Attention:
The frontline support team will be unavailable to answer questions on April 15th and 17th 2019. We will be back soon after. Thank you for your patience and we apologize for any inconvenience!

Is the HaplotypeCaller less cohort-dependent than UnifiedGenotyper?

JeremyLeipzigJeremyLeipzig Member
edited January 2015 in Ask the GATK team

Does the HaplotypeCaller method rely less on other samples to make calls (in the GenotypeGVCFs step) than the UnifiedGenotyper?

Answers

  • SheilaSheila Broad InstituteMember, Broadie, Moderator admin

    @JeremyLeipzig‌

    Hi,

    Haplotype Caller actually relies more on other samples' data to make calls than Unified Genotyper. Please read more about the joint genotyping here: https://www.broadinstitute.org/gatk/guide/article?id=4017

    -Sheila

  • JeremyLeipzigJeremyLeipzig Member
    edited January 2015

    Is there a way to run the GenotypeGVCFs tool in a way that makes no attempt to correct any calls coming from the HaplotypeCaller - basically to run it as a gVCF->VCF converter - for comparison's sake?

  • Geraldine_VdAuweraGeraldine_VdAuwera Cambridge, MAMember, Administrator, Broadie admin

    To elaborate slightly on Sheila's answer, the "reliance on other samples' data" (which we prefer to view as an empowerment rather than a dependence: you get good calls alone, but even better calls together) is essentially the same behavior between the HaplotypeCaller ref confidence/GVCF method and UnifiedGenotyper run in multisample mode. But the maximum potential extent of the empowerment is greater in the HC method because it enables you to use much larger cohorts. So it's not that either is more or less dependent on cohorts, but that HC can give you a bigger bang if you have a bigger cohort. (On top of making better indel calls whether you use a cohort or not, thanks to the reassembly step.)

    You can certainly run GenotypeGVCFs on a single-sample gVCF, in which case it will have no cohort information to use. That will mimic basic single-sample calling behavior. You may see marginal differences if you compare that to just HC run in basic mode on a single sample, but nothing that would make a real difference between the two callsets after filtering. If you try this, do let us know how it behaves in your hands of course.

Sign In or Register to comment.