EBSeq: improving mixing computations for multi-group differential expression analysis

EBSeq is a Bioconductor package designed to calculate empirical-Bayesian inference summaries from sequence-based gene-expression (RNA-Seq) data. It produces gene or isoform-specific scores that measure various patterns of differential expression among a set of sample groups, and is most commonly deployed to measure differential expression between two groups. Its use of local posterior probabilities from a fitted mixture model provides the data analyst a direct way to score the false discovery rate of any reported list of genes, and it is one of the only tools that can address local false discovery rates when analyzing multiple sample groups. Contemporary applications have increasing numbers of sample groups, and the algorithms deployed in EBSeq are neither space nor time efficient in this important case. We describe a version update utilizing code improvements and novel pruning and clustering algorithms in order to reduce the complexity of mixture computations. The algorithms are supported by a theoretical analysis and tested empirically on a variety of benchmark and synthetic data sets.