A new approach called RETRO SYNFLOW (RSF) has been introduced to address the challenges in single-step retrosynthesis in organic chemistry.
RSF employs a discrete flow-matching framework that creates a bridge between the target product molecule and the reactant molecule.
By using reaction center identification to generate synthons and Feynman-Kac steering with Sequential Monte Carlo, RSF achieves 60.0% top-1 accuracy, outperforming the previous state-of-the-art by 20%.
The approach improves top-5 round-trip accuracy by 19% over prior methods, showcasing competitive results in top-k accuracy while enhancing diversity and feasibility of generated samples.