Accounting for network noise in graph-guided Bayesian modeling of structured high-dimensional data

Area of Science:

• Bioinformatics
• Statistical Learning
• Genomics and Proteomics Data Analysis

Background:

• Knowledge-guided statistical learning methods analyze structured high-dimensional data (e.g., genomic, transcriptomic).
• Existing methods use potentially incomplete or erroneous network data from databases or expert knowledge.
• This limits variable selection, prediction accuracy, and interpretability.

Purpose of the Study:

• To propose a graph-guided Bayesian modeling framework to address network noise in regression models.
• To integrate multiple sources of network information, including database graphs and data-driven estimates.
• To improve the analysis of structured high-dimensional predictors in biological data.

Main Methods:

• Developed a Bayesian framework incorporating a latent scale modeling approach to handle network noise.
• Combined external network information with network structures estimated from observed data.
• Employed an adaptive structured shrinkage prior within a Bayesian regression model.
• Utilized an efficient Markov chain Monte Carlo algorithm for posterior inference.

Main Results:

• The proposed method demonstrated advantages over existing approaches in simulation studies.
• Applied the framework to analyze genomics and proteomics datasets related to Alzheimer's disease.
• Showcased improved performance in variable selection and prediction accuracy.

Conclusions:

• The graph-guided Bayesian framework effectively accounts for network noise in high-dimensional data analysis.
• This approach offers a more robust and interpretable alternative for integrating biological network information.
• The method shows promise for applications in complex disease research, such as Alzheimer's disease.