Retrosynthesis planning aims to discover synthesis routes that connect purchasable starting materials to a target molecule. Conventional approaches generate candidate paths solely based on predicted reactions; if any step is infeasible, the entire route is only recognized as invalid after the search finishes, wasting many resources.
To address this, we introduce Verification‑Guided Reflow (VGR), which treats reaction verification as a decision action within an AND/OR search. The verifier returns binary verdicts (accept/reject), and the search structure schedules queries accordingly:
- Rejected reactions are pruned immediately;
- Accepted reactions gain search priority only when they become part of a complete candidate route.
We evaluate performance with verified solvability (VS), defined as
$$\text{VS}=\frac{\text{Number of targets with a step‑wise verified route}}{\text{Total number of targets}}$$
On the Retro190 benchmark, feeding verification results back into the search raises VS from 51.1% to 62.6%. Controlled experiments reveal that both earlier verification and immediate prioritization of accepted reactions individually improve route discovery, yet their combination degrades performance.
Consequently, we propose an early pruning + delayed guidance strategy: rejected reactions are removed right away, while accepted reactions receive priority only when processed as part of a full candidate route. With a fixed verification scheduling rule, this approach lifts VS from 61.1% (immediate guidance) to 71.1%, using fewer proposal calls and unique verifier queries.
Review