For antibody sequence generative modeling, mixture models may be all you need.

Area of Science:

• Biotechnology
• Immunology
• Computational Biology

Background:

• Antibody therapeutics require high target affinity and favorable developability, including low immunogenicity.
• Assessing antibody sequence properties is crucial for therapeutic candidate selection.

Purpose of the Study:

• To develop a simple, accurate, and interpretable generative model for human antibody sequences.
• To evaluate the model's performance against existing methods, including large language models (LLMs).
• To create tools for antibody sequence humanization, generation, and scoring.

Main Methods:

• A simple generative model, SAM, was trained on a large dataset of human antibody sequences (60 million heavy, 70 million light chains).
• Model performance was benchmarked on datasets exceeding 400 million sequences.
• A new, highly efficient tool for antibody sequence numbering was developed.

Main Results:

• SAM accurately distinguishes human antibody sequences from other species, surpassing existing models and LLMs.
• SAM demonstrates capabilities in humanizing, generating, and scoring antibody sequences for humanness.
• The developed antibody sequence numbering tool is significantly faster than current methods.

Conclusions:

• Simple generative models can serve as effective baselines for protein engineering tasks.
• SAM offers a fast, interpretable, and high-performing solution for antibody sequence analysis.
• The developed tools, including SAM and the numbering tool, are readily available to the research community.