Steerable Single-step Retrosynthesis with Mechanism-based Discrete Flow Matching and Feynman-Kac steering

One of the main bottlenecks of data-driven Computer Assisted Synthesis Planning (CASP) tools is that single-step retrosynthetic models are limited to the routine reaction space composed of the patented reactions that form their training data. We argue that two properties are required for a model to recover complex out-of-distribution disconnections: it must reason in terms of mechanistic electron flow, as expert synthetic chemists do, and its inductive bias must allow the use of distribution steering methods. Therefore, we introduce Verso MAELLE, the first mechanism-based single-step retrosynthetic model which, moreover, is steerable through the Feynman-Kac (FK) method. Constructed upon MAELLE, Verso MAELLE models retrosynthesis as a Continuous Time Markov Chain (CTMC) trajectory comprised of edits in the electron space between products and synthons. The synthons are then completed to precursors using a Transformer-based model. Apart from achieving competitive \textit{top-k} synthon and roundtrip accuracies on a processed USPTO 480K version, we encode synthetic strategies into FK rewards, steering the model towards ring-breaking disconnections and recovering the key domino disconnection that was used to synthesize Punctaporonin U, a complex natural product of the sesquiterpene family.