Somatic Variant Pipeline Usage Guide

Overview

An end-to-end tumor-normal somatic pipeline chains a somatic caller (Mutect2 or Strelka2) with the four somatic-specific removers - panel of normals, gnomAD germline-resource prior, cross-sample contamination estimate, and the FFPE/oxoG orientation-bias model - then somatic SV/CNV, then interpretation. The governing reality: there is no universal somatic truth set (the DREAM/Alioto benchmark showed low cross-pipeline concordance), somatic variants sit at continuous sub-1.0 VAF set by purity/ploidy/clonality, and interpretation uses the AMP/ASCO/CAP tier system plus oncogenicity - never germline ACMG.

Prerequisites

conda install -c bioconda gatk4 strelka manta bcftools ensembl-vep

Resources: a genome reference (.fa + .dict + .fai), af-only-gnomAD VCF (germline prior), a small common biallelic-SNP VCF such as small_exac_common_3.vcf.gz (contamination), and a 40+ normal panel of normals matched to the assay.

Quick Start

Tell your AI agent what you want to do: - "Call somatic mutations from my tumor-normal BAM pair" - "Run Mutect2 with contamination and orientation-bias filtering" - "Build a panel of normals from my normal samples for the same assay" - "Call somatic variants tumor-only and warn me about germline leakage" - "Run Mutect2 and Strelka2 and take the consensus" - "How do purity and copy number change how I read the VAF?" - "Which tier and oncogenicity framework applies to my somatic variants?"

Example Prompts

Calling

"Call somatic SNVs and indels from tumor.bam vs normal.bam with Mutect2, including the full contamination and orientation-bias filtering chain."

"Set up a panel of normals from 40 normal BAMs sequenced on the same platform, then use it in Mutect2."

Tumor-only and reproducibility

"I only have the tumor (archival FFPE, no matched normal) - call somatic variants tumor-only and tell me what false positives to expect."

"Run Mutect2, Strelka2, and MuSE and give me the 2-of-3 consensus somatic callset."

Purity and interpretation

"My tumor is low purity - why are expected driver mutations missing, and how does purity/ploidy change VAF?"

"Annotate my somatic VCF and route it to AMP/ASCO/CAP tiers and oncogenicity, not germline ACMG."

What the Agent Will Do

1. Confirm inputs (tumor + matched normal BAMs, reference, af-only-gnomAD, common-SNP resource, PoN) or set up tumor-only with caveats.
2. Run Mutect2 (tumor+normal in one command) with the PoN and germline-resource priors, emitting --f1r2-tar-gz.
3. Learn the orientation-bias model (FFPE/oxoG) and estimate contamination on common biallelic SNPs.
4. Apply FilterMutectCalls with the contamination table, tumor segmentation, and orientation priors; extract PASS.
5. Optionally add a Strelka2 arm (with Manta candidateSmallIndels) and build a multi-caller consensus.
6. Reason about VAF against purity/ploidy and copy number; add somatic SV/CNV and TMB/MSI/signatures.
7. Normalize, annotate (VEP/Funcotator), and route to tier/oncogenicity interpretation.

Tips

• Always use a matched normal when available - tumor-only has a materially higher false-positive rate and cannot cleanly separate somatic from germline.
• Build the PoN from 40+ normals on the SAME platform/chemistry; a mismatched PoN imports the wrong artifact profile.
• GetPileupSummaries needs a COMMON biallelic-SNP resource, not the af-only-gnomAD used for the germline prior.
• Always run the orientation-bias model for FFPE/archival input - it removes the C>T/G>A (FFPE) and C>A/G>T (oxoG) low-VAF artifacts.
• Low-purity tumors need more depth to hit the same sensitivity; get purity/ploidy from copy-number analysis and correct VAF to cancer-cell fraction before calling a variant subclonal.
• Normalize every caller's VCF (bcftools norm -f ref.fa -m-) before consensus intersection or annotation, or shared calls silently drop.
• Never apply germline ACMG to a tumor variant - use AMP/ASCO/CAP tiers (tumor-type-specific) and oncogenicity.

Related Skills

• variant-calling/gatk-variant-calling - Mutect2 mechanism and germline context
• variant-calling/filtering-best-practices - FilterMutectCalls internals and normalization
• variant-calling/variant-annotation - VEP/Funcotator/SnpEff, transcript choice, COSMIC
• variant-calling/clinical-interpretation - AMP/ASCO/CAP tiers and oncogenicity
• variant-calling/structural-variant-calling - Somatic SV detection with Manta/GRIDSS
• copy-number/cnvkit-analysis - Somatic CNV and purity/ploidy