Gene networks in cancer are biased by aneuploidies and sample impurities

Abstract

Gene regulatory network inference is a standard technique for obtaining structured regulatory information from, among other data sources, gene expression measurements.Methods performing this task have been extensively evaluated on synthetic, and to a lesser extent real data sets. They are often applied to gene expression of human cancers.However, in contrast to the evaluations, these data sets often contain fewer samples, more potential regulatory links, and are biased by copy number aberrations as well as cell mixtures and sample impurities. Here, we take networks inferred from TCGA cohorts as an example to show that (1) transcription factor annotations are essential to obtaining reliable networks, and (2) even when taking these into account, we should expect between 20 and 80% of edges to be caused by copy number changes and cell mixtures rather than transcription factor regulation.

Co-expression analysis reveals interpretable gene modules controlled bytrans-acting genetic variants

“…These can lead to spurious trans-eQTL signals, because genetic variants associated with cell type composition changes would appear as trans-eQTLs for cell-type-specific genes [3]. Furthermore, multiple studies have demonstrated that also the co-expression signals in tissues are largely driven by cell type composition effects [44][45][46]. Thus, even though PLIER detected the ARHGEF3 trans-eQTL in whole blood, this could have been at least partially driven by the change in platelet proportion between individuals [9].…”

Gene networks in cancer are biased by aneuploidies and sample impurities

“…Classic methods that infer networks from multiple samples of unperturbed gene expression can roughly be divided in correlation-based and information-theoretic models. These and more methods have been reviewed in detail 8,10,28,29 and hence we only provide a brief overview. Information-theoretic approaches started out with relevance networks 12 , in which the pairwise mutual information (MI) is computed between all pairs of genes.…”

corto: a lightweight R package for Gene Network Inference and Master Regulator Analysis

“…CNV correction. As the presence of CNVs can influence and bias the generation of gene networks (Schubert et al, 2019), corto gives the optional possibility to use CNV data to correct target expression profiles via linear regression. An example of corto CNV-corrected network analysis in the TCGA Glioblastoma (GBM) dataset is provided in the package vignette and in Supp Figure S1 .…”