Abstract
The semantic integration of biological data is hindered by the vast heterogeneity of data sources and their limited semantic formalization. A crucial step in this process is mapping data elements to ontological concepts, which typically involves substantial manual effort. Large Language Models (LLMs) have demonstrated potential in automating complex language-related tasks and may offer a solution to streamline biological data annotation. This study investigates the utility of LLMs-specifically various base and fine-tuned GPT models-for the automatic assignment of ontological identifiers to biological sample labels. We evaluated model performance in annotating labels to four widely used ontologies: the Cell Line Ontology (CLO), Cell Ontology (CL), Uber-anatomy Ontology (UBERON), and BRENDA Tissue Ontology (BTO). Our dataset was compiled from publicly available, high-quality databases containing biologically relevant sequence information, which suffers from inconsistent annotation practices, complicating integrative analyses. Model outputs were compared against annotations generated by text2term, a state-of-the-art annotation tool. The fine-tuned GPT model outperformed both the base models and text2term in annotating cell lines and cell types, particularly for the CL and UBERON ontologies, achieving a precision of 47-64% and a recall of 88-97%. In contrast, base models exhibited significantly lower performance. These results suggest that fine-tuned LLMs can accelerate and improve the accuracy of biological data annotation. Nonetheless, our evaluation highlights persistent challenges, including variable precision across ontology categories and the continued need for expert curation to ensure annotation validity.