Why using too extensive gene sets (e.g. from bulk RNA-seq) can lead to artifacts

[halterc]

Related to question: https://support.bioconductor.org/p/9149618/
In summary:
A user saw that the nFeatures and cell number per sample seemed to confound his UCell scoring (see image).
The reason is likely that this may happen with huge gene sets (>1000 genes), where the # of detected genes can affect ranks just by chance.

The solution:

• Avoid using too extensive gene sets (e.g. gene-set enrichment analyses from bulk RNA-seq) and lowly/sparsly expressed genes because they can lead to artifacts when applying them to sparse scRNA-seq data.
• Use small but highly expressed gene sets that will be found across most cells and samples