Context Specific and Differential Gene Co-expression Networks via Bayesian Biclustering.

Area of Science:

• Genomics
• Bioinformatics
• Statistical Modeling

Background:

• Identifying latent structures in high-dimensional genomic data is crucial for understanding biological processes.
• Gene co-expression networks reveal genes regulated by shared biological mechanisms.

Purpose of the Study:

• To develop a Bayesian statistical model for biclustering to infer co-regulated gene subsets.
• To recover context-specific gene co-expression networks from genomic data.

Main Methods:

• Developed BicMix, a Bayesian biclustering model allowing overcomplete representations and joint modeling of confounders and biological signals.
• Implemented a method to recover context-specific gene co-expression networks from sparse biclustering matrices.

Main Results:

• BicMix demonstrated higher precision in recovering latent structure compared to state-of-the-art methods across simulations.
• Applied BicMix to breast cancer and cardiovascular data, revealing differential co-expression networks across sample subtypes (e.g., ER+ vs. ER-, male vs. female).
• Identified tissue-specific gene networks in the Genotype-Tissue Expression (GTEx) pilot data, validated by identifying tissue-specific trans-eQTLs.

Conclusions:

• BicMix is an effective tool for inferring gene co-expression networks and uncovering biological insights from genomic data.
• The method successfully identified clinically relevant and tissue-specific gene expression patterns, highlighting its potential for biomarker discovery and biological pathway elucidation.