The statistical stability of consensus independent component analysis for RNA-SEQ data in cancer research

Abstract

Independent component analysis (ICA) became a part of the standard machine learning pipeline for genomics data analysis. The approach allows to correct technical biases and batch effects in transcriptomics datasets. Separated signals are successfully used to characterize biological functions, their weights might be used for diagnostics (cancer subtypes classification) and prognostics (survival prediction). Using weights of independent components as features for downstream analysis requires high reproducibility of decomposition. Here we investigated the stability of extracted components depending on ICA parameters and validated the optimal number of parallel consensus ICA runs that provided reproducible deconvolution. Also, we estimated the effect of parallel runs on the quality of lung cancer type classification (LUSC/LUAD) and gene enrichment analysis results. Finally, we estimated the boundary values for the number of components that allows detecting biologically relevant signals in smaller patient cohorts.