Statistical inference on high-dimensional covariate-dependent Gaussian graphical regressions.

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• Gene co-expression graphs are crucial in genomic studies.
• Subject-level covariates, like single nucleotide polymorphisms (SNPs), influence these graphs.
• Traditional Gaussian graphical models (GGMs) overlook covariate effects, masking heterogeneity.

Purpose of the Study:

• To develop statistical inference methods for covariate-dependent Gaussian graphical models.
• To address the limitation of existing models that ignore subject-specific covariates.
• To enable accurate modeling of gene network structures that vary with covariates.

Main Methods:

• Proposed a multi-task learning approach for fitting covariate-dependent GGMs.
• Developed debiased estimators based on multi-task learners.
• Introduced a novel projection technique for inverse covariance matrix estimation, optimizing for sample size (n).

Main Results:

• The proposed multi-task learning approach yields lower error rates than node-wise regressions.
• Debiased estimators demonstrate fast convergence and asymptotic normality, facilitating valid statistical inference.
• Simulations confirmed the method's utility and an application to brain cancer data revealed significant biological insights.

Conclusions:

• The novel debiased estimators provide a computationally efficient and statistically valid framework for inference in covariate-dependent GGMs.
• This method enhances the understanding of gene co-expression networks by incorporating individual genetic variations.
• The approach has practical implications for analyzing complex biological data, such as gene expression in cancer.