Question

Strategy to perform DESeq, DEXSeq workflow

0

Entering edit mode

rohitsatyam102 ▴ 20

@rohitsatyam102-24390

Last seen 6 weeks ago

India

Hi Michael Love

I have RNASeq data for different layers of tissues that are stratified over one another in the same order. The library is SE unstranded data. I wish to perform DE and DEU RNAseq analysis to find out some novel patterns if present. (note: The library size for each tissue differs drastically from one another (even between replicates).

I obtained both genomic and transcriptomic bams using STAR and now I ran DESeq in the following way:

Samples	group_time	group	Time
ALrep1_16daf	AL_16_daf	AL	16_daf
ALrep2_16daf	AL_16_daf	AL	16_daf
ALrep3_16daf	AL_16_daf	AL	16_daf
CCrep1_16daf	CC_16_daf	CC	16_daf
CCrep2_4daf	CC_4_daf	CC	4_daf
NErep1_16daf	NE_16_daf	NE	16_daf
NErep1_8daf	NE_8_daf	NE	8_daf
NErep2_16daf	NE_16_daf	NE	16_daf
NErep2_4daf	NE_4_daf	NE	4_daf
NErep2_8daf	NE_8_daf	NE	8_daf

I wish to perform an all vs all comparison and though I know it's straight forward using contrast, I want to understand which is a valid approach:

Approach 1: Performing DESeq analysis on two samples at a time, say AL_16_daf.vs.NE_16_daf, which can be done by subsetting the above sample table for the replicates of these two samples, followed by DESeq analysis. This can be repeated for all pairwise comparisons that are desired.

Approach 2 The second approach could be the one described by Kevin here and by you here, where we take all the samples together, calculate the dispersion, and then using contrasts as described by Kevin can fish out results for each pairwise comparison.

On the same note, in DEXSeq, however, you can't perform such all vs all comparison using contrasts as mentioned here and here(also for DESeq) and therefore have to resort to something like I mentioned in Approach1 taking a pair at a time.

Having said that I have few queries:

So, to maintain uniformity in the analysis (both DESeq and DEXSeq), should I use Approach1 (wherein we estimate size factors and dispersion in a similar way for both DESeq and DEXseq, taking two tissues at a time and looping this over using, say, a FOR loop for all pairwise comparison). OR use Approach2 for DESeq2 and Approach1 for DEXSeq
What repercussions do you think one will have having used Approach1 over Approach2 and which one do you suggest. I did compare the list of DE genes generated by two approaches. All I could infer is the overlap between the DE genes obtained from both approaches is ~36% (based on two pairs that I compared ) and in some cases Approach 1 gives more DE genes and vice versa.

DEXSeq DESeq2 • 1.9k views

ADD COMMENT • link updated 3.9 years ago by Michael Love 43k • written 4.1 years ago by rohitsatyam102 ▴ 20

score 3 · Accepted Answer · 2020-11-06

3

Entering edit mode

Michael Love 43k

@mikelove

Last seen 5 days ago

United States

What does the PCA plot look like? This is usually our guiding plot for motivating "all together" or "split by groups".

ADD COMMENT • link 4.1 years ago Michael Love 43k

0

Entering edit mode

Hi Michael Love

So here are the pca plots for two design formulas:

## Accounting for differences due to different cell types and time of sampling. Couldn't add interaction term Sampletype:daf
# since it throws the following error: Error in checkFullRank(modelMatrix) : the model matrix is not full rank, so the model cannot be fit as specified.

ddsHTSeq_combined <- DESeqDataSetFromHTSeqCount(sampleTable = sampleTable, directory = ".", design = ~ Sampletype + daf)
keep <- rowSums(counts(ddsHTSeq_combined)) > 10
ddsHTSeq <- ddsHTSeq_combined[keep,]
dds <- DESeq(ddsHTSeq)
vsd <- vst(ddsHTSeq, blind = FALSE)

plotPCA(vsd, intgroup = c("Sampletype")) enter image description here plotPCA(vsd, intgroup = c("Sampletype", "daf")) plotPCA(vsd, intgroup = c("daf"))

<h2>Hardcoding the sample type and days information</h2>

Because I faced the error (mentioned below) when I tried introducing the interaction term Sampletype:daf, I separately ran DESeq2 for this condition to check if such grouping explained more variance than the above design formula:

Error in checkFullRank(modelMatrix) : 
  the model matrix is not full rank, so the model cannot be fit as specified.
  Levels or combinations of levels without any samples have resulted in
  column(s) of zeros in the model matrix.

  Please read the vignette section 'Model matrix not full rank':

  vignette('DESeq2')
In addition: Warning message:
In DESeqDataSet(se, design = design, ignoreRank) :
  some variables in design formula are characters, converting to factors

ddsHTSeq_combined <- DESeqDataSetFromHTSeqCount(sampleTable = sampleTable, directory = ".", design = ~ group)
keep <- rowSums(counts(ddsHTSeq_combined)) > 10
ddsHTSeq <- ddsHTSeq_combined[keep,]
dds <- DESeq(ddsHTSeq)
vsd <- vst(ddsHTSeq, blind = FALSE)

plotPCA(vsd2, intgroup = c("group")) enter image description here

ADD REPLY • link 4.1 years ago rohitsatyam102 ▴ 20

1

Entering edit mode

I would run these samples altogether I think.

Why do you say you can't use such models with DEXSeq?

"you can't perform such all vs all comparison using contrasts"

ADD REPLY • link 4.1 years ago Michael Love 43k

0

Entering edit mode

So Unlike DESeq wherein one can simply define contrast while using the results function (you can do any sample vs any sample comparison), in DEXSeq, DESeqResults() function does not have that functionality (it compares all samples against the sample that comes first lexicographically i.e. all against one comparison)

ADD REPLY • link 4.1 years ago rohitsatyam102 ▴ 20

1

Entering edit mode

I see (I'm not a DEXSeq developer so I didn't realize the interface. I think you would just use pair-wise for DEXSeq, while the contrast argument for DESeq2, unless there was a way to arrange it so it is an LRT within DEXSeq, but that's not obvious to me now.

ADD REPLY • link 4.1 years ago Michael Love 43k

0

Entering edit mode

Hi Michael,

Can you elaborate on your comment in regards to using a PCA plot as guidance for a "all together" vs. "split by groups" approach? Thank you in advance.

ADD REPLY • link 4.0 years ago Jonathan Giacomini • 0

1

Entering edit mode

See the FAQ in the vignette which elaborates on this, feel free to followup with questions as you need.

ADD REPLY • link 3.9 years ago Michael Love 43k