B1285
Title: Covariance assisted multivariate penalized additive regression (ComPAdRe)
Authors: Neel Desai - University of Pennsylvania (United States) [presenting]
Abstract: The aim is to propose a robust, computationally efficient solution for simultaneously selecting and estimating multiple sparse additive models with correlated errors. The proposed method (ComPAdRe) simultaneously selects between null, linear, and smooth nonlinear effects for each predictor while incorporating jointly estimated sparse residual structure among responses for potential gains both in selection accuracy and in statistical efficiency in a manner analogous to the principles of seemingly unrelated regressions (SUR). The method is constructed computationally efficient, allowing the selection and estimation of linear and non-linear covariates to be conducted in parallel across responses. Compared to single response approaches that marginally select linear and non-linear covariate effects, it is demonstrated with extensive designed simulations that this approach leads to gains in both statistical efficiency and selection accuracy, particularly in settings where the signal is moderate relative to the level of noise. The approach is applied to protein-mRNA expression levels from 8 known breast cancer pathways obtained from The Cancer Proteome Atlas (TCPA), and both mRNA protein associations and protein-protein subnetworks for each pathway are characterized. The non-linear mRNA-protein associations is found for the Core Reactive, EMT, PIK-AKT, and RTK pathways.
