Single-step retrosynthesis asks chemists to find reactant sets that produce a target molecule, but the problem is inherently one-to-many: many different reactant combinations can lead to the same product. Most current models, however, are trained and evaluated against only one recorded reactant set per target. The authors of a new arXiv preprint argue that this reference-based training provides no explicit positive signal for unrecorded but chemically valid alternatives.
To address this, they propose a reinforcement learning approach that uses feedback from forward reaction prediction. Instead of only penalizing deviations from the recorded answer, the model receives a reward when a predicted reactant set is confirmed to produce the target in the forward direction. This allows the model to learn from alternatives that are chemically sound even if they are not in the training database.
The work highlights a conceptual shift in how retrosynthesis models are trained: from matching a single historical record to exploring the broader space of feasible chemistry. The abstract does not provide experimental details or quantitative results, so the practical gains remain to be seen, but the direction addresses a known limitation in existing retrosynthesis benchmarks.