Learning Latent Variable Gaussian Graphical Model for Biomolecular Network with Low Sample Complexity.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Learning Gaussian graphical models with latent variables is challenging due to ill-posedness with insufficient sample complexity.
• Common convex regularizations (ℓ1 plus nuclear norm) may not yield optimal performance, particularly in low sample settings.

Purpose of the Study:

• To develop a novel concave regularization approach for Gaussian graphical models with latent variables.
• To address intrinsic estimation biases and improve performance in low sample complexity scenarios.
• To extend the method for fused structure-sparsity and low-rank decomposition, applicable to temporal data.

Main Methods:

• Introduced concave additive regularization to induce sparsity and low rankness, correcting estimation biases.
• Established proximity operators for the proposed concave regularizations.
• Developed a modified alternating direction method of multipliers (ADMM) with local convergence guarantees.
• Extended the framework to incorporate fused structure-sparsity for temporal network analysis.

Main Results:

• The proposed concave regularization outperforms traditional convex methods, especially in low sample complexity.
• The method effectively induces sparsity and low rankness in the estimated graphical models.
• The extended method successfully decomposes fused structure-sparsity and low rankness.
• Application to cancer network reconstruction revealed the 'Warburg effect'.

Conclusions:

• Concave regularization offers a powerful alternative for learning Gaussian graphical models with latent variables, particularly when sample size is limited.
• The method provides improved estimation accuracy and bias correction.
• The framework is versatile, handling complex structures like fused sparsity and low rankness, with demonstrated utility in biological network analysis.