FGTD: a desktop application for GEO dataset discovery with sample-level metadata extraction and local semantic relevance scoring

The Gene Expression Omnibus (GEO) contains a large amount of reusable functional genomics data, but finding the series relevant to a biological question still requires the manual screening of heterogeneous metadata and linked publications. We present FGTD (From GEO To Dataset), a desktop application that automates the retrieval, filtering and ranking of GEO series and associated publications. FGTD organizes the extracted metadata, down to the sample level, in one table, reconstructing links to the Sequence Read Archive and BioProject, and ranks the series with a score that combines semantic similarity (computed locally with sentence embeddings), keyword and Medical Subject Headings matches, and citation and recency metrics. In a breast cancer case study FGTD produced the metadata table of 15 series in about 38 seconds, while manual curation (timed on two series and extrapolated to the cohort) was estimated at about 3.8 hours. In known-item evaluations on two published meta-analyses, FGTD ranked the reference series closer to the top than the date-ordered GEO listing (mean reciprocal rank 0.571 versus 0.020), although the gain depended on the keywords supplied, and its semantic component alone and its fusion with a lexical baseline performed at least as well. FGTD is freely available at https://github.com/SorenKK/FGTD.