Great tool; doesn't catch -BrdU cells

[vincent-hanlon]

First off, I was very happy to see this automated QC tool published, and I now use it for the Strand-seq libraries we produce in Vancouver. By my own testing, it accords with my manual QC ~95% of the time, and it incorrectly marks libraries that I consider "good" as "poor" only ~0.5% of the time, which is great. Thank you!

With a recent round of libraries I noticed something a bit odd. We intentionally sequenced a mix of +BrdU and -BrdU cells, and I applied ASHLEYS to all of them. For some reason, ASHLEYS marked quite a few of the -BrdU cells as good-quality Strand-seq libraries, even though all chromosomes had WC strand state (see image for an example; each row is a cell). Even if Strand-seq libraries are made with BrdU, some cells may be sequenced that don't incorporate BrdU, and then this tool might call them good quality. Such libraries could be excluded using separate filters, but it would be nice if ASHLEYS did all the QC in one pass.

Here's the code I used to run ASHLEYS:

ashleys.py -j 12 features -f /path/to/dir/ -w 5000000 2000000 1000000 800000 600000 400000 200000 -o features.tsv
ashleys.py predict -p features.tsv -o quality.txt -m /path/to/ASHLEYS/models/svc_default.pkl

[ptrebert]

Hi Vincent,
I am excited to hear that ASHLEYS has found your interest. We are in the process of preparing "ASHLEYS 2.0", and will take your input into account (and may even approach you about some test data, if you can share them).

However, one remark about the table: the probabilities above 0.5 (with the exception of 75.8%) are all in the "uncertainty" margin, which is between 0.3 and 0.7 by default if I recall correctly. Libraries falling in that probability range should be dumped in a separate file (_critical.tsv) and it is strongly recommended that these libraries are examined by a human expert to determine their quality.
The current version of ASHLEYS is not yet as much "fire and forget" as we would like, but the next release will hopefully be closer to that goal.
Thanks a lot for your input!

[vincent-hanlon]

Thanks Peter, good to know that I should be manually QCing libraries in the 0.3-0.7 range. I'd intended to do that for sensitive applications, but the raw ASHLEYS scores are definitely a good first pass for evaluating whether a library prep experiment worked well in general. I imagine I could also share some more libraries for testing, as well, although I'd have to ask my PI.

Some more details on the -BrdU libraries: as you said, only ~4 of them scored 0.7 or higher.