I am running a notebook with example for the MAGIC algorithm.

In the data preprocessing step, a filtering operation is required to filter out cells with a small count of transcripts.

My question is twofold:

  • Why filtering out cells with a small count of transcripts?
  • How to decide the cutoff, i.e. the minimum number of transcripts each cell should have to not be removed from the dataset?

I attach some plots from the original notebook, which display some distributions of the single-cell rnaseq dataset.


  • $\begingroup$ In my experience it's helpful to explore the data (ie. make some heatmaps etc) and look for trends that are associated with low molecule counts. For example, cells with low umi's may look weird or cluster separately. You can use these to guide your cutoff. $\endgroup$
    – GWW
    Dec 20, 2017 at 20:48
  • $\begingroup$ @GWW can you elaborate more on that? Do you mean for example to plot the two principal components and see where the different cells cluster? And maybe if I have a cluster composed by only cells with limited library size, exclude that cluster? Otherwise, which plots will you use? Thanks $\endgroup$
    – gc5
    Dec 20, 2017 at 21:18
  • $\begingroup$ I suggest a heatmap of the most variable genes. PCA may be harder to see. $\endgroup$
    – GWW
    Dec 20, 2017 at 21:29

1 Answer 1


By reading this tutorial:

Wells with few reads/molecules are likely to have been broken or failed to capture a cell, and should thus be removed.

This answers the first part of the question.

For the second part, i.e. how to decide cutoff, it still seems a bit arbitrary to me, but some authors trim the cells on the left tail of the highest mode, in order to remove other modes (likely due to noise) on the left tail, considering the distribution of cells dependent on the sum of reads (molecules per cell, first plot).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.