Correct, and I agree -- reads lying fully within introns are usually not counted for either single-cell or bulk.

Inclusion of introns in a transcriptome "index" is what improves accuracy (but this is because mapping accuracy is improved). Reads lying fully within introns are usually only used in single-cell for things like RNA velocity analysis and splicing analysis -- not for typical quantification.

For nuclear preps, purely intronic reads are quantified simply because what we're interested in is the nascent (unspliced) RNA molecules which is mostly what's present in the nucleus.

swbarnes2 and dsull:

Thanks for the concern. I know that counting intronic reads in bulk RNA-seq is rather unorthodox and it's something I've never done before. In this usage case, however, I do have some justifications:

1. I'm looking at whether the gene in question (a human GENCODE gene) is expressed in a tissue type or not, without any intent for expression analyses.
2. Alternatively, I understand that intronic reads in bulk RNA-seq can potentially be genomic DNA contamination. However, I'm also looking at the capture rates of this kind of reads as well.

That said, this put the issue of 10X decision of by-default inclusion intronic reads in read matrix since Cell Ranger 7 into question, since most scRNA-seq are performed on whole cells rather than nuclei. I suppose their assumption is that in a microfluidics platform, the gDNA contamination should not affect the total count per barcode too much?