Systematic tissue annotations of genomics samples by modeling unstructured metadata.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Over 1.3 million human -omics samples are publicly available, representing a vast but underutilized resource.
• Discovering specific samples is challenging due to unstructured, varied terminology in sample descriptions.
• Current methods struggle to effectively extract meaningful annotations from free-text metadata.

Purpose of the Study:

• To develop and validate a novel natural-language-processing-based machine learning (NLP-ML) approach for inferring tissue and cell-type annotations.
• To enhance the accessibility and usability of public human -omics sample data.
• To outperform existing methods for sample annotation based on free-text metadata.

Main Methods:

• Implemented a natural-language-processing-based machine learning (NLP-ML) pipeline.
• Converted unstructured sample descriptions into numerical representations for feature extraction.
• Utilized a supervised learning classifier to predict tissue and cell-type annotations.

Main Results:

• NLP-ML significantly outperformed advanced graph-based reasoning (MetaSRA) and exact string matching (TAGGER) methods.
• NLP-ML models captured biologically meaningful signals, showing similarities between related tissues.
• Models accurately classified tissue-associated biological processes and diseases based solely on text descriptions.
• NLP-ML achieved accuracy comparable to gene-expression profile-based models for tissue annotation, independent of experiment type.

Conclusions:

• NLP-ML effectively annotates genomics samples using only free-text metadata, significantly improving data discoverability.
• This approach unlocks the potential of the vast public -omics data repository.
• The NLP-ML method offers a powerful, versatile tool for biological data analysis and interpretation.