Bayesian analysis of high-dimensional gene expression using semiparametric quantile regression.

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• High-dimensional gene expression data often presents challenges like heterogeneity, nonlinearities, and distributional shifts.
• Traditional mean regression methods struggle to capture these complex patterns effectively.

Purpose of the Study:

• To develop a novel Bayesian Semiparametric Quantile Regression (BSQR) framework.
• To integrate Gaussian process priors for nonlinear modeling and sparsity-inducing priors for variable selection.
• To enhance the analysis of high-dimensional genomic data by addressing its inherent complexities.

Main Methods:

• The proposed BSQR framework utilizes Gaussian process priors for nonlinear modeling.
• Sparsity-inducing priors are employed for effective variable selection.
• Gene relevance is determined using posterior inclusion probabilities (PIPs) and effect size thresholds with false discovery control.

Main Results:

• Extensive simulations demonstrate BSQR's superior performance over penalized B-spline quantile regression (B-SP).
• BSQR shows enhanced precision, recall, probability calibration, and robustness under nonlinear and heteroskedastic conditions.
• The method achieved higher AUCs, lower Brier scores, and more stable gene discovery across quantiles.

Conclusions:

• BSQR offers a flexible and powerful approach for quantile-specific genomic analysis in high-dimensional settings.
• Application to a breast cancer dataset confirmed BSQR's advantages in discrimination, calibration, and selection reliability.
• The framework provides principled inference with false discovery control for genomic studies.