Hello All,

I am using DESeq2 to perform something like a one-way ANOVA. When I tried running the code on the datasets I have, it seems to be working fine for all except for one dataset. This dataset is from the organism fly while the others are either mouse or human.When I  run deseq2 with 'LRT' for the fly dataset I get the following result.Below is the code I used. Since there were hidden batch effects for the dataset, I have included it as co-variates in the design.

dds.fc<-DESeqDataSetFromMatrix(data,colData=pheno,design=~ condition)

# if hidden batch is present

#compute null and full model
mod =model.matrix(~condition, data=colData(dds.fc))
mod0 = model.matrix(~1,data=colData(dds.fc))

#Get normalized data
dds.norm <- dds.fc
dds.norm=estimateSizeFactors(dds.norm)
data_normalized <- counts(dds.norm, normalized=TRUE) #normalizing data

n.sv = num.sv(data_normalized,mod,method="be")

#as default remove 3 surrogate variables

rld<-rlogTransformation(dds.fc)

svobj = sva(assay(rld), mod, mod0, n.sv=3)

ddssva <- dds.fc
ddssva$SV1 <- svobj$sv[,1]
ddssva$SV2 <- svobj$sv[,2]
ddssva$SV3 <- svobj$sv[,3]
design(ddssva) <- ~ SV1 + SV2 + SV3 + condition

#execute if no batch effect   dds<-DESeq(dds.fc,test = "LRT",reduced = ~1)

#if hidden batch effect present then

dds<-DESeq(ddssva,test = "LRT",reduced = ~1)

res<-results(dds)

http://imgur.com/DYSR8LZ
 

Your last likelihood ratio test is testing whether any of the variables (SV1 + SV2 + SV3 + condition) are informative for describing differences in counts across samples, because you used a reduced design of ~1.

The surrogate variables should be included in the reduced design.

The typical use of an LRT would be, e.g., if you had full=~batch + condition, you would have reduced=~batch.