In this study we presented a comprehensive library of 1,889 non-redundant SVs identified by breakpoint-resolution mapping in eight studies. Our approach, BreakSeq, leverages a breakpoint junction library for SV detection. While other computational approaches for SV detection (such as paired-end mapping (PEM)5, 37
, DNA read depth analysis38–40
and split-read alignment analysis41
) remain essential for identifying previously unknown SVs (a process that typically involves targeted PCR and sequencing), our approach serves as a reliable tool for rapidly identifying specific SV alleles in personal genomics data. Specifically, by mining personal genomes for sequences present in the breakpoint junction library, BreakSeq leverages alternative, non-reference genomic sequence data to rapidly detect previously described SVs that short-read based personal genomics surveys commonly fail to ascertain. As such, BreakSeq enables a step towards overcoming reference biases, which is the favoring in ascertainment of SV alleles present in the human reference genome sequence.
We foresee that BreakSeq will further gain in utility as datasets grow (e.g., when SV calls from the 1000 Genomes Project are published). As our approach has a linear time complexity (Online Methods), it is easily extendable to larger datasets. In this regard, the size of our junction library currently comprises 0.004% of the reference genome in terms of nucleotide bases, and even a 100-fold increase of its size (>0.2 million SVs; ~10 times of DGV) will result in a dataset considerably smaller than the reference genome. Thus, applying BreakSeq in personal genomics studies adds negligible computing efforts (compared to SNP genotyping) and at the same time dramatically improves SV calling. The library will be updated regularly to serve the personal genomics community in enabling precise SV detection with various next-generation sequencing platforms.
In addition to enabling accurate SV mapping, our junction library allows characterizing SV ancestral states. While the ancestral states of SNPs and small indels have been inferred according to ancestral alignments in earlier studies42, 43
, we here report systematic ancestral state inference for SVs. When applying our new classification approach to 1,281 SVs, we found that overall there is a balance of insertions and deletions, unlike most currently published SV sets that display a considerable bias towards deletions. It should be noted that the non-human primate genomes used in our ancestral state inference correspond to single animals, which certainly do not represent idealized ancestral genomes. Nonetheless, here we reasonably assume that SV loci can be classified at high confidence when ancestral states can be consistently inferred across three distinct primates.
Furthermore, we have developed a computational pipeline for classifying SVs according to their formation mechanisms, and for analyzing various DNA sequence characteristics of the affected genomic loci. Together with the ancestral state analysis, this allowed us to analyze SV formation processes with respect to likely ancestral loci, an analysis that revealed some insights into SV formation. For example, our analyses suggest that the physical properties of the underlying DNA sequence influence locus-specific propensities for different SV formation mechanisms. We observed that NAHR-based SVs are associated with a relatively high GC content and with recombination hotspots, indicating that double-strand breaks (DSBs) occurring specifically during meiotic recombination contribute to NAHR-associated SV formation. On the other hand, NHR breakpoint regions appear to have lower DNA stability and higher flexibility, features that may increase the chance of DSBs in general. Overall, our analysis reveals formational biases underlying SV formation and conforms to the fact that NAHR is driven by recombination between repeat sequences, whereas NHR is likely driven by DNA repair and replication errors.
By applying BreakSeq on a large scale, we envisage that it could be used for genotyping and determining SV allele frequencies. In fact, it should be possible to put each of the breakpoint sequences in our library directly onto a commercially available ‘SNP chip’, which could be used to precisely assess SV genotypes simultaneously with all of the SNPs in an individual. (This should add only a small number of probes to the approximately 1M probes already on the commercial chips.)
Lastly, we note that as our approach depends on the current SV lists, it is inevitably affected by their existing biases owing to presently applied technologies. Likely biases include the difficulty in mapping insertions relative to the reference genome and in ascertaining SVs in repetitive regions, e.g. segmentally duplicated sequences. We anticipate that in the near future, as technologies advance in terms of read lengths, inherent biases against repeat-rich sequences will be further reduced and the mapping of SVs onto our junction library will further improve, making it essentially comparable to SNP-genotyping. In this regard, as thousands of human genomes will be sequenced in the coming years, there will be a huge demand for a reliable and accurate SV-mapping and SV-genotyping.