DESeq: question about with replicates and without any replicates.

[nb509]

hi
I have 4 samples, A1,A2,B1,B2.
two conditions:A,B. each conditions have two replicates.

when I feed the reads count to the DESeq with replicates and without any replicates, the results vary greatly.

1.With replicates(A1A2-VS-B1B2), I get about 12000 DEGs(padj<0.1).
2.Without any replicates(A1-VS-B1), I get about only 200 DEGs(padj<0.1).

Both of the MA-plot are attached.

When working without any replicates, it sames that many informations are lossed, even the fold change is very high. Such as the gene "23416":

"id"	"baseMean"	"baseMeanA"	"baseMeanB"	"foldChange"	"log2FoldChange"	"pval"	"padj"
"without replicates"	292.791254092328	9.3250242940964	576.25748389056	61.7968882135121	5.94946228783961	0.0039474188223427	0.179640885315504
"with replicates" 285.324742802128        12.0227636114328        558.626721992823        46.4640859661922        5.53804412255897        5.31049903576043e-59    6.44978249349437e-58

how to explain this?

PS:

head(countsTable)
gene	A1	A2	B1	B2
57573	104	67	233	224
8563	357	346	110	104
8434	88	94	94	120
10309	33	23	11	14
7652	17	16	1	0

#With replicates
countsTable <- read.delim("A1A2B1B2",header=TRUE, stringsAsFactors=TRUE, row.names="gene")
conds <- factor(c( "A", "A","B","B"))
cds <- newCountDataSet( countsTable, conds )
cds <- estimateSizeFactors( cds, locfunc=shorth )
cds <- estimateDispersions( cds )
res <- nbinomTest( cds, "A", "B" )

#Without any replicates
cds11 <- cds[,c("A1","B1")]
cds11 <- estimateDispersions(cds,method="blind", fitType ="local",sharingMode="fit-only") #is right here?
res112 <- nbinomTest(cds112,"A","B")

[nb509]

The two images were uploaded.

http://www.flickr.com/photos/69065766@N06/6279366385/
http://www.flickr.com/photos/69065766@N06/6279366363/

[Simon Anders]

I'm not sure what your question is.

Of course you hardly get any usable results without replicates. This is, after all, why it is so important to have them.