Model selection for factorial Gaussian graphical models with an application to dynamic regulatory networks

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Factorial Gaussian graphical models (fGGMs) are emerging tools for dynamic gene regulatory network inference from high-throughput genomic data.
• These models incorporate restrictions like low-order Markov dependencies and equal time-lag dependency strengths to simplify network structures.
• Estimating the precision matrix using l1-penalized maximum likelihood leads to sparse networks, but optimal sparsity selection remains a challenge.

Purpose of the Study:

• To evaluate various model selection criteria for fGGMs in the context of dynamic gene regulatory network inference.
• To assess the performance of fGGMs against alternative methods for inferring dynamic networks.
• To demonstrate the utility of fGGMs and the KLCV criterion using simulated and real biological data.

Main Methods:

• Utilized two simulated regulatory networks based on realistic biological processes to test model selection criteria.
• Employed l1-penalized maximum likelihood estimation for precision matrix estimation in fGGMs.
• Applied the KLCV (Kullback-Leibler Cross-Validation) criterion for model selection.

Main Results:

• fGGMs showed good performance in inferring dynamic networks compared to other methods.
• The KLCV criterion proved particularly effective for selecting the optimal sparsity level in fGGMs.
• The methodology was successfully applied to high-resolution time-course microarray data from Neisseria meningitidis.

Conclusions:

• fGGMs are a valuable approach for inferring dynamic gene regulatory networks.
• The KLCV criterion is a reliable method for model selection in fGGMs.
• The R package sglasso provides a practical implementation of the described methodology.