Ancestral Sequence Reconstruction Meets Machine Learning: Ene Reductase Thermostabilization Yields Enzymes with Improved Reactivity Profiles

Abstract

Ene reductases (EREDs) are enzymes that catalyze the asymmetric reduction of C═C bonds. EREDs are potentially useful in the large-scale synthesis of pharmaceutical compounds, but their application as biocatalysts is limited because they are often unstable under process conditions. Previous work addressed this limitation by identifying stabilized EREDs with ancestral sequence reconstruction (ASR), a bioinformatic method that predicts evolutionary ancestors based on a set of homologous sequences. In this work, we sought to apply ASR to design enzyme libraries and leverage machine learning to predict the most stable library variants. We generated an ERED library that targeted residues based on uncertainty in the ASR prediction. Screening data from a portion of the library were used to build a machine learning model that could accurately predict variants with improved thermostability. The most stabilized enzyme outperformed the wild-type and ancestral parent enzymes under process-like conditions with a panel of substrates. We envision that the combination of ASR and machine learning could be generally applied to other classes of enzymes, facilitating the development of high-quality industrial biocatalysts.