PlantGPT: An Arabidopsis-Based Intelligent Agent that Answers Questions about Plant Functional Genomics.

Area of Science:

• Genomics
• Bioinformatics
• Plant Science

Background:

• Advancing crop yields requires understanding plant gene function.
• Large language models (LLMs) can process vast research data but are prone to hallucinations.
• High-quality, reliable AI outputs are needed for scientific research.

Purpose of the Study:

• To develop a specialized LLM, PlantGPT, for plant functional genomics research.
• To minimize hallucinations in AI-generated scientific knowledge.
• To create an accessible tool for researchers investigating plant gene-phenotype relationships.

Main Methods:

• Compiled abstracts from over 60,000 plant research articles into a Chroma database for retrieval-augmented generation (RAG).
• Fine-tuned the Llama3-8B LLM using linguistic data from 13,993 Arabidopsis phenotypes and 23,323 gene functions.
• Developed an online tool (http://www.plantgpt.icu) for broader researcher access.

Main Results:

• PlantGPT demonstrated superior performance compared to general LLMs in answering specialized questions on Arabidopsis phenotype-gene research.
• The RAG approach effectively minimized hallucinations in the AI's output.
• The fine-tuned model provides accurate and reliable information.

Conclusions:

• PlantGPT serves as a virtual expert in Arabidopsis functional genomics.
• The study provides a blueprint for developing LLMs in crop functional genomics research.
• The developed online tool facilitates the adoption of AI in plant science.