I have recently come across some false positive variant calls, for which visual inspection revealed obvious strand bias. I looked at the FS and SOR scores for these variants but they do not seem to provide clear reason to filter, when compared with FS and SOR scores for other variants that passed our filters. I can't seem to find a straightforward guideline for interpreting these scores, either. Does GATK provide any such guideline?
The only useful differentiation I found between them was that SOR "is better at taking into account large amounts of data in high coverage situations" but doesn't define what the threshold might be to define "high coverage."
I can post more details if necessary (e.g. screen shots) but for now, here are the FS and SOR scores for two groups of variants, true positives and false positives, in our data:
True positives (FS, SOR):
0, 0.169
0, 0.705
0, 2.211
0, 2.303
5.65, 11.106
17.459, 14.404
False positives (FS, SOR):
1.307, 0.509
1.876, 0.33
Is it relevant that HaplotypeCaller produces one FS score and one SOR score for a variant, regardless of how many samples the variant is called in? We are running on 30+ samples, so having one score to determine the strand bias across all samples seems like it could lead to errors in variant calls.