Epistatic protein models trained on homologous protein sequences found in nature have proven to be highly informative in protein design. However, these epistatic models, which predict how combinations of amino acid mutations affect evolutionary fitness, generally need 103-104 natural sequences to achieve strong predictive performance. Additionally, a protein’s evolutionary fitness is not always correlative with functions of interest, such as catalytic activity. Thus, it is desirable to investigate how such powerful models can be adapted to proteins with too few natural homologs and whether functional information gathered from high throughput assays can improve model performance. We proposed that endolysins are an ideal test case for this work as they are compelling selective antimicrobial enzymes which degrade the peptidoglycan layer of targeted bacteria, would benefit therapeutically from stability and activity engineering, and have &gt;104 natural homologous sequences, allowing rigorous comparisons to epistatic models trained on subsets of these natural sequences. Towards this goal, we generated six libraries of the endolysin LysEFm5-V7 using a systematic set of protein diversification strategies and evaluated variants for proteolytic stability on the surface of yeast and catalytic activity in a novel depletion-based assay in E. coli. Stability and activity results from &gt;105 tested variants trained sitewise and pairwise models to assess which diversification strategies and functional data most accurately model performance. Our results elucidated efficient diversification and evaluation strategies for improving protein and enzyme function, yielded a model which achieved higher predictive performance than epistatic models in the field, and led to the identification of a compelling endolysin variant with enhanced stability.

Engineering endolysins via epistatic protein models trained on deep functional data

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.