To celebrate the release of GATK 4.0, we are giving away free credits for running the GATK4 Best Practices pipelines in FireCloud, our secure online analysis portal. It’s first come first serve, so sign up now to claim your free credits worth $250. Sponsored by Google Cloud. Learn more at

How to make results from GAT4.beta.3 "HaplotypeCaller" comparable to GATK2.7's "UnifiedGenotyper"?

We have been using “UnifiedGenotyper” of GATK2.7 for SNV calling, with "EMIT_ALL_SITES" mode, which always generate great results. We recently learnt GATK4 is in-development, with UnifiedGenotyper discontinued & HaplotypeCaller recommended. We thus test performance of HaplotypeCaller on our data with “-ERC GVCF” (didn't include the GenotypeGVCFs step). We found the amount of SNV identified decreased dramatically, with ~80%-90% reduction, in comparison with what's found by “UnifiedGenotyper”.

Our samples are from single cells, shallow sequenced. Paired-reads are 150bp each. Reads are supposed to align with short regions of 30-200bp across human genome, thus 99% of genome won’t be covered with reads. We’re not interested in arbitrary SNV, and don’t have target region or any window; we only care for mapping SNV across our samples. Based on quite a few experiments analyzed with UnifiedGenotyper, we found that even with low coverage, the short regions we aligned to always have highly reproducible base calls, and we could always identify SNV within these regions. We usually process five single cell samples each time, thus most region should have identical sequence and thus only 1 or 2 major alleles.

To our understanding, the “HaplotypeCaller” call variant based on de-novo assembly of active regions; if there are large amount of missing data in surrounding regions of our 30-200bp alignments, will it result in failure of haplotype identification, and lead to failure of SNV calling? Is the algorithm required certain sample size to work well? Is there any “HaplotypeCaller” parameters or discovery mode we could use to serve SNV calling with our current experiment design, or at least bring SNV calling rate to a level comparable to what identified by “UnifiedGenotyper”?

Greatly appreciate your advice!


  • SheilaSheila Broad InstituteMember, Broadie, Moderator


    I don't think we can comment before you run GenotypeGVCFs. The GVCF is just an intermediate file not to be used in final analysis. What happens when you run GenotypeGVCFs on your GVCFs? The sensitivity should improve.


    P.S. Have a look at this article.

  • Geraldine_VdAuweraGeraldine_VdAuwera Cambridge, MAMember, Administrator, Broadie

    @cdsc HaplotypeCaller's expectations in terms of sequence depth and assembly are indeed a bit more stringent than UnifiedGenotyper's. Although Sheila is correct that we wouldn't perform the evaluation on the basis of the GVCF intermediate, it's not unlikely that the shallow/sparse coverage of your data might be causing HC some trouble. You may have some success mitigating this by relaxing the assembly parameters (including the min pruning parameter).

    That being said it's not clear to me what you mean by 80 to 90 % reduction -- if you mean you're losing 80 to 90 % of expected variants, then the problem goes deeper than that. Would you be able to post some screenshots of some typical regions where you expect to see variant calls but HC does not emit anything? And please post the corresponding GVCF records.

Sign In or Register to comment.