singscore

In singscore: Rank-based single-sample gene set scoring method

Introduction

This vignette demonstrates how to use the functions in R package 'singscore' to score a gene expression dataset against a gene set at a single-sample level and provides visualisation functions to improve interpretation of the results.

singscore implements a simple single-sample gene-set (gene-signature) scoring method which scores individual samples independently without relying on other samples in gene expression datasets. It provides stable scores which are less likely to be affected by varying sample and gene sizes in datasets and unwanted variations across samples. The scoring method uses a rank-based statistics and is quick to compute. For details of the methods please refer to the paper [@Foroutan231217]. It also provides various visualisation functions to further explore results of the analysis.

Install "singscore" R package

Install 'singscore' from Bioconductor

if (!requireNamespace("BiocManager", quietly=TRUE))
    install.packages("BiocManager")
BiocManager::install("singscore")

The most updated version of 'singscore' is hosted on GitHub and can be easily installed using devtools::install_github() function provided by devtools, (https://cran.r-project.org/package=devtools)

# You would need to install 'devtools' package first.
 install.packages("devtools")

# And install the 'singscore' package from the GitHub repository
# 'singscore' requires these packages to be installed: methods, stats, graphics, ggplot2, ggsci, grDevices,
#  ggrepel, plotly, tidyr, plyr, magrittr, reshape, edgeR, RColorBrewer, Biobase, GSEABase, BiocParallel
 devtools::install_github('DavisLaboratory/singscore')
# Set build_vignette = TRUE if would like to browseVignette()

Scoring samples against a gene-set

Load datasets

To illustrate the usage of 'simpleScore()', we first need to load the example datasets. The datasets used in this vignette have been built within the package. You can use the following scripts to load them into your R environment. Detailed steps of obtaining the full datasets are included at the end of the vignette. The 'tgfb_expr_10_se' dataset was obtained from [@Foroutanmolcanres.0313.2016] and it is a ten-sample subset of the original dataset. We are going to score the integrated TGFb-treated gene expression dataset (4 cases and 6 controls) against a TGFb gene signature with an up-regulated and down-regulated gene-set pair [@Foroutanmolcanres.0313.2016].

library(singscore)
library(GSEABase)
# The example expression dataset and gene signatures are included in the package
# distribution, one can directly access them using the variable names

# To see the description of 'tgfb_expr_10_se','tgfb_gs_up','tgfb_gs_dn', look at 
# their help pages using:

# ?tgfb_expr_10_se
# ?tgfb_gs_up
# ?tgfb_gs_dn

# Have a look at the object tgfb_expr_10_se containing gene expression data
# for 10 samples 
tgfb_expr_10_se

# Get the sample names by
colnames(tgfb_expr_10_se)

# View what tgfb_gs_up/dn contains
tgfb_gs_up
tgfb_gs_dn

# Get the size of the gene sets
length(GSEABase::geneIds(tgfb_gs_up))
length(GSEABase::geneIds(tgfb_gs_dn))

Sample scoring

To score samples, the gene expression dataset first needs to be ranked using the rankGenes() function which returns a rank matrix. This matrix along with the signatures are then passed to the simpleScore() function which returns a data.frame containing the scores for each sample. When only a single gene-set is available (i.e. not an up- and down- regulated pair), the same function can be called by setting the upSet argument to the gene-set.

# The recommended method for dealing with ties in ranking is 'min', you can
# change by specifying 'tiesMethod' parameter for rankGenes function.
rankData <- rankGenes(tgfb_expr_10_se)

# Given the ranked data and gene signature, simpleScore returns the scores and 
# dispersions for each sample
scoredf <- simpleScore(rankData, upSet = tgfb_gs_up, downSet = tgfb_gs_dn)
scoredf

# To view more details of the simpleScore, use ?simpleScore
# Note that, when only one gene set is available in a gene signature, one can 
# only input values for the upSet argument. In addition, a knownDirection 
# argument can be set to FALSE if the direction of the gene set is unknown.

# simpleScore(rankData, upSet = tgfb_gs_up, knownDirection = FALSE)

The returned data.frame consists of the scores for the up- and down- regulated gene-sets along with the combined score (TotalScore). Dispersion is calculated using the mad function by default and can be substituted by passing another function to the dispersionFun argument in simpleScore() such as IQR to calculate the inter-quartile range.

Sample scoring with a reduced number of measurements

Singscore requires transcriptome-wide measurements whereby all or most transcripts are measured. This is required to assess the expression of a gene in relation to the transcriptome. Panel-based transcriptomic assays measure a much smaller selection of transcripts therefore assessing the relative expression of each gene becomes a challenge. In such a setting, stably expressed genes can be used to assess relative expression. Expression of such genes is invariable across samples therefore they can be used to calibrate sample-wise gene expression measurements. This property allows for estimation of transcriptome-wide ranks with a small set of measurements inclusing a small set of stable genes. Detailed explanation of the rank estimation procedure using this logic is available in Bhuva et al. (2020) (manuscript in preparation).

Stably expressed genes in carcinoma transcriptomes and blood can be obtained using the getStableGenes function.

#get the top 5 stable genes in carcinomas
getStableGenes(5, type = 'carcinoma')

#get the top 5 stable genes in blood
getStableGenes(5, type = 'blood')

#get ensembl IDs instead of gene symbold
getStableGenes(5, type = 'carcinoma', id = 'ensembl')

This list can be used to score samples using a variant of singscore called stingscore.

#here we specify a custom set of genes (Entrez IDs)
stable_genes <-  c('8315', '9391', '23435', '3190')

#create a dataset with a reduced set of genes (signature genes and stable genes)
measured <-  unique(c(stable_genes, geneIds(tgfb_gs_up), geneIds(tgfb_gs_dn)))
small_tgfb_expr_10 <-  tgfb_expr_10_se[measured, ]
dim(small_tgfb_expr_10)

#rank genes using stable genes
rankData_st <-  rankGenes(small_tgfb_expr_10, stableGenes = stable_genes)
head(rankData_st)

#score samples using stingscore
#simpleScore invoked with the new rank matrix will execute the stingscore
#   algorithm
scoredf_st <- simpleScore(rankData_st, upSet = tgfb_gs_up, downSet = tgfb_gs_dn)
scoredf_st

#plot the two scores against each other
plot(scoredf$TotalScore, scoredf_st$TotalScore)
abline(coef = c(0, 1), lty = 2)

Scores computed using the classic singscore and the modified stingscore will be similar (correlated if not equal) if the selection of stable genes is good. Ideal stable genes will have stable gene expression across the samples being investigated and should cover a wide range of the expression spectrum. The genes produced by getStableGenes fulfill these conditions. Our approach using the product of ranks can be used to determine a set of stable genes for a new context (manuscript in preparation).

Visualisation and diagnostic functions

In this section, we show example usages of the visualisation functions included in this package.

Plot rank densities

Scores of each sample are stored in scoredf. We can use the plotRankDensity function to plot the ranks of genes in the gene-sets for a specific sample. We plot the rank distribution for the second sample in rankData which combines a density plot (densities calculated using KDE) with a barcode plot. Please note that since we are subsetting the data.frame rankData by one column, we set drop = FALSE to maintain the structure of the data.frame/matrix.

#  You can provide the upSet alone when working with unpaired gene-sets 
# We plot the second sample in rankData, view it by 
head(rankData[,2,drop = FALSE])

plotRankDensity(rankData[,2,drop = FALSE], upSet = tgfb_gs_up, 
                downSet = tgfb_gs_dn, isInteractive = FALSE)

Setting isInteractive = TRUE generates an interactive plot using the plotly package. Hovering over the bars in the interactive plot allows you to get information such as the normalised rank (between 0 and 1) and ID of the gene represented by the bar. For the rest of the plotting functions, the isInteractive = TRUE argument has the same behavior.

Plot dispersions of scores

Function plotDispersion generates the scatter plots of the 'score VS. dispersions' for the total scores, the up scores and the down score of samples. It requires the scored data.frame from simpleScore function and annotations (via annot parameter) can be used for coloring the points.

#  Get the annotations of samples by their sample names
tgfbAnnot <- data.frame(SampleID = colnames(tgfb_expr_10_se),
                       Type = NA)
tgfbAnnot$Type[grepl("Ctrl", tgfbAnnot$SampleID)] = "Control"
tgfbAnnot$Type[grepl("TGFb", tgfbAnnot$SampleID)] = "TGFb"

# Sample annotations
tgfbAnnot$Type

plotDispersion(scoredf,annot = tgfbAnnot$Type,isInteractive = FALSE)
# To see an interactive version powered by 'plotly', simply set the 
# 'isInteractive' = TRUE, i.e :
#
# plotDispersion(scoredf,annot = tgfbAnnot$Type,isInteractive = TRUE)

Plot score landscape

plotScoreLandscape plots the scores of the samples against two different gene signatures in a landscape for exploring their relationships.

There are two styles of the landscape plot (i.e scatter and hexBin plot). When the number of samples in the gene expression dataset is above the default threshold (100), plotScoreLandscape generates a hex bin plot otherwise a scatter plot. The threshold can be modified by changing the hexMin.

In order to better demonstrate the usage of plotScoreLandscape, we load some additional datasets consisting of pre-computed scores of larger public datasets. scoredf_ccle_epi and scoredf_ccle_mes are two scored results of a CCLE dataset [@barretina2012cancer] against an epithelial gene signature and mesenchymal gene signature [@Ep-MES] respectively. For details on how to obtain the dataset please see the section at the end of the vignette.

plotScoreLandscape(scoredf_ccle_epi, scoredf_ccle_mes, 
                   scorenames = c('ccle-EPI','ccle-MES'),hexMin = 10)

Similarly, pre-computed scores for the TCGA breast cancer RNA-seq dataset against epithelial and mesenchymal gene signatures are stored in scoredf_tcga_epi and scoredf_tcga_mes respectively [@Ep-MES]. The utility of this function is enhanced when the number of samples is large.

tcgaLandscape <- plotScoreLandscape(scoredf_tcga_epi, scoredf_tcga_mes, 
                   scorenames = c('tcga_EPI','tcga_MES'), isInteractive = FALSE)

tcgaLandscape

# To get an interactive version of plot, set isInteractive = TRUE

You can also project new data points onto the landscape plot generated above by using the projectScoreLandscape function. For example, the plot below overlays 3 CCLE samples onto the TCGA epithelial-mesenchymal landscape. Points are labeled with their sample names by default.

# Project on the above generated 'tcgaLandscape' plot
projectScoreLandscape(plotObj = tcgaLandscape, scoredf_ccle_epi, 
                      scoredf_ccle_mes,
                      subSamples = rownames(scoredf_ccle_epi)[c(1,4,5)],
                      annot = rownames(scoredf_ccle_epi)[c(1,4,5)], 
                      sampleLabels = NULL,
                      isInteractive = FALSE)

Custom labels can be provided by passing a character vector to the sampleLabels argument.

projectScoreLandscape(plotObj = tcgaLandscape, scoredf_ccle_epi, scoredf_ccle_mes,
                      subSamples = rownames(scoredf_ccle_epi)[c(1,4,5,8,9)],
                      sampleLabels = c('l1','l2','l3','l4','l5'),
                      annot = rownames(scoredf_ccle_epi)[c(1,4,5,8,9)], 
                      isInteractive = FALSE)

Estimate empirical p-values for the obtained scores in individual samples and plot null distributions

Permutation test

Hypothesis testing of the calculated scores is performed using a permutation test. The null hypothesis is that the gene-set is not enriched in the sample. For each sample, gene labels are randomly shuffled and scores computed against the gene-set. This is done $B$ times to generate the null distribution. The generateNull() function computes these for multiple samples ($n$) simultaneously resulting in an $n \times B$ matrix with permuted scores along the columns for each sample.

The permutation function has parallel computing features provided by using BiocParallel

# tgfb_gs_up : up regulated gene set in the tgfb gene signature
# tgfb_gs_dn : down regulated gene set in the tgfb gene signature

# This permutation function uses BiocParallel::bplapply() parallel function, by 
# supplying the first 5 columns of rankData, we generate Null distribution for 
# the first 5 samples.

# detect how many CPU cores are available for your machine
# parallel::detectCores()

ncores <- 1

# Provide the generateNull() function the number of cores you would like
# the function to use by passing the ncore parameter

permuteResult <-
  generateNull(
    upSet = tgfb_gs_up,
    downSet = tgfb_gs_dn,
    rankData = rankData,
    subSamples = 1:5,
    centerScore = TRUE,
    knownDirection = TRUE,
    B = 1000,
    ncores = ncores,
    seed = 1,
    useBPPARAM = NULL
  )
# Note here, the useBPPARAM parameter allows user to supply a BPPARAM variable 
# as a parameter which decides the type of parallel ends to use.
# such as 
# snow <-  BiocParallel::SnowParam(type = "SOCK")
# permuteResult <-  generateNull(upSet = tgfb_gs_up, downSet = tgfb_gs_dn,
# rankData[,1:5],  B = 1000, seed = 1,ncores = ncores, useBPPARAM = snow)

# If you are not sure about this, just leave the value as NULL and set how many
# CPU cores to use via the ncores argument. It will use the default parallel 
# backend for your machine.

# For more information, please see the help page for ?generateNull()
# Please also note that one should pass the same values to the upSet, 
# downSet, centerScore and knownDirection arguments as what they provide 
# for the simpleScore() function to generate a proper null distribution.

head(permuteResult)

Estimate empirical p-values

$p$-values can be estimated using the getPvals() function by providing the null distributions calculated above. Unless all permutations are exhausted (mostly infeasible), the minimum $p$-value obtained is $\frac{1}{B}$.

pvals <- getPvals(permuteResult, scoredf, subSamples = 1:5)

# getPval returns p-values for each individual sample.
# show the p-values for first 5 samples
pvals

Plot null distribution

Plot the null distributions for the first sample with the estimated $p$-value labelled. The function uses the sampleNames parameter to decide which samples to plot.

plotNull(permuteResult, scoredf, pvals, sampleNames = names(pvals)[1])

You can provide multiple sample names to plot these samples in one plot. For example, plot the first 2 samples.

# plot the null distributions for the first 2 samples and the p-values
# We can see from the plot, the control samples are not significant and TGFb 
# samples are very significant with very low p-values
plotNull(permuteResult, scoredf, pvals, sampleNames = names(pvals)[1:2])

More on the datasets

1. TGFb-EMT data

In the examples above, we loaded a gene expression matrix data tgfb_expr_10_se. This dataset is a ten-sample subset of a full dataset originally from the integrated TGFb-EMT data published by [@Foroutanmolcanres.0313.2016]. The full dataset can be accessed here https://figshare.com/articles/comBat_corrected_Foroutan_et_al_2017/5682862. The tgfb_gs_up and tgfb_gs_dn are the derived up-regulated/down-regulated genes of the TGFb-induced EMT gene signature by [@Foroutanmolcanres.0313.2016] (see Table S1. TGFβ-EMT signature).

1. CCLE dataset

scoredf_ccle_epi and scoredf_ccle_mes are two data frames of the pre-computed scores using The Cancer Cell Line Encyclopaedia (CCLE) breast cancer cell line RNA-seq dataset [@barretina2012cancer]. The CCLE dataset was normalised by TPM and can be downloaded from https://www.synapse.org/#!Synapse:syn5612998 Cell lines were scored against the epithelial and mesenchymal gene signatures, which were obtained from [@Ep-MES] and can be found in the 'Table S1B. Generic EMT signature for cell line' in the supplementary files.

1. TCGA cancer samples dataset

scoredf_tcga_epi and scoredf_tcga_mes are two data frames of the pre-computed scores using The Cancer Genome Atlas (TCGA) breast cancer RNA-seq data (RSEM normalised) [@TCGAcite] against the epithelial and mesenchymal gene signatures respectively. The gene signatures were obtained from [@Ep-MES] and can be found in the 'Table S1A. Generic EMT signature for tumour'. The TCGA dataset was downloaded from The UCSC Cancer Genomics Browser in February 2016 (https://genome-cancer.ucsc.edu)

# preview the scored CCLE samples
head(scoredf_ccle_epi)

# preview the scored TCGA samples
head(scoredf_tcga_epi)

Session Info

sessionInfo()

References

singscore documentation built on Nov. 8, 2020, 8:27 p.m.