PGCopilot: an AI-assisted research agent for plant functional genomics

The exponential expansion of multi-omic data presents a major bottleneck for researchers seeking to derive biological insights from heterogeneous genomic resources. Although large language models (LLMs) enable natural language interaction with scientific knowledge, general-purpose LLMs are not designed to reason over plant-specific biological issues and frequently produce hallucinated or oversimplified outputs. We developed PGCopilot, an AI-assisted research agent tailored for plant functional genomics. It integrated task planning, identifier disambiguation, and tool orchestration to actively retrieve and synthesize evidence across heterogeneous biological databases using the Model Context Protocol. Built on a knowledge base of 1.83 million records, together with online access to PubMed titles and available abstracts covering over 33 million indexed articles and STRING protein-protein interaction data, it could plan retrieval strategies and integrate evidence into grounded responses, reducing unsupported outputs and addressing data fragmentation and identifier discrepancies. In expert-guided benchmarking, PGCopilot achieved higher performance than the evaluated baseline LLMs, demonstrating advantages in task relevance, evidence integration, hallucination control, and research utility, and providing more verifiable citations and explicit reporting of cross-database inconsistencies. As an autonomous agent supporting multiturn tool interaction, it streamlined cross-database evidence integration from labor-intensive search-and-retrieval to automated analysis-and-synthesis. This approach streamlines cross-database navigation and evidence integration, allowing researchers to focus more on hypothesis formulation, experimental design, and critical interpretation of comprehensive evidence.