RetroPrime: A Chemistry-Inspired and Transformer-based Method for Retrosynthesis Predictions

Retrosynthesis prediction is a crucial task for organic synthesis. In this work, we propose a template-free and Transformer-based method dubbed RetroPrime, integrating chemists’ retrosynthetic strategy of (1) decomposing a molecule into synthons then (2) generating reactants by attaching leaving groups. These two steps are accomplished with versatile Transformer models, respectively. While RetroPrime performs competitively against all state-of-the art models on the standard USPTO-50K dataset, it manifests remarkable generalizability and outperforms the only published result by a non-trivial margin of 4.8% for the Top-1 accuracy on the large-scale USPTO-full dataset. It is known that outputs of Transformer-based retrosynthesis model tend to suffer from insufficient diversity and high invalidity. These problems may limit the potential of Transformer-based methods in real practice, yet no prior works address both issues simultaneously. RetroPrime is designed to tackle these challenges. Finally, we provide convincing results to support the claim that RetromPrime can more effectively generalize across chemical space.