Differential expression bias in Seurat

Question

[sleepystudent]

Hello everyone! I am dealing with a scRNA-seq dataset having cells (SKOV3 cell culture) in 4 different conditions (2 levels of treatment). There are some biases: one of the conditions has 10 times more cells than others, whereas 3 others had doublets which have been filtered out.

Basically, clustering gives me the same result as sample annotation. My primary interest was detecting marker genes between clusters and/or conditions. However, running FindAllMarkers gives me a table where for each cluster there is a strong bias towards either upregulated, either downregulated genes . I tried to sample the biggest condition so all of them would be of the same size, but it didn't fix my problem.

Is it a common problem with single-cell data? Is it possible to fix it? Any advices would be appreciated.

Thanks!!

Answer 1

[jared.andrews07]

This is like due at least in part due to double dipping resulting in deflated p-values, which you can read more about in the OSCA book. As such, it's often more effective and robust to focus on effect size metrics when identifying cluster markers.

Your large sample could also be biasing things, and you should think about whether integration makes any sense in this case. Normally, I'd recommend pseudobulking and traditional bulk RNA-seq methods for cross-condition comparisons, but you need replicates for that.