Generation of antigen-specific paired-chain antibodies using large language models.

Summary

This study introduces MAGE, a protein language model that generates paired heavy/light-chain human antibodies with validated binding to SARS‑CoV‑2, H5N1, and RSV-A. The approach bypasses template dependence and demonstrates broad, target-specific antibody design capability.

Key Findings

• A sequence-based PLM (MAGE) generated paired heavy/light-chain human antibodies with experimental binding to SARS‑CoV‑2, H5N1, and RSV‑A.
• Antibody design was accomplished without a starting template, indicating de novo capability.
• Generated antibodies were novel and diverse, demonstrating breadth across multiple antigens.

Clinical Implications

While not yet clinically validated, this platform could accelerate generation of therapeutic and prophylactic antibodies for emerging respiratory viruses (e.g., pandemic influenza, RSV, novel coronaviruses), enabling faster translation to trials.

Limitations

• Lacks in vivo neutralization and efficacy data; binding does not guarantee therapeutic activity.
• Manufacturability, stability, and immunogenicity were not assessed.

Future Directions

Prospectively validate neutralization and protection in relevant animal models and early-phase trials; integrate developability filters and multi-objective optimization (affinity, specificity, stability) into the LLM pipeline.