p-IgGen: a paired antibody generative language model.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Antibody drug discovery faces challenges in designing sequences with optimal developability.
• Issues such as aggregation, polyspecificity, poor expression, and low solubility hinder therapeutic antibody development.

Purpose of the Study:

• To present p-IgGen, a novel protein language model for paired heavy-light chain antibody generation.
• To develop a fine-tuned version of p-IgGen for generating antibodies with desirable 3D biophysical properties.

Main Methods:

• Utilized a protein language model for antibody sequence generation.
• Trained and fine-tuned the p-IgGen model on antibody sequence data.
• Evaluated generated sequences for antibody-like properties and biophysical characteristics.

Main Results:

• p-IgGen generates diverse, antibody-like sequences with natural pairing properties.
• The fine-tuned model biases generation towards antibodies with clinical-stage biophysical distributions.
• The model addresses key developability challenges in antibody design.

Conclusions:

• p-IgGen offers a powerful tool for designing novel antibody sequences with enhanced developability.
• The model facilitates the generation of therapeutic antibodies with improved biophysical properties.
• This approach aids in overcoming critical hurdles in antibody drug discovery.