DESeq2 multiple comparison

[dazhudou1122]

Dear Seqanswers community:

I ran into a question when I was doing RNA-seq analysis using DeSEQ2. I have 4 groups, 3 samples per group, and set my groups using the following code:

(condition <- factor(c(rep("ctl", 3), rep("A", 3), rep("B", 3), rep("C", 3))))

Can anyone tell me how the comparison in DeSEQ2 is done? I am not sure what the p value and log2FC mean in the default output. Supposedly it should be outputting ctl vs A in the default output. But the data is different when I have only ctl and A as input (2 groups). Also, is there a way to specify how the program performs the comparison? I read some previous threads but the answers were not very clear. Any help will be highly appreciated!

Best,

Wenhan

Below is the code I used:

countdata <- read.table("B6_B6MHV_B6MHVWY.txt", header=TRUE, row.names=1)
countdata <- countdata[ ,6:ncol(countdata)]
colnames(countdata) <- gsub("\\.[sb]am$", "", colnames(countdata))
countdata <- as.matrix(countdata)
head(countdata)
(condition <- factor(c(rep("ctl", 3), rep("A", 3), rep("B", 3), rep("C", 3))))
library(DESeq2)
(coldata <- data.frame(row.names=colnames(countdata), condition))
dds <- DESeqDataSetFromMatrix(countData=countdata, colData=coldata, design=~condition)
dds
dds <- DESeq(dds)
# Plot
dispersions
png("qc-dispersions.png", 2000, 2000, pointsize=20)
plotDispEsts(dds, main="Dispersion plot")
dev.off()

# Regularized log transformation for clustering/heatmaps, etc
rld <- rlogTransformation(dds)
head(assay(rld))
hist(assay(rld))

# Colors for plots below
## Ugly:
## (mycols <- 1:length(unique(condition)))
## Use RColorBrewer, better
library(RColorBrewer)
(mycols <- brewer.pal(8, "Dark2")[1:length(unique(condition))])

# Sample distance heatmap
sampleDists <- as.matrix(dist(t(assay(rld))))
library(gplots)
png("qc-heatmap-samples.png", w=1500, h=2500, pointsize=1500)
heatmap.2(as.matrix(sampleDists), key=F, trace="none",
col=colorpanel(100, "black", "white"),
ColSideColors=mycols[condition], RowSideColors=mycols[condition],
margin=c(10, 10), main="Sample Distance Matrix")
dev.off()

# Principal components analysis
## Could do with built-in DESeq2 function:
## DESeq2:lotPCA(rld, intgroup="condition")
## I like mine better:
rld_pca <- function (rld, intgroup = "condition", ntop = 500, colors=NULL, legendpos="bottomleft", main="PCA Biplot", textcx=1, ...) {
require(genefilter)
require(calibrate)
require(RColorBrewer)
rv = rowVars(assay(rld))
select = order(rv, decreasing = TRUE)[seq_len(min(ntop, length(rv)))]
pca = prcomp(t(assay(rld)[select, ]))
fac = factor(apply(as.data.frame(colData(rld)[, intgroup, drop = FALSE]), 1, paste, collapse = " : "))
if (is.null(colors)) {
if (nlevels(fac) >= 3) {
colors = brewer.pal(nlevels(fac), "Paired")
} else {
colors = c("black", "red")
}
}
pc1var <- round(summary(pca)$importance[2,1]*100, digits=1)
pc2var <- round(summary(pca)$importance[2,2]*100, digits=1)
pc1lab <- paste0("PC1 (",as.character(pc1var),"%)")
pc2lab <- paste0("PC1 (",as.character(pc2var),"%)")
plot(PC2~PC1, data=as.data.frame(pca$x), bg=colors[fac], pch=21, xlab=pc1lab, ylab=pc2lab, main=main, ...)
with(as.data.frame(pca$x), textxy(PC1, PC2, labs=rownames(as.data.frame(pca$x)), cex=textcx))
legend(legendpos, legend=levels(fac), col=colors, pch=20)
# rldyplot(PC2 ~ PC1, groups = fac, data = as.data.frame(pca$rld),
# pch = 16, cerld = 2, aspect = "iso", col = colours, main = draw.key(key = list(rect = list(col = colours),
# terldt = list(levels(fac)), rep = FALSE)))
}
png("qc-pca.png", 1500, 1500, pointsize=25)
rld_pca(rld, colors=mycols, intgroup="condition", xlim=c(-75, 35))
dev.off()


# Get differential expression results
res <- results(dds)
table(res$padj<0.05)
## Order by adjusted p-value
res <- res[order(res$padj), ]
## Merge with normalized count data
resdata <- merge(as.data.frame(res), as.data.frame(counts(dds, normalized=TRUE)), by="row.names", sort=FALSE)
names(resdata)[1] <- "Gene"
head(resdata)
## Write results
write.csv(resdata, file="diffexpr-results.csv")

[luminasapientiae]

There is a problem with this kind of analysis from statistical point of view. (more than two groups)
In practice, DESEq2 use a chi-squared like strategy and a binominal test, the analysis is better when you have one vs another group. In the other hand, edgeR is cappable to make this strategy works, but in their own manual this is not advisable and still in test phase, since you can have an "inflation" and false tendency of result. In other words: the logic in RNA-Seq still working in a 2D universe (difference of A-B) and you are talking about 3D (difference of A-B-C).

In my opinion, you should produce different DGEs crossing the groups (with the same control group if possible) and use a Venn diagram to identify the common findings.