A Global-Local Approach for Detecting Hotspots in Multiple-Response Regression.

Area of Science:

• Genetics
• Statistical Modeling
• Bioinformatics

Background:

• Variable selection in regression with many predictors and responses is challenging.
• Identifying 'hotspots' (predictors linked to multiple responses) is critical in statistical genetics.
• Existing methods for hotspot detection lack scalability and are sensitive to parameter choices.

Purpose of the Study:

• To develop a flexible, scalable hierarchical regression framework for hotspot detection.
• To address limitations of existing methods in large-scale genetic applications.
• To accurately identify genetic variants that control multiple gene expressions.

Main Methods:

• Implemented a fully Bayesian hierarchical regression model.
• Utilized the horseshoe shrinkage prior for its global-local formulation.
• Employed a fast variational algorithm with simulated annealing for efficient inference.

Main Results:

• The proposed framework effectively detects hotspots in large datasets.
• The horseshoe prior successfully shrinks noise globally while preserving hotspot signals.
• The method demonstrates robustness and scalability for high-dimensional genetic data.

Conclusions:

• The novel Bayesian framework provides a scalable and robust solution for hotspot detection in genetic studies.
• This approach enhances the understanding of genomic architecture and disease mechanisms.
• Efficient inference algorithms enable application to large-scale genetic datasets.