Filter rarely expressed genes/features?

[kloot]

Question
... and prompt for discussion ;-)

Is it a good idea to QC features (genes) that are detected in only a small number of barcodes (cells)- i.e. features rarely detected across the population?

I can't recall seeing this QC'd in typical publications; however these features could complicate highly variable gene lists etc.
Their number is not small - the dataset in front of me (1,040 barcodes x 23,700 features; filtered according to the principles in chapter 6, plus others) contains 5,600 features that were detected in less than 10 barcodes.

If their QC is a good idea, we'd have to remove (or set to zero) these features from the count matrices (and possible re-normalize) - or can someone think of a better way?

Note this is different from filtering barcodes with very few features.

Keen to hear everyone's thoughts on the matter - thanks!
Please let me know if there's a better mechanism for discussion