New Benchmark Reveals the Best Pipelines for Finding Structural Variants in Plant Genomes

Structural variants—large-scale rearrangements of DNA such as deletions, insertions, duplications, inversions, and translocations—are among the most consequential sources of genetic diversity in flowering plants. They can silence genes, duplicate entire pathways, reshape chromosomes, and drive the dramatic phenotypic differences that separate a wild grass from a high-yielding crop variety. Yet detecting these variants reliably has remained a stubborn technical challenge, because plant genomes are typically enormous, riddled with repetitive sequences, and structurally complex in ways that confound even the most sophisticated analytical tools. Now, a team of researchers at the Chinese Academy of Agricultural Sciences has delivered what many plant genomicists have long been waiting for: a rigorous, cross-species benchmark that reveals exactly which computational pipelines can be trusted to find structural variants in plant genomes sequenced with PacBio high-fidelity long reads.

The study, published in Genome Biology, was led by Zhiliang Zhang and Jun Xu, who contributed equally, under the corresponding authorship of Huihui Li at the State Key Laboratory of Crop Gene Resources and Breeding in Beijing. The research team recognized a critical gap in the field. While structural variant detection has been extensively benchmarked on human data—where the tools were largely developed in the first place—the performance and cross-species generalizability of these pipelines in plants had been almost entirely overlooked. Plant genomes present a fundamentally different landscape: they can span many gigabases, harbor extraordinarily high repeat content, and carry structural complexity that far exceeds what is found in most mammalian genomes. A pipeline that performs beautifully on human chromosome 1 might falter badly when confronted with the repetitive expanses of a maize or wheat genome.

To address this problem, the researchers constructed something genuinely valuable: high-confidence structural variant truth sets at base-level resolution for ten representative angiosperm species. Rather than relying on a single variant-calling strategy, they used multiple assembly-based approaches to build these truth sets, triangulating across methods to identify variants that could be trusted with confidence. This multi-strategy approach is important because assembly-based calling itself carries method-specific biases, and by combining several strategies the team could filter out artifacts and retain only variants supported by independent evidence. The truth sets they produced are not merely an internal benchmarking resource—they represent a community asset that other researchers can use to evaluate their own pipelines and validate their own findings.

With truth sets in hand, the team then undertook a systematic evaluation of 91 different HiFi-based structural variant calling pipelines. Each pipeline is a combination of two components: a read aligner, which maps long sequencing reads onto a reference genome, and a variant caller, which interprets the alignment patterns to identify structural variants. The combinatorial space is large—aligners such as Minimap2, NGMLR, and pbmm2 can each be paired with callers such as Sniffles1, Sniffles2, cuteSV1, cuteSV2, and DeBreak, among others—and the choice of pairing matters enormously. The researchers evaluated every combination across all ten species, producing a comprehensive performance matrix that reveals how sensitive and precise each pipeline is for detecting different types of structural variants in different genomic contexts.

The results expose a striking heterogeneity in performance across both pipelines and species. No single pipeline emerged as universally superior, and the ranking of tools shifted depending on which plant genome was being analyzed. This cross-species variability is precisely why the team introduced a novel metric: a stability score that quantifies how well a pipeline generalizes across species. A pipeline that achieves excellent results on rice but mediocre results on barley might score high on individual species but poorly on stability. Conversely, a pipeline that performs consistently well—even if not spectacularly—across all ten species earns a high stability score and becomes a more reliable choice for researchers working with diverse or uncharacterized plant genomes.

Using this framework, the team identified clear winners for different variant classes. For deletion detection—arguably the most common and biologically significant class of structural variant—the combinations NGMLR/Sniffles2 and NGMLR/DeBreak emerged as the optimal cross-species pipelines. These pairings combine the NGMLR aligner, which was designed specifically to handle the characteristic error profiles of long reads, with callers that are particularly adept at identifying missing segments of DNA. For insertion detection, a different set of pipelines proved most robust: Minimap2/Sniffles1, Minimap2/cuteSV2, and pbmm2/cuteSV1. The fact that different aligner-caller combinations excel at different variant types underscores a key insight of the study—there is no single best pipeline, and researchers must choose their tools based on the specific biological questions they are asking.

Beyond the pipeline rankings, the study revealed several important patterns about where and when structural variant detection breaks down. Detection accuracy declined sharply for large structural variants and for variants located within repetitive regions, particularly simple sequence repeats. This finding has profound implications for plant genomics, because many of the most interesting structural variants in crops—those affecting agronomic traits—are precisely the large ones, and they often reside in the repetitive, gene-poor regions that are hardest to analyze. The researchers also discovered that detection precision follows a nonlinear relationship with sequencing depth for each individual haplotype, with F1 scores peaking at approximately 15-fold coverage. This is a critical practical finding: sequencing more deeply does not necessarily improve structural variant calling, and beyond the optimal depth, accuracy may actually decline as the sheer volume of data introduces more opportunities for alignment errors and false positive calls.

The implications of this work extend well beyond the immediate benchmarking results. For plant breeders and geneticists, the availability of validated pipeline recommendations means that structural variant discovery can now be conducted with much greater confidence, enabling more accurate association studies, more reliable pan-genome construction, and better-informed breeding decisions. Structural variants are increasingly recognized as major drivers of crop domestication and adaptation, and having the right analytical tools to detect them is essential for harnessing that variation in breeding programs. The truth sets generated in this study provide a standardized reference against which new tools can be evaluated as they emerge, ensuring that the field does not simply accumulate more data but also improves its ability to interpret it correctly.

For the broader community, this study also serves as a cautionary tale about the dangers of uncritically applying tools developed for one organism to another. The plant genomics community has sometimes borrowed human-focused analytical frameworks without fully validating them for plant-specific challenges, and this benchmark demonstrates that such borrowing can lead to substantially suboptimal results. The stability score framework introduced here could readily be adapted to other non-model organisms, providing a template for rigorous cross-species tool evaluation in fields ranging from fungal genomics to vertebrate biology. As long-read sequencing becomes increasingly affordable and accessible, the bottleneck in structural variant discovery is shifting from data generation to data interpretation—and studies like this one are essential for ensuring that the computational tools researchers rely upon are actually up to the task.

The research was supported by the Biological Breeding-National Science and Technology Major Project, the National Natural Science Foundation of China, the Project of Sanya Yazhou Bay Science and Technology City, the Research Program of Zhongyuan Research Center, and the Agricultural Science and Technology Innovation Program. The full article, including supplementary figures and tables, is available open access in Genome Biology, allowing researchers worldwide to consult the detailed performance data and select the pipelines most appropriate for their own species of interest.

Subject of Research: Benchmarking of long-read structural variant calling pipelines across ten angiosperm species using PacBio HiFi sequencing

Article Title: Benchmarking long-read aligner-caller pipelines for structural variant detection across ten angiosperm species with PacBio high-fidelity sequencing

Article References: Benchmarking long-read aligner-caller pipelines for structural variant detection across ten angiosperm species with PacBio high-fidelity sequencing. (n.d.). https://doi.org/10.1186/s13059-026-04300-x

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04300-x

Keywords: structural variants, PacBio HiFi, long-read sequencing, plant genomics, benchmarking, angiosperms, variant calling, NGMLR, Sniffles2, Minimap2, Genome Biology, crop genetics