The article details the generation of antigen-specific paired-chain antibodies using large language models (LLMs). The authors, Wasdin et al., introduce MAGE (monoclonal antibody generator), a sequence-based protein language model fine-tuned to efficiently design diverse human variable heavy- and light-chain antibody sequences targeting specific pathogens. MAGE utilizes an extensive antibody-antigen sequence database and demonstrated its capability by generating functional antibodies with experimentally validated binding specificity against SARS-CoV-2 (RBD), H5N1 (avian influenza), and RSV-A (respiratory syncytial virus A). This AI-driven approach overcomes the limitations of traditional, template-dependent antibody discovery, even showing zero-shot learning capabilities against an antigen not explicitly present in its training data. Structural analysis confirmed that the MAGE-generated antibodies exhibit diverse binding modes and high novelty compared to known sequences.

References:

• Wasdin P T, Johnson N V, Janke A K, et al. Generation of antigen-specific paired-chain antibodies using large language models[J]. Cell, 2025.