A Local LLM Normalized Metadata for 804 Thousand GEO Samples
A local AI system annotated 804 thousand samples in GEO, a public archive of gene activity studies
On August 31, the authors released a preprint describing a local pipeline that converts free-text sample descriptions into labels for tissue, condition, and treatment. The code, saved results, and validation scripts are available in the project repository.
A transcriptome is the set of RNA molecules in a sample. It indicates which genes were active in that sample. Comparing results across studies requires groups of samples from the same tissue, with the same condition, or after the same treatment. In GEO, researchers record this information as free text. A diagnosis, drug, or tissue may appear in a title, protocol, abbreviation, or group code.
The system first selects phrases from the record of a specific sample to support each label. When a group code or incomplete description requires interpretation, it reads the description of the study that groups the samples, known as a GEO Series (GSE). This context includes the title, summary, study design, and labels of related samples. The model instructions give an explicit characteristic of the individual sample priority over the broader study topic. A control sample in a disease study therefore retains the control label.
The pipeline then maps different names for the same tissue, condition, or treatment to a shared vocabulary: MeSH, a controlled vocabulary of medical terms, or Cellosaurus, a catalog of cell lines. It retains values absent from these vocabularies as separate concepts derived from the corpus itself. In this way, the system mapped 21 254 tissue name variants to 10 548 concepts. Each final label remained linked to the phrase in the original description.
Adding study context increased the proportion of samples with a specified condition from 64,01% to 80,43%. The authors manually reviewed annotations for 1 500 samples for each of the three fields: tissue, condition, and treatment. The F1 score, which combines extraction precision and recall, was 0,997 for tissue, 0,957 for condition, and 0,907 for treatment.
When sample information is lost, researchers cannot readily compare open datasets again. An essay on functional genomics explains how this problem can make datasets effectively single-use. The released corpus allows researchers to select comparable sample groups and trace each label back to the supporting phrase in the original description.