ABSTRACT Enzymatic biosynthesis has become increasingly crucial in green chemistry and biosynthesis. However, current computational tools struggle to effectively integrate enzyme identification with pathway synthesis due to the specificity of enzymes and their complex interactions with substrates. Here, we propose EnzRetro, a novel framework for enzymatic retrosynthesis that provides an end‐to‐end solution bridging retrosynthesis planning with enzymatic engineering. The core innovative concept of EnzRetro is site‐specific reaction edits (SSREdits), a dynamic approach to representing structural transformations at specific enzyme active sites and forging a direct link between enzyme identification and reaction patterns. To enable the model to learn meaningful representations of SSREdits, we developed three pretraining tasks and then fine‐tuned two specialized models: (1) the SSREdits generation model for pathway synthesis, which translates the target product into a sequence of reaction edits, and (2) the EC generation model for enzyme identification, which focuses on precise transformation sites within SSREdits, enabling generalization across diverse reactions. Extensive experiments demonstrate the superior performance of EnzRetro, with a promising 56.1% and 97.7% Top‐1 accuracy on USPTO‐50k dataset for retrosynthesis and ECREACT dataset for enzyme identification, respectively. Finally, EnzRetro combines the two‐stage processes of retrosynthesis and enzyme identification into one‐pot learning, enhancing both computational efficiency and interpretability, and bridging the gap between pathway synthesis and enzyme identification. The accuracy of EnzRetro has been validated by three enzymatic pathways. We developed a web platform for multi‐step retrosynthesis planning that reconstructs multiple enzymatic pathways for putrescine biosynthesis with substantial diversity and significantly outperforms state‐of‐the‐art baselines.
Cao et al. (Thu,) studied this question.