Variational Bayes for High-Dimensional Multi-Source Heterogeneous Data With Sparse Priors

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• High-dimensional data is common in genomics, economics, and medicine.
• Existing methods struggle with joint analysis of multi-source, heterogeneous datasets.
• A gap exists in modeling shared features while accounting for subpopulation heterogeneity.

Purpose of the Study:

• To develop a Bayesian method for estimating parameters in high-dimensional multi-source heterogeneous linear data.
• To extract shared features across subpopulations and identify unique heterogeneity within each.
• To provide a computationally efficient and scalable solution for complex data integration.

Main Methods:

• A Bayesian model using a sparsity-inducing spike-and-slab prior (Laplace slab, Dirac spike).
• Mean-field variational approximation for efficient posterior computation, overcoming Gibbs sampling limitations.
• Variable selection through posterior inclusion probabilities.

Main Results:

• The variational Bayesian approach demonstrates effectiveness on simulated data and TCGA cancer datasets.
• Achieved superior computational efficiency and scalability compared to Gibbs sampling and penalized frequentist methods.
• Successfully identified shared and unique features in multi-source heterogeneous data.

Conclusions:

• The proposed variational Bayesian method offers an effective and efficient solution for analyzing high-dimensional multi-source heterogeneous data.
• The VBMS R package provides a publicly available tool for implementing this approach.
• This method facilitates integrative analysis in fields like genomics and medicine.