0
$\begingroup$

I have RNAseq data from 4 samples with 3 biological replicates per sample. I am currently trying to do the differential expression analysis with DESeq2 but the biological replicates will not cluster together when I make the PCA plot or correlation heatmap. This is my first time with RNASeq analysis and so am not sure what the best route forward is? I would like to avoid repeating the experiment with new samples if possible!

My pipeline prior to DESeq2 was the following:

FastQC quality check -> Trimmomatic -> Kallisto

I used tximport to convert kallisto files into suitable format for DESeq2

PCA plot with rlog transformed data

$\endgroup$
  • 1
    $\begingroup$ Why do you say they "don't cluster by replicate"? On the PCA, most colours are grouped together, no? Anyway, batch effects among biological replicates are common- there are many methods to correct for batch effects of course $\endgroup$ – Chris_Rands Jul 23 at 15:12
  • $\begingroup$ Did you use plotPCA from DESeq2 for this? Note that it takes the 500 most variable genes. If that's the case, I suppose midgut and SG are two different tissues and the most variable genes are dominated by DEGs in this context $\endgroup$ – StupidWolf Aug 24 at 10:24
  • $\begingroup$ Some of your replicates do cluster together (e.g M_wt_SG) so depending on the question you want to answer, it might be ok. I suggest running DESeq2 first, and see whether you can detect differentially expressed genes for the comparison you are interested in. You can then look at those counts and see whether the variation is too large to be trusted $\endgroup$ – StupidWolf Aug 24 at 10:25
0
$\begingroup$

PC1 is 81% of the variance?

This PCA plot confirms that different tissues are different. You probably already knew that. I'd make more PCA plots that are tissue specific. That will be more informative than these.

Personally, I'd also not do DESeq on all of these samples together, unless the goal of your experiment really is to learn the differences between these two tissues.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks, I tried again after separating the tissues into two different plots but these showed very little clustering at all. My comparison is between genotypes so made a lot of sense to separate them in the end but still can't manage to make them group together $\endgroup$ – nmp116 Jul 27 at 11:00
  • $\begingroup$ Your samples are what they are. You can't make them be more uniform than they are. $\endgroup$ – swbarnes2 Jul 27 at 15:35
0
$\begingroup$

Clustering of replicates looks decent enough to me, so you should be abl to push ahead, but I agree the tissues are grouping, which could mask any differences based on sex or genotype.

You might consider the EdgeR package for DE analysis here. It allows for flexibility when making complex comparisons while accounting for tissue/batch effects. I've had good luck with using it to compare across batch effects from complex experiments like this.

| improve this answer | |
$\endgroup$
0
$\begingroup$

You used Kallisto for alignment. I think Kallisto reports TPM values, are you using this value? DEseq2 uses count data, so I am not sure whether these two methods are compatible.

Also, I agree with previous answers that your PCA actually looks OK. One possible way to improve is to choose top variable genes. For example, you can try top 3,000, 5,000, 7,000 genes and so on. The idea is that for the genes that do not show much variation between samples, including them in PCA may just introduce noise. You can also try to color samples in your PCA by some other variables, like batch, sequencing depth and so on to trouble-shooting.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Yes I used the raw counts instead of the TPM to create the DESeq2 tables using tximport. I will try using the most variable genes as you suggested thanks! $\endgroup$ – nmp116 Aug 24 at 16:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.