Evaluating H5N1 Vaccine Durability using Computationally-Designed Proteins

Navami Jain, Noor Youssef, Sarah Gurev, Debora Susan Marks

Keywords: Vaccine, viral evolution, deep learning, protein language models

TL;DR: We use deep learning models to design a panel of proteins to stress-test the protection by a flu vaccine

Abstract: The ongoing outbreaks of highly pathogenic avian influenza H5N1 viruses, particularly genotype D1.1 in clade 2.3.4.4b, and associated spillover events pose a concerning pandemic threat to humans. While immunity from seasonal flu exposure may confer limited cross-protection, H5N1 viruses are antigenically distinct and can evade this immunity. Moreover, while H5N1 vaccine stockpiles exist, they were developed against older clades and are a high mutational distance from currently circulating strains. Thus, they would confer only partial protection. We must ensure a new H5N1 vaccine provides sufficient protection against standing and future antigenic diversity likely to develop in a human population. To this end, we computationally design a panel of H5N1 proteins, VaxVal, to evaluate a candidate vaccine's durability and breadth. We show that deep learning models trained on historical Influenza hemagglutinin sequences can forecast close to 80% of the mutations that have occurred in clade 2.3.4.4b. Using these models, we successfully design 22 hemagglutinin variants, each carrying 2-4 mutations, that reflect antigenic changes across the protein. We show that constructs can easily escape protection by D1.1 vaccination and escape known broadly neutralizing monoclonal antibodies, sometimes close to 10-fold more than the wildtype. In forecasting immune escape, our pipeline can guide the design of broadly protective, long-lasting vaccines.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 161