Variant Calling (GATK4) — Pipeline Skill (Thin Scaffold)

Overview

Adds decision logic for VQSR vs hard filters, WES vs WGS parameter differences, and GATK4-specific gotchas that LLMs frequently get wrong (filter thresholds, GVCF requirements, reference consistency).

Usage

• Activate when choosing between VQSR and hard filters for a given cohort size
• Activate when setting up WES vs WGS pipeline parameters
• Activate when running Mutect2 tumor/normal or tumor-only somatic calling

Core Concepts

Decision Logic

Germline calling strategy:
├── Single sample, no future joint calling → HaplotypeCaller direct VCF
└── Cohort (≥2 samples) → HaplotypeCaller -ERC GVCF → GenomicsDBImport → GenotypeGVCFs

Filtering strategy:
├── ≥30 WGS samples OR ≥30 WES exomes → VQSR
└── <30 samples OR single sample → Hard filters

Somatic calling:
├── Tumor + matched normal → Mutect2 with both BAMs
└── Tumor-only → Mutect2 + PoN (MANDATORY) + gnomAD germline resource

WES vs WGS parameter differences:

Critical Parameters

Hard filter thresholds (GATK Best Practices):

VQSR truth sensitivity levels:

• SNPs: 99.7
• Indels: 99.0
• Indel --max-gaussians 4 (fewer training variants than SNPs)

VQSR annotations: -an QD -an FS -an MQ -an MQRankSum -an ReadPosRankSum -an SOR

Mutect2 essentials:

• --germline-resource af-only-gnomad.hg38.vcf.gz
• --panel-of-normals pon.vcf.gz (essential for tumor-only)
• --f1r2-tar-gz → LearnReadOrientationModel (FFPE/OxoG artifacts)
• GetPileupSummaries + CalculateContamination before FilterMutectCalls

Common Mistakes

• Wrong: Aligning without proper @RG headers (missing ID, SM, PL, LB) Right: Always specify at alignment: -R '@RG\tID:x\tSM:x\tPL:ILLUMINA\tLB:x' Why: GATK refuses to run or silently merges samples when SM tags are wrong
• Wrong: Running HaplotypeCaller without -ERC GVCF for cohort analysis Right: Always produce GVCFs when joint genotyping will be performed Why: Regular VCFs cannot be joint-genotyped; must re-call from BAM
• Wrong: Reusing SNP hard-filter thresholds for indels (e.g., FS > 60 for indels) Right: Use FS > 200 for indels, FS > 60 for SNPs Why: Indels tolerate higher strand bias; SNP thresholds over-filter real indels
• Wrong: Applying VQSR to <30 samples Right: Use hard filters for small cohorts Why: VQSR needs many variants to train its Gaussian mixture model
• Wrong: Subsetting --known-sites VCFs to WES capture BED for BQSR Right: Use genome-wide known-sites; only restrict analysis intervals via -L Why: BQSR needs genome-wide known sites to model base quality errors
• Wrong: Skipping samtools index between GATK steps Right: Index after every BAM-producing step Why: GATK requires BAM indices; missing them causes immediate failure
• Wrong: Mixing reference files from different genome builds Right: Ensure ref.dict, ref.fa.fai, BWA index, and all VCFs match the same build Why: Mismatched dictionaries cause silent failures or cryptic errors
• Wrong: Running Mutect2 tumor-only without a panel of normals Right: Always provide --panel-of-normals for tumor-only calling Why: Without PoN, recurrent sequencing artifacts are called as somatic mutations

Response Format

• Lead with the command or code the user needs — explain after
• Structure as: confirm inputs → working code → key parameters explained → gotchas
• One complete working example per task; do not show every alternative
• Keep code comments minimal and functional (what, not why-it-exists)
• Target: 50-100 lines of code with brief surrounding explanation