Molecular Group and Correlation Guided Structural Learning for Multi-Phenotype Prediction

We propose a supervised learning algorithm to perform feature selection and outcome prediction for genomic data with multi-phenotypic responses. Our algorithm particularly incorporates the genome and/or phenotype grouping structures and phenotype correlation structures in feature selection, effect estimation, and outcome prediction under a penalized multi-response linear regression model. Extensive simulations demonstrate its superior performance over its competing methods. We apply the proposed algorithm to two omics studies. In the first study, we identified novel association signals between multivariate gene expressions and high-dimensional DNA methylation profiles, providing biological insights into how CpG sites regulate gene expressions. The second study is for cell type deconvolution. Using the proposed algorithm, we were able to achieve better cell type fraction predictions using high-dimensional gene expression data.