A feature selection method for vertical integrative analysis of multi-assay genomic data.
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.
Our method implement the MultiAssayExperiment R/Bioconductor S4 class for multi-assay data. Feature selection methods are implemented via a high-level machine learning approach with the mlr pacakge.
# Install development version from GitHub
devtools::install_github("FredHutch/JDRFCAV")
# TBA: Install release version from CRAN
# install.packages("JDRFCAV")A. NCBI/GEO raw files -> processed data -> MultiAssayExperiment (MAE) -> mlr's task foramt 
B. Biomarker discovery: B.1 Feature selection: Fun_lrn_univ_only_makePrep_MaG, Fun_lrn_univ_Clusters_All_makePrep_MaG B.2 Sensitivity analysis
C. (TBA) UnivCPO, UnivClustCPO (refactoring the above makePreprocWrapper() )
Fred Hutch Benaroya Research Institute JDRF
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent