Using semantic search to find publicly available gene-expression datasets

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Vast numbers of high-throughput molecular datasets are publicly available in repositories like Gene Expression Omnibus (GEO).
• Reusing these datasets is crucial for validating findings and exploring new research questions.
• Discovering relevant datasets is challenging due to sheer volume, inconsistent descriptions, and lack of semantic annotations, hindering FAIR data principles.

Purpose of the Study:

• To evaluate the effectiveness of language models in improving dataset discovery within the Gene Expression Omnibus (GEO).
• To assess if language model-generated embeddings can identify relevant datasets more efficiently than traditional search methods.

Main Methods:

• Utilized 30 language models to generate numerical representations (embeddings) of dataset descriptions from GEO.
• Focused on six human medical conditions, using datasets previously curated by humans.
• Compared the performance of language model-based similarity searches against GEO's built-in search engine.

Main Results:

• Language models, particularly those trained on general corpora using contrastive learning with large embeddings, often outperformed GEO's search engine in identifying relevant datasets.
• The effectiveness varied, indicating that this approach is promising but not universally superior.
• Identified specific model characteristics that correlate with better performance in dataset discovery.

Conclusions:

• Language models show significant potential to improve the discovery of scientific datasets, complementing existing search tools.
• This approach can aid researchers in efficiently finding and reusing valuable molecular data.
• Further development and integration of language models could streamline data discovery and enhance scientific reproducibility.