|
Abstract:
|
Naive approaches to prior specification for models of high-throughput RNA sequencing (RNA-seq) data often fail to adequately account for the strong correlations within each gene. As a computationally efficient alternative to allowing for an unstructured variance-covariance matrix for regression parameters, we use the spectral decomposition of the sample variance-covariance matrix of the maximum likelihood estimates to obtain an orthogonal reparameterization. This approach makes use of the high dimensionality of RNA-seq data to specify a nearly uncorrelated set of parameters, and allows for simplified prior specification in a Bayesian or empirical Bayesian approach. Using the R package ShrinkBayes, we demonstrate that, compared to a naive approach, the empirical reparameterization leads to more accurate estimation of hyperparameters and improved performance in tests for differential expression.
|
