Deep Learning Tools for Protein Binder Design

Abstract

The ability to design protein-binding proteins is broadly useful. In this dissertation I will show our work to develop a deep learning-based pipeline for protein binder design. I will show how we configured AlphaFold2 to classify in silico designs which are likely to bind from those which are not likely to bind. I will then demonstrate how we can use the ProteinMPNN model, in combination with classical Rosetta protocols, to perform efficient sequence design on binder backbones. Finally, I will show how we trained a denoising diffusion model to generate protein backbones and how this can be used to massively accelerate the binder design pipeline. This deep learning-based pipeline is faster, easier to use, and has much higher experimental success rates than the previous Rosetta-based pipeline.