Potential adverse effect of small chromosome coverage limit

[ljwharbers]

Hi Andrey,

Thanks for the continuous updates, it's much appreciated!

We were running the latest version of isoquant (3.11.1) and you have included a default of 1 million coverage for chrM, which solves a lot of the processing time issues that were present in the older versions.

However, we now run in to the issue that typically the % of MT genes is used in downstream scRNAseq analyses as a quality control. How I understand it, with the current implementation the %MT is artificially downsampled and thus not really reflective anymore of a cell quality.

Do you have any possible solution/guidelines for this? Is it for instance possible to perform isoform predicition (which if I understand correctly is the problematic step on small high coverage regions) on a subset of chromosomes (for example, human chr1-22XY) but keep the gene-level and/or (known) transcript-level quantification on all chromosomes?

Many thanks,
Luuk

[andrewprzh]

Dear @ljwharbers

Thanks for the bringing this up, I did anticipate that this strategy might not be suitable for all pipelines.

At the moment you set --max_coverage_small_chr -1, this will process small chromosomes normally. It may not necessary slow down the run, it's just more probable compared to other chromosomes. However, I've seen other loci slowing down the run because of extreme coverage.

Extreme coverage may affect both read assignment and transcript discovery, but the former is more typical.
You can also, for example, run IsoQuant with --max_coverage_small_chr -1 and --no_model_construction to fully quantify all chromosomes for your QC, and separately run transcript discovery with the default options.

I will think how this can be done in the future - I like the idea of downsampling reads only for the second stage.

Best
Andrey

[ljwharbers]

Thanks! We will run it with --max_coverage_small_chr -1 in our coming run and see how long it will take for chrM to go through. Will keep you updated!

I think if transcript discovery is the problematic one in the end, downsampling only for that step would be a good solution. If it is indeed read assignment I'm not sure what the best way to solve this would be.