Selecting methods for performing pairwise RNAseq analysis without replicates

Question

[bioinforesearchquestions]

Hi friends,

I need some help in deciding the tools and steps in the following analyses. I got 6 RNAseq samples, there are no replicates among the six samples. These six samples are from three different tissues. Three of them are wild-type (WT) samples and the remaining three of them are knock-out (KO) samples.

I checked with the coordinator, they don't have any replicates for the 6 samples. But he also mentioned that tissue1 and tissue2 are somewhat similar.

I have used the following tools for this pairwise analysis

a) Tissue1_WT and Tissue2_WT vs Tissue1_KO and Tissue2_KO (i.e Tissue1_WT is replicate1 and Tissue2_WT is replicate2 for WT and similarly Tissue1_KO is replicate1 and Tissue2_KO is replicate2 for KO groups).

step 1) Trimmed using Trimmomatic

step 2) Aligned using HiSAT2

step 3) HTSeq count from the BAM files

step 4) Differential gene expression analysis using DESeq2

Steps 1-3 are the same for the below pair-wise analyses, but can I still use DESeq2 for samples without replicates?

2) Tissue1_WT vs Tissue1_KO

3) Tissue2_WT vs Tissue2_KO

4) Tissue3_WT vs Tissue3_KO

Any suggestions are appreciated.

Answer 1

[i.sudbery]

No. Those comparisons in particular are comparisons it is not valid to make without replicates.

You can make the WT vs KO comparison across tissues as long as you use tissue as a blocking variable (effectively doing something comparable to a paired t-test).

You can also ask if there is a genotype independent difference between tissues (do an ANNOVA like full model=~tissue+genotype, reduced model =~genotype). But I'm pretty sure you are not interested in this.

The final comparison is the interaction term. It may be possible to do this by calculating pooled variance across all the conditions. DESeq may or may not agree to do this, but such comparisons are often low powered even with 3 or 4 replicates. The changes of getting anything meaningful here is very slim.

[i.sudbery]

Of course you could also do as you suggested in a. That is valid, if, I guess, underpowered.

[bioinforesearchquestions]

Yeah, I already explained to them that pair-wise analyses (2, 3 and 4) without replicates we won't be able to identify DE genes. The p-value which we get is meaningless without replicates. But they are persistently asking me to provide a list of genes based on the log2 fold change values.

[i.sudbery]

Well, thats probably better than using meaningless p-values. Better to be honest about having no statisitcal confidence than to pretend to have one that is false. I guess if they want to do downstream validation on every gene you give them, so be it.

Perhaps to the best way to scare them would be to tell them that no reviewer would let it through.

[bioinforesearchquestions]

Thanks, Sudbery. For 2,3,4 pair-wise analyses, even If I do paired t-test for all ~45,000 genes, I will get a single p-value for each pair-wise. How do I get the list of DE genes.

[i.sudbery]

No, if you do a paired t-test you will not get something for each pair-wise, you'll get just a single p-value.

But I'm not suggeting you actually do a paired t-test, I'm suggsted you do a paired t-test-like DESeq2 test.

your design would be ~tissue + genotype, and the contrast without be just on the genotype coefficient. Thus you would effectively be doing a single test of WT for Mut with 3 replicates in each category, but removing tissue specific differences.

[bioinforesearchquestions]

Sure, I will do it as you recommended for tissue 1 and tissue2. Am I correct?

Metadata

Sample Tissue Genotype

Sample1 Tissue1 Wildtype

Sample2 Tissue2 Wildtype

Sample3 Tissue1 Knockout

Sample4 Tissue2 Knockout

Blockquote

[i.sudbery]

NO real reason you could put tissue 3 in as well, unless you think that the differences will be in a different direction in tissue 3.