Substrate-Aware Zero-Shot Predictors for Non-Native Enzyme Activities

Abstract

Enzymes can be engineered to catalyze reactions with non-native substrates or even perform entirely new reactions unknown in nature. However, developing such novel activities through wet-lab engineering is time- and resource-intensive. By estimating enzyme activity without new experimental data, zero-shot (ZS) predictors can accelerate enzyme engineering. While ZS predictors have been demonstrated in various contexts, they have yet to be evaluated on non-native substrates and $\textit{new-to-nature}$ chemistry. Critically, many existing predictors do not explicitly encode substrate or transition-state properties, which we propose are essential for predicting $\textit{new-to-nature}$ chemistry. Here, we systematically studied two types of mechanistically distinct enzymes using 16 ZS predictors—including six general and ten substrate-aware scores derived from generative modeling, molecular docking, and active-site heuristics. We curated new experimental and literature mutation datasets spanning 11 non-native substrates and three $\textit{new-to-nature}$ reactions with 11 additional substrates. The six general ZS predictors could generalize to non-native substrates, but failed for $\textit{new-to-nature}$ chemistries. In contrast, certain substrate-aware approaches could predict $\textit{new-to-nature}$ chemistries, with AlphaFold 3's chain-predicted aligned error being the most predictive of both activity and stereoselectivity. A weighted ensemble of AlphaFold 3 and EVmutation scores generalized to all chemistries that we tested. Our findings highlight the potential of ZS predictors to accelerate enzyme engineering, advancing the expansion of the chemical universe beyond nature’s repertoire.