Scaling accurate genetic variant discovery to tens of thousands of samples
Ryan E. Poplin, Valentín Ruano-Rubio, Mark A. DePristoiD, Tim J. FennelliD, Mauricio O. Carneiro, Geraldine A. Van der AuweraiD, David E. KlingiD, Laura Doyle Gauthier, Ami Levy‐Moonshine, David Roazen, Khalid Shakir, Joel Thibault, Sheila Chandran, Chris Whelan, Monkol LekiD, Stacey GabrieliD, Mark J. DalyiD, Benjamin M. NealeiD, Daniel G. MacArthuriD, Eric BanksiD
Abstract
Comprehensive disease gene discovery in both common and rare diseases will require the efficient and accurate detection of all classes of genetic variation across tens to hundreds of thousands of human samples. We describe here a novel assembly-based approach to variant calling, the GATK HaplotypeCaller (HC) and Reference Confidence Model (RCM), that determines genotype likelihoods independently per-sample but performs joint calling across all samples within a project simultaneously. We show by calling over 90,000 samples from the Exome Aggregation Consortium (ExAC) that, in contrast to other algorithms, the HC-RCM scales efficiently to very large sample sizes without loss in accuracy; and that the accuracy of indel variant calling is superior in comparison to other algorithms. More importantly, the HC-RCM produces a fully squared-off matrix of genotypes across all samples at every genomic position being investigated. The HC-RCM is a novel, scalable, assembly-based algorithm with abundant applications for population genetics and clinical studies.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 1,941
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 3 (24 Jul 2018) · CC BY 4.0
Per-sample assembly with gVCF output, followed by joint genotyping, lets GATK HaplotypeCaller scale linearly with cohort size while calling indels more accurately than SAMtools, UnifiedGenotyper or Platypus.
Estimating population allele frequencies needs a genotype, including homozygous reference, for every sample at every site, yet assembly-based callers, which handle indels better than pileup callers, become intractable when one graph must describe many samples. The Reference Confidence Model assembles each sample on its own and records likelihoods at all positions, including a catch-all non-reference allele, so only allele-frequency estimation and genotype assignment are done jointly across the cohort.
Key results
ShownPast 250 samples, HaplotypeCaller with the Reference Confidence Model keeps a runtime that grows linearly with cohort size, while UnifiedGenotyper, Platypus and SAMtools grow superlinearly.
CPU-hours versus sample number on 20X exomes: HaplotypeCaller needed slightly more CPU-hours than the other three callers up to 250 samples, then rose linearly while the others rose superlinearly. The underlying figure values are not reproduced in this text. · 2.3 Scaling; 3 Discussion
ShownHaplotypeCaller was the most sensitive indel caller on both whole-genome and exome data and had the lowest indel false discovery rate (FDR) on whole genomes.
CEU trio calls (2×250bp PCR-free genomes, 2×76bp exomes) scored against Genome in a Bottle v2.18 for NA12878, versus SAMtools 1.1, UnifiedGenotyper and Platypus 0.7.8. The authors explain the disagreement with Rimmer et al. by that study's older HaplotypeCaller (v2.5 vs v3.4) and fosmid truth set. Per-tool values sit in a table or figure absent from this text. · 2.2 Comparative Analyses; 3 Discussion; 4.1 Data; 4.2 Software
ShownIndel genotypes were assigned much more accurately by HaplotypeCaller than by the other callers, whereas all four callers genotyped SNPs with similar accuracy.
Genotype concordance (correctly assigned genotypes divided by true-positive variants, judged against GiaB) was described as exceptionally high for HaplotypeCaller indels in genome and exome data; the authors credit local assembly and pair-HMM likelihoods being more robust in repeats, an explanation not tested directly. · 2.2 Comparative Analyses; 3 Discussion; 4.3 Evaluation Metrics
ShownSAMtools called the most SNPs and indels and was the most sensitive SNP caller, but its FDR was substantially higher than that of the other callers, especially for indels.
Same GiaB-based comparison on whole-genome and exome data; HaplotypeCaller combined high sensitivity with low FDR on both data types, which the authors read as high sensitivity plus high specificity. · 2.2 Comparative Analyses; 3 Discussion
ClaimedThe approach was used to jointly call more than 90,000 exomes for the Exome Aggregation Consortium (ExAC), giving genotypes for every sample at every examined position.
Stated in the abstract, the scaling results and the discussion; no runtime, cost or accuracy figures for the ExAC call set are given, so the claim of no accuracy loss at that scale is not backed by data in this text. · Abstract; 2.3 Scaling; 3 Discussion
Methods, data and reproducibility
HaplotypeCaller reassembles reads in active regions into candidate haplotypes (de Bruijn-like graphs) and scores reads against them with a pair hidden Markov model; the Reference Confidence Model writes per-sample likelihoods, including a symbolic <NON_REF> allele, to gVCFs that are genotyped jointly. Accuracy and runtime were compared with SAMtools, UnifiedGenotyper and Platypus.
- CEU HapMap trio high-coverage whole genomes (2×250bp PCR-free BAMs) from the 1000 Genomes repository · human · n = trio (NA12891, NA12892, NA12878)
- CEU trio exome-capture data (2×76bp BAMs) from the 1000 Genomes repository · human · n = trio
- Genome in a Bottle truth set v2.18 for NA12878, used as the accuracy reference · human
- 20X exomes used to measure runtime against sample number
- Exome Aggregation Consortium (ExAC) exomes jointly called with HC-RCM · human · n = over 90,000 exomes
- Methods & tools
- GATK HaplotypeCaller (HC), Reference Confidence Model (RCM), GATK UnifiedGenotyper (UG), SAMtools, Platypus, BWA, Picard Tools, GATK Base Quality Score Recalibrator (BQSR), GATK GenotypeGVCFs, GATK CombineGVCFs, VQSR, Genome in a Bottle (GiaB)
- Reagents & environment
- Platypus 0.7.8, SAMtools 1.1, GATK 3.4 (HaplotypeCaller-RCM and UnifiedGenotyper); reads aligned with BWA; key HaplotypeCaller defaults (active-region size 50-300bp, k-mers of 10 and 25, gap extension penalty 10) are given in the methods.
- Reproducibility gaps
- No code or data links appear in this text: exact command lines and GATK documentation are cited but the links are missing, and the 1000 Genomes inputs have no accession. Table and figure values behind the comparisons are not reproduced here.
Limitations (6)
- Raw HC-RCM output deliberately favours sensitivity and keeps false positives; gVCFs must pass through GenotypeGVCFs and variant recalibration before any analysis. · stated by the authors
- Up to 250 samples, HaplotypeCaller costs slightly more CPU-hours than UnifiedGenotyper, Platypus and SAMtools. · stated by the authors
- Platypus showed unexpectedly high whole-genome FDRs for SNPs and indels, and the authors ran the competing tools themselves, following each tool's published recommendations. · stated by the authors
- Assembly is abandoned in active regions whose reference sequence contains non-unique 65-mers, so the most repetitive sequence is not locally assembled. · stated by the authors
- Accuracy comparisons rest on one truth sample (NA12878, GiaB v2.18) in a benchmark run by the tool's own developers, with no external replication; an orthogonal Illumina Platinum Genomes evaluation is mentioned only in passing, with no results given in this text. · noted in this summary
- Neither accuracy nor runtime of the 90,000-exome ExAC call set is reported, so the claim of no accuracy loss at scale is extrapolated from the trio benchmark. · noted in this summary
Who should read:Read this if you run germline short-variant calling on large cohorts and want the design, defaults and benchmark behind GATK's per-sample gVCF plus joint-genotyping workflow.
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
- GW-CALL: Accurate Genome-Wide Variant Caller
Maryam Ghareghani, Seyed Abolfazl Motahari, … Mostafa Tavassolipour·bioRxiv·Oct 9, 2016
- A bioinformatic panel to interrogate thousands of ExAC variants with minor reference allele that are missed by conventional variant calling
Mahmoud Koko, Mohammed O. E. Abdallah, … Muntaser Eltayeb Ibrahim·BMC GenomicsPeer-reviewed·Jan 15, 2018
- HPC-based genome variant calling workflow (HPC-GVCW)
Yong Zhou, Nagarajan Kathiresan, … Rod A. Wing·bioRxiv·Jun 25, 2023
- Accurate, scalable cohort variant calls using DeepVariant and GLnexus
Taedong Yun, Helen Li, … Cory Y. McLean·BioinformaticsPeer-reviewed·Dec 15, 2020
- Rovaca: A highly optimized variant caller for rapid and robust germline DNA analysis
Qiwen Zheng, Longhui Yin·bioRxiv·Oct 19, 2025
- DNAscope: High accuracy small variant calling using machine learning
Donald N. Freed, Renke Pan, … Rafael Aldana·bioRxiv·May 20, 2022