Enzyme optimization and design with deep learning

Abstract

Enzymes are valuable tools for the catalysis of reactions under mild conditions. However, they often misfold when applied outside of their natural contexts. Thus, methods to optimize natural enzymes and design completely bespoke enzymes would be highly valuable. Recent advances in deep learning-guided protein design have made tractable what have historically been grand challenges in the field. Here, I apply these tools to enzyme design, both in the optimization of natural enzymes and the generation of completely novel hydrolases and plastic-degrading enzymes. In the former, I developed a simple and accessible method of sequence redesign to optimize the physical properties of valuable natural proteins. As a proof of concept, I applied this method to the protease from Tobacco etch virus (TEV protease). The designed variants showed increased expression, stability, and even function. In the latter project, I applied de novo protein design and modeling tools to create bespoke hydrolases for the model reaction of simple ester hydrolysis, and achieved catalytic efficiencies of up to 105 M-1 s-1. Applying these tools to the hydrolysis of polyethylene terephthalate (PET), I created, to our knowledge, the first completely de novo designed plastic-degrading enzyme. These data represent significant advances in the field of enzyme design and establish methods for the development of new catalysts.