Improving Vaccine Design with an AI Booster

A critical enabler of the 100 Days Mission is the establishment of a global “vaccine library”— a globally accessible store of scientific knowledge, data and prototype rapid-response vaccine candidates against selected viruses from the 25 priority virus families. CEPI’s aim is to store AI-generated, lab-tested and verified antigen designs developed by the HMRI consortium in the vaccine library so they can be quickly used to develop vaccine candidates in the event of an outbreak of a novel pathogenic threat. In this scenario, after sequencing the offending virus, these cataloged antigen designs could then be inserted into an appropriate rapid-response vaccine platform to start the production of vaccines for clinical testing. Gollihar’s group is leading immunogen design, but their work encompasses more than that. The team has been developing therapeutic monoclonal antibodies for SARS-CoV-2, and, as variants of concern emerged, they wanted to determine what those mutations were doing to their monoclonals. So, they developed a mammalian display of viral glycoproteins—called a spike display— to allow them to study virus mutations and source code in real time. “We started using spike display to dissect escape mechanisms etc., and realized it was also an engineering tool. We played around with rational design-based approaches with individual variants and then moved to library approaches where we could use millions of variants in human cells and mammalian cell lines and then sort, seek and find out who's binding what. The confluence of AI and directed evolution is outright, but protein engineering is really what allows us to do this rapidly,” said Gollihar. AI experts from the HMRI, University of Texas at Austin, La Jolla Institute and Argonne National Laboratory will use machine-learning approaches to optimize the design of potential epitopes. The University of Texas Medical Branch will validate their immunogenicity in established preclinical models. When a new pathogen emerges, vaccine developers could quickly respond by selecting AI-identified epitopes that would have already been validated in preclinical tests, enabling vaccine candidates to be moved quickly into clinical testing. It would provide a significant strategic advantage when battling the next pandemic threat, but Gollihar is looking even further into the future. “The next generation of models for protein engineering will require data sets that don’t yet exist, so we’re really interested in deep mutational scanning,” noted Gollihar. “Because of our engineering platform, we can now take every amino acid and put it into every single position of a protein and ask what that does to expression, to antibody binding or to host receptor binding. We’ll learn things we couldn’t otherwise learn. This is the next frontier and we're leading it.”