GPCRAgent: an intelligent multi-agent platform integrating knowledge graphs and large language models for GPCR-targeted drug discovery

G protein-coupled receptors (GPCRs) are major therapeutic targets in drug discovery. However, systematic characterization of GPCR functions and ligand interactions remains constrained by the high cost and complexity of experimental approaches. Although large language models (LLMs) have shown strong performance in biomedical question answering, their direct application to GPCR research is limited by insufficient integration of structured domain knowledge and poor interpretability of generated inferences.

This study aimed to develop an interpretable artificial intelligence framework for automated and traceable analysis of complex GPCR-related problems, and to evaluate its utility in GPCR knowledge reasoning, structural property interpretation, disease target prioritization and ligand recommendation.

We developed GPCRAgent, a multi-agent artificial intelligence framework enabled by VecKG-CoReason, which integrates vector-based semantic retrieval with knowledge graph-guided collaborative reasoning. GPCRAgent was evaluated on benchmark tasks covering GPCR knowledge question answering, multiple-choice reasoning and structural property interpretation. Its practical utility was further assessed in an idiopathic pulmonary fibrosis case study, in which prioritized targets and ligand candidates were supported by molecular docking and molecular dynamics simulations.

GPCRAgent achieved accuracies of 85 % in GPCR knowledge question answering, 90 % in multiple-choice reasoning and 83 % in structural property interpretation, outperforming general-purpose LLMs across benchmark tasks. In an idiopathic pulmonary fibrosis case study, GPCRAgent identified LPAR1 as the top candidate and prioritized potential ligands. In a separate renal fibrosis study, experimental evaluation of the optimized LPAR1 ligand in kidney organoids showed suppressed fibrosis-associated marker induction during IL-1β-induced fibrosis. These results illustrate how the platform integrates evidence for target prioritization and structure-based ligand assessment. To facilitate broader adoption, an online platform (https://cbcb.cdutcm.edu.cn/GPCRAgent/) was developed to support GPCR-related drug discovery.

By integrating structured knowledge retrieval with collaborative multi-agent reasoning, GPCRAgent provides an interpretable AI framework for evidence-based GPCR target prioritization, ligand recommendation, and decision support.