1 Installing ctsGE

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

2 Workflow of clustering with ctsGE

1. Build the expression matrix from expression data

2. Define an expression index (i.e., a sequence of 1,-1, and 0) for each gene

3. Cluster the gene set with the same index, applying K-means: \[kmeans(genesInIndex,k)\]

4. Graphic visualization of expression patterns

5. Interactive visualization and exploration of gene-expression data

3 Building the expression matrix

As input, the ctsGE package expects a normalized expression table, where rows are genes and columns are samples. This can consist of count data as obtained, e. g., from RNA-Seq or other high-throughput sequencing experiment or microarray experiment. Example data from the [Gene Expression Omnibus (GEO)] (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE2077) are used here illustrate ctsGE’s potential. The data are the expression profile of Cryptosporidium parvum-infected human ileocecal adenocarcinoma cells (HCT-8),33 Deng M, Lancto CA, Abrahamsen MS (2004). Cryptosporidium parvum regulation of human epithelial cell gene expression. Int. J. Parasitol. 34, 73–82.) comprised of 12,625 genes over 18 samples (three replicates of six developmental stages in human cancer). For tutorial purposes and for simplification, only one replicate out of three is used, for six overall time points.

library(GEOquery)

gse2077 <- getGEO('GSE2077')
gseAssays <- Biobase::assayData(gse2077)
gseExprs <- Biobase::assayDataElement(gseAssays[[1]][,c(1:6)],'exprs')

# list of the time series tables use only 6 samples
gseList <- lapply(1:6,function(x){data.frame(Genes = rownames(gseExprs),Value = gseExprs[,x])}) 
names(gseList) <- colnames(gseExprs)

rts <- readTSGE(gseList,labels = c("0h","6h","12h","24h","48h","72h")) 

data_dir <- system.file("extdata", package = "ctsGE")
files <- dir(path=data_dir,pattern = "\\.xls$")

rts <- readTSGE(files, path = data_dir, labels = c("0h","6h","12h","24h","48h","72h") )

names(rts)

## [1] "tsTable"    "samples"    "tags"       "timePoints"

rts$timePoints

## [1] 6

head(rts$samples)

## [1] "0h"  "6h"  "12h" "24h" "48h" "72h"

head(rts$tags)

## [1] "1000_at"   "1001_at"   "1002_f_at" "1003_s_at" "1004_at"   "1005_at"

4 Defining the expression index

1. First, the expression matrix is standardized. The function default standardizing method is a median-based scaling; alternatively, a mean-based scaling can be used. The new scaled values represent the distance of each gene at a certain time point from its center, median or mean, in median absolute deviation (MAD) units or standard deviation (SD) units, respectively.

2. Next, the standardized values are converted to index values that indicate whether gene expression is above, below or within the limits around the center of the time series, i.e., 1 / -1 / 0, respectively. The function defines a parameter cutoff (see Section 4.1) that determines the limits around the gene-expression center.Then the function calculates the index value at each time point according to:

   • 0: standardized value is within the limits (+/- cutoff)
   • 1: standardized value exceeds the upper limit (+ cutoff)
   • -1: standardized value exceeds the lower limit (- cutoff)

3. The +/- cutoff parameter defines a reference range to which the data are compared. When the range is too big, more time-series points will fall into it and will get an index value of 0, and this may be misleading. Too small range can result in too many index groups that will be too sensitive to small fluctuations in the time-series index. The function chooses the optimal cutoff (see Section 4.1) after testing different cutoff values from 0.5 to 0.7 in increments of 0.05.

##Checking the optimal cutoff The function PreparingTheIndexes generates an expression index (i.e., a sequence of 1,-1, and 0) that represents the expression pattern along time points for each gene. Setting different limits for the center era (with the parameter cutoff) will change the index for each gene-expression profile and consequently, the number of genes in each index group. Following the idea that instead of filtering “irrelevant” genes to reduce the noise, the clustering will be performed on small gene groups, one would like to choose a cutoff value, that will minimize the number of genes in each group, i.e., generate index groups of equal size. The test for equality is performed by calculating the chi-squared value from a comparison between the number of genes in the index groups and the null hypothesis that all index groups are equal. The test is performed for all the cutoffs in the range and the cutoff that gives the minimal chi-squared value is the most likely to generate equal index groups. The range of the cutoff values is given by min_cutoff and max_cutoff arguments. However, by setting the same value to min and max parameters one can define a cutoff regardless of what was suggested by the function.

prts <- PreparingTheIndexes(x = rts, min_cutoff=0.5, max_cutoff=0.7, mad.scale = TRUE)

cutoff = 0.55 is the optimal cutoff with the lowest chi-squared value

prts$cutoff

## [1] 0.55

With cutoff = 
0.55, most of the genes were assigned to indexes with three or two zeros, 
indicating a variety of expression patterns.

prts <- PreparingTheIndexes(x = rts, mad.scale = TRUE) 
names(prts)

## [1] "tsTable"    "samples"    "tags"       "timePoints" "scaled"    
## [6] "index"      "cutoff"

ClustIndexes <- ClustIndexes(prts, scaling = TRUE)
names(ClustIndexes)
# table of the index and the recommended k that were found by the function 
head(ClustIndexes$optimalK)

# Table of clusters index for each gene
head(ClustIndexes$ClusteredIdxTable) 

5 Graphic visualization of an index

The PlotIndexesClust() function generates graphs and tables of a specific index and its clusters. The user decides whether to supply the k or let the function calculate the k for the selected index.

indexPlot <- PlotIndexesClust(prts,idx = "1100-1-1",scaling = TRUE)
names(indexPlot)

## [1] "Clust_4_1100-1-1" "graphs"

Genes in ‘1100-1-1’ index and their clusters (k was chosen by the function):
Number of clusters (k) for ‘1100-1-1’ is: length(indexPlot$graphs)

## [1] 4

Table of genes in ‘1100-1-1’ index, seperated to clusters: head(indexPlot[[1]])

For this example, the index 1100-1-1 is used. by Looking at this index, it can be assumed that the expression of the genes belonging to it was downregulated along the time points. Since the index only states whether gene expression is upregulated (1), downregulated (-1) or stays the same (0), gene subsets of the same profile will usually show more than one expression pattern. K-means helps distinguish these patterns from one another.

Line graphs of the genes’ expression patterns in index ‘1100-1-1’ separated into clusters:

indexPlot$graphs

## $`Clust_1_1100-1-1`

## 
## $`Clust_2_1100-1-1`

## 
## $`Clust_3_1100-1-1`

## 
## $`Clust_4_1100-1-1`

##Export gene expression table to file Genes expression data in index ‘1100-1-1’ separated into clsuter can be exported with R function write.table:

write.table(indexPlot[[1]], file, sep = "\t")

6 GUI for interactive exploration of gene-expression data

The ctsGEShinyApp function the ctsGE object,and opens an html page as a GUI. On the web page, the user chooses the profile to visualize and the number of clusters (k parameter for K-means) to show. The line graph of the profile separated into the clusters will show in the main panel, and a list of the genes and their expressions will also be available. The tables and figures can be downloaded.

*Note that running the Shiny GUI block the current session while running.

library(shiny)
library(DT)
ctsGEShinyApp(rts)