Chapter Six - Bootstrapped Sparse Canonical Correlation Analysis: Mining Stable Imaging and Genetic Associations With Implicit Structure Learning

Abstract

Sparse canonical correlation analysis (SCCA) based on lasso and structured lasso has been widely studied to explore the complex associations between brain imaging and genetics features. Although those based on lasso have a better control of overall sparsity, they capture only a small portion of signals because of competition within correlated features. Advanced structure-based models provide a partial solution, but final patterns mostly depend on the prior structures applied. In this work, we propose a new framework, bootstrapped sparse canonical correlation analysis (BoSCCA), to explore the stable associations between correlated imaging and genetic data sets and to implicitly reconstruct the hidden structures. We compare the performances of BoSCCA and traditional SCCA using both synthetic and real data. In synthetic data, BoSCCA outperforms traditional SCCA in both association identification and group structure extraction, especially when the signal proportion goes below 5%. In real data, BoSCCA better captures the group structure within regions of interest and linkage disequilibrium blocks among single-nucleotide polymorphisms and yielded more biologically meaningful results.