Pathway-based analysis of multivariate traits using classical dimensionality reduction methods

2015-02-02Dimensionality reduction methods (DRMs) can capture important information of the data with fewer components, and testing those components can improve power in association testing. My work compared three classical DRMs: principal components analysis (PCA), partial least squares (PLS), and canonical correlations analysis (CCA), and demonstrated their power advantage for testing genetic associations between multiple genes and multiple quantitative traits. I found that the PCA-based approach is not as powerful as generally recognized. When the covariance of genetic markers and traits are not (or less) related to the correlations between them, PCA-based approaches showed relatively poor power in the simulations. Theoretical insights and the supportive evidence from the simulation results indicated testing CCA components has the best performance in joint association tests among all three DRMs. CCA-based approach and multivariate multiple regression both showed good power in the scenarios of correlated genetic markers and correlated traits, and the power performance is sensitive to the magnitude and direction of the correlations. Based on the findings, I applied an adaptation method motivated by the Adaptive Rank Truncated Product (ARTP) for addressing the issue of selecting CCA components in order to achieve optimal power. The same settings of simulations were followed, and I concluded that adaptive CCA-based approach (ACCA) is a powerful association testing approach for pathway-based analysis