GEO-BERT: A Retrieval augmented LLM Framework for Intelligent Dataset Retrieval from the Gene Expression Omnibus

The Gene Expression Omnibus (GEO) serves as a widely utilized data repository, housing high throughput gene expression and other functional genomics datasets. Researchers rely on GEO for a variety of purposes, including submitting experimental datasets, accessing them for hypothesis validation and reuse, and generating ideas for new research avenues.

Despite its extensive usage, the process of searching and identifying pertinent datasets within GEO poses challenges, primarily due to the platform's reliance on a keyword-based search system, which necessitates manual intervention for the retrieval of desired datasets. GEO's inconsistent metadata and varying data annotation practices across submissions, coupled with variable data quality, further complicate the development of structured query tools.

In light of recent advancements in Large Language Models (LLMs), there is a timely opportunity to reassess the GEO datasets and their query system. This article introduces GEO-BERT, an AI chat-bot designed to leverage GEO's metadata through a retrieval augmented generator framework utilizing a novel hierarchical hybrid field-aware retrieval system.

GEO-BERT is crafted to be locally installable, catering to the needs of programmers and project developers, and is readily accessible for regular users through a subset of test data. Demonstrations indicate that GEO-BERT significantly enhances semantic search performance, a feature currently absent in leading AI-powered chatbots.

Furthermore, GEO-BERT improves GEO's traditional search performance, benefiting project development efforts. Considering GEO's global usage and the latest strides made in large language models, GEO-BERT addresses a critical gap by facilitating context-aware extraction of GEO datasets.