Balanced mediated pathway detection in genomic data

Researchers are increasingly interested in identifying different parts of the genome which work together to influence a phenotypic trait. A major objective in bioinformatics involves finding groups of variables determined from omics technologies such as DNA methylation sites, transcriptome profiling, etc. Given one set of variables, one could determine how variables within work together to influence an outcome. These groups of variables are called functional modules and previous work has identified them through sparse matrix decomposition techniques such as sparse principal components analysis. To determine how different parts of the genome work together, we present methods to extend functional modules and identify variables that influence an outcome variable through a stepwise mediating fashion. Traditionally, module discovery involves sparse matrix decomposition accomplished through tuning regularization constraints. In this paper, we efficiently tune a cardinality-based sparse singular value decomposition to discover balanced mediated functional modules. These methods will be tested on simulated stepwise functional modules that contain several signal and non-signal variables and applied to real omics data collected in The Cancer Genome Atlas.