Covariate-Assisted Bayesian Graph Learning for Heterogeneous Data.

Area of Science:

• Computational Biology and Bioinformatics
• Statistical Genetics
• Network Analysis

Background:

• Traditional Gaussian graphical models assume data homogeneity, often under-utilizing auxiliary information in genomic datasets.
• Genomic data frequently contains rich covariate information that can refine the understanding of joint dependency structures.
• Existing methods struggle to integrate heterogeneous observations with covariate-specific network structures.

Purpose of the Study:

• To develop a novel Bayesian covariate-dependent Gaussian graphical model for heterogeneous multivariate observations.
• To leverage auxiliary information (covariates) to allow undirected graphs to vary, improving dependency structure analysis.
• To enhance the modeling of biological networks, such as protein-protein interactions, using gene expression data.

Main Methods:

• Proposed a Bayesian product partition model framework for covariate-dependent Gaussian graphical models.
• Explored Gaussian likelihood with G-Wishart prior and pseudo-likelihood with product of Gaussian conditionals for model embedding.
• Utilized fractional likelihood theory to establish minimax optimal posterior contraction rates for Gaussian mixture densities.

Main Results:

• The proposed model flexibly approximates various conditional variance-covariance matrices.
• Demonstrated minimax optimal posterior contraction rates under specific density assumptions.
• Simulation studies confirmed the model's efficacy in capturing covariate-influenced network structures.

Conclusions:

• The covariate-dependent Gaussian graphical model effectively integrates auxiliary information for improved network inference in heterogeneous genomic data.
• The Bayesian approach provides a flexible and theoretically sound framework for analyzing complex biological networks.
• Application to breast cancer protein network analysis using mRNA gene expression highlights practical utility.