UW Department of Physics

Machine Learning for Immune Receptor Discovery and Design

The adaptive immune system can mount a specific response against a multitude of pathogens, through molecular recognition grounded in complex biophysics of protein-protein interaction between immune B-cell and T-cell receptors and pathogenic proteins. In the past few years, machine learning has revolutionized protein science, opening new avenues to study immune-pathogen interactions. This project aims to leverage these innovative approaches to develop a versatile suit of tools for modeling immune function and designing novel immune receptors and antigens, with vast therapeutic applications. Specifically, this student team will work to build upon the geometric deep learning algorithms for protein science, from Nourmohammad's lab, and develop software platforms for analysis of immune receptor response to pathogens. This student team will work to create a flexible software that can incorporate a versatile suit of ML tools for this purpose, and that can be interfaced with easily by non-programmers. This student team will work to implement existing ML algorithms (developed by Nourmohammad's lab and others) and partially improve the algorithms to predict function of immune receptors and ultimately design novel immune receptors and antigens for a given target. The team will work to develop a well-functioning software interface for this purpose. The outcomes this student team will work to achieve are: 1. Transfer learning model for B-cell (antibody) and T-cell receptor function prediction and design, built upon prior ML developments in Nourmohammad's lab for protein universe (as a foundation model) for proteins structures, and further incorporating other work on protein sequence language models. 2. Development of a software with ML model on the back end to predict the effect of mutations in immune receptors and antigens on the specificity and affinity of interactions, given a immune-antigen protein complex. In the same vein, having a component of the software to design new immunogens for a given immune receptor. Ideally, the final tool can be interacted with via a web interface.