SCtransfrom before or after merging if I have to look at 4 different conditions

[AnupaSP]

Hi!

I just had a doubt regarding SCTransform and the order in which I should do my analysis.

I have samples from 4 different conditions I want to look at:

1. Non-diseased male (n=3)
2. Non-diseased female (n=6)
3. Diseased male (n=4)
4. Diseased female (n=5)

My end goal is to look at the DEG between diseased female vs male + non-diseased female vs male.

For this, should I do:

Option 1: QC on individual samples > merge all samples > run SCTransform > Run downstream analysis (such as PCA & Determine dimensionality, clustering etc)

OR

Option 2: QC on individual samples > run SCTransform on individual samples > merge > integrate> Run downstream analysis (such as PCA & Determine dimensionality, clustering etc)

OR

Option 3: QC on individual samples > merge the samples of group (so in total, there would be 4 groups: Non-diseased male, Non-diseased female, Diseased male, Diseased female) > run SCTransform on each group > merge > integrate > Run downstream analysis (such as PCA & Determine dimensionality, clustering etc)

for example, integrate using: 
obj.features <- SelectIntegrationFeatures(object.list = sample.list, nfeatures = 3000)
VariableFeatures(seurat[["SCT"]]) <- obj.features_

I am so sorry if this is a silly question! But really appreciate the help!

Have a great day!

[mccormickceci]

Hello,
My thought is that it might be most appropriate to run SCTransform on each of the datasets on their own before going through the steps for integration. I would then suggest that for running DE you might try pseudobulking (perhaps also across different cell types, condition, sample) and using DESeq2 for differential expression analysis.

[AnupaSP]

Hi!

Thanks a million for your answer! Just to clarify, do you mean SCTransform on each sample (eg patient 1, patient 2) or on each dataset (all samples from one paper)? I get quite confused when talking about datasets or samples (sorry!)

Also yes from all the papers I was reading, it looked like pseudo bulking + DESeq2 was the best option, since wilcox has been shown to have a lot of false positives.

Thank you again for replying! I really appreciate it!