Machine learning enables engineering of potent, specific, and therapeutically developable proteases

Jung-Eun Shin, Nathan J Rollins, Purvi Mande, Jordan M Anderson, Allison Colthart, Soumya Bengeri, Emily Hoyt, Alex Pellerin, Ivan Mascanfroni, Jyothsna Visweswaraiah, Yi Xing, Kevin L. Otipoby, Nathan Higginson-Scott, Ryan Peckner

Keywords: Drug design, Immunogenicity, T cell epitope prediction, Rational design, Enzyme therapeutics

TL;DR: We apply machine learning and rational design to re-engineer naturally occurring, immunoglobulin-cleaving proteases as therapeutics.

Abstract: The development of biologic therapeutics is a labor-intensive, time-consuming, and expensive process. From the discovery of a protein with the desired function to optimizing the protein for manufacturability and removing immunogenicity, each step requires many iterations and large screening efforts. In order to accelerate this process, efforts have been increasingly focused on methods for protein engineering and design, both experimentally and computationally. Here we show that the careful application of machine learning methods in combination with structure-based and data-driven rational design allows for successful engineering of proteases with desired target specificity, low immunogenicity, and potent cleavage potential. We present two case studies describing the impact of machine learning in the development of two potential therapeutics. First, we apply machine learning models to optimize an IgG-degrading protease for drug-like properties while minimizing immunogenicity and other liabilities and maintaining therapeutic activity. Second, we leverage both machine learning and rational design methods to engineer an IgE-degrading protease with high specificity and activity. These results demonstrate the power of applying machine learning methods to accelerate the discovery and development of protein therapeutics.

Submission Number: 86