De novo protein design with deep learning

Abstract

The ability to predict the exact set of atoms which possess a desired chemical function lies at the heart of modern day problems in energy, health, and sustainability. A particularly useful medium for constructing these sets of atoms is proteins due to their diverse chemistry, ease of synthesis, and broad compatibility with materials and molecules from across the periodic table. The grand challenge of designing proteins to possess chemical functions is the problem of mapping a desired chemical function to the amino acid sequence which would encode it. Here I describe my efforts (in close collaboration with many colleagues) to uncover this mapping using deep learning. Specifically, through the training and deployment of various deep neural networks that generate protein structures in an intuitive and useful manner, we uncover design principles and algorithms that enable a new generation of custom-designed molecules.