gatk4-germline-snps-indels

Purpose :

Workflows for germline short variant discovery with GATK4.

haplotypecaller-gvcf-gatk :

The haplotypecaller-gvcf-gatk4 workflow runs HaplotypeCaller from GATK4 in GVCF mode on a single sample according to the GATK Best Practices (June 2016), scattered across intervals.

Requirements/expectations

• One analysis-ready BAM file for a single sample (as identified in RG:SM)
• Set of variant calling intervals lists for the scatter, provided in a file

Outputs

• One GVCF file and its index

joint-discovery-gatk :

The second WDL implements the joint discovery and VQSR filtering portion of the GATK Best Practices (June 2016) for germline SNP and Indel discovery in human whole-genome sequencing (WGS) and exome sequencing data.

NOTE: joint-discovery-gatk4-local.wdl is a slightly modified version of the original to support users interested in running the workflow locally.

Requirements/expectations

• One or more GVCFs produced by HaplotypeCaller in GVCF mode
• Bare minimum 1 WGS sample or 30 Exome samples. Gene panels are not supported.
• When deteriming disk size in the json, use the guideline below
  • small_disk = (num_gvcfs / 10) + 10
  • medium_disk = (num_gvcfs * 15) + 10
  • huge_disk = num_gvcfs + 10

Outputs

• A VCF file and its index, filtered using variant quality score recalibration
  (VQSR) with genotypes for all samples present in the input VCF. All sites that
  are present in the input VCF are retained; filtered sites are annotated as such
  in the FILTER field.

Software version requirements :

• GATK 4 or later
• Samtools (see gotc docker)
• Python 2.7

Cromwell version support

• Successfully tested on v31
• Does not work on versions < v23 due to output syntax

IMPORTANT NOTE :

• Runtime parameters are optimized for Broad's Google Cloud Platform implementation.
• VQSR wiring. The SNP and INDEL models are built in parallel, but then the corresponding recalibrations are applied in series. Because the INDEL model is generally ready first (because there are fewer indels than SNPs) we set INDEL recalibration to be applied first to the input VCF, while the SNP model is still being built. By the time the SNP model is available, the indel-recalibrated file is available to serve as input to apply the SNP recalibration. If we did it the other way around, we would have to wait until the SNP recal file was available despite the INDEL recal file being there already, then apply SNP recalibration, then apply INDEL recalibration. This would lead to a longer wall clock time for complete workflow execution. Wiring the INDEL recalibration to be applied first solves the problem.
• The current version of the posted "Generic germline short variant joint genotyping" is derived from the Broad production version of the workflow, which was adapted for large WGS callsets of up to 20K samples. We believe the results of this workflow run on a single WGS sample are equally accurate, but there may be some shortcomings when the workflow is modified and run on small cohorts. Specifically, modifying the SNP ApplyRecalibration step for higher specificity may not be effective. The user can verify if this is an issue by consulting the gathered SNP tranches file. If the listed truthSensitivity in the rightmost column is not well matched to the targetTruthSensitivity in the leftmost column, then requesting that targetTruthSensitivity from ApplyVQSR will not use an accurate filtering threshold. This workflow has not been tested on exomes.
  The dynamic scatter interval creating was optimized for genomes. The scattered SNP VariantRecalibration may fail because of two few "bad" variants to build the negative model. Also, apologies that the logging for SNP recalibration is overly verbose.