AI-Powered Generation of Human Monoclonal Antibodies Against Novel Viruses

Human monoclonal antibodies are a diverse class of therapeutic drugs that, in theory, can target any protein with extremely high specificity, making them highly promising candidates for treating a wide range of diseases. Until recently, antibody development primarily relied on discovery-based experimental methods, typically involving screening human or animal samples previously exposed to the antigen target of interest. Even with significant improvements in the throughput of antibody discovery methods, this process remains laborious, slow, and cost-ineffective. The continued expansion of the therapeutic market and the scope of monoclonal antibody applications has increased the demand for computational tools that can accelerate and expand antibody discovery capabilities.

Recently, significant breakthroughs in the field of artificial intelligence (AI), particularly the unparalleled performance of Transformer-based large language models (LLMs) and diffusion models in various tasks, have driven the rapid development of computational methods for antibody-related design tasks, including affinity maturation, antibody redesign, and the generation of single-domain antibodies.

However, no published method has yet demonstrated the ability to design template-free, antigen-specific antibodies. Existing methods are limited to antibody redesign, focusing on the generation of complementarity-determining regions (CDRs), which requires an initial antibody template to provide the variable genes and framework regions of the antibody. Furthermore, these antibody design models are primarily based on structural information and require antibody-antigen complex data during training. This presents significant limitations in situations with insufficient data.

Recently, researchers at Vanderbilt University Medical Center published a research paper titled "Generation of antigen-specific paired-chain antibodies using large language models" in the top international academic journal Cell. This study developed a protein language model (PLM)-based monoclonal antibody generator-MAGE (Monoclonal Antibody Generator)-which can generate paired human antibody variable heavy and light chain sequences targeting a specific antigen. Experimental validation showed that MAGE can generate novel and diverse antibody sequences targeting SARS-CoV-2, the emerging avian influenza virus H5N1, and respiratory syncytial virus A (RSV-A) with experimentally validated binding specificity.

MAGE is a first-in-class model that designs and generates human antibodies targeting multiple targets without the need for an initial template. MAGE is a Protein Language Model (PLM) that generates paired heavy and light chain antibody variable region sequences with binding specificity based on the input antigen sequence.

Figure 1. MAGE is a sequence-based protein language model designed to generate diverse human antibody sequences targeting various pathogens, including SARS-CoV-2, H5N1, and RSV-A. (Wasdin P T, et al., 2025)

MAGE's development is based on fine-tuning Progen2-an autoregressive decoder large language model pre-trained on general protein sequences. This model captures complex dependencies in the input sequence through a self-attention mechanism and learns patterns from observed amino acid sequences using a next-token prediction strategy.

This study leverages the pre-trained model's representation knowledge of amino acid sequences as a foundation to further learn human antibody sequence features related to binding specificity to diverse antigen targets. Experiments show that MAGE can generate antibodies with diverse sequence characteristics, including different heavy/light chain variable region gene usage patterns, somatic hypermutation (SHM) levels, and novel complementarity-determining regions (CDRs) not present in the training data.

When prompted with the SARS-CoV-2 wild-type receptor-binding domain (RBD), 9 out of 20 experimentally validated MAGE-generated antibodies (45%) successfully confirmed binding specificity, with one antibody exhibiting in vitro neutralization potency against SARS-CoV-2 better than 10 ng/ml. Furthermore, against the respiratory syncytial virus A (RSV-A) pre-fusion F protein, which was significantly underrepresented in the training data, 7 out of 23 MAGE-generated antibodies (30%) were experimentally validated for binding activity. The research team performed cryo-electron microscopy (cryo-EM) structural analysis of two MAGE-generated antibodies in complex with the RSV F protein. The results show that MAGE-generated antibodies exhibit diverse binding modes and can introduce key amino acid residues affecting function at the critical binding interface. Finally, for the H5/TX/24 influenza virus hemagglutinin (HA) antigen, which was not present in the training data, 5 out of 18 MAGE-designed antibodies (28%) were validated, demonstrating the model's zero-shot learning capability for novel antigens.

Therefore, MAGE is a first-in-class, landmark AI model that can design novel human antibodies with defined target binding functions without requiring antibody sequence templates.

Reference

Wasdin P T, et al. Generation of antigen-specific paired-chain antibodies using large language models. Cell, 2025.