Introduction

Aims

ConfoundR is an interactive web application developed in R with Shiny. The goals of this app are to enable users to:

1. Compare the expression of an individual gene in tumour epithelium with the expression of the same gene in stromal/non-epithelial cells
2. Compare the expression of multiple genes in tumour epithelium with the expression of the same genes in stromal/non-epithelial cells
3. Compare the expression/enrichment of a gene set/gene signature in tumour epithelium with the expression/enrichment of the same gene set/gene signature in stromal/non-epithelial cells

By enabling the above comparisons the app enables users to identify individual genes, or gene sets/gene signatures which show differential expression/enrichment across tumour cell populations and therefore may be confounded when examined in bulk transcriptmomic tumour samples as a consequence of different proportions of tumour cell populations.

If you are ready to carry out this analysis, you can skip the below introductions, and examine the expression of individual genes using the Expression Boxplots page, or examine the expression of multiple genes using either a heatmap on the Expression Heatmap page or the Gene Set Enrichment Analysis (GSEA) method on the GSEA page. For more information on the datasets used in the ConfoundR app or for more details on the analysis modules available in the app see below.

Datasets

Eight datasets, available from Gene Expression Omnibus (GEO), are used in the ConfoundR app:

• three from colorectal cancer (CRC) (GSE39396, GSE35602 & GSE31279)
• two from breast cancer (GSE8183 & GSE14548) one of which is from triple negative breast cancer (TNBC) (GSE81838)
• one from pancreatic ductal adenocarcinoma (PDAC) (GSE164665)
• one from ovarian cancer (GSE9899)
• one from prostate cancer (GSE97284)

Each of these datasets consists of matched tumour epithelium and stromal/non-epithelial samples that have been obtained from bulk tumour specimens using either laser capture microdissection (LCM) or fluorescence-activated cell sorting (FACS). To view more detailed descriptions of each of the datasets and the pre-processing steps applied click here .

Expression Boxplots

The Expression Boxplots module allows users to compare the expression of a single gene between epithelial and stromal/non-epithelial cells in each of the datasets. This module enables the user to enter a gene symbol and the analysis module will draw boxplots for the expression of the chosen gene in epithelial and non-epithelial cells allowing a visual comparison of gene expresssion (for each dataset in which the chosen gene is found). In addition, a Mann-Whitney U test is performed comparing the expression of the chosen gene in epithelial and stromal/non-epithelial cells and the p-value is displayed on the plots. Boxplots are made using the ggplot2 package and the Mann-Whitney U tests are performed via the ggpubr package.

Expression Heatmap

The Expression Heatmap module allows users to visually compare the expression of multiple genes between epithelial and stromal/non-epithelial cells in each of the datasets. This module enables the user to enter a list of gene symbols (each gene symbol should be on a new line) and the analysis module will draw a heatmap of the expression of the chosen genes (which are present in each of the datasets). For heatmaps the expression values for each gene are converted to Z-scores for plotting, enabling visual comparison of the expression of each gene across samples. For the RNA-seq dataset (GSE164665) the transformed, normalised counts produced by the DESeq2 function, vst, are used as the expression values and are converted to Z-scores for plotting of the heatmap. In the heatmaps, the samples are split into epithelial and stromal/non-epithelial samples to aid comparison between these populations. Heatmaps are made using the ComplexHeatmap R package.

GSEA

The GSEA module enables users to perform gene set enrichment analysis (GSEA). Users can select existing gene sets from the Hallmark, KEGG (Kyoto Encyclopedia of Genes and Genomes), BioCarta, Reactome and PID (Pathway Interactions Database) gene set collections, using the dropdown menus provided, or use their own custom gene set by entering a list of gene symbols (each gene symbol should be on a line).

DE analysis methods

To perform GSEA it is necessary to first perform differential expression (DE) analysis. This app implements two different methods to conduct DE analysis, depending on the profiling technology used for the transcriptomic profiling of each dataset. For datasets that were profiled using microarray technology DE analysis was carried out using the limma R package while for datasets that were profiled using RNA-Seq technology DE analysis was carried out using the DESeq2 package. For the LCM experiments DE analysis is conducted comparing stroma to tumour epithelium. For the FACS sorted dataset (GSE39396), by default the DE analysis is performed comparing fibroblasts to tumour epithelial cells. However, the user can change which cell populations they wish to compare, for this FACS sorted dataset, using the “Compare” and “to” selection menus provided within the GSE39396 panel on the GSEA page. Multiple cell populations can be selected as one group and compared to another individual cell population or another group of multiple cell populations (e.g. Fibroblasts & Leukocytes compared to Epithelial cells).

GSEA analysis methods

To perform pre-ranked GSEA first DE analysis is performed in each dataset as described above comparing gene expression in stromal/non-epithelial cells with expression in epithelial cells. Following DE analysis for each dataset genes are ranked based on the t-statistic (limma) or Wald statistic (DESeq2) calculated by the DE analysis package. This ranked gene list is used, in combination with the gene set selected/uploaded by the user, to perform pre-ranked GESA using the GSEA function from the clusterProfiler R package using the fgsea method (specifically the fgseaSimple method) with 10000 permutations and a random seed of 123.

GSEA statistics

GSEA provides three statistics: Enrichment Score (ES), Normalised Enrichment Score (NES) and p-value (p). The meaning of each of these is explained below.

Enrichment Score (ES)

For the LCM datasets, a positive ES indicates that the gene set is more enriched in the stroma compared to the tumour epithelium. For the FACS dataset, a positive ES indicates that the gene set is more enriched in the cell populations selected in the “Compare” box relative to the cell populations selected in the “to” box (by default this is fibroblasts relative to epithelial cells but this can be changed by the user). On the other hand, a negative ES in the LCM datasets indicates that the gene set is less enriched in the stroma compared to the tumour epithelium (or in other words the gene set is more enriched in the tumour epithelium compared to the stroma). For the FACS dataset, a negative ES indicates that the gene set is less enriched in the cell populations selected in the “Compare” box relative to the cell populations selected in the “to” box (or in other words more enriched in the cell populations selected in the “to” box relative to the cell populations selected in the “Compare” box).

Normalised Enrichment Score (NES)

To calculate the NES, enrichment scores are calculated for 10000 random gene sets, with the same number of genes as the user selected gene set. The ES obtained by the user selected gene set is then divided by the mean of the enrichment scores for the random gene sets to obtain the NES.

p-value (p)

Using the enrichment scores of the 10000 random gene sets, the probability of obtaining an ES as extreme or more extreme than the ES obtained for the user selected gene set is calculated and this is the p-value (p) for the gene set. The p-value can then be used in conjunction with the NES to determine if the gene set is significantly positively/negatively enriched.

Expression Boxplots

Go

Expression Heatmap

Make Heatmap

GSEA

Perform GSEA