New algorithm HitSV sharpens structural variant detection across sequencing technologies

…

…

The challenge of structural variant detection stems from a fundamental tension in modern genomics: the technologies that make sequencing affordable are often the same technologies that struggle most with the repetitive, duplicated, and rearranged regions where structural variants concentrate. Short-read sequencers produce hundreds of millions of reads per run at remarkably low cost, but each read covers only a few hundred bases, leaving most variants larger than that window invisible to direct observation. Long-read sequencers can span entire repeat arrays and complex junctions in single molecules, yet the per-base cost remains high enough that most large-scale population studies cannot afford the coverage depths needed for confident variant calling. This economic and technical divide has shaped the field for over a decade, and the introduction of methods that extract more information from whichever data type is available represents a meaningful step toward closing the gap between what can be sequenced and what can be reliably discovered.

One of the more subtle problems in structural variant calling is that the alignment signals used to detect variants are not uniformly distributed across the genome. In unique, single-copy regions, a deletion or insertion produces a clean, characteristic pattern: split reads that map to two distant locations, discordant read pairs whose insert size deviates from expectation, or local changes in read depth. In repetitive regions, however, these same signals become ambiguous. Reads may align equally well to multiple paralogous loci, split alignments may reflect mapping errors rather than true breakpoints, and the density of clipping events can rise simply because the underlying sequence is difficult to map, not because a rearrangement occurred. Traditional callers often apply uniform thresholds across the genome, which means they are simultaneously too insensitive in repeat-rich regions, where genuine signals are diluted, and too permissive in unique regions, where background noise can be mistaken for evidence. The approach described in this work addresses this by building empirical models of how signal density varies with local repetitiveness, using population-scale reference data to calibrate what normal background looks like in different genomic contexts.

The use of shared reference samples from both the Human Pangenome Reference Consortium and the 1000 Genomes Project to train these empirical distributions is noteworthy. Rather than relying on theoretical assumptions about how alignment signals should behave, the method learns from real data what the actual distribution of sequence divergence, clipping, and discordant pair signals looks like across regions of low, medium, and high repetitiveness. This data-driven calibration allows the caller to adjust its sensitivity thresholds contextually, becoming more permissive in regions where signals are known to be attenuated while maintaining specificity elsewhere. The result is a detection strategy that adapts to local genomic complexity rather than applying a one-size-fits-all criterion that inevitably fails at the extremes.

Progressive clustering of SV signatures represents another conceptual advance. In conventional pipelines, reads supporting the same variant are typically grouped by proximity of their inferred breakpoints, with fixed windows determining which reads belong together. This works reasonably well for simple heterozygous deletions in unique sequence but breaks down for large events, complex rearrangements, or variants in repetitive contexts where breakpoint estimates scatter widely. By treating breakpoints as latent variables to be inferred probabilistically, and by iteratively refining the cluster assignments as evidence accumulates, the method can aggregate signals that would otherwise remain fragmented across multiple weak clusters. This is particularly important for heterozygous variants, where only half the reads carry the alternate allele, and for events that disrupt repetitive sequence, where individual reads may provide only partial or ambiguous evidence of the underlying rearrangement.

The haplotype-resolved local assembly component deserves particular attention because it addresses a persistent weakness in alignment-based calling: the inability to reconstruct the actual sequence of a variant allele. Most callers report the coordinates and type of an event but not its precise nucleotide sequence, which limits downstream analyses that depend on knowing exactly what changed. By constructing a de Bruijn graph from both variant-supporting and background reads in the affected region, and then traversing that graph in a coverage-informed manner, the method can generate candidate contigs that represent the actual alternate haplotypes. For short-read data, this traversal can extend beyond the intrinsic read length, effectively recovering sequence information that no individual read contains. For long-read data, clustering of graph paths followed by iterative refinement produces contigs that capture the full complexity of the rearranged allele. In hybrid datasets, long-read contigs serve as structural backbones that are then polished with short-read consensus to correct residual base-level errors.

The benchmarking against thirteen existing callers across multiple platforms and species provides a rigorous test of these design choices. On the GIAB HG002 sample, which serves as a community standard for variant calling evaluation, the reported genotype F1 scores of 85.00% on PacBio HiFi and 88.48% on Oxford Nanopore data represent substantial improvements over the best competing methods. The fact that the advantage grows with more stringent evaluation parameters suggests that the improvements are concentrated in precisely the areas where existing methods struggle most: complex events and repeat-proximate breakpoints where precise base-level agreement is hardest to achieve. The extension to seventy long-read datasets from thirty-five diverse human samples, spanning all five 1000 Genomes superpopulations, demonstrates that the performance gains are not specific to a single sample or population but generalize across human genetic diversity.

Cross-species transferability is an often-overlooked dimension of method evaluation. Many bioinformatic tools are developed and tuned exclusively on human data, with implicit assumptions about genome size, repeat content, and variant architecture that may not hold in other organisms. The evaluation across six species spanning mammals, insects, fish, and plants, with F1 advantages ranging from roughly 3% to over 21% depending on platform and species, suggests that the underlying principles of repeat-aware signal recognition and progressive clustering are not artifacts of human genome architecture but reflect general properties of how alignment signals behave in the presence of structural variation. This is encouraging for researchers working in non-model organisms, where structural variants are increasingly recognized as important drivers of adaptation and domestication but where the bioinformatic toolkit has lagged behind what is available for human genetics.

The application to the 1000 Genomes Project phase 4 cohort, expanding the structural variant callset by 31.5%, has implications that extend beyond simple variant discovery. Allele frequency estimates for structural variants are foundational for association studies, clinical interpretation, and population genetics analyses. When a substantial fraction of variants are missed or incorrectly genotyped, the resulting frequency spectra are distorted, rare variants may be misclassified as common or vice versa, and the power of association studies is diminished. A 31.5% expansion of the callset suggests that a meaningful proportion of structural variation in the human population has been invisible to previous methods, particularly in repetitive regions where the new approach shows its greatest advantage. The reshaping of allele-frequency landscapes mentioned in the abstract likely reflects both the addition of entirely new variants and the improved genotyping accuracy for variants that were previously detected but poorly characterized.

The discovery of tandem-repeat-mobile-element composite arrays in a large Chinese long-read cohort represents perhaps the most biologically intriguing finding. Composite structural variants, where a mobile element insertion is embedded within or adjacent to a tandem repeat expansion, have been individually described but not previously recognized as a prevalent, multi-allelic class. The characterization of these as common contributors to human structural variation suggests that the mutational mechanisms generating them, likely involving combined activity of retrotransposition and replication-based repeat expansion, may be more active than previously appreciated. These composite architectures are precisely the kind of variants that would be systematically missed by methods relying on simple alignment signatures, because the combination of repetitive sequence and mobile element insertion creates alignment ambiguity that confounds standard detection approaches. Their identification as a prevalent class rather than rare curiosities underscores how improvements in detection methodology can reveal entirely new categories of variation rather than simply finding more instances of known types.

The reduction in long-read data requirements mentioned for this method has practical implications for study design. Long-read sequencing at sufficient coverage for confident structural variant calling remains expensive enough that many research groups must choose between sequencing fewer samples deeply or more samples shallowly. If a caller can achieve high accuracy at lower coverage by extracting more information from each read and by being more robust to the noise that increases at lower depths, then the effective cost per sample decreases. This could enable long-read structural variant analysis in studies that would otherwise be limited to short-read approaches, particularly in clinical settings where individual patient budgets are constrained but the diagnostic value of comprehensive structural variant detection is increasingly recognized.

The integration of alignment-based evidence with local assembly, rather than treating these as competing paradigms, reflects a maturing understanding of the structural variant detection problem. Purely assembly-based approaches, while capable of resolving very complex rearrangements, require high coverage and are vulnerable to misassemblies that can create false variant calls. Purely alignment-based approaches scale well but sacrifice sequence-level resolution and struggle in repetitive contexts. The hybrid strategy implemented here, where alignment signals guide where to look and local assembly determines what is actually there, leverages the complementary strengths of both paradigms. The k-mer statistical modeling used to evaluate candidate contigs and select the most probable allele pairs provides a principled framework for this integration, ensuring that assembly output is not accepted uncritically but is validated against the underlying read evidence.

Looking forward, the principles embodied in this approach, particularly the explicit modeling of how genomic context affects alignment signal reliability, are likely to influence future method development across the variant calling spectrum. As pangenome references become more widely adopted and as sequencing technologies continue to evolve, the specific implementation details may change, but the fundamental insight that detection methods must account for the interaction between sequence context and signal quality will remain relevant. The demonstration that a single method can handle long-read, short-read, and hybrid data with consistent performance suggests a path toward more unified analytical pipelines that do not require researchers to choose different tools depending on their sequencing strategy, reducing the analytical complexity that currently fragments the structural variant detection landscape.

Subject of Research: …

Article Title: Structural variant detection across sequencing technologies

Article References: Li, G., Liu, Y., Zhang, A., Lin, X., Hu, Z., Liu, B., Qian, L., & Wang, Y. (2026). Structural variant detection across sequencing technologies. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03332-9

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03332-9

Keywords: Structural, variant, detection, across, sequencing, technologies, scientific research