Construction of deep learning model to predict the sugar donor specificity of UDP-glycosyltransferase

Uridine diphosphate (UDP)-glycosyltransferase (UGT) is representative enzyme for catalyzing the glycosylation of natural product or oligo- and polysaccharides , which can transfer sugar from UDP-sugar to acceptors. The decorated sugar types determine the function of glycosides. Currently, most reported plant UGTs have preference to UDP-glucose (Glc), while UGT showing activity towards non-UDP-Glc are rather rare, but they are also in urgent need in the glycosylation. It is difficult to obtain UGTs towards non-UDP-Glc with the current try-and-error method for mining UGTs, since most UGTs inherently show activity towards UDP-Glc. Herein, in this study, we established a dataset collecting 310 plant UGTs with clear UDP-sugar characterization, and based on this, a deep learning model, called UGT-donorCNN, was established to predict the UDP-sugar type of a given UGT. Finally, we showcased that UGT-donorCNN was effective in mining UGT with rare UDP-sugar preference other than UDP-Glc.