HMFGraph: Novel Bayesian approach for recovering biological networks

Area of Science:

• Computational Biology
• Statistical Genetics
• Bioinformatics

Background:

• Gaussian graphical models (GGMs) are essential for analyzing partial correlation structures in high-dimensional omics data.
• Bayesian implementations of GGMs are gaining traction but face challenges in hyperparameter tuning, edge selection, scalability, and prior selection.

Purpose of the Study:

• To introduce a novel Bayesian GGM with a hierarchical matrix-F prior and a fast implementation.
• To address limitations in existing Bayesian GGM methods, including computational efficiency and network recovery.

Main Methods:

• Developed a novel Bayesian GGM utilizing a hierarchical matrix-F prior.
• Implemented a fast computational approach using a generalized expectation-maximization algorithm.
• Introduced a new shrinkage hyperparameter tuning method via precision matrix condition number constraints and edge selection using false discovery rate-controlled credible intervals.

Main Results:

• The proposed prior demonstrates competitive network recovery capabilities compared to state-of-the-art methods.
• The method shows good properties for recovering biologically meaningful networks.
• The generalized expectation-maximization algorithm offers significant computational advantages over Markov chain Monte Carlo methods.

Conclusions:

• The novel Bayesian GGM with a hierarchical matrix-F prior provides an efficient and effective tool for omics data analysis.
• The method enhances network recovery, offers improved computational performance, and facilitates community detection.
• The R package HMFGraph is available for practical application.