Minor bias against low coverage cells

[vincent-hanlon]

ASHLEYS QC tends to give score <0.3 for libraries with <200k aligned reads, even though they contain good-quality Strand-seq data according to my manual QC. To be clear, this is a relatively minor issue and I will continue to happily use ASHLEYS QC as is, but I will probably do manual QC from now on for low coverage libraries. Overall---thanks for making such a great QC tool!

To confirm that this issue wasn't a quirk of the particular library pool, I also looked at the 28 libraries with between 5k and 200k aligned reads from a completely separate library prep experiment. Half of them (14) scored <0.5 with ASHLEYS QC but looked perfectly fine to me.

As an aside, it could also be argued that we don't want libraries in the 100-200k range for aligned reads anyway (most of the errors below are for libraries in that range). I personally think they are valuable, because they still show SCEs and contribute reads towards inversion calls and phasing.

The command:

ashleys.py -j $1 features -f ./output/bam -w 5000000 2000000 1000000 800000 600000 400000 200000 -o ./output/bam/features.tsv
ashleys.py predict -p ./output/bam/features.tsv -o ./output/bam/quality.txt -m scripts/tools/svc_default.pkl

Manual vs automated QC for 79 libraries. I did this blind to the ASHLEYS QC scores on libraries with 75 bp reads:

Some example of disagreements:

ASHLEYS QC score	Manual QC	Aligned reads
27.72%	good	140946
2.53%	good	123526
2.63%	good	136970
2.07%	good	153016
9.91%	good	127418
2.37%	good	124374
14.97%	good	140564
4.82%	good	121376
11.46%	good	427612
20.49%	good	153288

BreakpointR plots for two example libraries that I thought were good but ASHLEYS QC did not: