Computational tools for studying viral antigenic evolution

Abstract

RNA viruses like influenza and SARS-CoV-2 evolve rapidly under selection from host immunity, accumulating mutations that enable escape from antibodies elicited by prior infection or vaccination. However, viral evolution is constrained by the requirement that surface proteins must remain functional: they must fold properly, express on the virion, and bind host cell receptors to mediate entry. Understanding how mutations affect these molecular phenotypes alongside immune evasion is essential for predicting viral evolution and designing effective vaccines. This dissertation presents three computational tools for modeling genotype-phenotype relationships inviral proteins and for simulating viral evolution at the population level. These tools share a common emphasis on biological interpretability while leveraging modern statistical and machine learning methods to analyze high-throughput experimental data or benchmark viral surveillance methods. The first tool, torchdms, suggests a framework to jointly model multiple phenotypes from deep mutational scanning experiments using biophysically motivated neural network architectures. We apply torchdms models to data from a deep mutational scan of the SARS-CoV-2 receptor binding domain, demonstrating accurate prediction of variant phenotypes and recovery of biologically meaningful mutational effects. The second tool, polyclonal, models how mutations enable viral escape from polyclonal antibody responses. The model decomposes serum activity into epitope-specific components and quantifies mutation effects at each epitope, capturing the nonlinear patterns of escape that arise when antibodies target multiple sites on a viral antigen. The third tool, antigen-prime, is a forward-time epidemic simulator that models the coupled genetic and antigenic evolution of viruses under selection from host population immunity. We use antigen-prime to simulate 30 years of influenza-like evolution and benchmark methods for variant assignment and growth rate inference, revealing a previously undocumented failure mode in growth rate inference models.