test gVCF merging on Dragen 4.3

[marc-sturm]

• test command:

  dragen --enable-gvcf-genotyper-iterative true --ht-reference …/GRCh38.fa --variant …/NA12878x2_93_dragen.gvcf.gz --variant …/NA12891_03_dragen.gvcf.gz --variant …/NA12892_03_dragen.gvcf.gz --output-directory dragen_gvcf_merge_on_dragen/ --gg-remove-nonref true

  However, AF, DP and MQ are missing in the output with Dragen 4.2.

  Dragen 4.3 should have an option to add them with --gg-msvcf-format-fields.

• also test why the ML model gives worse results for the Twist exom kit in our benchmarks and not when Illumina does it:

  Please find below Faezeh’s analysis:

  We ran the same DRAGEN cmd as you provided us, and processed the resulting VCF through our benchmarking tools.
  We used dragen 4.2.4 (default ML model), and rtg tools 3.12.1 (latest version).
  The rtg command line to run the vcf comparison is the following:

  rtg vcfeval --ref-overlap -b /staging/tmp/suite_def/prod/10362103/002_MLon/03_SIMPLE_vcf_compare/vcf_compare_output_vcf1_vs_truth_set_vcf__high_confidence_v4.2.1/vcf-compare-tmp/HG001_GRCh38_1_22_draft_3_v4.2.1_benchmark.vcf.gz -c /staging/tmp/suite_def/prod/10362103/002_MLon/03_SIMPLE_vcf_compare/vcf_compare_output_vcf1_vs_truth_set_vcf__high_confidence_v4.2.1/vcf-compare-tmp/Tuebingen.vcf.gz -t /staging/tmp/suite_def/prod/10362103/002_MLon/03_SIMPLE_vcf_compare/vcf_compare_output_vcf1_vs_truth_set_vcf__high_confidence_v4.2.1/vcf-compare-tmp/temp_SDF -o /staging/tmp/suite_def/prod/10362103/002_MLon/03_SIMPLE_vcf_compare/vcf_compare_output_vcf1_vs_truth_set_vcf__high_confidence_v4.2.1/vcf-compare-results --output-mode annotate --vcf-score-field QUAL --sample HG001,NA12878x2_80 --bed-regions /staging/tmp/suite_def/prod/10362103/002_MLon/03_SIMPLE_vcf_compare/vcf_compare_output_vcf1_vs_truth_set_vcf__high_confidence_v4.2.1/vcf-compare-tmp/bed_used.bed.gz.gz

  Here are the results as well as the ROC curves for SNP and indels:
  As you can see from the ROC, ML enabled provides a significant improvement in precision (curve gets shifted to the left).
  We then show in the table comparisons at the end point (i.e., using the default QUAL threshold cut off) and at the best Fmeas point.
  The comparison does take genotypes into account.
  Hethom counts for genotyping errors and vardiff counts for variant allele errors.

  Overall our results show that ML ON provides improvement over ML OFF.