Agentic AI for screening ligand-target binding affinity

Advances in high-throughput discovery technologies, computational design, and sequencing methods are generating unprecedented numbers of candidate ligands—including small molecules, antibodies, and other molecules—binding to disease-relevant targets. However, validating binding activity through wet-lab experiments remains costly and time-consuming. To address this bottleneck, we investigate the application of fine-tuned foundation models and agentic AI to computationally prioritize ligand candidates based on their predicted binding affinity to target molecules.

Our system leverages MAMMAL, an open-source biomedical foundation model trained on billions of biological samples spanning multiple modalities, including proteins, small molecules, and single-cell gene expression data. To evaluate its ability to prioritize ligand–target interactions across diverse modalities, we fine-tuned MAMMAL on multiple binding prediction tasks. For antibody binding prediction, we fine-tuned the model on the SAbDab dataset, achieving an AUC-ROC of 0.90 on the test set and outperforming baseline approaches based on ESM-2 embeddings coupled with classifiers. For small-molecule binding prediction, we fine-tuned MAMMAL on the SGC DEL dataset, achieving an AUC-ROC of 0.95 on the test set. These results demonstrate the ability of foundation models to learn binding-relevant representations across diverse ligand classes.

To make these models accessible to researchers without machine-learning expertise, we integrated them into an agentic AI system. The fine-tuned models were exposed as tools within agentic workflows, enabling interactive research dialogues in which AI agents select tools, perform inference, and interpret results. For selected models, we additionally provide optimized inference through vLLM plugins to support efficient, scalable deployment. This framework lowers technical barriers while enabling efficient triage of ligand–target pairs and reducing experimental workload. The architecture is designed to accommodate additional ligand modalities and binding prediction tasks, supporting scalable virtual screening and therapeutic discovery. This work was conducted in part within the scope of the LIGAND-AI EU project.