Calling Somatic SNVs and Indels with Mutect2
David I. Benjamin, Takuto Sato, Kristian CibulskisiD, Gad A. GetziD, Chip StewartiD, Lee Lichtenstein
Abstract
Mutect2 is a somatic variant caller that uses local assembly and realignment to detect SNVs and indels. Assembly implies whole haplotypes and read pairs, rather than single bases, as the atomic units of biological variation and sequencing evidence, improving variant calling. Beyond local assembly and alignment, Mutect2 is based on several probabilistic models for genotyping and filtering that work well with and without a matched normal sample and for all sequencing depths.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 720
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (2 Dec 2019) · CC BY-ND 4.0
Mutect2 calls somatic SNVs and indels via local haplotype assembly, Bayesian genotyping and F-score-optimised filtering, with roughly one false call per 5 Mb in exome replicate-versus-replicate tests.
Somatic mutations can sit at low allele fractions because of normal-cell contamination, subclonality and copy-number changes, which makes them hard to separate from library, sequencing and mapping artifacts. Mutect2 uses whole haplotypes and read pairs, not individual bases, as its units of evidence, adding a multiallelic Bayesian genotyping model, adaptive assembly-graph pruning and probabilistic artifact filters that the authors say work with or without a matched normal.
Key results
SuggestedWhen one exome replicate is treated as the tumour and another as its normal, so that every call is false, default Mutect2 plus FilterMutectCalls yields about one false-positive call per 5 Mb.
12 ordered pairs built from four NA12878 exome replicates sequenced at the Broad Institute; with about 37 Mb of covered territory, the text equates a total of 7 false positives to this rate. The per-pair table is not reproduced in this text. · Normal-Normal Calls
ClaimedThe genotyping and filtering models are presented as working well with or without a matched normal and across sequencing depths, but no validation in this text is explicitly described as tumour-only.
Default settings except germline filter and panel of normals off for mixtures: first four DREAM Challenge synthetic WGS tumour-normal pairs (leaderboard metrics), WES mixtures of 5, 10 and 20 1000 Genomes samples in triplicate (sensitivity by allele fraction and depth bins), normal-normal pairs and TCGA. Sensitivity and precision values are in tables and plots not included in this text. · Abstract; Results; DREAM Challenge; Normal mixtures
ClaimedAgainst the MC3 callers on TCGA exomes, the authors report that mapping errors in the WGS truth data systematically understate Mutect2's sensitivity and overstate other callers' precision.
65 TCGA tumour-normal pairs with both WES and WGS; candidate calls validated in the matched WGS (at least two supporting reads; unseen variants with power above 0.8 counted as false) against VarScan 2, SomaticSniper, Mutect, RADIA, MuSE, Pindel and Indelocator. No comparison numbers are given in this text, and the authors call the mapping-error point informal. · TCGA WGS
ClaimedA likelihood-based, adaptive pruning of the assembly graph replaces HaplotypeCaller's fixed edge-multiplicity cutoff and is said to help most with high or uneven depth, as in mitochondria, exomes and RNA.
A first pass learns a global error rate from low-likelihood chains, and a second pass uses it to prune chains; no quantitative comparison with constant-threshold pruning appears in this text. · Assembling Haplotypes
Methods, data and reproducibility
Mutect2 flags active sites with a simplified somatic log-odds model, assembles candidate haplotypes in a de Bruijn-like graph with adaptive pruning, aligns reads to haplotypes with pair-HMM and genotypes fragments under a Dirichlet allele-fraction model (mean-field variational Bayes). FilterMutectCalls combines hard filters with probabilistic artifact models, thresholding at the maximum estimated F score.
- ICGC-TCGA-DREAM Somatic Mutation Calling Challenge synthetic WGS tumour-normal BAMs with variants spiked in by BamSurgeon · n = first four tumour-normal pairs
- NA12878 exome technical replicates prepared and sequenced at the Broad Institute, run as ordered tumour/normal pairs · n = four replicates, 12 ordered pairs
- WES libraries from in vitro mixtures of 1000 Genomes Project samples as simulated subclonal tumours; truth from HaplotypeCaller genotypes, keeping variants with gnomAD frequency of at least 0.001 · n = mixtures of 5, 10 and 20 samples, three replicates each
- TCGA MC3 tumour-normal pairs having both WES and WGS, with WGS used to validate exome calls · n = 65 samples (extension to 929 planned)
- Methods & tools
- Mutect2, FilterMutectCalls, GATK 4, HaplotypeCaller, CreateSomaticPanelOfNormals, Calculate-Contamination, ValidateBasicSomaticShortMutations, BamSurgeon, WDL, gnomAD, MC3 callers: VarScan 2, SomaticSniper, Mutect, RADIA, MuSE, Pindel, Indelocator
- Reagents & environment
- Runs as a GATK 4 tool on local or cloud data with a full pipeline in WDL; command lines for Mutect2 and FilterMutectCalls are said to be in the supplement, which is not part of this text.
- Reproducibility gaps
- No code or data links appear in this text; methods shared with HaplotypeCaller and the filtering details are deferred to a supplement, and the validation results tables and plots are not reproduced.
Limitations (5)
- The TCGA comparison so far covers 65 samples, with extension to 929 promised for later versions. · stated by the authors
- Mapping errors in the WGS truth set are thought to penalise Mutect2's sensitivity and inflate other callers' precision, an observation the authors have not yet rigorously justified. · stated by the authors
- In the mixture experiment, dropping variants rarer than 0.001 in gnomAD from the truth set means precision cannot be measured, and the germline filter and panel of normals had to be switched off. · stated by the authors
- No validation in this text is explicitly run tumour-only (the mixture runs do not say whether a normal was used), so good performance without a matched normal is not directly demonstrated here. · noted in this summary
- The false-positive rate comes from exome replicates of a single sample (NA12878), so it may not reflect artifact levels in FFPE or low-purity tumour samples. · noted in this summary
Who should read:Read this if you call somatic SNVs or indels with GATK and want to understand Mutect2's assembly, multiallelic genotyping and artifact-filtering models and how they were validated.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- SV2: Accurate Structural Variation Genotyping and De Novo Mutation Detection
Danny Antaki, William M. Brandler, Jonathan L. Sebat·BioinformaticsPeer-reviewed·Dec 28, 2017
- Lancet2: Improved and accelerated somatic variant calling with joint multi-sample local assembly graph
Rajeeva Musunuri, Bryan Zhu, … Giuseppe Narzisi·NAR Genomics and BioinformaticsPeer-reviewed·Mar 25, 2026
- SVIM-asm: Structural variant detection from haploid and diploid genome assemblies
David Heller, Martin Vingron·BioinformaticsPeer-reviewed·Dec 1, 2020
- Lancet: genome-wide somatic variant calling using localized colored DeBruijn graphs
Giuseppe Narzisi, André Corvelo, … Michael C. Zody·Communications BiologyPeer-reviewed·Mar 14, 2018
- Comprehensive benchmarking of somatic single-nucleotide variant and indel detection at ultra-low allele fractions using short- and long-read data
Yoo-Jin Jiny Ha, Dominika Maziec, … Peter J. Park·bioRxiv·Oct 13, 2025
- A unified haplotype-based method for accurate and comprehensive variant calling
Daniel P. Cooke, David C. Wedge, Gerton A. Lunter·Nature BiotechnologyPeer-reviewed·Mar 29, 2021