Quality Control for single cell RNA-seq -- removing highly expressed cells

[icanwinwyz]

I am using Seurat for scRNA-seq analysis. In the QC step, I found they simply removed the cells with number of gene expressed over 2500 (http://satijalab.org/seurat/pbmc-tutorial.html). Also, when I read this nature paper (Kelley S. Yan et, al 2017), they simply set the cutoff to remove the cells with expressed gene more than 4400 (http://www.nature.com/nature/journal...ture22313.html). I understand it is necessary to remove the cells with low number of genes expressed since they are likely to be dying cells. But I don't understand why peopel remove the cells with high number of genes expressed? It is very possible that these cells are some rare populations. Probably, people may claim the cells are doublets, but is there any evidence for that?

I am building up the QC pipeline for scRNA-seq and eager to know the answer for this. Many thanks in advance!

[nucacidhunter]

I can think of following evidence for massively parallel techniques capturing polyA transcripts such as 10x Chromium:

1- Observation that after clustering cells to different types (for instance in t-SNE plots) usually cells with high number of expressed genes cluster with others with low number of detected genes and do not form their own clusters
2- Correlation between high number of un-detached cells and also higher concentration of cells used for capture and detection of cells with high number of genes
3- The fact that reverses transcription efficiency is low (around 20%) and converting estimated polyA RNA content of cells does not add up

QC pipeline should be able to distinguish differences among techniques. Some such as Smart(er)-seq and Fluidigm target whole transcript body but others such as Drop-seq, Cell-seq and 10x target only 3' end of transcripts.

Edit: you can also check human-mouse mix experiment data sets from 10x or Drop-seq. I would expect to see more detected transcripts from droplets containing both cell types in comparison to ones containing only one cell type