A Large Language Model Guides the Affinity Maturation of Variant Antibodies Generated by Combinatorial Optimization

Machine learning-based antibody design and optimization by computational affinity maturation is emerging as a promising approach to combating infectious diseases. This has been possible because of significant advances in artificial intelligence methods and a surge in experimental datasets on antigen-antibody interaction. The ability of an antibody to bind an antigen with sufficient strength (measured by binding affinity , the inverse of the equilibrium dissociation constant) and specificity are critical properties in the design of neutralizing antibodies. Here we introduce Ab-Affinity, a new large language model in conjunction with a genetic algorithm and simulated annealing for diversity generation and fitness optimization, which can accurately predict the binding affinity of specific antibodies against a target peptide within the SARS-CoV-2 spike protein. When trained on large datasets of existing antibodies that bind to certain antigens, we show that Ab-Affinity can generate novel antibodies with more than a 160-fold enhancement in binding affinities over those obtained experimentally. The predicted biophysical properties of the synthetic antibodies demonstrate their robustness. Molecular docking and molecular dynamics simulation of the binding interactions of the best candidate synthetic antibodies showed enhanced interactions and stability on the target peptide epitope. In general, antibodies generated by Ab-Affinity appear to be superior to those obtained with other existing computational methods.