GLnexus: joint variant calling for large cohort sequencing

Abstract

As ever-larger cohorts of human genomes are collected in pursuit of genotype/phenotype associations, sequencing informatics must scale up to yield complete and accurate genotypes from vast raw datasets. Joint variant calling, a data processing step entailing simultaneous analysis of all participants sequenced, exhibits this scaling challenge acutely. We present GLnexus (GL, Genotype Likelihood), a system for joint variant calling designed to scale up to the largest foreseeable human cohorts. GLnexus combines scalable joint calling algorithms with a persistent database that grows efficiently as additional participants are sequenced. We validate GLnexus using 50,000 exomes to show it produces comparable or better results than existing methods, at a fraction of the computational cost with better scaling. We provide a standalone opensource version of GLnexus and a DNAnexus cloudnative deployment supporting very large projects, which has been employed for cohorts of >240,000 exomes and >22,000 whole-genomes.

Accurate, scalable cohort variant calls using DeepVariant and GLnexus

“…As sequencing projects have grown to include hundreds of thousands of samples, the need for highly accurate variant calls and computationally efficient merging algorithms is increasingly acute. By optimizing GLnexus 18 to merge single-sample DeepVariant calls, we demonstrated that the superior accuracy 15 and generalizability across sequencing methods 36 of DeepVariant can generate more accurate cohort callsets at large scale. The callset quality metrics of the optimized pipeline consistently outperformed the GATK Best Practices across a range of cohort sizes and sequence coverages.…”

“…We adapted GLnexus 18 for merging DeepVariant gVCFs because of its computational scalability to large cohorts, access to relevant parameters, performance on allele normalization, and open-source license. To identify optimal parameters for merging DeepVariant gVCFs, we created four custom WGS cohorts of 3, 100, 333, and 1,247 samples at both high coverage (40-50x) and low coverage (15x) on chromosome 2, resulting in eight total cohorts ( Supplementary Table 1 ).…”

“…Many variant callers support the generation of Genome Variant Call Format (gVCF) outputs, which supplement the variant calls with block records of non-variant regions annotated with confidence estimates that the regions match the reference genome. Joint genotyping tools such as GATK GenotypeGVCFs 17 and GLnexus 18 transform a cohort of gVCFs into a project-level VCF that contains a complete matrix of every variant in a cohort with a call for each sample. Compared to a full joint-calling strategy, joint genotyping both substantially reduces the size of required input data and avoids the need to fully reprocess all samples when adding samples to an existing cohort.…”

“…Here we introduce a framework to generate highly accurate and scalable cohort callsets with DeepVariant, using its superior calibration of variant confidences and high single-sample accuracy 15 . We adapt the scalable joint genotyper GLnexus 18 to DeepVariant gVCFs and tune filtering and genotyping parameters to optimize performance for whole-genome sequences (WGS) and whole-exome sequences (WES) across a range of sequence coverages and cohort sizes. We compare the resulting callsets to analogous callsets generated by the broadly-used GATK Best Practices 20 which serve as current state-of-the-art benchmarks.…”

Rare variant analysis of 4,241 pulmonary arterial hypertension cases from an international consortium implicateFBLN2,PDGFDand rarede novovariants in PAH

“…Specifically, we used BWA-MEM(32) to map and align pairedend reads to the human reference genome (version GRCh38/hg38, accession GCA 000001405.15), Picard v1.93 MarkDuplicates to identify and flag PCR duplicates and GATK v4.1 (33,34) HaplotypeCaller in Reference Confidence Model mode to generate individual-level gVCF files from the aligned sequence data. We then performed joint calling of variants from all three datasets using GLnexus (35). We used the following inclusion rules to select variants for downstream analysis: AF<0.05% in the cohort, <0.01% in gnomAD exome_ALL (all ancestries), >90% target region with dp>=10; mappability=1; allele balance>=0.25; and "PASS" from DeepVariants (36).…”