GOFlowLLM-curating miRNA literature with large language models and flowcharts

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Exponential growth in non-coding RNA research presents a knowledge management challenge.
• Manual curation of microRNA (miRNA) data for the Gene Ontology (GO) is a significant bottleneck.
• Automated systems are needed to accelerate biocuration while maintaining quality.

Purpose of the Study:

• To develop an automated curation pipeline using Large Language Models (LLMs) for miRNA-mediated gene silencing data.
• To improve the efficiency and scale of Gene Ontology (GO) annotation for miRNAs.
• To integrate transparent reasoning traces for human curator review.

Main Methods:

• Developed GOFlowLLM, an automated pipeline powered by reasoning-enabled LLMs.
• Pipeline follows established GO curation flowcharts for data extraction and structuring.
• Utilized the Qwen QwQ-32B model for processing uncurated articles.

Main Results:

• GOFlowLLM achieved 90% accuracy in selecting correct GO terms on existing data, with high curator agreement on reasoning (95%) and evidence (90%).
• Processed 6996 uncurated articles, identifying 2538 new candidate GO annotations in 58 hours.
• Manual review of AI-generated annotations showed high curator agreement (87% for terms, 92% for reasoning, 93% for evidence).

Conclusions:

• GOFlowLLM effectively automates miRNA GO curation, addressing the knowledge management crisis.
• The system demonstrates high accuracy and efficiency, potentially doubling available miRNA GO annotations.
• Reasoning traces enhance transparency and facilitate human curator review, promoting AI adoption in scientific curation.