Antibody Language Models Point to 'Evolutionarily Plausible' Mutations

In Nature Biotechnology, a Stanford University-led team considers protein language models as a method for designing artificial antibodies based on knowledge of realistic natural protein evolution trajectories. Using a language model-guided strategy to screen a select number of laboratory evolution-related variants in seven antibodies known for binding coronavirus, ebolavirus, or influenza A virus antigens, for example, the investigators were able to narrow in on changes that bolstered binding affinities for mature or unmatured forms of the antibodies. "We improved the affinity of all antibodies after measuring only 20 or fewer new variants of each antibody across just two rounds of evolution, which, to our knowledge, represents unprecedented efficiency for machine learning-guided evolution," they report. "We also demonstrate that the same general protein language models that we sued to affinity mature antibodies can also enrich for high-fitness substitutions to diverse proteins beyond antibodies."