AntibodyGPT: Language Modeling for fast ex novo Monoclonal Antibody Generation and Evolution

Antibodies are incredibly useful molecules with a broad spectrum of clinical applications, ranging from immunotherapy against viruses and toxins to the treatment of cancer. However, discovery and development of antibodies binding to a desired target is time-consuming and challenging, as it requires interrogation of the antibody repertoire of convalescent individuals or immunization of animals followed by extensive screening to find those rare B lymphocytes expressing molecules that bind to the desired antigen.Since the approval of the first monoclonal antibody (Palivizumab) against respiratory syncytial virus (RSV) by the FDA in 1998, prophylactic or therapeutic antibodies have been or are being developed against a broad range of infectious diseases, including Ebola, HIV, Yellow Fever, Zika and COVID-19. However, while antiviral antibodies are overall potent and safe, viruses can develop resistance to treatment. For example, SARS-CoV-2 evolves rapidly, mutating the epitopes that are recognized by antibodies and within a few years already made clinically approved monoclonal antibody therapies obsolete or of strongly diminished effectiveness. It is therefore crucial to continue exploring innovative approaches to overcome the challenges faced by antibody therapies, including those caused by the rapid evolution of viruses with considerable potential of causing future epidemics of great societal impact, such as coronaviruses.Artificial Intelligence (AI) made remarkable progress in recent months, especially in the areas of molecular structure prediction (e.g. Alphafold) and language modeling (e.g. ChatGPT). These advances bear great potential for synergistic application to biomedicine, for example through their application to the development of next-generation antibody-based immunotherapies that remain efficacious despite virus evolution. The overall goal of this proposal is to harness and apply recent advances in AI to the field of antibody-antigen recognition in the context of infectious diseases with epidemic potential. To achieve this, we will combine the complementary expertise of three collaborating laboratories, headed by Andrea Cavalli (applicant; computational structural modeling and AI), Davide F. Robbiani (co-applicant; discovery and characterization of monoclonal antibodies), and Daniel Ružek (partner; virology and preclinical development of antivirals). Specifically, we will develop and apply Large Language Modeling (LLM)-based methods to generate, in silico, antibodies with a determined specificity, or to evolve antibodies towards acquiring novel specificities, with the ambitious goal of preclinically advancing the most promising antibody against the coronavirus and its variants that was generated computationally. Ultimately, this work may pave the way to innovative methods, entirely in silico, for the rapid discovery of antibodies for prophylaxis or immunotherapy during response against the next pandemic threat.