What GATK Is and How to Use It on Your Own Genome
GATK, the Genome Analysis Toolkit, is a set of command-line programs from the Broad Institute that turn aligned sequencing reads into a list of genetic variants. You give it a BAM or CRAM file of reads mapped to a reference genome, and it gives you a VCF file listing the positions where your DNA differs from that reference, with a genotype and a confidence score at each position. It is the de facto standard for calling single-nucleotide variants (SNVs) and short insertions and deletions (indels) in human data, and if you have a 30x whole genome from a consumer or clinical provider, GATK or a GATK-compatible pipeline almost certainly produced your VCF.
The original toolkit was published in 2010 as a MapReduce-style framework for traversing sequencing data efficiently.1 Modern GATK4 keeps the traversal engine but the part you care about is the set of tools built on top: HaplotypeCaller for germline variants, Mutect2 for somatic variants, GenotypeGVCFs for joint calling across samples, and VQSR or hard filtering for separating real variants from artifacts. The recommended sequence of these steps is published as the GATK Best Practices, first described in detail in 2013 and revised many times since.2
What each GATK tool does
The core germline caller is HaplotypeCaller. Rather than looking at each base column independently, it identifies active regions where the reads suggest variation, then locally reassembles the reads in that region into a De Bruijn graph and infers the most likely pair of haplotypes. This local reassembly is why GATK handles indels and clustered variants better than simple pileup callers. Comparison studies of callers on next-generation sequencing data consistently show that choice of caller and filtering strategy changes the variant set substantially, especially for indels.3
HaplotypeCaller has two output modes. In normal mode it emits a VCF with only the variant sites. In -ERC GVCF mode it emits a genomic VCF, which records confidence information at every position including the homozygous-reference ones, compressed into blocks. GVCF mode matters when you plan to combine samples later, because it lets you distinguish “this person matches the reference here” from “we had no coverage here.” For a single genome analyzed once, normal mode is simpler.
Mutect2 is the somatic caller. It uses the same local reassembly machinery as HaplotypeCaller but drops the assumption that you are looking at a diploid genome with two clean haplotypes. Somatic mutations in a tumor or in a clonal population of blood cells appear at low allele fractions, sometimes 2% or less, which a germline caller would dismiss as sequencing error. Mutect2 models that case explicitly and, when given a matched normal sample and a panel of normals, filters out germline variants and recurrent technical artifacts.
Beyond calling, GATK includes tools for other variant classes. GATK-gCNV infers rare copy number variants from exome read depth using a probabilistic model of coverage bias, which extends a GATK pipeline to a class of variation that HaplotypeCaller cannot see.4 There is also a documented workflow for calling variants from RNA sequencing data, which requires splice-aware alignment and a step to split reads at N-cigar elements before calling.5
The pipeline we would run on a 30x human genome
Start from FASTQ if you have it. Align with BWA-MEM to GRCh38, keeping the alt contigs, and pipe directly into sorting to avoid writing an unsorted intermediate:
bwa mem -t 16 -R '@RG\tID:L1\tSM:sample1\tPL:ILLUMINA\tLB:lib1' \
GRCh38.fa r1.fq.gz r2.fq.gz \
| samtools sort -@ 4 -o sample1.bam -
The read group line is not optional. GATK refuses to run without @RG tags, and the SM field becomes the sample name in your VCF.
Next, mark duplicates. PCR and optical duplicates inflate apparent depth and let a single amplification error look like strong evidence:
gatk MarkDuplicates -I sample1.bam -O sample1.md.bam \
-M sample1.dup_metrics.txt --CREATE_INDEX true
Base quality score recalibration comes next. The sequencer’s reported quality scores are systematically biased by machine cycle, sequence context, and position. BQSR fits an empirical model of those biases using known variant sites as a mask, then rewrites the quality scores:
gatk BaseRecalibrator -R GRCh38.fa -I sample1.md.bam \
--known-sites Homo_sapiens_assembly38.dbsnp138.vcf \
--known-sites Mills_and_1000G_gold_standard.indels.hg38.vcf.gz \
-O recal.table
gatk ApplyBQSR -R GRCh38.fa -I sample1.md.bam \
--bqsr-recal-file recal.table -O sample1.bqsr.bam
Then call variants. On a whole genome, run HaplotypeCaller in parallel across intervals rather than as one process, because a single-threaded run over 3.1 gigabases takes a day or more:
gatk --java-options "-Xmx8g" HaplotypeCaller \
-R GRCh38.fa -I sample1.bqsr.bam \
-L chr1 -ERC GVCF -O sample1.chr1.g.vcf.gz
Finally, filter. With a single sample the standard machine-learning filter, Variant Quality Score Recalibration, is unreliable because it needs many variants across many samples to fit a Gaussian mixture. Use hard filters instead. The published GATK thresholds are a starting point, and a careful study of targeted panel data showed that tuning the individual annotation cutoffs meaningfully changes sensitivity and specificity.6 For SNVs we start with QD < 2.0, FS > 60.0, MQ < 40.0, SOR > 3.0, and ReadPosRankSum < -8.0, then inspect the distribution of each annotation in your own data before accepting them.
Where it fails
Reference choice propagates through everything. A VCF called against GRCh37 and one called against GRCh38 are not directly comparable, and lifting coordinates over with CrossMap or Picard LiftoverVcf fails silently in regions where the assembly changed. Check which reference your provider used before you interpret a single coordinate.
Coverage is not uniform. At nominal 30x, GC-rich promoters, segmental duplications, and the pseudoautosomal regions can drop below 10x, and HaplotypeCaller will either emit a low-confidence call or nothing at all. Absence of a variant in your VCF is not evidence of absence in your genome unless you check the depth at that position. Reading it as evidence is the single most common misreading of a personal VCF.
Repetitive and paralogous regions produce false positives that no filter fully removes. Genes with close pseudogenes are the usual offenders, and short-read data cannot resolve them. Long reads help substantially here, and early work with Pacific Biosciences data showed that a different error profile changes which variants are discoverable rather than simply reducing error.7 If you care about a specific difficult gene, targeted long-read or Sanger confirmation is the answer.
Exome and panel data are worse than whole genome at the edges of capture baits, where coverage falls off sharply. Published clinical exome studies using GATK routinely confirm their key findings by an independent method before drawing any conclusion.8 Some disease-relevant genes are also genuinely hard to sequence for structural reasons, and targeted panel designs exist precisely because generic pipelines miss them.9
Finally, GATK produces a variant list and nothing more. Interpretation requires annotation with tools like VEP or Annovar, population frequency filtering, and consultation of curated databases such as ClinVar. A pipeline can tell you a variant exists. Whether it means anything for your health is a clinical question, and any variant you would act on needs confirmation in a certified laboratory and discussion with a clinician or genetic counselor. Do not treat a research-grade VCF as a diagnosis.
Questions people also ask
What is GATK in bioinformatics? GATK is a toolkit for discovering genetic variants from high-throughput sequencing data. It reads aligned reads in BAM or CRAM format and writes variants in VCF format, and it includes supporting tools for duplicate marking, quality score recalibration, joint genotyping, and filtering. It was originally described as a programming framework for building such tools efficiently.1
What is GATK Mutect2? Mutect2 is GATK’s somatic variant caller, designed to find mutations present in only a fraction of the cells in a sample. It uses the same local haplotype assembly as HaplotypeCaller but replaces the diploid genotype model with one that allows arbitrary allele fractions, and it filters germline and artifact calls using a matched normal sample and a panel of normals.
Do I need to run BQSR? At 30x whole-genome depth with modern Illumina chemistry, BQSR changes the final call set modestly. We still run it because it is cheap relative to alignment and because the Best Practices thresholds for hard filtering were calibrated on recalibrated data.2
How long does this take on one machine? On 16 cores, alignment of a 30x genome takes roughly a day, and HaplotypeCaller takes a comparable time unless you split by interval. Splitting into 24 chromosome-level jobs brings calling down to a few hours. Budget around 500 GB of disk for intermediates, and delete the pre-BQSR BAM once you have verified the recalibrated one.
Are there alternatives worth using? DeepVariant, which replaces the statistical model with a convolutional neural network over pileup images, matches or beats GATK on SNVs for single samples and is simpler to run. We run both on the same BAM and treat the intersection as high confidence, since caller disagreement is informative about which sites deserve scrutiny.3
Footnotes
-
Aaron McKenna, Matthew Hanna, Eric Banks, et al. The Genome Analysis Toolkit: A MapReduce framework for analyzing next-generation DNA sequencing data. Genome Research, 2010. https://doi.org/10.1101/gr.107524.110 ↩ ↩2
-
Geraldine A. Van der Auwera, Mauricio O. Carneiro, Christopher Hartl, et al. From FastQ Data to High‐Confidence Variant Calls: The Genome Analysis Toolkit Best Practices Pipeline. Current Protocols in Bioinformatics, 2013. https://doi.org/10.1002/0471250953.bi1110s43 ↩ ↩2
-
Xiangtao Liu, Shizhong Han, Zuoheng Wang, et al. Variant Callers for Next-Generation Sequencing Data: A Comparison Study. PLoS ONE, 2013. https://doi.org/10.1371/journal.pone.0075619 ↩ ↩2
-
Mehrtash Babadi, Jack M. Fu, Samuel K. Lee, et al. GATK-gCNV enables the discovery of rare copy number variants from exome sequencing data. Nature Genetics, 2023. https://doi.org/10.1038/s41588-023-01449-0 ↩
-
Jean-Simon Brouard, Nathalie Bissonnette. Variant Calling from RNA-seq Data Using the GATK Joint Genotyping Workflow. Methods in Molecular Biology, 2012. https://doi.org/10.1007/978-1-0716-2293-3_13 ↩
-
Simona De Summa, Giovanni Malerba, Rosamaria Pinto, et al. GATK hard filtering: tunable parameters to improve variant calling for next generation sequencing targeted gene panel data. BMC Bioinformatics, 2017. https://doi.org/10.1186/s12859-017-1537-8 ↩
-
Mauricio O Carneiro, Carsten Russ, Michael G Ross, et al. Pacific biosciences sequencing technology for genotyping and variation discovery in human data. BMC Genomics, 2012. https://doi.org/10.1186/1471-2164-13-375 ↩
-
Johann Greil, Tobias Rausch, Thomas Giese, et al. Whole-exome sequencing links caspase recruitment domain 11 (CARD11) inactivation to severe combined immunodeficiency. Journal of Allergy and Clinical Immunology, 2013. https://doi.org/10.1016/j.jaci.2013.02.012 ↩
-
G. Inal Gültekin, B. Toptaş Hekimoğlu, Z. Görmez, et al. Myophosphorylase (PYGM) mutations in Turkish patients with McArdle disease: A next generation sequencing study. Neuromuscular Disorders, 2015. https://doi.org/10.1016/j.nmd.2015.06.129 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.