Detecting outlier DESeq2

Question: Detecting outlier DESeq2

15 months ago by

maria2019 • 130

maria2019 • 130 wrote:

I have RNA-seq data for 20 samples with 2 condition and 2 sex (male, female : control, treatment). I am very new to RNA-seq analysis and am trying to find the DEGs using DESeq2. Since I want to have the normalization to be calculated based on all samples, I get the rld based on all and then I will use contrast to find the DEG for 1. tratment vs. control male 2. tratment vs. control female.

For the QC, the PCA plot separates the male and female but the control and treatment is not separated very well on the PC2 for male C1 (Figure 1) enter image description here

Then I decided plot PCA only for male samples. again PC1 does not separate the control and treatment male samples but PC2 is kinda separating them. (Figure2)

enter image description here

My questions are:

What numbers on the PCA plot (x and y axis) decide the separation? I read in a website that samples that are at PC1>0 are outlier. is that true?
Can I just look at figure 1 and remove male c1 and continue DEG analysis with 9 samples? or I should definately plot figure 2 ?
If I need to consider Figure 2, can I rely on only the PC2 which separates control and treatment and continue the DEG analysis or I should remove C1, tr4, tr5 samples and then work on DEG analysis based on remaining 7 samples?

bulk rna-seq deg pca deseq2 • 909 views

ADD COMMENT • link •

modified 15 months ago by Kevin Blighe ♦ 69k • written 15 months ago by maria2019 • 130

15 months ago by

Kevin Blighe ♦ 69k

Republic of Ireland

Kevin Blighe ♦ 69k wrote:

Hey,

Regarding the website to which you are referring, that was written by Wolfgang Huber. He is not saying, in his tutorial, that PC1>0 is a global definition of an outlier from a PCA bi-plot. It is just in his example that he has decided to select 2 outlier samples via a vertical cut-off line drawn at PC1>0. In another experiment, the cut-off for outliers may be at PC1 < -500, or PC1 > 42. The units on these axes are, well, essentially unitless.

Looking at your first plot, you cannot really remove male sample C1 just because it groups away from the main group of samples. You should first investigate why it may be grouping in the way that it is [grouping]. Is it a sample mix-up?; is it a male with Klinefelter Syndrome?; is the clinical information for this sample incorrect?

For your third question, you need to consider the fact that sex represents a substantial amount of variation in your dataset, i.e., it is greater than the very condition of interest that you are studying - was this expected?; did you include sex in your design formula as a covariate? If you include it as a covariate in your design formula, normalise your data again, and then run rld function with blind=FALSE, do you see the same PCA bi-plot?

You can also segregate your dataset based on male | female and then cross-compare results in a sort of meta-analysis, if you wish. sex is a clear and major source of variation among your samples. If you proceed with this route (sample segregation based on sex), then I see no reason to exclude c1.

Kevin

ADD COMMENT • link written 15 months ago by Kevin Blighe ♦ 69k

Hi Kevin,

Thanks for your response. I believe it is not a sample mix up (but I am gonna check it again). I do not know it is a male with Klinefelter Syndrome. How can I check it?

For the normalization, I actually made a 3rd column in my sample table and grouped together the sex and condition.

column group of sample table: control_male, treatment_male, control_female, treatment_female

then the dds object as:

dds <- DESeqDataSetFromHTSeqCount( sampleTable = sampleTable, directory = directory, design= ~group)

rld <- rlog(dds, blind=TRUE)

data <- plotPCA(rld, intgroup=c("treatment", "sex"), returnData=TRUE )
percentVar <- round(100 * attr(data, "percentVar"))
ggplot(data, aes(PC1, PC2, color=treatment, shape=sex)) +
xlab(paste0("PC1: ",percentVar[1],"% variance")) +
ylab(paste0("PC2: ",percentVar[2],"% variance")) + 
+coord_fixed()

Then I used contrasts for further analysis:

 contrast_male <- c("group", "treatment_male", "control_male")
 contrast_female <- c("group", "treatment_female", "control_female")

res_male_unshrunken <- results(dds, contrast = contrast_male, alpha = 0.05)
res_male <- lfcShrink(dds, contrast = contrast_male, res=res_male_unshrunken)
summary_res_male <- summary(res_male, alpha=0.05)

The result of the pca plot is from the above code. I think making the "group" column is including the sex in the design and is similar to ~sex+condition design. is that right? is the code above what you meant by including sex?

By segregating male | female, do you mean to run the analysis for male and female samples separately due to the major difference between the male/female samples?

ADD REPLY • link modified 15 months ago • written 15 months ago by maria2019 • 130

Ah, I see... sex is already included in group. When you look at the results of the differential expression comparison, do they make sense to you?

Out of curiosity, take a look at the output of this:

rld <- rlog(dds, blind = FALSE)
plotPCA(rld, intgroup=c("treatment", "sex"))

ADD REPLY • link written 15 months ago by Kevin Blighe ♦ 69k

The results actually make sense with the 10 samples but when I remove C1, I get 2times more DEG genes.

Changing to Blind=FALSE, I get the following PCA plot:

enter image description here C1,tr4,tr5 are kinda different compared to other samples.

ADD REPLY • link written 15 months ago by maria2019 • 130

Thanks - in that case, it just seems like a genuine result with regard to that sample (C1), unless there was indeed a mix-up. As I don't know your area of study, I cannot comment much further; however, occasionally, controls are not quite what they seem!

There is no right or wrong answer here. You can choose to leave it in your dataset or exclude it. Either way, you have to record what you do.

ADD REPLY • link modified 15 months ago • written 15 months ago by Kevin Blighe ♦ 69k

Thank you Kevin, I checked and it is not a sample mix up. I will continue with further analysis and see what makes sense more based on the results.

ADD REPLY • link written 15 months ago by maria2019 • 130

Similar posts • Search »