Sparse Estimation of Conditional Graphical Models With Application to Gene Networks.

Area of Science:

• Statistics
• Network Analysis
• Bioinformatics

Background:

• Network structures often result from both inherent connections and external influences.
• Distinguishing these sources is vital for accurate network analysis.

Purpose of the Study:

• To develop a sparse estimation procedure for graphical models that isolates intrinsic connections by removing external effects.
• To formulate this as a conditional graphical model where external effects act as predictors.

Main Methods:

• Introduced two sparse estimators for the conditional precision matrix using reproduced kernel Hilbert space with lasso and adaptive lasso.
• Established theoretical properties including sparsity, variable selection consistency, oracle property, and asymptotic distributions.
• Developed convergence rates for high-dimensional conditional precision matrices.

Main Results:

• The proposed estimators demonstrate strong theoretical guarantees for sparsity and consistency.
• Performance was compared against unconditional graphical model estimators and constrained maximum likelihood estimates.
• The methods were successfully applied to genetic data for gene network construction.

Conclusions:

• The conditional graphical model approach effectively isolates intrinsic network connections.
• The developed sparse estimators provide a robust tool for network analysis in the presence of external effects.
• This method has practical applications in fields like genetic network inference.