(118g) Deep Reinforcement Machine Learning Driven Vaccine Design Against Highly Mutable Pathogens

Made abundantly clear by the current COVID-19 pandemic, there is a pressing need for new approaches to rapidly develop robust vaccines against highly mutable pathogens. SARS-CoV-2 has demonstrated the worst-case scenario in this regard, however, HIV and influenza continue to pose serious threats annually because of the lack of universal vaccines against these pathogens. To this end, computational methods have the potential to revolutionize the development of such vaccines. Recent advances in agent-based models of affinity maturation (AM) – the Darwinian process by which antibodies evolve against a pathogen – have enabled the use of real nucleotide and amino acid sequences of B cell receptors and pathogenic proteins (antigens; Ags), respectively, in rational vaccine design. However, the vast sequence space of the vaccine-candidate Ags does not feasibly allow for iterative Ag sequence design due to the high computational demands. Here, we present the development of an efficient and robust computational pipeline for designing vaccines against highly mutable pathogens. This is facilitated by coupling deep reinforcement learning (DRL) algorithms to agent-based models of AM to effectively ‘steer’ the AM process towards an optimal solution, namely, towards evolving large quantities of cross-reactive antibodies against the administered Ags. We first describe coupling off-the-shelf DRL algorithms to a previously published coarse-grained model of AM to determine optimal coupling performance parameters, considering both speed and accuracy against previously published results. We then describe coupling DRL to the more realistic AM model described earlier, exploring the optimal way in which to administer four HIV-based Ags with fixed sequences across three sequential immunizations. Our results serve as a proof-of-concept for efficient and biologically-relevant Ag sequence design using next generation computational and data-driven methods. Future work will include development of a computational pipeline to explore the Ag sequence space and determine the optimal amino acid sequences of HIV-based Ags for evolving cross-reactive antibodies across various numbers of immunizations. Furthermore, we plan to optimize this pipeline to encompass a larger repertoire of pathogens, such as influenza and malaria. Using this approach, we hope to accelerate the development of universal vaccines against a variety of highly mutable pathogens that remain significant burdens on global health today.