Expert-Augmented Learning for Synthesis Route Evaluation

Computer-aided synthesis planning can generate many multistep routes, but deciding which routes are chemically credible remains difficult to automate. The central obstacle is a supervision mismatch: literature routes provide scalable positive examples but do not define a unique optimum, whereas informative failure labels and detailed expert assessments are scarce. We introduce a two-stage, expert-augmented route evaluator that separates scalable proxy learning from chemical-feasibility alignment. A permutation-invariant route model is first trained on 1.72 million candidate routes for 47,303 targets to predict tree edit distance to a target-specific literature route. Limited reaction-level expert supervision is then introduced with low-rank adaptation to produce five-point feasibility predictions, three actionable route categories, and stepwise explanations. The structural-difference-fingerprint model recovers the literature route at rank one for 60.3% of held-out targets, compared with 17.2% for a learned ranking baseline. Yet disagreement with the single literature route does not imply failure: experts endorse 75.4% of novel rank-one proposals as feasible. On an independent 50-route multi-expert set, majority agreement rises from 0.48 before fine-tuning to 0.64 afterward, comparable to 0.65 expert–peer agreement under the same protocol. These results show how small, high-value expert datasets can align scalable route scoring without treating one reported synthesis as a complete definition of route quality.