Application of fused graphical lasso to statistical inference for multiple sparse precision matrices

Area of Science:

• Statistics
• Computational Biology
• Bioinformatics

Background:

• Estimating precision matrices is crucial for understanding complex biological systems.
• Existing methods often struggle with high-dimensional data and multiple populations.

Purpose of the Study:

• To introduce and validate the fused graphical lasso (FGL) method for simultaneous estimation of multiple precision matrices.
• To develop robust statistical inference and hypothesis testing for high-dimensional, multi-population data.

Main Methods:

• Utilizing the lasso penalty for sparsity and a moderate penalty for structural similarity across groups.
• Developing a de-biasing technique for consistent estimation and asymptotic theory.
• Providing an oracle inequality for FGL estimators in high-dimensional settings.

Main Results:

• The FGL method effectively estimates multiple precision matrices with controlled sparsity and similarity.
• A novel de-biased FGL estimator with known distribution was developed for statistical inference.
• The proposed hypothesis testing method demonstrates strong performance in high-dimensional simulations and real-world data.

Conclusions:

• The FGL method offers a powerful tool for analyzing complex, high-dimensional data across multiple populations.
• The developed statistical inference framework enhances the reliability of findings in bioinformatics and related fields.
• The approach is validated by simulation studies and application to diffuse large B-cell lymphoma data.