Accurate prediction of antibody function and structure using bio-inspired antibody language model

Area of Science:

• * Computational biology
• * Structural biology
• * Immunology

Background:

• * Antibodies are crucial therapeutics, but their development is limited by scarce structural data and complex engineering.
• * Deep learning advances protein structure prediction, yet antibody conformation remains challenging due to unique evolution and flexible binding regions.

Purpose of the Study:

• * To introduce a novel deep learning model, the Bio-inspired Antibody Language Model (BALM), for antibody-specific sequence analysis.
• * To present BALMFold, an end-to-end method for rapid, atomic-level antibody structure prediction from single sequences.
• * To enhance therapeutic antibody discovery and engineering.

Main Methods:

• * Training BALM on a large dataset of 336 million unlabeled antibody sequences.
• * Developing BALMFold as an end-to-end structure prediction pipeline derived from BALM.
• * Evaluating BALMFold's performance against established methods like AlphaFold2, IgFold, ESMFold, and OmegaFold on antibody benchmarks.

Main Results:

• * BALM demonstrated high performance in four antigen-binding prediction tasks.
• * BALMFold achieved superior accuracy in predicting full atomic antibody structures compared to existing state-of-the-art methods.
• * The model successfully predicts antibody structures from individual sequences rapidly.

Conclusions:

• * BALMFold significantly advances antibody structure prediction, outperforming current leading tools.
• * This method has the potential to streamline therapeutic antibody development and reduce experimental costs.
• * The BALMFold structure prediction server is publicly accessible for research and development.