The 'un-shrunk' partial correlation in Gaussian graphical models.

Area of Science:

• Systems Biology
• Bioinformatics
• Statistical Modeling

Background:

• Reconstructing regulatory networks from molecular profiles is crucial in systems biology.
• Gaussian graphical models (GGMs) are widely used for network inference, representing relationships via partial correlations.
• The high-dimensional problem (fewer samples than nodes) necessitates shrinkage methods for GGM regularization.

Purpose of the Study:

• To investigate the impact of shrinkage on partial correlations in GGMs.
• To address the non-linear bias introduced by shrinkage in network reconstruction.
• To develop a method for correcting shrinkage bias to improve network interpretability.

Main Methods:

• Analysis of non-linear bias introduced by shrinkage in partial correlation estimation.
• Development and application of an 'un-shrinking' method to correct for this bias.
• Validation using gene expression datasets from Escherichia coli and Mus musculus.

Main Results:

• Shrinkage non-linearly biases partial correlations, altering their magnitudes and order.
• This bias hinders network comparability and biological interpretation across experiments.
• The proposed 'un-shrinking' method yields more accurate and interpretable partial correlations.

Conclusions:

• Shrinkage in GGMs, while addressing high-dimensionality, introduces significant non-linear bias.
• Ignoring this bias can lead to misinterpretation of regulatory networks and impede result validation.
• The 'un-shrinking' approach offers a more reliable method for GGM-based network reconstruction.