Agentic AI for screening antibody:antigen binding affinity

Recent progress in humanized and transgenic antibody mouse models coupled with single-cell BCR RNA sequencing has generated unprecedented numbers of antibody candidates targeting disease-relevant antigens. Antibodies are an attractive and well-established therapeutic modality, offering high binding specificity, strong target affinity, and well-characterized developability and manufacturing pipelines. However, validating binding activity through wet-lab experiments is costly and time-consuming. To address these challenges, we investigate the application of fine-tunned foundation models and agentic AI to computationally prioritize antibody candidates for target binding. Our system leverages MAMMAL, an open-source biomedical foundation model trained on billions of protein sequences, which we fine-tuned for antibody binding prediction using the SabDab dataset. We systematically evaluated fine-tuning strategies varying in input type (immunoglobulins versus heavy-chain–only), input representation (full antibody sequence versus CDR3 region) and training approach (full-model fine-tuning versus low-rank adaptation, LoRA). The top-performing models, selected based on AUC-ROC performance, were subsequently used to generate a ranked list of candidate binders. To democratize access for biomedical researchers and make interaction with predictive models more intuitive, we developed an agentic system. Fine-tuned models were integrated as MCP-based tools within agentic workflows, enabling interactive research dialogues in which AI agents autonomously select tools, execute inference, interpret results, and iteratively refine outputs. This integration lowers technical barriers by embedding specialized models into accessible interfaces, thereby broadening adoption. Our results demonstrate that even with limited fine-tuning data, agentic AI can efficiently triage thousands of antibody–antigen pairs, reducing experimental workload, accelerating discovery, and improving the process of lead candidate selection. Collectively, this system illustrates the potential of agentic AI to transform virtual antibody screening into a scalable, efficient, and widely accessible approach for biomedical research and therapeutic development.