|
Abstract:
|
Vertical integrative multi-assay genomic data analysis combines multiple sources of genomic assays (e.g. RNA-seq, CNV, genotyping, methylation) for the same set of samples. This design increases statistical power and accuracy. For a model-based supervised task with either continuous or categorical dependent variable, penalized regression approaches are usually employed to handle both the feature selection step, and the model fitting evaluation. Here we introduce an early screening step, employed prior to the model-fitting step. For each assay top analytes are selected based on their univariate association with the dependent variable, followed by an unsupervised hierarchical clustering across all selected analytes from all assays. For each cluster, a single representative analyte is selected. Then a penalized regression based model is fitted to measure the model's performance. This two-step approach was successfully validated on previously published multi-assay studies such as the HIV RV144 study. It allows balanced use of data from assays of different sizes (number of analytes), and ease interpretation of the results.
|
