Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RETRO SYNFLOW: Discrete Flow Matching for Accurate and Diverse Single-Step Retrosynthesis

Robin Yadav, Qi Yan, Guy Wolf, Avishek Joey Bose, Renjie Liao

Abstract

A fundamental problem in organic chemistry is identifying and predicting the series of reactions that synthesize a desired target product molecule. Due to the combinatorial nature of the chemical search space, single-step reactant prediction -- i.e. single-step retrosynthesis -- remains challenging even for existing state-of-the-art template-free generative approaches to produce an accurate yet diverse set of feasible reactions. In this paper, we model single-step retrosynthesis planning and introduce RETRO SYNFLOW (RSF) a discrete flow-matching framework that builds a Markov bridge between the prescribed target product molecule and the reactant molecule. In contrast to past approaches, RSF employs a reaction center identification step to produce intermediate structures known as synthons as a more informative source distribution for the discrete flow. To further enhance diversity and feasibility of generated samples, we employ Feynman-Kac steering with Sequential Monte Carlo based resampling to steer promising generations at inference using a new reward oracle that relies on a forward-synthesis model. Empirically, we demonstrate \nameshort achieves $60.0 \%$ top-1 accuracy, which outperforms the previous SOTA by $20 \%$. We also substantiate the benefits of steering at inference and demonstrate that FK-steering improves top-$5$ round-trip accuracy by $19 \%$ over prior template-free SOTA methods, all while preserving competitive top-$k$ accuracy results.

RETRO SYNFLOW: Discrete Flow Matching for Accurate and Diverse Single-Step Retrosynthesis

Abstract

A fundamental problem in organic chemistry is identifying and predicting the series of reactions that synthesize a desired target product molecule. Due to the combinatorial nature of the chemical search space, single-step reactant prediction -- i.e. single-step retrosynthesis -- remains challenging even for existing state-of-the-art template-free generative approaches to produce an accurate yet diverse set of feasible reactions. In this paper, we model single-step retrosynthesis planning and introduce RETRO SYNFLOW (RSF) a discrete flow-matching framework that builds a Markov bridge between the prescribed target product molecule and the reactant molecule. In contrast to past approaches, RSF employs a reaction center identification step to produce intermediate structures known as synthons as a more informative source distribution for the discrete flow. To further enhance diversity and feasibility of generated samples, we employ Feynman-Kac steering with Sequential Monte Carlo based resampling to steer promising generations at inference using a new reward oracle that relies on a forward-synthesis model. Empirically, we demonstrate \nameshort achieves top-1 accuracy, which outperforms the previous SOTA by . We also substantiate the benefits of steering at inference and demonstrate that FK-steering improves top- round-trip accuracy by over prior template-free SOTA methods, all while preserving competitive top- accuracy results.

Paper Structure

This paper contains 25 sections, 10 equations, 12 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Background and preliminaries
  3. Discrete Flow Matching
  4. Feynman-Kac Steering
  5. Discrete Flow Matching for Retrosynthesis
  6. Retro SynFlow
  7. Reward-based steering
  8. Experiments and Results
  9. Experimental Setup
  10. Main Results
  11. Ablation Studies
  12. Related Works
  13. Conclusion
  14. Acknowledgements
  15. Experimental details
  16. ...and 10 more sections

Figures (12)

  • Figure 1: An overview of our Retro ProdFlow (RPF) and Retro SynFlow (RSF) framework. RPF directly maps a product molecule to reactants via discrete flow. RSF first predicts synthons from the product using a reaction center predictor, then maps these synthons to reactants via discrete flow.
  • Figure 2: Inference time steering with a forward-synthesis reward model.
  • Figure 3: Overview of flow matching denoiser $p_\theta$.
  • Figure 4: Top-$5$ reactants selected by each method. A star indicates an exact match, a checkmark indicates a round-trip match but not an exact match, and a cross means neither.
  • Figure 5: Performance of Retro ProdFlow-RS on the USPTO-50k validation set as we vary the number of particles for SMC resampling. We sample $N = 50$ reactants per product.
  • ...and 7 more figures