A new AI framework designs molecules by searching through entire synthesis routes, not just chemical structures.
Researchers built RouteFlow, which treats synthesizable molecular design as a search over a continuous route latent space. Each point in that space decodes back into a full synthesis route, not just a target molecule, so synthesizability is baked in rather than checked afterward. To explore the space efficiently, the team uses reward-guided flow matching to steer the search toward molecules with desired properties, plus a cycle-consistency check that keeps the optimizer from drifting into regions that do not decode into valid routes. Tested across 16 optimization tasks from the Therapeutic Data Commons benchmark, RouteFlow beat other synthesizability-aware methods on sample efficiency, synthetic accessibility, and retrosynthesis success rate.
Most molecule-generating AI chases potency or binding affinity first and treats synthesizability as a filter applied later, which is a big reason so many computationally discovered compounds never make it past a chemist's bench. By making the synthesis route itself the object of search, RouteFlow builds buildability into every candidate from the start. That is a real shift for drug-discovery pipelines, where the gap between an AI's shortlist and something a lab can actually make has been a persistent bottleneck.
The gains here are measured against other synthesizability-aware baselines, not against how these molecules fare once a chemist tries to make them at the bench - the real test a benchmark score cannot fully capture.