Research map / Probabilistic and scientific modeling / Molecules and proteins
SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints
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.
Introduces SynFlowNet, a GFlowNet whose action space is chemical reactions plus purchasable building blocks, with a separately trained backward policy, to generate diverse, synthesizable molecules.
Problem
De novo molecular generation often yields synthetically inaccessible molecules, and reaction-based MDPs make backward GFlowNet trajectories fail to return to the initial state.
Approach
Formulate molecule generation as a GFlowNet over a reaction-and-building-block action space so that synthesis routes are sampled with products, train the backward policy separately (maximum likelihood or REINFORCE toward s0) to keep backward trajectories inside the MDP, and scale to 200k+ building blocks via fixed Morgan-fingerprint action embeddings.
Key results
- SynFlowNet (L=3) reaches sEH proxy 0.92 with diversity 0.79 and SA 2.92, versus REINVENT 0.91 / 0.68 / 2.19 and FragGFN 0.77 / 0.83 / 6.28 in the baseline table.
- A maximum-likelihood or REINFORCE backward policy solves 99.3% and 100% of training routes, versus 11.0% for a uniform backward policy.
- AiZynthFinder retrosynthesis success is 62% for L=3 versus 40% for L=4, while both FragGFN models score 0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a reaction-based synthesis action space rather than fragment actions | SynFlowNet vs FragGFN · sEH molecule generation | AiZynth retrosynthesis success: 62% vs 0% | Helped |
| increase maximum trajectory length from 3 to 4 (ablation) | SynFlowNet · sEH molecule generation | AiZynth retrosynthesis success: 62% -> 40% | Hurt |
| use GFlowNet sampling instead of entropy-regularized Soft Q-learning | SynFlowNet · sEH, GSK3β, and DRD2 | number of high-reward modes | Helped |
| add synthetic accessibility score to the reward (ablation) | SynFlowNet · sEH proxy | sEH proxy: 0.92±0.01 -> 0.94±0.01 | Helped |
| add synthetic accessibility score to the reward (ablation) | SynFlowNet · sEH proxy | Diversity: 0.79±0.01 -> 0.75±0.02 | Hurt |
| train the backward policy with maximum likelihood (ablation) | SynFlowNet · backward route retrieval (train) | Solved routes: 11.0±3.7% -> 99.3±0.5% | Helped |
| train the backward policy with REINFORCE (ablation) | SynFlowNet · backward route retrieval (test) | Solved routes: 11.0±4.1% -> 44.3±2.6% | Helped |
| use a free backward policy instead of the trained backward policy (ablation) | SynFlowNet · backward route retrieval (test) | Solved routes: 44.3±2.6% -> 1.0±0.8% | Hurt |
| use Morgan-fingerprint-based building-block selection instead of learnable building-block logits (ablation) | SynFlowNet · large building-block sets | average reward | Helped |
| curate the building-block library using experimentally validated target fragments (ablation) | SynFlowNet · Mpro | reward | Helped |
| use SynFlowNet instead of REINVENT | sEH molecule generation · sEH proxy | ChEMBL similarity: 0.64±0.02 -> 0.43±0.01 | Helped |
| use SynFlowNet rather than SynNet | molecule optimization · GSK3β | top-100 score: 0.815 -> 0.862 | Helped |
Surprising findings
- A free (unconstrained) backward policy performs worse than the rest, with only 1.0% solved test routes.
- Soft Q-learning collapses to narrow high-reward distributions, with far lower diversity (0.42) than GFlowNet sampling.
Limitations
- Reaction set limited to 105 templates; no non-linear pathways, no reactions with more than two reactants, and no selectivity or reaction-condition handling.
- Rediscovery is imperfect; celecoxib is not reachable through the building-block space.
- Synthesizability is assessed by SA score and AiZynthFinder, which are proxies rather than real synthesis.
Experimental setup
Datasets: Enamine building blocks Enamine REAL reactions ChEMBL sEH proxy training data PMO oracles
Benchmarks: sEH proxy GSK3beta DRD2 KRAS Vina docking PMO sample efficiency SynNet optimization tasks
Metrics: reward diversity (Tanimoto) synthetic accessibility (SA) AiZynthFinder retrosynthesis success QED molecular weight ChEMBL similarity AUC top-10 solved routes
Baselines: REINVENT FragGFN FragGFN SA SyntheMol Soft Q-Learning SynNet
Builds on: GFlowNet trajectory balance FragGFN REINVENT SyntheMol
Closest papers by meaning
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.94)
- RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · similarity 0.93)
- Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.91)
- 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 0.90)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (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.