Protein redesign by learning from data

Protein redesign methods aim to improve a desired property by carefully selecting mutations in relevant regions guided by protein structure. However, often protein structural requirements underlying biological characteristics are not well understood. Here, we introduce a methodology that learns relevant mutations from a set of proteins that have the desired property and demonstrate it by successfully improving production levels of two enzymes by Aspergillus niger, a relevant host organism for industrial enzyme production. We validated our method on two enzymes, an esterase and an inulinase, creating four redesigns with 5-45 mutations. Up to 10-fold increase in production was obtained with preserved enzyme activity for small numbers of mutations, whereas production levels and activities dropped for too aggressive redesigns. Our results demonstrate the feasibility of protein redesign by learning. Such an approach has great potential for improving production levels of many industrial enzymes and could potentially be employed for other design goals.