Coupling Machine Learning with Agent-Based Modeling to Design a Universal Influenza Vaccine

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Influenza is a devasting illness that causes up to 61,000 deaths each year in the United States, and up to 650,000 deaths each year globally. While influenza vaccines exist, they must be modified annually due to rapid sequence evolution of hemagglutinin (HA), the oft-targeted influenza surface protein (antigen, Ag). A vaccine formulated to be effective against diverse HA sequences would have greatly increased efficacy and would constitute a ‘universal’ influenza vaccine that could save many lives. To this end, there exist multiple groups of residues on the surface of HA that do not mutate as readily, due to their functional role in allowing the virus to attach to and enter host cells. The receptor binding site (RBS) contains such ‘conserved’ residues, which would be ideal targets for a universal influenza vaccine; however, the high sequence diversity of the surrounding ‘variable’ residues renders it difficult for antibodies (Abs) to bind to the conserved residues with high affinity. We hypothesize a universal influenza vaccine should comprise multiple immunizations of HA-based Ags with increasingly diverse sequences at variable positions surrounding conserved sites. We expect this approach to provide a continuous driving force for Abs to target conserved HA residues while simultaneously coaching them on how to tolerate or altogether avoid binding to variable residues. To test this hypothesis, we will adapt our computational model of affinity maturation (AM) – the process by which antibodies mature in vivo – geared towards evolving anti-HIV Abs, into a robust tool for Ag design against the conserved residues of the RBS. This model will incorporate important disease features, such as the crucial role of stabilizing framework mutations in the evolution of anti-influenza Abs. To efficiently traverse the vast sequence landscape of the HA- based Ags, we will employ deep reinforcement learning (DRL) to steer the AM process towards the optimal Ag sequences. We will first test this unique coupling of machine learning with stochastic biological modeling on our recently developed AM model with coarse-grained resolution to enable efficient optimization of algorithmic parameters. We will then apply this framework to our realistic AM model towards the design of real HA-based sequences for a universal influenza vaccine. Optimized HA sequences will be directly compared against naturally evolved anti-influenza Ab sequences with high potency and neutralization breadth against multiple influenza subtypes.