DEXSeq multiple time points

Question

[clausndh]

Dear all,

we try to analyze a time series with DEXSeq (DEXSeq_1.12.1). We have 4 time points and at each time point we have 3 replicates, so in total 12 samples.

But we don't know how to design the fullModel and the reducedModel?

Or with other words, we want do similar comparisons as with the --time-series flag of cuffdiff.

Thanks for your help.

tps<-c("TP5","TP10","TP15","TP20")
sampleTable = data.frame(
      row.names =  c( paste(sep=".",tps[1],c(1,2,3)),  paste(sep=".",tps[2],c(1,2,3)) ,
                      paste(sep=".",tps[3],c(1,2,3)),  paste(sep=".",tps[4],c(1,2,3)) ),
      condition = c(rep(tps[1],3),rep(tps[2],3),rep(tps[3],3),rep(tps[4],3) )
    )

dxd2<-DEXSeq::DEXSeqDataSetFromHTSeq(
      countfiles=countfiles, sampleData=sampleTable,
      design= ~ sample + exon + condition:exon, flattenedfile=flattenedfile )

BPPARAM <- MulticoreParam(workers = 32)
    dxd2 = DEXSeq::estimateSizeFactors( dxd2 )
    dxd2 = DEXSeq::estimateDispersions( dxd2, BPPARAM=BPPARAM)
    dxd2 = DEXSeq::testForDEU( dxd2, BPPARAM=BPPARAM)
    dxd2 = DEXSeq::estimateExonFoldChanges(dxd2, BPPARAM=BPPARAM,)
    dxr2 = DEXSeq::DEXSeqResults( dxd2 )

[Devon Ryan]

Do you want to compare the individual time points against each other or just see if time has an effect? In the latter case, the reduced model is ~1. In the former case, it'd make more sense to uses contrasts (not sure if DEXseq supports that), though you can always subset (this isn't ideal, but if it's the only option...).

[clausndh]

Thanks for your fast reply.

1. We want to compare the time points. Currently, we compare:

   T5 vs T10, T10 vs T15,and T15 vs T20.

   Here, we make the three comparisons individual ( with individual normalization), but do you think it would be better to make an "overall" normalization ( calculating the library sizes ones) and afterwards the three comparisons?

2. How should we define the contrasts if we decide to use DESeq or edgeR for detecting differentially expressed genes?

[Devon Ryan]

1. In theory, fitting the entire model at once should provide better results.

2. Assuming a base level of T5, then T5 vs anything doesn't need a contrast, T10 vs T15 would be c(0,1,-1,0) and T15 vs T20 would be c(0,0,1,-1). In edgeR you wouldn't specify an intercept (~0+time), in deseq2 you can leave it in.