A modeling framework for detecting and leveraging node-level information in Bayesian network inference

Area of Science:

• Computational Biology
• Statistical Genetics
• Network Science

Background:

• Bayesian graphical models are powerful for high-dimensional data but face computational and statistical challenges.
• Leveraging auxiliary information, like genetic variant data, can enhance the inference of dependence structures.

Purpose of the Study:

• To develop a novel Gaussian graphical modeling framework that integrates node-level information to improve network inference.
• To simultaneously infer sparse precision matrices and the relevance of auxiliary variables for uncovering network structures.

Main Methods:

• A fully joint hierarchical model incorporating a spike-and-slab submodel for hub propensity.
• Development of a variational expectation-conditional maximization algorithm for scalable inference.
• Application to simulations and a gene network study.

Main Results:

• The framework effectively identifies and leverages node centrality information for network structure detection.
• The developed algorithm scales inference to hundreds of samples, nodes, and auxiliary variables.
• Identification of hub genes in biological pathways relevant to immune-mediated diseases.

Conclusions:

• The proposed Bayesian graphical modeling framework offers a computationally efficient and statistically robust approach to gene network inference.
• Integrating auxiliary information significantly improves the detection of complex relationships and identification of key biological drivers.
• The method has practical implications for understanding the genetic architecture of immune-mediated diseases.