DeepGEOSearch: LLM-Powered Schemaless Retrieval for Biomedical Data Discovery

ABSTRACT Public biomedical repositories contain extensive, valuable datasets, yet identifying datasets that precisely match specific study requirements remains inefficient. Conventional keyword- and schema-based systems frequently fall short when queries encompass multiple biological and experimental facets. To address this, we developed DeepGEOSearch, an LLM-powered, schema-less retrieval system that interprets dataset text directly rather than relying solely on predefined metadata fields. The system extracts key study attributes, harmonizes terminology across sources, and ranks datasets by contextual alignment with natural-language queries while providing verifiable evidence for each match. In this work, we applied DeepGEOSearch for the Gene Expression Omnibus (GEO). The framework is repository-agnostic and can integrate new metadata sources without manual relabeling. This design enables complex compositional queries that existing tools cannot support. In evaluations against strong baselines using a curated evaluation benchmark comprising a mix of query complexities, DeepGEOSearch achieved more than 90% precision and best recall, with the largest performance gains observed on complex, real-world queries. DeepGEOSearch consistently identifies relevant datasets overlooked by conventional search tools, accelerating dataset discovery and improving reuse of public biomedical data. Competing Interest Statement The authors have declared no competing interest.