PCBA USPTO reaction templates

Benchmarks: LIT-PCBA CrossDocked2020

Metrics: hit ratio Vina docking score AiZynthFinder synthesizability synthetic complexity (number of steps) QED diversity generation success

Baselines: SynNet BBAR SynFlowNet RGFN FragGFN Pocket2Mol TargetDiff DecompDiff DiffSBDD MolCRAFT TacoGFN

Builds on: GFlowNet SynFlowNet RGFN TacoGFN trajectory balance

Closest papers by meaning

• SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.94)
• RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · similarity 0.92)
• Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.91)
• Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.91)
• Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract) (AAAI 2025 · similarity 0.90)
• Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.89)