Machine learning optimizes antibody yields, yielding sub-nanomolar affinity libraries with high diversity.

Scientists have developed a cutting-edge machine learning technique to create high-affinity sub-nanomolar antibodies in large libraries, according to a recent study published in the journal Nature Communications. The technique combines language models, Bayesian optimization, and extensive experimentation to generate strong-variable fragments of Fabs which are affixed with targets. These fragments are known as scFvs. The new technique works by integrating information from large natural protein sequences with target-specific binding affinities, ultimately leading to the rapid and efficient engineering of stronger scFvs in comparison to traditional directed evolution methods. The process involves five stages starting from generating supervised training data using an engineered yeast mating assay to the Bayesian-based fitness landscape that thoroughly maps the scFv sequence to posterior probability while iterating through multiple sampling algorithms to find the best candidate. Although the process generates more diverse and stronger-tailored antibodies than traditional methods, scientists advise that a balance is necessary to successfully manage trade-offs between performance and variance.