Pre-Meta: priors-augmented retrieval for LLM-based metadata generation

Area of Science:

• Genomics
• Bioinformatics
• Data Science

Background:

• High-throughput sequencing generates vast genomic data, but manual annotation and metadata creation hinder discovery and publication.
• Large language models (LLMs) show promise for streamlining dataset profiling, yet struggle with specialized domains like biomedical genomics.
• Current limitations in LLM generalization impede efficient use of genomic data resources.

Purpose of the Study:

• To present Pre-Meta, an LLM-agnostic and domain-independent data annotation pipeline.
• To improve automated metadata generation accuracy by leveraging related priors like metadata tags and ontologies.
• To enhance the discovery and publication of genomic data resources.

Main Methods:

• Developed Pre-Meta, a data annotation pipeline.
• Implemented an enriched retrieval procedure using auxiliary information (metadata tags, ontologies).
• Validated the pipeline on five metadata fields across 1500 papers.

Main Results:

• Pre-Meta demonstrated systemic improvement in the annotation task without finetuning or prompt optimization.
• Achieved accuracy gains of 23% (GPT-4o mini), 72% (Llama 8B), and 75% (Mistral 7B) over conventional RAG.
• Validated the effectiveness of leveraging related priors for metadata generation.

Conclusions:

• Pre-Meta significantly enhances automated metadata generation for genomic datasets.
• The pipeline improves LLM performance in specialized domains by utilizing prior knowledge.
• Facilitates more efficient discovery and publication of genomic data resources.