Enhanced semantic classification of microbiome sample origins using large language models (LLMs).

Area of Science:

• Bioinformatics
• Computational Biology
• Environmental Science

Background:

• Central sequence repositories have grown substantially, but heterogeneous metadata quality hinders data re-use.
• Well-annotated sequencing data is crucial for downstream studies, especially in microbiome research.
• Improving data FAIRness (Findability, Accessibility, Interoperability, Reusability) and AI readiness is essential for complex biological metadata.

Purpose of the Study:

• To evaluate the cost-effectiveness and performance of large language models (LLMs) for automating the re-annotation of environmental sequencing records.
• To assess the potential of LLMs to improve the FAIRness and AI readiness of microbiome sequencing metadata without fine-tuning.
• To compare the performance of proprietary LLMs with open-weight alternatives for ecological metadata re-annotation.

Main Methods:

• Utilized OpenAI Generative Pre-trained Transformer (GPT) models for automated re-annotation of sequencing records against a simplified ecological environment classification scheme.
• Developed and tested a pipeline for re-annotating environmental sequencing metadata, focusing on scalability, time, and cost-effectiveness.
• Compared LLM performance against a manually curated, non-machine learning (ML) keyword-based baseline and assessed proprietary vs. open-weight models (Qwen, meta-Llama, microsoft-phi-4).

Main Results:

• LLM-based annotation significantly outperformed the baseline keyword approach.
• Prompt design was found to be critical for task matching, while model choice had minor effects on performance.
• Open-weight LLMs demonstrated comparable accuracy to proprietary models for biome and sub-biome classification.
• The optimized pipeline was applied to 2 million environmental sequencing records, yielding standardized global sample origin annotations.

Conclusions:

• LLMs offer a cost-effective solution for automating the standardization and simplification of complex biological metadata from environmental sequencing records.
• The study validates the use of LLMs to enhance the FAIRness and AI readiness of large-scale sequencing datasets.
• Open-weight LLMs present a viable and accurate alternative for ecological metadata re-annotation, promoting broader accessibility and application.