3.Differential Expression

Differential Expression for bulk RNAs

Pipeline

Data Structure

Inputs

File format	Information contained in file	File description	Notes
txt	miRNA.homer.ct.mx	Raw counts matrix	Used for edgeR and DESeq2 analysis
txt	miRNA.homer.rpm.mx	RPM matrix	Used for wilcoxon test
txt	miRNA.NCvsHCC.DESeq2.ids.rpkm.mx	RPKM matrix	Used for clustering plot
txt	design.txt	Sample info.	Classification information

Outputs

File format	Information contained in file	File description	Notes
tsv	miRNA.NCvsHCC.edgeR.tsv	EdgeR result	LogFC, PValue and FDR
txt	miRNA.homer.DESeq2.rlog.mx	DESeq2 result	Rlog normalized counts
tsv	miRNA.NCvsHCC.DESeq2.tsv	DESeq2 result	LogFC, PValue and padj
tsv	miRNA.NCvsHCC.wilcox.tsv	Wilcox result	LogF, PVaule and FDR
pdf	miRNA.NCvsHCC.DESeq2.norm.pdf	Heatmap result	Rlog normalized count
pdf	miRNA.NCvsHCC.DESeq2.pca.pdf	Scatter plot	PCA analysis
pdf	miRNA.NCvsHCC.DESeq2.MAplot.pdf	Scatter plot	MAplot analysis
pdf	miRNA.NCvsHCC.DESeq2.Dist.pdf	Heatmap result	Sample distance
pdf	miRNA.NCvsHCC.DESeq2.ids.DE.pdf	Heatmap result	DE genes

Running Scripts

Software/Tools

Assumption for most normalization and differential expression analysis tools: The expression levels of most genes are similar, i.e., not differentially expressed.

a) DEseq: defines scaling factor (also known as size factor) estimates based on a pseudoreferencesample, which is built with the geometric mean of gene counts across all cells (samples).

b) EdgeR (TMM): trimmed mean of M values

c) Wilcox Test using RPM: Read counts Per Million of total mapped reads; alternatives: RPKM, TPM

Performance:

Normalizing single-cell RNA sequencing data: challenges and opportunities, Nature Methods, 2017

Example of single case

Preprocess

# experimential design
design <- read.table("design.txt",sep="\t",header=T)

# expression matrix
mx <- read.table("miRNA.homer.ct.mx",sep="\t",header=T)

# filter genes
filter_genes <- apply(
    mx[,2:ncol(mx)],
    1,
    function(x) length(x[x > 2]) >= 2
)
mx_filterGenes <- mx[filter_genes,]

edgeR

library(edgeR)

#Read Data in
countData <- mx_filterGenes[,-1]
rownames(countData) <- mx_filterGenes[,1]
design <- read.table("design.txt",sep="\t",header=T)
colData <- design

# generate DGE object
y <- DGEList(countData, samples=colData, group=colData$Treatment)
y <- calcNormFactors(y)

#Estimate Error Model
design <-model.matrix(~Treatment, data=colData)
y <- estimateDisp(y, design)

# classic methods: compute p-values, then output
et <- exactTest(y)
res <- topTags(et,Inf)
tidyResult <- data.frame(Gene=rownames(res$table), res$table)
#write.table(tidyResult,file="hcc_example.miRNA.NCvsHCC.edgeR.classic.tsv", quote=F, sep="\t",row.names=FALSE)

# Generalized linear models 
fit <- glmFit(y,design)
# likelihood ratio test
lrt <- glmLRT(fit,contrast = c(1,-1))
FDR <- p.adjust(lrt$table$PValue, method="BH")
padj_lrt <- cbind(lrt$table,FDR)
fit_df <- lrt$fitted.values
#write.table(fit_df,file = "hcc_example.miRNA.homer.edgeR.TMM.mx",row.names = T, sep="\t", quote=F)
merged_lrt <- merge(fit_df,padj_lrt,by="row.names")
colnames(merged_lrt)[1] <- "Genes"
write.table(merged_lrt,file = "miRNA.NCvsHCC.edgeR.tsv",row.names = F, sep="\t", quote=F)

DESeq2

library(DESeq2)
#Read Data in
countData <- mx_filterGenes
colData <- design

# generate Dataset object 
dds <- DESeqDataSetFromMatrix(countData, colData, design=~Treatment, tidy=TRUE)

# normlize using rlog method
norm <- rlog(dds,blind=FALSE)
norm_matrix <- assay(norm)
norm_df <- data.frame(Gene=rownames(norm_matrix), norm_matrix)
write.table(norm_df, "miRNA.homer.DESeq2.rlog.mx", quote=F, row.names = FALSE,sep="\t")

deg <- DESeq(dds)
res <- results(deg,tidy=TRUE)
merged_res <- merge(norm_df,res,by.x="Gene",by.y="row")
write.table(merged_res,file="miRNA.NCvsHCC.DESeq2.tsv",quote=F, sep="\t",row.names=FALSE)

Wilcoxon

cpmMx <- read.table("miRNA.homer.rpm.mx",sep="\t",header=T)
filter_cpm <- apply(
    mx[,2:ncol(cpmMx)],
    1,
    function(x) length(x[x > 0]) >= 2
)
mx_filterCPM <- cpmMx[filter_cpm,]

myFun <- function(x){
  x = as.numeric(x)
  v1 = x[2:4]
  v2 = x[5:10]
  out <- wilcox.test(v1,v2)
  out <- out$p.value
}
p_value <- apply(mx_filterCPM,1,myFun)
p_value[is.nan(p_value)] <- 1
FDR <- p.adjust(p_value,method = "BH")
mx_filterCPM$avgNC <- apply(mx_filterCPM[,2:4],1,mean)
mx_filterCPM$avgHCC <- apply(mx_filterCPM[,5:10],1,mean)
mx_filterCPM$log2fc <- log2((mx_filterCPM$avgNC+0.25)/(mx_filterCPM$avgHCC+0.25))
results <- cbind(mx_filterCPM,p_value,FDR)
write.table(results,file = "miRNA.NCvsHCC.wilcox.tsv",row.names = F, sep="\t", quote=F)

Plot

# 1.heatmap for DESeq2 rlog normalized count matrix
pdf("miRNA.NCvsHCC.DESeq2.norm.pdf")
library("pheatmap")
select <- order(rowMeans(counts(deg,normalized=TRUE)), decreasing=TRUE)[1:25]
df <- as.data.frame(colData(deg)[,c("Treatment","Sample")])
pheatmap(assay(norm)[select,], cluster_rows=FALSE, show_rownames=FALSE, cluster_cols=TRUE, annotation_col=df)
dev.off()

# 2.PCA analysis
pdf("miRNA.NCvsHCC.DESeq2.pca.pdf")
plotPCA(norm, intgroup=c("Treatment"))
dev.off()

# 3.MA plot
pdf("miRNA.NCvsHCC.DESeq2.MAplot.pdf")
plotMA(deg, ylim=c(-5,5))
dev.off()

# 4.distance between samples
pdf("miRNA.NCvsHCC.DESeq2.Dist.pdf")
sampleDists <- dist(t(assay(norm)))
library("RColorBrewer")
sampleDistMatrix <- as.matrix(sampleDists)
rownames(sampleDistMatrix) <- paste(norm$Treatment, norm$Sample, sep="-")
colnames(sampleDistMatrix) <- paste(norm$Treatment, norm$Sample, sep="-")
colors <- colorRampPalette( rev(brewer.pal(9, "Blues")) )(255)
pheatmap(sampleDistMatrix,
         clustering_distance_rows=sampleDists,
         clustering_distance_cols=sampleDists,
         col=colors)
dev.off()


# 5.Hierarchical clustering for differential expressed genes
awk 'BEGIN{FS=OFS="\t"}($12>1 && $15<=0.05){print $1}' hcc_example.miRNA.NCvsHCC.DESeq2.tsv > hcc_example.miRNA.NCvsHCC.DESeq2.NC_high.ids
awk 'BEGIN{FS=OFS="\t"}($12<-1 && $15<=0.05){print $1}' hcc_example.miRNA.NCvsHCC.DESeq2.tsv > hcc_example.miRNA.NCvsHCC.DESeq2.HCC_high.ids

mx <- read.table("miRNA.NCvsHCC.DESeq2.ids.DE.mx",sep="\t",head=T)
# log2 transfer the RPKM value
log2mx <- log2(mx+0.05)
pdf("miRNA.NCvsHCC.DESeq2.ids.rpkm.heatmap.pdf",width=4,height=6)
pheatmap(as.matrix(log2mx), Rowv=T, Colv=T, dendrogram = "both", trace = "none", density.info = c("none"), scale="row")
dev.off()

Example of batch job

Draw Plots

1. Heatmap for DESeq2 normalized count matrix

2. PCA analysis

3. MA plot

4. Distance between samples

5. Hierarchical clustering for differential expressed genes

Tips/Utilities

edgeR User Guide

DESeq2 User Guide

Homework and more

1. Identify differential expressed genes for other RNA types. between differential conditions, i.e. Normal Control (NC) V.S. HCC using three methods: edgeR, DESeq2 and Wilcox/Mann-Whitney-U Test.

2. Draw Venn plot to show the difference among the above three methods.

More Reading and Practice

• Additional Tutorial : 3. Differential Expression Analysis

Video

1. Differential expression

@Youtube

@Bilibili