Abstract
Antibodies are versatile therapeutic molecules that use combinatorial sequence diversity to cover a vast fitness landscape. Designing optimal antibody sequences, however, remains a major challenge. Recent advances in deep learning provide opportunities to address this challenge by learning sequence-function relationships to accurately predict fitness landscapes. These models enable efficient in silico prescreening and optimization of antibody candidates. By focusing experimental efforts on the most promising candidates guided by deep learning predictions, antibodies with optimal properties can be designed more quickly and effectively. Here we present AlphaBind, a domain-specific model that uses protein language model embeddings and pre-training on millions of quantitative laboratory measurements of antibody-antigen binding strength to achieve state-of-the-art performance for guided affinity optimization of parental antibodies. We demonstrate that an AlphaBind-powered antibody optimization pipeline can deliver candidates with substantially improved binding affinity across four parental antibodies (some of which were already affinity-matured) and using two different types of training data. The resulting candidates, which include up to 11 mutations from parental sequence, yield a sequence diversity that allows optimization of other biophysical characteristics, all while using only a single round of data generation for each parental antibody. AlphaBind weights and code are publicly available at: https://github.com/A-Alpha-Bio/alphabind.