Regularized estimation of large-scale gene association networks using graphical Gaussian models.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Graphical Gaussian models are crucial for estimating gene association networks from microarray data.
• Estimating partial correlation matrices is challenging with more variables than samples.
• Standard methods fail, necessitating regularization techniques like Lasso and Partial Least Squares.

Purpose of the Study:

• To investigate a general framework for combining regularized regression methods with Graphical Gaussian model estimation.
• To introduce and evaluate new methods based on ridge regression and adaptive Lasso.
• To compare existing and novel methods using simulations and real data.

Main Methods:

• Framework combining regularized regression with Graphical Gaussian models.
• Implementation of ridge regression and adaptive Lasso approaches.
• Extensive comparison via simulation studies and analysis of six real gene expression datasets.

Main Results:

• Non-sparse methods (Ridge, PLS) are conservative with FDR control for network edge detection.
• Lasso tends to select too many edges in sparse networks; adaptive Lasso provides sparser, more accurate solutions.
• Partial Least Squares often selects very dense networks; Lasso/adaptive Lasso struggle with correlated observations.

Conclusions:

• Adaptive Lasso is a promising alternative for sparse network reconstruction.
• Method performance varies with network density and data characteristics.
• The R package 'parcor' is available for implementing these algorithms.