Module 5: Somatic Variant Calling Paradigms: Mutect2, Panel of Normals & Low-VAF Detection

1. Why Germline Callers Fail on Somatic Oncology Data

Traditional germline variant callers (such as GATK HaplotypeCaller or FreeBayes) operate under the rigid biological assumption of diploidy: an individual inherits either 0, 1, or 2 copies of an allele, producing expected Variant Allele Fractions of 0.0 (homozygous reference), 0.50 (heterozygous), or 1.0 (homozygous alternate). Any call with a VAF of 4% or 12% is discarded by germline algorithms as sequencing noise or PCR error.

In cancer genomics, tumors exhibit severe aneuploidy, copy-number gains/losses, variable non-neoplastic stromal contamination, and subclonal heterogeneity. A true, clinically actionable somatic driver mutation may present at 2% to 10% VAF. Somatic callers utilize specialized probabilistic mixture models to detect ultra-low VAF alterations while controlling false discoveries.

2. GATK Mutect2 Algorithmic Architecture

Broad Institute's Mutect2 is the gold-standard Bayesian somatic variant caller for SNVs and small indels. The Mutect2 calling engine executes through four computational stages:

1. Active Region Detection: Scans BAM alignments for genomic regions showing evidence of non-reference variation using a lightweight profile model.
2. Local De Novo Assembly: Constructs local de Bruijn graphs from reads spanning active regions, resolving complex insertions, deletions, and multinucleotide variants (MNVs) without relying on initial alignment gaps.
3. Pairwise Alignment & Read Likelihoods: Aligns individual reads against assembled candidate haplotypes using a PairHMM algorithm to compute likelihood matrices: $$L_{i,j} = P(\text{Read}_i \mid \text{Haplotype}_j)$$
4. Bayesian Somatic Likelihood Model: Evaluates whether observed reads supporting a non-reference haplotype in the tumor are consistent with true somatic variation versus sequencing error, germline contamination, or Panel of Normals artifacts.

3. The Panel of Normals (PoN) in Oncology

Even with high-fidelity chemistry, sequencing instruments and library preparation protocols exhibit recurrent technical artifacts at specific genomic coordinates (e.g. library preparation hairpins, sequence-specific polymerase stalling, alignment ambiguities around pseudogenes). To eliminate these recurring false positives, clinical workflows generate a Panel of Normals (PoN):

4. Production Pipeline: Mutect2 Execution & FilterMutectCalls

#!/usr/bin/env bash
# Module 5 Practical: Running Somatic Variant Calling with GATK4 Mutect2

REF_FASTA="GRCh38.d1.vd1.fa"
TUMOR_BAM="Case1_Lung_Tumor.bam"
NORMAL_BAM="Case1_Normal_Blood.bam"
PON_VCF="1000g_pon.hg38.vcf.gz"
GERMLINE_RESOURCE="af-only-gnomad.hg38.vcf.gz"
TARGET_BED="Lung_CGP_Panel_Targets.bed"

echo "[1/3] Executing GATK Mutect2 in paired tumor-normal mode..."
gatk Mutect2     -R ${REF_FASTA}     -I ${TUMOR_BAM} -tumor Case1_Tumor     -I ${NORMAL_BAM} -normal Case1_Normal     --germline-resource ${GERMLINE_RESOURCE}     --panel-of-normals ${PON_VCF}     -L ${TARGET_BED}     -O raw_somatic_variants.vcf.gz     --f1r2-tar-gz orientation_model_data.tar.gz

echo "[2/3] Learning read orientation artifacts (FFPE deamination model)..."
gatk LearnReadOrientationModel     -I orientation_model_data.tar.gz     -O read_orientation_model.tar.gz

echo "[3/3] Applying FilterMutectCalls with orientation bias correction..."
gatk FilterMutectCalls     -R ${REF_FASTA}     -V raw_somatic_variants.vcf.gz     --ob-priors read_orientation_model.tar.gz     -O filtered_somatic_variants.vcf.gz

echo "Somatic calling completed. Filtered calls written to filtered_somatic_variants.vcf.gz"