Using semantic search to find publicly available gene-expression datasets | JoVE Visualize

Area of Science:

Bioinformatics
Computational Biology
Data Science

Background:

Millions of high-throughput molecular datasets are publicly available, but finding relevant ones is challenging due to description inconsistencies and scale.
Discovering relevant datasets is crucial for data reuse and validation, yet current methods in repositories like Gene Expression Omnibus (GEO) are time-consuming and may miss important data.
This challenge directly impacts the FAIR data principles, emphasizing the need for improved dataset discovery mechanisms.

Purpose of the Study:

To evaluate the effectiveness of language models in enhancing dataset discovery within the Gene Expression Omnibus (GEO) repository.
To determine if language model-generated embeddings can identify relevant datasets more efficiently than existing search tools.
To explore the potential of natural language processing techniques for improving access to large-scale molecular data.

Main Methods:

Utilized 30 different language models to generate numerical representations (embeddings) of dataset descriptions from the Gene Expression Omnibus (GEO).
Focused on six human medical conditions, using datasets curated by humans as a reference.
Compared the performance of language model-based similarity searches against GEO's native search engine to find related datasets.

Main Results:

Language models, particularly those trained on general corpora using contrastive learning with large embeddings, often improved dataset discovery compared to GEO's search.
The approach successfully identified additional relevant datasets for specific medical conditions.
While effective, the language model approach was not universally superior to existing search methods in all tested scenarios.

Conclusions:

Language models show significant potential to enhance the discovery of scientific datasets by providing semantic summaries (embeddings).
This methodology can complement existing search tools, offering a more efficient way to find relevant molecular data.
Further development and integration of language models could greatly improve data accessibility and reuse in scientific research.