A new genome-wide method to identify genes with bimodal gene expression

Abstract

A new method is presented to detect bimodality in gene expression data using the Gaussian Mixture Models to cluster samples in each mode. We have used the method to search for bimodal genes in data from 25 tumor types available from The Cancer Genome Atlas. The method identified 554 genes with bimodal gene expression, of which 46 were identified in more than one cancer type. To further illustrate the impact of the method, we show that 96 out of the 554 genes with bimodal expression patterns presented different prognosis when patients belonging to the two expression peaks are compared. The software to execute the method and the corresponding documentation are available at https://github.com/LabBiosystemUFRN/Bimodality_Genes.

Modelling Bimodal Data Using a Multivariate Triangular-Linked Distribution

“…A mixture of bivariate negative binomial-normal distributions has also been considered for the same purpose [11]. In genetics, Gaussian mixture models have been used to identify genes with bimodal expression patterns in tumors [12]. Other applications where mixture models have been used for bimodal data include modelling ratings from Tripadvisor.com [13],…”

“…If bimodality in the gene expressions are identified, then this can be used to extract interesting insights into biological attributes of a particular cancer associated with the tumor. The data we will use are extracted using developed software [12], which is written in R. These data consist of expression and clinical data from 25 different tumor types which is in turn harvested from the Cancer Genome Atlas (https://portal.gdc.cancer.gov/ accessed on 5 September 2021). In total the authors of [12], get expression values measured in Fragments by Exon Kilobase per Millions of Mapped Fragments values (FPKM), for just under 25,000 genes which yields more than 10 million observations.…”