Context-aware LLM extraction and controlled-vocabulary normalization of GEO transcriptomic sample metadata

Motivation Public transcriptomic repositories contain millions of samples, but reuse is limited by inconsistent depositor-provided metadata. In the Gene Expression Omnibus (GEO), biologically critical fields such as tissue, disease state and treatment are frequently encoded within heterogeneous free text descriptions, undefined abbreviations/acronyms, or implicit experimental context. Results We developed a locally deployable large language model (LLM) pipeline converting GEO sample metadata into structured Tissue, Condition and Treatment labels. It performs sample-level extraction (Phase 1), study-level contextual inference of missing labels from GSE context and sibling-sample label distributions (Phase 1b), and controlled-vocabulary entity-linking, also called concept normalization, to Medical Subject Headings (MeSH) and Cellosaurus cell-line identifiers (Phase 2), with out-of-vocabulary clustering for concepts not within these vocabularies. We evaluated 804,427 human GEO samples against other automated annotations and manually curated benchmarks for Tissue, Condition and Treatment. On the manual benchmark, the extractor labelled Tissue and Condition at 99.67% and 95.0% accuracy. Entity linking reached accuracies of 0.985, 0.993 and 0.982 (F1 = 0.993, 0.996 and 0.991) on the labels for which a controlled-vocabulary answer exists, and separately assigned an identifier at all to 80.69%, 62.91% and 23.13% of extracted labels against 31.68%, 35.01% and 7.46% for deterministic exact MeSH matching. Availability All code, evaluation and results are provided as Supplementary Data. An interactive program built from the pipeline is available at https://github.com/SciSpectator/LLM-GEO-Label-Extractor. Contact Jonathan-Wren@OMRF.org.