Machine learning-guided engineering of phosphocholine cytidylyltransferase for improved activity and halotolerance in citicoline production

Phosphocholine cytidylyltransferase (CCT) is an important biocatalyst for citicoline (CDP-choline) production. However, it suffers from relatively low catalytic activity and halotolerance when it was applied in the one-pot catalytic system coupling with the acetylphosphate based ATP regeneration system. A machine learning guided directed evolution approach was applied to simultaneously improve activity and halotolerance of Saccharomyces cerevisiae CCT (ScCCT), through random mutation on “hot-spot” regions predicted by selected models trained on different combinatory datasets generated by site-directed mutagenesis and random mutagenesis. The most desirable variant M3 (P347S/P365L/K340T) exhibited an approximately 1.4 folds improvement in final titer of CDP-choline (182 mM) comparing with WT. This research proves that machine learning can improve effectiveness of random mutagenesis, and provides a possible solution to engineering membrane protein with complicated evolutionary fitness landscapes, which can be difficult for classic enzyme engineering approaches.