Advances in Generative and Discriminative Models for De Novo DNA Binder Design

Abstract

In organisms across domains of life, the process of gene regulation, or the expression of information stored in the genetic code, is dominated by interactions between proteins and DNA. In particular, various families of DNA binding proteins are responsible for: scaffolding, winding, unwinding of the DNA strand; promotion and suppression of specific genes; damage detection and repair; and more. Although numerous engineering tools have been developed in recent years to manipulate DNA, these tools are derived from naturally-occurring biological matter, and are therefore limited in scope and function — at best, we can slightly modify or mix and match functions that nature has evolved. Recent advances in the field of de novo protein design, where proteins are rationally designed by humans from scratch, promises to unlock new, powerful genetic engineering tools, with novel function wholly outside of the scope of existing biological matter. In this thesis I describe efforts to advance the field of de novo protein design for DNA binding, starting with 1. a set of increasingly powerful generative machine learning models for protein design, applied to DNA binders; 2. a computational pipeline for designing DNA binding proteins with the ability to discover sequence selective DNA binding proteins from as few as 96 designs tested in the wet lab, and 3. discriminative models for DNA specificity prediction, to enhance the ability to filter candidate DNA binding proteins.