A0550
Title: Sequential canonical variate regression for unified correlation analysis and outcome prediction of multi-omics data
Authors: Chongliang Luo - Washington University in St Louis (United States) [presenting]
Abstract: When integrating multiple sets of data in biomedical research, the objectives are often two-fold: to learn the correlation structure between the sets and to optimize the prediction of clinical outcomes. The correlation structure can be achieved by canonical correlation analysis (CCA), and the extracted canonical variates (CVs) can be used to predict the outcome, hence the canonical variate regression (CVR). On the other hand, besides the shared joint structure across sets, set-specific structures for individual sets may also exist, which are predictive. This joint and individual structure motivates us to develop a sequential CVR (CVR-seq) method, which can further improve the prediction. The CVR-seq adaptively extracts layers of CVs that balance the correlation between sets and the prediction of the interested outcome. The CVR-seq method provides a flexible integration of multiple sets of data as well as interpretable outcome prediction. The usage of CVR-seq is demonstrated by simulation and the METABRIC study that integrates copy number alteration (CNA) and gene expression (GEXP) data to predict breast cancer risk-free survival.
