GRACKLE: an interpretable matrix factorization approach for biomedical representation learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression disruptions are linked to diseases.
• Identifying disease-specific gene signatures is challenging due to co-occurring conditions and small sample sizes.
• Existing unsupervised learning methods lack clear biological explanations and do not integrate prior biological knowledge with sample labels.

Purpose of the Study:

• To develop a novel method for identifying disease-specific gene signatures by integrating prior biological knowledge.
• To improve the interpretability and accuracy of unsupervised learning in high-dimensional biological data.
• To address the limitations of current models in jointly considering molecular interactions and sample labels.

Main Methods:

• Introduced GRACKLE (Graph Regularization Across Contextual KnowLedgE), a nonnegative matrix factorization approach.
• Integrated sample similarity and gene similarity matrices using sample metadata and molecular relationships.
• Validated GRACKLE through simulation studies and application to breast tumor and Down syndrome datasets.

Main Results:

• GRACKLE outperformed other nonnegative matrix factorization algorithms, particularly under high background noise.
• Successfully stratified breast tumor samples and identified condition-enriched subgroups in individuals with Down syndrome.
• Latent representations generated by GRACKLE aligned with known biological patterns, including autoimmune conditions and sleep apnea in Down syndrome.

Conclusions:

• GRACKLE provides a robust solution for identifying context-specific molecular mechanisms in biomedical research.
• The model's flexibility allows application across various data modalities.
• GRACKLE enhances the understanding of gene expression in complex diseases and small sample settings.