Authored by: Robert M Flight <rflight79@gmail.com> on 2022-11-01 17:02:48
Current high-throughput molecular biology experiments are generating larger and larger amounts of data. Although there are many different methods to analyze individual experiments, methods that allow the comparison of different data sets are sorely lacking. This is important due to the number of experiments that have been carried out on biological systems that may be amenable to either fusion or comparison. Most of the current tools available focus on finding those genes in experiments that are listed as the same, or that can be shown statistically that it is significant that the gene was listed in the results of both experiments.
However, what many of these tools do not do is consider the similarities (and just as importantly, the differences) between experimental results at the categorical level. Categoical data includes any gene annotation, such as Gene Ontologies, KEGG pathways, chromosome location, etc. categoryCompare has been developed to allow the comparison of high-throughput experiments at a categorical level, and to explore those results in an intuitive fashion.
To make the concept more concrete, we will examine data from the microarray data set estrogen available from Bioconductor. This data set contains 8 samples, with 2 levels of estrogen therapy (present vs absent), and two time points (10 and 48 hours). A pre-processed version of the data is available with this package, the commands used to generate it are below. Note: the preprocessed one keeps only the top 100 genes, if you use it the results will be slightly different than those shown in the vignette.
library("affy")
library("hgu95av2.db")
library("genefilter")
library("estrogen")
library("limma")datadir <- system.file("extdata", package = "estrogen")
pd <- read.AnnotatedDataFrame(file.path(datadir,"estrogen.txt"),
header = TRUE, sep = "", row.names = 1)
pData(pd)## estrogen time.h
## low10-1.cel absent 10
## low10-2.cel absent 10
## high10-1.cel present 10
## high10-2.cel present 10
## low48-1.cel absent 48
## low48-2.cel absent 48
## high48-1.cel present 48
## high48-2.cel present 48Here you can see the descriptions for each of the arrays. First, we will read in the cel files, and then normalize the data using RMA.
filePaths <- file.path(datadir, rownames(pData(pd)))
a <- ReadAffy(filenames=filePaths, phenoData = pd, verbose = TRUE)## 1 reading /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/low10-1.cel ...instantiating an AffyBatch (intensity a 409600x8 matrix)...done.
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/low10-1.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/low10-2.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/high10-1.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/high10-2.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/low48-1.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/low48-2.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/high48-1.cel
## Reading in : /home/biocbuild/bbs-3.16-bioc/R/library/estrogen/extdata/high48-2.celeData <- mas5(a)## background correction: mas
## PM/MM correction : mas
## expression values: mas
## background correcting...## Warning: replacing previous import 'AnnotationDbi::tail' by 'utils::tail' when
## loading 'hgu95av2cdf'## Warning: replacing previous import 'AnnotationDbi::head' by 'utils::head' when
## loading 'hgu95av2cdf'## done.
## 12625 ids to be processed
## | |
## |####################|To make it easier to conceptualize, we will split the data up into two eSet objects by time, and perform all of the manipulations for calculating significantly differentially expressed genes on each eSet object.
So for the 10 hour samples:
e10 <- eData[, eData$time.h == 10]
e10 <- nsFilter(e10, remove.dupEntrez=TRUE, var.filter=FALSE,
feature.exclude="^AFFX")$eset
e10$estrogen <- factor(e10$estrogen)
d10 <- model.matrix(~0 + e10$estrogen)
colnames(d10) <- unique(e10$estrogen)
fit10 <- lmFit(e10, d10)
c10 <- makeContrasts(present - absent, levels=d10)
fit10_2 <- contrasts.fit(fit10, c10)
eB10 <- eBayes(fit10_2)
table10 <- topTable(eB10, number=nrow(e10), p.value=1, adjust.method="BH")
table10$Entrez <- unlist(mget(rownames(table10), hgu95av2ENTREZID, ifnotfound=NA))And the 48 hour samples we do the same thing:
e48 <- eData[, eData$time.h == 48]
e48 <- nsFilter(e48, remove.dupEntrez=TRUE, var.filter=FALSE,
feature.exclude="^AFFX" )$eset
e48$estrogen <- factor(e48$estrogen)
d48 <- model.matrix(~0 + e48$estrogen)
colnames(d48) <- unique(e48$estrogen)
fit48 <- lmFit(e48, d48)
c48 <- makeContrasts(present - absent, levels=d48)
fit48_2 <- contrasts.fit(fit48, c48)
eB48 <- eBayes(fit48_2)
table48 <- topTable(eB48, number=nrow(e48), p.value=1, adjust.method="BH")
table48$Entrez <- unlist(mget(rownames(table48), hgu95av2ENTREZID, ifnotfound=NA))And grab all the genes on the array to have a background set.
gUniverse <- unique(union(table10$Entrez, table48$Entrez))For both time points we have generated a list of genes that are differentially expressed in the present vs absent samples. To compare the time-points, we could find the common and discordant genes from both experiments, and then try to interpret those lists. This is commonly done in many meta-analysis studies that attempt to combine the results of many different experiments.
An alternative approach, used in categoryCompare, would be to compare the significantly enriched categories from the two gene lists. Currently the package supports two category classes, Gene Ontology, and KEGG pathways. Both are used below.
Note 1: I am not proposing that this is the best way to analyse this particular data, it is a sample data set that merely serves to illustrate the functionality of this package. However, there are many different experiments where this type of approach is definitely appropriate, and it is up to the user to determine if their data fits the analytical paradigm advocated here.
library("categoryCompare")
library("GO.db")
library("KEGGREST")
g10 <- unique(table10$Entrez[table10$adj.P.Val < 0.05])
g48 <- unique(table48$Entrez[table48$adj.P.Val < 0.05])For each list the genes of interest, and a background must be defined. Here we are using those genes with an adjusted P-value of less than 0.05 as the genes of interest, and all of the genes on the chip as the background.
list10 <- list(genes=g10, universe=gUniverse, annotation='org.Hs.eg.db')
list48 <- list(genes=g48, universe=gUniverse, annotation='org.Hs.eg.db')
geneLists <- list(T10=list10, T48=list48)
geneLists <- new("ccGeneList", geneLists, ccType=c("BP","KEGG"))## Warning in makeValidccLists(.Object): NAs introduced by coercionfdr(geneLists) <- 0 # this speeds up the calculations for demonstration
geneLists## List: T10
## Size of gene list: 400
## Size of gene universe: 8776
## Annotation: org.Hs.eg.db
##
## List: T48
## Size of gene list: 75
## Size of gene universe: 8776
## Annotation: org.Hs.eg.db
##
## Types of annotations to examine: BP; KEGG
## Number of FDR runs to perform: 0
## pValue Cutoff to decide significantly enriched annotations: 0.05
## Testdirection: over representedNow run the enrichment calculations on each list. In this case enrichment will be performed using the biological process (BP) Gene Ontology, and KEGG Pathways.
enrichLists <- ccEnrich(geneLists)## Performing Enrichment Calculations ....
## T10 : BP
## T48 : BP
## T10 : KEGG
## T48 : KEGG
## Done!!enrichLists## Annotation category: GO BP
## FDR runs: 0
## Default p-values to use: pval
## pCutoff: 0.05
##
## List: T10
## Gene to GO BP test for over-representation
## 6114 GO BP ids tested (898 have p <= 0.05 & count >= 0)
## Selected gene set size: 394
## Gene universe size: 8365
## Annotation package: org.Hs.eg
##
## List: T48
## Gene to GO BP test for over-representation
## 2858 GO BP ids tested (740 have p <= 0.05 & count >= 0)
## Selected gene set size: 74
## Gene universe size: 8365
## Annotation package: org.Hs.eg
##
##
## Annotation category: KEGG
## FDR runs: 0
## Default p-values to use: pval
## pCutoff: 0.05
##
## List: T10
## Gene to KEGG test for over-representation
## 167 KEGG ids tested (16 have p <= 0.05 & count >= 0)
## Selected gene set size: 222
## Gene universe size: 3499
## Annotation package: org.Hs.eg
##
## List: T48
## Gene to KEGG test for over-representation
## 61 KEGG ids tested (6 have p <= 0.05 & count >= 0)
## Selected gene set size: 43
## Gene universe size: 2917
## Annotation package: org.Hs.egThere are a lot of GO BP processes enriched using the p-value cutoff of 0.05, so lets make that more stringent (0.001). This is done here merely for speed, in a usual analysis you should choose this number, and the type of cutoff (p-value or fdr) carefully.
pvalueCutoff(enrichLists$BP) <- 0.001
enrichLists## Annotation category: GO BP
## FDR runs: 0
## Default p-values to use: pval
## pCutoff: 0.001
##
## List: T10
## Gene to GO BP test for over-representation
## 6114 GO BP ids tested (147 have p <= 0.001 & count >= 0)
## Selected gene set size: 394
## Gene universe size: 8365
## Annotation package: org.Hs.eg
##
## List: T48
## Gene to GO BP test for over-representation
## 2858 GO BP ids tested (70 have p <= 0.001 & count >= 0)
## Selected gene set size: 74
## Gene universe size: 8365
## Annotation package: org.Hs.eg
##
##
## Annotation category: KEGG
## FDR runs: 0
## Default p-values to use: pval
## pCutoff: 0.05
##
## List: T10
## Gene to KEGG test for over-representation
## 167 KEGG ids tested (16 have p <= 0.05 & count >= 0)
## Selected gene set size: 222
## Gene universe size: 3499
## Annotation package: org.Hs.eg
##
## List: T48
## Gene to KEGG test for over-representation
## 61 KEGG ids tested (6 have p <= 0.05 & count >= 0)
## Selected gene set size: 43
## Gene universe size: 2917
## Annotation package: org.Hs.egCurrently you can see that for T10, there are 147 processes enriched, and 70 for T48. For KEGG, there are 16 and 6 for T10 and T48, respectively.
To see which processes and pathways are enriched, and to compare them, we will run ccCompare, after generating a ccOptions object to tell the software exactly which comparisons to do.
ccOpts <- new("ccOptions", listNames=names(geneLists), outType='none')
ccOpts## List Names: T10; T48
## Comparisons: T10; T48; T10,T48
## Colors: #FF7A9E; #89BC00; #00C8EA
## Output Types: noneccResults <- ccCompare(enrichLists,ccOpts)
ccResults## ccCompare results for:
##
## Annotation category: GO BP
## Main graph: A graphNEL graph with directed edges
## Number of Nodes = 99
## Number of Edges = 2360
##
## Annotation category: KEGG
## Main graph: A graphNEL graph with directed edges
## Number of Nodes = 17
## Number of Edges = 37The ccResults is a list object where for each type of category (Gene Ontologies, KEGG pathways, etc) there are ccCompareResult objects containing various pieces, including the output of the enrichments in table form (mainTable) with a designation as to which of the geneLists they originated from, a graph that shows how the annotations are connected to each other (mainGraph), and which genes belong to which annotation, and which list they originated from (allAnnotation).
Currently the easiest way to visualize and interact with this data is by using Cytoscape and the RCy3 package. To set up RCy3, see the RCy3 webiste .
## Oh no! I can't find Cytoscape. RCy3 can not continue!
## Please check that Cytoscape is running, CyREST is installed and your base.url parameter is correct.Error in .cyFinally(res) : object ‘res’ not found [1] “No connection to Cytoscape available, subsequent visualizations were not run”
Once you have Cytoscape up and running, then we can examine the results from each category of annotations. First up, GO Biological Process.
Note on deleting edges in Cytoscape: As of RCy3 v 1.7.0 and Cytoscape 3.5.1, deleting nodes after putting the graph in Cytoscape is slow. This network contains ~ 20,000 edges to start, most with very low weights. Therefore, I recommend removing some low weight edges before putting the graph into Cytoscape!
ccBP <- breakEdges(ccResults$BP, 0.2)
cw.BP <- ccOutCyt(ccBP, ccOpts)