Multi-platform discovery of haplotype-resolved structural variation in human genomes

Abstract

The incomplete identification of structural variants from whole-genome sequencing data limits studies of human genetic diversity and disease association. Here, we apply a suite of long-and short-read, strand-specific sequencing technologies, optical mapping, and variant discovery algorithms to comprehensively analyze three human parent-child trios to define the full spectrum of human genetic variation in a haplotype-resolved manner. We identify 818,181 indel variants (<50 bp) and 31,599 structural variants (≥50 bp) per human genome, a sevenfold increase in structural variation compared to previous reports, including from the 1000 Genomes Project. We also discovered 156 inversions per genome-most of which previously escaped detection-as well as large unbalanced chromosomal rearrangements. We provide nearcomplete, haplotype-resolved structural variation for three genomes that can now be used as a gold standard for the scientific community and we make specific recommendations for maximizing structural variation sensitivity for future large-scale genome sequencing studies. INTRODUCTIONStructural variants (SVs) contribute greater diversity at the nucleotide level between two human genomes than any other form of genetic variation (Conrad et al. 2010;Kidd et al. 2010;Korbel et al. 2007;Sudmant et al. 2015). To date, such variation has been difficult to identify and characterize from the large number of human genomes that have been sequenced using shortread, high-throughput sequencing technologies. The methods to detect SVs in these datasets are dependent, in part, on indirect inferences (e.g., read-depth and discordant read-pair mapping). The limited number of SVs observed directly using split-read approaches (Rausch et al. 2012;Kronenberg et al. 2015;Ye et al. 2009) is constrained by the short length of these sequencing reads. Moreover, while larger copy number variants (CNVs) could be identified using microarray and read-depth approaches, smaller events (<5 kbp) and balanced events, such as inversions, remain poorly ascertained .

GraphAligner: Rapid and Versatile Sequence-to-Graph Alignment

“…In this experiment we built a variation graph of the human chromosome 22 and compared GraphAligner and vg [16] on it. To build the graph, we took the GRCh38 reference and variants [33] called by the Human Genome Structural Variation Consortium [33], and used vg to build the graph from the reference and the variants. We first randomly subsampled the reads [2] to 15x coverage.…”

“…In addition, we use whole human genome PacBio Sequel [3] and Illumina [4] data from HG00733, randomly subsampled to 15x coverage for PacBio and 30x for Illumina. We use the diploid assembly from [33] as the ground truth to evaluate against for HG00733. We did not include LoRDEC in the fruit fly or HG00733 experiments as the results in [34] show that FMLRC outperforms it in both speed and accuracy.…”

Paragraph: A graph-based structural variant genotyper for short-read sequence data

“…This is chiefly because, compared to short reads, long (10-50kbp) reads can be more reliably mapped to such regions and are more likely to span entire SVs [8][9][10] . These technologies combined with data generated by population studies using multiple sequencing platforms, are leading to a rapid and ongoing expansion of the reference SV databases in a variety of species [11][12][13] .…”

A fully phased accurate assembly of an individual human genome

“…This technique is particularly powerful when combined with other data types like linked reads or long reads to create dense long-range haplotypes 18 . Previously, we used this approach for partitioning reads prior to local assembly to improve structural variation sensitivity 19 but read partitioning required mapping to a reference genome as an intermediate step, which can entail biases towards reference alleles and alignment artifacts. Here, we show how this limitation can be removed by exploiting Strand-seq's additional ability to assign contigs to chromosomes in order to phase them and how this linking technology can be coupled with recent advances in highly accurate long-read sequencing.…”

“…We initially assembled HiFi reads for HG00733 using Canu 26 , into a haplotype-unaware ("collapsed") assembly with contig N50 values of 14.9 Mbp. To scaffold the genome, we aligned 115 single-cell Strand-seq libraries generated for HG00733 in the context of the Human Genome Structural Variation Consortium (HGSVC) 19 to the collapsed assembly. The cumulative depth of Strand-seq reads was 2.87-fold and covered 73% of genomic positions in the assembly.…”