RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models

Retrosynthesis prediction is a central task in computer-aided synthesis planning, and synthesizability remains a persistent bottleneck in AI-driven molecular and materials discovery. We introduce RetroMol, a family of compact text-based models that improve starting-material generation through energy-based reranking and knowledge distillation. RetroBase, an 11M-parameter transformer encoder--decoder trained on USPTO reaction data, achieves Top-1 and Top-5 accuracies of 46.8\% and 67.8\%. We then train RetroERM, an energy-based model that scores candidate reactions using molecular fingerprint count features, and distill its ranking preferences back into the generator. The distilled model, RetroDistil, reaches Top-1 and Top-5 accuracies of 52.6\% and 73.5\%, improving on RetroBase by roughly six points at every $K$ and outperforming both an 8B-parameter reasoning LLM and a 110M-parameter string-editing model on USPTO-full. An energy-based critic is therefore a practical mechanism for transferring chemistry-informed ranking knowledge into a generator, improving retrosynthesis predictions without larger or more complicated generative architectures.