Research map / Probabilistic and scientific modeling / Molecules and proteins
RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis
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 Retro SynFlow, a discrete flow-matching model for single-step retrosynthesis that uses reaction-center-derived synthons as the source distribution, plus Feynman-Kac/SMC steering with a forward-synthesis reward to improve round-trip accuracy.
Problem
Single-step retrosynthesis requires generating accurate yet diverse, feasible reactant sets for a product molecule; existing template-free generative methods struggle with this combinatorial space.
Approach
Builds a discrete flow-matching Markov bridge from synthons (product with reaction-center bond removed) to reactant graphs, rather than from products, and steers sampling at inference with Feynman-Kac SMC resampling using a round-trip reward computed from a forward-synthesis model and the denoiser's expected reactant.
Key results
- RSF achieves 60.0% top-1 exact-match accuracy on USPTO-50k, the best among template-free non-ensemble methods.
- RPF-RS reaches 80.1% top-5 round-trip accuracy, about 19% above template-free SOTA.
- Greedy sampling baseline reaches 89.4% top-1 round-trip accuracy vs 91.4% for RPF-RS (k=1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use synthons as the source distribution instead of product molecules | Retro SynFlow vs Retro ProdFlow · USPTO-50k test | top-1 exact-match accuracy: 50.0 ± 0.15 -> 60.0 ± 0.22 | Helped |
| add product molecule as context to synthon completion (ablation) | Retro SynFlow · USPTO-50k validation | top-1 exact-match accuracy: 59.7 -> 67.7 | Helped |
| add FK steering with SMC resampling (ablation) | Retro ProdFlow · USPTO-50k test | top-5 round-trip accuracy: 66.8 -> 80.1 | Helped |
| use SMC resampling during steering rather than greedy selection at the end (ablation) | Retro ProdFlow-RS · USPTO-50k test | top-5 round-trip accuracy: 73.0 -> 80.1 | Helped |
| increase sampling to 400 reactants per product without reward steering (ablation) | Retro ProdFlow · USPTO-50k test | top-5 round-trip accuracy: 68.5 -> 80.1 | Hurt |
| use RK sampling with 25 steps instead of Euler sampling with 50 steps (ablation) | Retro ProdFlow · USPTO-50k test, N=50 reactants | top-1 exact-match accuracy: 49.6 -> 49.3 | Hurt |
| use RK sampling with 50 steps instead of Euler sampling with 50 steps (ablation) | Retro ProdFlow · USPTO-50k test, N=50 reactants | top-1 exact-match accuracy: 49.6 -> 49.3 | Hurt |
| allocate 50 rather than 70 reactant samples to the top-ranked synthon (ablation) | Retro SynFlow, M=2 · USPTO-50k validation | top-1 exact-match accuracy: 58.1 -> 48.5 | Hurt |
| increase the number of SMC particles (ablation) | Retro ProdFlow-RS · USPTO-50k validation | round-trip accuracy | Helped |
| increase the number of SMC particles (ablation) | Retro ProdFlow-RS · USPTO-50k validation | exact-match accuracy | Hurt |
| steer predictions toward forward-model round-trip rewards (ablation) | Retro ProdFlow-RS · USPTO-50k test | top-1 exact-match accuracy: 257 correct examples ... converted to incorrect; 251 incorrect examples ... converted to correct | Hurt |
Surprising findings
- Steering with a round-trip reward lowers top-k exact-match accuracy because it trades off exact matches for feasible alternatives.
Limitations
- Exact-match accuracy drops as steering strength increases.
- Evaluation limited to USPTO-50k.
- Round-trip reward depends on the accuracy of the forward-synthesis model.
Experimental setup
Datasets: USPTO-50k
Benchmarks: USPTO-50k
Metrics: top-k exact match accuracy top-k round-trip accuracy round-trip coverage
Baselines: GLN GraphRetro LocalRetro RetroGFN Chimera RetroBridge G2G MEGAN Graph2SMILES Retroformer
Builds on: discrete flow matching RetroBridge Feynman-Kac steering Molecular Transformer
Closest papers by meaning
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.91)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.90)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.89)
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.89)
- Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · 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.