Annotation of biological samples data to standard ontologies with support from large language models

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Biological data integration is challenging due to diverse sources and poor semantic formalization.
• Mapping data to ontologies requires significant manual effort, hindering large-scale analysis.
• Large Language Models (LLMs) offer potential for automating complex language tasks like biological data annotation.

Purpose of the Study:

• To evaluate the effectiveness of various Large Language Models (LLMs), including base and fine-tuned GPT models, for automatically assigning ontological identifiers to biological sample labels.
• To compare LLM performance against a state-of-the-art tool (text2term) using established biological ontologies.

Main Methods:

• LLMs (base and fine-tuned GPT models) were used to annotate biological sample labels.
• Annotations were mapped to four ontologies: Cell Line Ontology (CLO), Cell Ontology (CL), Uber-anatomy Ontology (UBERON), and BRENDA Tissue Ontology (BTO).
• Model outputs were compared against annotations from the text2term tool using a dataset from public biological databases.

Main Results:

• A fine-tuned GPT model significantly outperformed base models and text2term in annotating cell lines and cell types.
• The fine-tuned model achieved high recall (88-97%) and moderate precision (47-64%) for CL and UBERON ontologies.
• Base LLMs demonstrated considerably lower performance in ontological annotation tasks.

Conclusions:

• Fine-tuned LLMs can accelerate and enhance the accuracy of biological data annotation, aiding semantic integration.
• Challenges remain, including variable precision across different ontology categories.
• Expert curation is still essential to validate LLM-generated annotations and ensure data integrity.