Novel backbone methods for de novo protein design

Abstract

Solving a protein design problem first requires sampling suitable backbones given the needs and constraints of the problem. The available structural space of backbones is vast, containing countless potential solution backbones for a given problem. Previous methods to explore this space include parametric sampling, fragment assembly, and more recently, generative deeplearning methods. With more advanced methods and algorithms, we can more effectively sample this space to solve new protein design problems. In this work, I present three protein design projects focused on backbone sampling methods. First, I describe a traditional parametric approach to redesign heterodimers for synthetic protein logic. Second, I describe a generative reinforcement learning approach I developed to design protein architectures from the top–down. This method can fill arbitrary volumes and enables the design of capsids for vaccine antigen presentation. Third, I describe a collection of methods used to accomplish helical peptide recognition. The resulting high-affinity binders are useful as capture reagents for disease diagnosis, and can be engineered into biosensors.