Rapid in silico design of antibodies targeting SARS-CoV-2 using machine learning and supercomputing

Thomas Desautels; Adam Zemla; Edmond Lau; Magdalena Franco; Daniel Faissol

Lawrence Livermore National Laboratory (LLNL) and GlaxoSmithKline (GSK) Vaccines Research have developed a combined computational-experimental platform for vaccine antigen design over the last two years. This combined platform combines experiment-driven, data-driven, and theory-driven approaches to leverage the strength of each approach, while mitigating their limitations. The initial predictions from this platform are driven by a computational component based on integrating existing experimental data, structural biology/bioinformatic modeling, and molecular dynamics simulations on high-performance computing systems. An active machine learning model aims to optimize binding behavior by iteratively proposing mutations to the amino acid sequence of an initial antigen. Proposed mutant antigens are evaluated with existing computational binding estimation tools using known or estimated antibody-antigen structures. The platform leverages a feature representation of the three-dimensional antigenantibody interface and a Bayesian optimization algorithm to propose computational evaluation of mutants with high predicted performance and mutants that improve the machine learning model itself. The computational platform further improves its predictions by proposing designs that are evaluated with a high throughput experimental evaluation component, the results of which are incorporated into the machine learning model in a feedback loop. This combined computational-experimental antigen design platform, and results generated by it, will be described in greater detail in a future joint publication with GSK, including details of the machine learning model.