Machine learning optimization of candidate antibody yields highly diverse sub-nanomolar affinity antibody libraries

Area of Science:

• Biotechnology
• Computational Biology
• Drug Discovery

Background:

• Therapeutic antibodies are a key drug class, but their design is time-consuming and expensive.
• Developing high-affinity antibody fragments, such as single-chain variable fragments (scFvs), is crucial for effective therapeutics.

Purpose of the Study:

• To present a novel, end-to-end Bayesian language model for designing large and diverse libraries of high-affinity scFvs.
• To compare the performance of this AI-driven method against traditional directed evolution approaches.

Main Methods:

• Utilized a Bayesian, language model-based approach for the de novo design of scFv libraries.
• Empirically measured the binding affinity of designed scFvs.
• Compared AI-designed scFvs against those generated through directed evolution.

Main Results:

• The best scFv designed by the AI method showed a 28.7-fold improvement in binding compared to the best from directed evolution.
• 99% of designed scFvs in the most successful library demonstrated improved binding over the initial candidate.
• The method successfully predicted library success and explored diversity-affinity tradeoffs.

Conclusions:

• Machine learning models, particularly language models, can significantly accelerate and enhance scFv development.
• This AI-driven method offers a powerful, broadly applicable tool for protein engineering and therapeutic antibody design.