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

We used the machine learning module of our computational design platform to iteratively propose mutations to the original antibody (M396) and run FoldX [21] calculations on LLNL HPC to estimate free energies to the SARS-CoV-2 spike protein homology model. FoldX binding calculations estimate the change in free energy (ddG) of the antibody-antigen protein complex resulting from the proposed mutations (to M396). We used this calculation as part of an objective function to drive the antibody optimization via the machine learning model. We evaluated 89,263 mutants with FoldX during the course of our search for improved antibodies with estimated ddG values ranging from -10.1 to 19.2 kcal/mole. Figure 3 plots improvements in the FoldX-estimated ddG values over the course of the machine learning-driven optimization process. These results indicate that the machine learning model was effective in searching the combinatorial space of possible antibody mutants to perform design optimization and identify increasingly improved predicted antibody designs. We then selected a subset of these antibody mutants for additional calculations using Rosetta [22] , STATIUM [24] , and molecular dynamics calculations [23] , as well as antibody developability estimates via the Therapeutic Antibody Profiler [25] (see Methods section). While all in silico calculations performed were used to select our 20 first round antibodies, the molecular mechanics/generalized Born solvent accessible surface area (MM/GBSA) molecular dynamics calculations [23] , described in the Methods section, are considered to be our most accurate estimate. MM/GBSA calculates antibody/antigen interaction free energies using fully solvated molecular dynamics (MD) for conformational sampling of the protein complex, but estimates free energy by a computationally less expensive implicit solvent model (GBSA). M396, our starting template for design, is known to neutralize SARS-CoV-1 by binding its RBD and preventing the virus from binding and entering the human ACE2 receptor; our MM/GBSA calculations for M396 in complex with the SARS-CoV-1 spike protein yield -52.2 kcal/mole (±7.2). M396 is known not to bind to the SARS-CoV-2 spike protein [18] and yields -48.