RxnMol couples chemical-space search with forward reaction prediction so each generated molecule comes with a route
- A global search over chemical space optimizes molecular properties while sequential forward-reaction prediction builds each proposal, so the output is a molecule plus a synthetic route.
- Routes start from purchasable building blocks, tying generation directly to what a chemist can order and run rather than to a post-hoc synthetic-accessibility score.
💡 Points Based on the abstract
Generating the route with the molecule beats scoring synthetic accessibility afterwards, but the route is only as trustworthy as the forward-reaction predictor, and those still produce plausible-looking steps that fail at the bench. Before trusting it I'd have a chemist review routes for a handful of top designs, and run it against an SA-filtered generator on identical property objectives.
RxnMol: synthesis-aware molecular generation via reaction-driven global chemical space search ↗