DBCLS BioHackathon 2026 report: An interactive, harmonised BioSample metadata resource for biological sample discovery

Abstract

The BioSample database contains tens of millions of records describing biological samples, but their metadata are submitted largely as key--value pairs with inconsistent field names, terminology and free-text descriptions, limiting search and reuse. Before the DBCLS BioHackathon 2026, we used a pipeline based on locally hosted large language models (LLMs) to extract nine types of biological information (cell line, cell type, tissue, disease, drug, knockout gene, knockdown gene, overexpressed gene and ChIP antigen) from 4.2 million human and mouse BioSample records and map them to ontology terms, producing a harmonised collection of 8.2 million extracted values. During the BioHackathon, we extended this work in three directions. First, we developed an alpha version of a web application for searching this collection by ontology terms and visualising the experimental conditions of selected samples. Second, we developed a provenance-tracing procedure, combining deterministic matching with LLM-based checks, that linked 99% of extracted values to supporting text in the original records. The resulting provenance links showed that although 32 submitted field names accounted for 90% of the evidence, the remaining 10% was spread across more than 6,000 other field names. Third, we developed a procedure in which an LLM assesses mapping correctness and classifies the causes of errors, offering a scalable alternative to building new manually curated gold standards. Our preliminary results show the potential of locally hosted LLMs to harmonise existing BioSample metadata at scale. Future work will validate the LLM-based evaluations with human experts and extend the approach to additional species.