GRIDSS: sensitive and specific genomic rearrangement detection using positional de Bruijn graph assembly

Abstract

The identification of genomic rearrangements with high sensitivity and specificity using massively parallel sequencing remains a major challenge, particularly in precision medicine and cancer research. Here, we describe a new method for detecting rearrangements, GRIDSS (Genome Rearrangement IDentification Software Suite). GRIDSS is a multithreaded structural variant (SV) caller that performs efficient genome-wide break-end assembly prior to variant calling using a novel positional de Bruijn graph-based assembler. By combining assembly, split read, and read pair evidence using a probabilistic scoring, GRIDSS achieves high sensitivity and specificity on simulated, cell line, and patient tumor data, recently winning SV subchallenge #5 of the ICGC-TCGA DREAM8.5 Somatic Mutation Calling Challenge. On human cell line data, GRIDSS halves the false discovery rate compared to other recent methods while matching or exceeding their sensitivity. GRIDSS identifies nontemplate sequence insertions, microhomologies, and large imperfect homologies, estimates a quality score for each breakpoint, stratifies calls into high or low confidence, and supports multisample analysis.

Figure 1.

Outline of the GRIDSS pipeline. (A) Soft clipped and indel-containing reads as well as discordant and one-ended anchored read pairs are extracted from input BAM files. Split reads are identified through realignment of soft clipped read bases. (B) Extracted reads are added to a positional de Bruijn graph in all positions consistent with an anchoring alignment. Break-end contigs are identified by iterative identification of the highest weighted unanchored graph path followed by removal of supporting reads. Unanchored contig bases are aligned to the reference genome to identify all breakpoints spanned by the assembly. (C) Variants are called from assembly, split read, and read pair evidence using a probabilistic model to score and prioritize variants.

Figure 2.

Variant caller performance on simulated heterozygous genomic rearrangements. Different classes of genomic rearrangement were randomly generated against human Chr 12 (hg19), and 60× coverage of 2×100-bp sequencing data was simulated. (A) The sensitivity of each method (rows) for each event type (columns) is plotted against event size. (B) Receiver operating characteristic (ROC) curves for all breakpoints (left) and breakpoints located in SINE/Alus (right).

Figure 3.

Performance of different SV callers on deletion detection in NA12878 at 50× coverage. Multiple variant calls were compared to both the Mills et al. (2011) validation call set (A,B) and PacBio/Illumina Tru-Seq Synthetic Long-Read (Moleculo) orthogonal validation data (C,D). Plots show the number of true positives versus false positives (A,C) and the precision versus true positives (B,D). Long-read validation required three split, or seven spanning long reads supporting the call.

Figure 4.

Performance of GRIDSS variant calling and assembly on NA12878 deletions events using long-read orthogonal validation data. Precision versus the number of true positives for different types of support (A) and for different k-mer sizes (B). Assembly of both split reads and read pairs improves both sensitivity and specificity to levels not achievable by either evidence source. Scoring only assembly-supported variants and varying the type of assembly and k-mer size demonstrates that robust small k-mer break-end assembly can be achieved with positional de Bruijn graph assembly but not windowed de Bruijn assembly.

Figure 5.

A tandem duplication identified in a var gene region of the AT-rich Plasmodium falciparum. Coverage is shown for two samples of P. falciparum—a genetically modified line (C5), which was derived from the parental laboratory strain (3D7). The AT-rich genome shows high coverage in genes, which drops to very low levels in the AT-rich nonexonic regions. A change in copy number is apparent in the C5 coverage. GRIDSS detected the underlying tandem duplication in the C5 vaccine candidate (indicated). The supporting discordant read pair (DP) evidence is shown for both strains. Weak evidence (one read pair) for this rearrangement was also detected in the parental population, indicating that the SV was subclonal in this population. This evidence contributed to the positional de Bruijn graph assembly.