Optimizing large language models for ontology-based annotation: a study on gene ontology in biomedical texts

Description

Pratik Devkota, Somya Mohanty, and Prashanti Manda present a 2025 study on using large language models to tag biomedical papers with Gene Ontology terms. Accurate tagging powers search and knowledge linking in biology, but older neural models struggle with complex terms. The team fine-tuned four language models on a curated corpus and compared them with a recurrent network baseline on accuracy, meaning-level agreement, memory use, and speed.

The older baseline stayed competitive on raw accuracy. The language models sometimes produced tags that matched the intended meaning better, especially for long multi-word terms, though the edge was not uniform. They also demanded far more computing resources. Tricks like training only small adapter weights cut the cost while keeping performance. The authors conclude deployment is a trade: richer tagging against heavier machinery. They call for more domain training and optimization before routine use.

Catalog record

Source: https://europepmc.org/articles/PMC12781589