0
$\begingroup$

In order to perform an integration analysis of single cell RNA seq data, is it important to check the percentage expression of mitochondrial genes of cells as well as the feature counts to exclude dying or dead cells as well as multiplets from further analyses?

I ask because the standard Seurat work flow to perform integration as outlined in Integration and Label Transfer does not mention filtering steps prior to integration.

In this publication, Eleanor Alenzi et al. 2018 also did a canonical correlation analysis with Seurat but did not mention steps to filter out poor quality cells before doing the integrated analysis. This second publication does same. Both publications do not mention if their visualisations or differential gene expression analyses are done on integrated (discouraged) or raw uncorrected RNA data (recommended).

Following the styles of the publications linked above (many more do same), is it also efficient to perform integration without filtering out poor quality cells?

$\endgroup$
2
$\begingroup$

It is absolutely necessary to remove low quality cells: In the case of CCA (and this applies to other "integration" or "data alignment" methods as well), one would need to use "anchors", basically same type/state/kind of cells from the samples to be "integrated" and are used to "align" the different samples. In a scenario where you have a set of cells whose distinguished feature is high mitochondrial content due to apoptosis irrespective of their origin (epithelial, T cell, ...) and such cells are picked up as so called anchors, the whole integration effort will be off. Even when not using integration methods, bad quality cells quite often form their own clusters (highly skewed biology) not contributing anything to analyses and are often removed at the downstream steps anyway.

And when it comes to published work, unfortunately it comes down to the backgrounds of the reviewers. I tend to think that there are way too many single cell studies than experts with required bioinformatics and statistics so that a thorough review can be conducted. And lack of a paragraph on QC does not mean no QC has been performed anyway.

$\endgroup$
2
  • $\begingroup$ I usually run the usual standard Seurat - Guided Clustering Tutorial protocol to filter out poor quality cells (cells with high mitochondrial content and as well as multiplets). In the papers cell filtering as a quality control process is not mentioned at all. If I understood you well, does it mean that the presence of dead cells has no effect on the clustering process or clustering outcome? $\endgroup$ – Charles May 12 '20 at 7:51
  • 1
    $\begingroup$ Depending on what makes the cells bad quality and how different these bad quality cells from the rest, the overall clustering pattern might be effected (highly or partially) or not. I would nevertheless remove the bad cells and also would use SCTransform (or its scran counterpart) over crude log normalization in the downstream. $\endgroup$ – haci May 12 '20 at 8:09

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.