A novel joint analysis framework improves identification of differentially expressed genes in cross disease transcriptomic analysis

Motivation: Detecting differentially expressed (DE) genes between disease and normal control group is one of the most common analyses in genome-wide transcriptomic data. Since most studies don't have a lot of samples, researchers have used meta-analysis to group different datasets for the same disease. Even then, in many cases the statistical power is still not enough. Taking into account the fact that many diseases share the same disease genes, it is desirable to design a statistical framework that can identify diseases' common and specific DE genes simultaneously to improve the identification power. Results: We developed a novel empirical Bayes based mixture model to identify DE genes in specific study by leveraging the shared information across multiple different disease expression data sets. The effectiveness of joint analysis was demonstrated through comprehensive simulation studies and two real data applications. The simulation results showed that our method consistently outperformed single data set analysis and two other meta-analysis methods in identification power. In real data analysis, overall our method demonstrated better identification power in detecting DE genes and prioritized more disease related genes and disease related pathways than single data set analysis. Over 150% more disease related genes are identified by our method in application to Huntington's disease. We expect that our method would provide researchers a new way of utilizing available data sets from different diseases when sample size of the focused disease is limited.