Enzyformer: a two-stage pretrained model for enzymatic retrosynthesis

Although enzymatic retrosynthesis planning is widely used to accelerate drug synthesis and discovery, it still faces key challenges, including limited knowledge transfer from general chemical space and incomplete EC number prediction. Here, we introduce Enzyformer, an innovative model that integrates the pretraining of molecules and reactions to capture both the syntax of SMILES and the transformation rules of organic reactions. With this two-stage pretraining strategy, Enzyformer improves top-1 and top-10 accuracies in retrosynthesis prediction by 7.5% and 11.7%, respectively, compared with the baseline R-SMILES. In addition, our pipeline also includes EC number prediction as a sequential task, using a contrastive learning model to improve performance on ECREACT dataset. With the contrastive learning model, we achieve the highest average F1-score of 0.924 for first-level EC number prediction. In conclusion, Enzyformer offers a promising solution for improving the accuracy and interpretability of enzymatic retrosynthesis prediction, which may help to reduce researchers' workload and cost while accelerating drug design and mitigating the environmental impact of drug synthesis.