DESeq2 for subset of (4115) genes

[mnoon]

I have a question about how reliable results would I get if I ran DESeq2 on about 4000 genes instead of all 17000 genes?

Basically I have two conditions with 2 and 3 replicates each respectively. I ran DESeq2 and got about 4115 differentially expressed up-regulated genes (with Padj < 0.05). Would it make sense if I subset my counts-matrix for these 4000 genes and run DESeq2 on this subset (dataMatrix for 4000 genes only) for other samples and/or conditions?

I did and I got about 51 genes (with Padj < 0.05). Should I not get more genes? what I am trying to ask is that would it make any statistical sense if I only take subset of these genes for other conditions (samples) and do differential expression? would my results be statistically reliable?

here is my code:

countData <- read.table("/Users/mnoon/Desktop/fruit_fly/analysis2/ugr_utgr_gfp/ugr_utgr_gfp_merged_counts.txt",header=TRUE)
head(countData)

gene_4114_list <- read.table("/Users/mnoon/Desktop/fruit_fly/analysis2/ugrInput_ugrvg_upregulated_4114genes.csv",header=TRUE)
head(gene_4114_list)

#Subset 4114 genes to get Read-Counts
subLIST <- merge(gene_4114_list, countData, by.x= "fly_id", by.y= "gene")
head(subLIST)
rownames(subLIST) <- subLIST$fly_id
subLIST <- subLIST[-1]
head(subLIST)

colData<-data.frame(condition=factor(c("UGRgfp", "UGRgfp", "UTGRgfp", "UTGRgfp", "UTGRgfp")))


dds <- DESeqDataSetFromMatrix(countData = subLIST, colData = colData, design = ~ condition)
colData(dds)$condition<-factor(colData(dds)$condition,levels=c("UGRgfp","UTGRgfp"))

dds <- DESeq(dds)

res<-results(dds)
res<-res[order(res$padj),]
head(res)
mcols(res, use.names=TRUE)
#get number of genes filtered with this threshold
sum( res$padj < 0.05, na.rm=TRUE )

[wetSEQer]

Why would you want to do that?

[mnoon]

I did that for two reasons;

1. I know this set of genes represent CNS and I wanted to see how there are expressed in CNS

2. this could be a silly reason since some biologist suggested that if I have small set of genes for Differential expression it might give me more fold changes which doesnt make sense at all to me but I gave it a try and dont see any change in the fold change.

[dpryan]

The biologist that suggested that should be ignored, he/she hasn't the remotest clue.

1. Taking a subset that you know to be up-regulated and reanalysing that from scratch should yield few if any significant results. The actual changes will get normalized out.

2. Even if (1) weren't the case, doing what's suggested would be statistically impermissible and frankly a breach of scientific ethics. Any attempt to publish such results should result in immediate rejection. You would be well advised to never believe any results produced by the person that suggested this.

[mnoon]

Thanks so much Dpryan for shedding light on the topic. I am sorry for not being clear earlier that I took subset of up-regulated (4115) genes from one comparison and ran DESeq2 for different samples and/or conditions (not for the same comparison/data matrix). Actually the idea was not to violate science ethics but taking a subset of 4000 genes (which are genes of interest) was an attempt to get a reasonable/doable set of targets for further validation.
I was a little afraid of losing some effects/changes. So you are saying it is absolutely a wrong approach to take subset of counts-matrix for differential expression or is it a matter of losing some statistical power??