Bayesian Network Marker Selection via the Thresholded Graph Laplacian Gaussian Prior.

Area of Science:

• Computational Biology
• Statistical Genetics
• Network Analysis

Background:

• Selecting informative nodes in large networks is crucial for many research areas.
• Existing methods often focus on local structures and struggle with computational costs on large datasets.
• Bayesian network marker selection within the generalized linear model (GLM) framework requires efficient and scalable approaches.

Purpose of the Study:

• To propose a novel prior model, the Thresholded Graph Laplacian Gaussian (TGLG) prior, for Bayesian network marker selection.
• To account for global network structure in marker selection using the graph Laplacian matrix.
• To develop a computationally efficient algorithm for posterior computation in large-scale networks.

Main Methods:

• The proposed Thresholded Graph Laplacian Gaussian (TGLG) prior utilizes the graph Laplacian matrix to model conditional dependence between neighboring markers.
• The method is integrated within the generalized linear model (GLM) framework.
• A Metropolis-adjusted Langevin algorithm (MALA) is developed for efficient posterior computation.

Main Results:

• The TGLG prior demonstrates posterior consistency under mild conditions, even with a growing number of network nodes and edges.
• The MALA algorithm provides scalability for analyzing large-scale networks.
• Simulations and analysis of breast cancer gene expression data show superior performance compared to existing methods.

Conclusions:

• The TGLG prior offers an effective and scalable solution for Bayesian network marker selection in large-scale networks.
• The method successfully incorporates global network structure, leading to improved performance.
• This approach has significant implications for biological data analysis, such as in the Cancer Genome Atlas (TCGA).