Filtering multisample vcf files on DP

[DavyK]

Hi All,
I have a multi-sample VCF file produced by the GATK Unified Genotyper. I need to now filter these variants for SNPs that have a DP < 10. However the DP entry in the info field for a multi-sample VCF is the depth across all samples. so very few variants will fail this filter and there will be many variants with low depth marked as a pass. Does anyone know of a GATK option to filter on the depth of the samples themselves. Of course, this presents another problem. How to resolve a variant that has low coverage in one individual and high coverage in another? The following is probably not possible in GATK but perhaps one could say if the largest depth across all samples is <10 or maybe if more than 10% of the samples have DP less than 10? Does it even make sense to do this?

Cheers,
Davy.

[jkerouac]

Yes it makes a lot of sense to do this. For example I have a modestly sized exome sequencing project of an extreme phenotype (50 and 50 samples). We filtered for highly deleterious variants, then compared what is in one group versus the other. What we found were a fair number false positives which fit a scenario where there was low coverage for this area (in general) but a few samples got up to a coverage depth (6-10 reads) where they were called. So there wasn't really a variant existing in one group that wasn't in the other, rather just stochastic calling of low coverage variants that gave a false positive. I think the slight unevenness of coverage from sample to sample in low coverage areas is a big problem.

I don't have an answer right now, I just started working on this problem today (hence my finding your question) but I will repost if I work a solution out. And if anyone else knows how to filter VCF files such that you only select variants that were at least "callable" in all or a defined proportion of samples, I would much appreciate it.

[DavyK]

Yes, filtering is clearly important, although you should always filter before comparing a two sample groups. Then you get into the issue of adjusting your filters based on what you see in a case vs control sample.

In any case, on more thorough reading of the GATK documentation website, filtering on READ depth is no longer recommended. Instead they suggest a number of filters that might (emphasis on might) help to rule out FPs.

For SNPs:

QD < 2.0
MQ < 40.0
FS > 60.0
HaplotypeScore > 13.0
MQRankSum < -12.5
ReadPosRankSum < -8.0

I added another filter though from the seqanswers exome sequencing analysis wiki

MQ0 >= 4 && ((MQ0 / (1.0 * DP)) > .01)

However your project sounds like it's adequately powered for you to run the variant quality score recalibration tool from the GATK. Whole-exome of more than 30 samples is stated as being the minimum, and it's shown to be better than hard filtering.

[jkerouac]

Thanks for the reply that is helpful.

Yes we used the VQSR tool, and by manual inspection of hundreds of calls it did a nice job. But it doesn't get around the false positive problem I described (which I thought you were describing also): that is, low coverage areas that vary in their ability to be called from one sample to the next.