Trustworthy Retrosynthesis: Mitigating Hallucinations with Reaction Plausibility Filtering and Retrieval-Augmented Scoring

While Artificial Intelligence (AI) has lead to significant improvements in Computer-Aided Synthesis Planning (CASP), its credibility within the chemical community is fragile. AI retrosynthesis models frequently “hallucinate” chemically implausible reactions, which undermines trust. To address this, we propose a framework that integrates three orthogonal validation strategies to ensure reaction plausibility. This key insight, combining reaction validation strategies which cover different error patterns, was the basis of our winning solution to the \textit{[Anonymized for the sake of blind review]} Retrosynthesis Challange. Our approach combines: (1) a novel Transformer-based model, called Reaction Prior, that estimates reaction likelihood from large-scale experimental data, mimicking chemical reasoning (2) a Graph Neural Network trained on a reaction dataset augmented with synthetically generated incorrect reactions, and(3) a retrieval-based scoring system that leverages chemical databases and grounds suggestions in known chemical literature. The framework was validated on unseen targets through a novel human evaluation process, successfully rejecting the most hallucinated reactions. In this evaluation, chemical experts manually reviewed reactions within the synthetic paths, providing a more reliable and trustworthy form of verification compared to purely automatic methods.