An empirical Bayes approach to network recovery using external knowledge

Area of Science:

• Computational Biology
• Network Science
• Bioinformatics

Background:

• High-dimensional network reconstruction, particularly for gene interaction networks, benefits from incorporating prior topological knowledge.
• Prior knowledge can be sourced from pathway databases (e.g., KEGG) or pilot studies.
• Bayesian frameworks offer a natural approach for integrating prior information into network analysis.

Purpose of the Study:

• To develop an Empirical Bayes (EB) procedure for network reconstruction that assesses the agreement between prior knowledge and observed data.
• To approximate posterior densities using variational Bayes and compare its accuracy against Gibbs sampling.
• To demonstrate the utility of the EB method in reconstructing gene expression networks.

Main Methods:

• Development of an Empirical Bayes (EB) procedure based on a Bayesian Simultaneous Equation Model.
• Utilizing variational Bayes for approximating posterior densities.
• Comparison of EB method's accuracy and computational speed against existing network reconstruction techniques.

Main Results:

• The EB method automatically assesses the concordance of prior knowledge with the data.
• Variational Bayes approximation is compared for accuracy against Gibbs sampling.
• Accurate prior data significantly enhances network reconstruction, while inaccurate priors do not substantially harm the results.
• The proposed method is computationally efficient and outperforms competitors.

Conclusions:

• The developed Empirical Bayes method effectively integrates prior knowledge for high-dimensional network reconstruction.
• The method demonstrates superior reproducibility of reconstructed network edges compared to competing methods, as shown in gene expression data analysis.
• Accurate prior information is crucial for optimal network reconstruction, but the method remains robust to inaccurate priors.