SeedLLM·Rice: A large language model integrated with rice biological knowledge graph.

Area of Science:

• Agricultural Science
• Bioinformatics
• Computational Biology

Background:

• Rice biology research faces challenges due to vast literature and multiomics data.
• Existing large language models (LLMs) lack specialization for rice and struggle with multimodal data synthesis.
• Standardized evaluation frameworks for domain-specific LLM tasks in rice biology are needed.

Purpose of the Study:

• To develop a specialized LLM for rice biology research.
• To create a novel human-centric evaluation framework for LLM performance in this domain.
• To integrate the LLM with a comprehensive biological knowledge graph for enhanced data fusion.

Main Methods:

• Trained a 7-billion-parameter model (SeedLLM) on 1.4 million rice-related publications.
• Developed a human-centric evaluation framework for rice biology LLM tasks.
• Integrated SeedLLM with the Rice Biological Knowledge Graph (RBKG), incorporating genome annotations and multiomics data from over 1800 studies.

Main Results:

• SeedLLM demonstrated superior performance on rice-specific tasks, achieving win rates of 57%–88% against general-purpose models like GPT-4o and DeepSeek-R1.
• The integration with RBKG enabled SeedLLM to address complex research questions requiring fused textual and multiomics data.
• Free web-based access to SeedLLM and RBKG was provided to facilitate global collaboration.

Conclusions:

• SeedLLM is a transformative tool for rice biology, enhancing knowledge retrieval and data integration.
• The model facilitates discoveries in crop improvement and climate adaptation through advanced reasoning.
• The developed framework and accessible resources support the advancement of global rice research.