GeMI

GeMI extracts structured metadata from plain-text Gene Expression Omnibus (GEO) experiment descriptions to enable standardized indexing and searchable metadata across functional genomics datasets.

Key Features:

• GEO free-text processing: Processes unstructured experiment descriptions from the Gene Expression Omnibus (GEO).
• Transformer-based NLP (fine-tuned GPT-2): Uses a fine-tuned Generative Pre-trained Transformer 2 (GPT-2) model for natural language processing of metadata text.
• Structured key-value extraction: Converts unstructured free-text into structured key-value pairs for metadata standardization.
• Indexing for search: Produces outputs that can be indexed to support efficient and precise search across datasets.
• Active learning with feedback: Incorporates an active learning framework to refine model predictions based on user feedback.
• Model interpretation (saliency maps): Provides saliency-map-based interpretation to highlight text regions driving model predictions.
• Implicit attribute inference: Infers attributes not explicitly mentioned in text, including sex, tissue type, cell type, ethnicity, and disease.

Scientific Applications:

• Metadata curation and augmentation: Automates extraction and harmonization of metadata to reduce redundancy, inconsistency, and incompleteness in GEO records.
• Attribute-based search and retrieval: Enables indexed, attribute-driven search across functional genomics datasets for targeted queries.
• Experiment classification and cohort selection: Supports classification of experiments and selection of sample cohorts using inferred attributes such as sex, tissue, cell type, ethnicity, and disease.
• Downstream data mining and integrative analysis: Facilitates large-scale data mining and integrative analyses by supplying structured metadata for functional genomics studies.

Methodology:

Applies a fine-tuned GPT-2 transformer-based NLP model to convert GEO free-text experiment descriptions into structured key-value pairs that can be indexed for search, employs an active learning framework for model refinement based on feedback, and uses saliency maps for model interpretation.