Research map / Probabilistic and scientific modeling / Molecules and proteins
RGFN: Synthesizable Molecular Generation Using GFlowNets
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 Reaction-GFlowNet (RGFN), a GFlowNet that builds molecules by applying curated high-yield reactions to low-cost building blocks, yielding synthesizable candidates with search spaces orders of magnitude larger than Enamine REAL-scale libraries.
Problem
Generative molecular design often produces candidates that are hard or impossible to synthesize, and existing synthesizability-aware approaches either rely on unreliable scoring or on reaction spaces that are not shared or reproducible.
Approach
Extends GFlowNets to act in the space of chemical reactions: each step selects a reaction template and a compatible reactant from a curated set of 17 reactions and 350 cheap building blocks, so every generated molecule has a defined synthetic route. It also embeds building-block actions with MACCS fingerprints so the policy can scale to larger fragment libraries.
Key results
- On the senolytic proxy task RGFN finds high-reward molecules where FGFN fails to discover any, and RGFN outperforms FGFN and FGFN+SA on average reward across tasks, though it trails GraphGA.
- RGFN's top modes achieve AiZynth scores comparable to SyntheMol and casVAE (e.g., 0.56 vs 0.80 and 0.82 on sEH) while far exceeding GraphGA (0.04) and FGFN (0.14) on sEH.
- With 350 building blocks and up to four reactions the state space exceeds Enamine REAL (6.5B) by about an order of magnitude, per the paper's trajectory-based estimate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RGFN instead of FGFN | Molecular generation · sEH top modes | AiZynth: 0.14 -> 0.56 | Helped |
| Use RGFN instead of FGFN | Molecular generation · Senolytics top modes | AiZynth: 0.02 -> 0.58 | Helped |
| Use RGFN instead of FGFN | Molecular generation · ClpP top modes | AiZynth: 0.25 -> 0.65 | Helped |
| Use RGFN instead of FGFN | Molecular generation · DRD2 top modes | AiZynth: 0.76 -> 0.87 | Helped |
| Use fingerprint-based molecular embeddings for building-block selection instead of independent action embeddings | RGFN with larger fragment libraries · sEH proxy; larger fragment libraries | number of discovered Murcko scaffolds | Helped |
| Add SAScore as a reward term | FGFN · Synthesizability evaluation | SAScore | Helped |
| Add SAScore as a reward term | FGFN · Synthesizability evaluation | AiZynth | No clear effect |
Surprising findings
- Adding SAScore as a reward (FGFN+SA) improves the SAScore metric but barely changes AiZynthFinder scores, suggesting SAScore is insufficient for guaranteeing synthesizability.
Limitations
- Only 17 reactions and 350 building blocks, so the chemical space covered is limited and generated molecules tend to be linear and flat.
- Reliance on molecular docking as the oracle, which correlates with molecular weight and weakly with experimental affinity.
- Diversity is lower than FGFN, possibly due to the small fragment and reaction sets; no experimental validation is reported.
Experimental setup
Datasets: Enamine building blocks Enamine REAL ZINC15 USPTO
Benchmarks: sEH proxy senolytic proxy DRD2 proxy ClpP docking Mpro docking TBLR1 docking
Metrics: average reward number of discovered modes molecular weight QED SAScore AiZynthFinder solved fraction
Baselines: GraphGA SyntheMol casVAE FGFN FGFN+SA
Builds on: GFlowNet trajectory balance GuacaMol Graph GA FGFN
Closest papers by meaning
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.93)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.92)
- Scalable and Cost-Efficient de Novo Template-Based Molecular Generation (NeurIPS 2025 · similarity 0.90)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.89)
- Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract) (AAAI 2025 · similarity 0.88)
- Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.88)
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.