gSV: a general structural variant detector using the third-generation sequencing data Authors: Jingyu Hao, Jiandong Shi, Sheng Lian, Zhen Zhang, Yongyi Luo, Taobo Hu, Toyotaka Ishibashi, Depeng Wang, Shu Wang, Xiaodan Fan, Weichuan Yu Published: 2026 ABSTRACT AbstractStructural variants (SVs) are major contributors to genome diversity and disease susceptibility, particularly in cancer. Although third-generation sequencing technologies have substantially improved SV detection sensitivity, accurate detection of complex SVs remains challenging due to fragmented and heterogeneous alignment signals, as well as the dependence of many existing methods on predefined variant models. In this paper, we propose gSV, a general SV detector that integrates alignment-based and assembly-based approaches with the maximum exact match strategy, with particular emphasis on resolving SVs with complex or atypical alignment signatures. Without predefined assumptions about SV types, gSV captures diverse variant signals, enabling the detection of SVs that are usually missed by conventional tools. Benchmarking using both simulated datasets and real long-read sequencing data demonstrates that gSV achieves improved sensitivity and overall detection performance compared with current state-of-the-art SV callers, particularly for simple and complex SV events with complex alignment patterns. Unique SV discoveries in four breast cancer cell lines, particularly in cancer-associated genes, demonstrate the potential biological relevance of gSV-enabled discoveries. Furthermore, analysis of a breast cancer cohort from the Chinese population highlights the utility of gSV for population-scale genomic studies. Collectively, gSV provides a unified framework for comprehensive SV discovery in both research and clinical genomics settings.Keywords: structural variant detection, complex structural variant, third-generation sequencing Introduction The advent of the third-generation sequencing technologies, such as PacBio and Nanopore, has enabled the acquisition of longer reads spanning thousands of base pairs (bps). Consequently, longer genomic mutations such as structural variants (SVs) are of increasing interest. SVs may be of different levels of complexity. Simple SVs can be categorized as balanced (e.g. inversions (INVs)) or unbalanced (e.g. deletions (DELs), insertions (INSs), and duplications (DUPs)). Complex SVs are usually regarded as combinations of simple SVs in various orders and numbers. These events, which often involve multiple breakpoints, nested rearrangements, or complex alignment patterns that are difficult to resolve using standard SV representations, have been widely observed in real datasets and are of increasing research interest [1]. This offers a new angle to study genetic abnormalities associated with diseases. Many tools have been developed to detect SVs. Most of them focus on detecting simple SVs. They can be broadly categorized into alignment-based and assembly-based methods. Alignment-based methods typically involve three steps: capturing SV signatures from alignment results; clustering the signatures in terms of positions, lengths, and types; and finalizing the clustered signatures to call SVs and filtering using quality metrics (e.g. supporting reads) [2–6]. Alignment-based methods are popular due to their simplicity and high sensitivity, particularly in detecting simple SVs. However, existing methods can only detect potential SVs that conform to their predefined patterns at the initial step, thereby limiting their ability to resolve complex SVs with novel structural patterns that deviate from these predefined categories. Assembly-based methods, on the other hand, first assemble reads into longer consensus and subsequently align the consensus to the reference genome to detect SVs [7]. Assembly-based methods are computationally expensive and are usually not primarily optimized for resolving complex SVs. Recently, Denti et al. [8] have proposed to use both alignment-based and assembly-based methods. But their method can only detect INSs and DELs. Lin et al. [9] have applied a deep learning method to detect both simple and complex SVs, and Wang et al. [10] have further improved it. However, the issues of overfitting and weak interpretability are not tackled when applying the deep learning method. To date, accurately detecting both simple and complex SVs in an interpretable manner remains challenging, particularly for complex SVs where achieving satisfactory sensitivity and specificity (recall and precision) is still difficult. In this paper, we propose a general SV detection framework, gSV, which integrates alignment-based and assembly-based strategies with a maximum exact match (MEM), with a particular focus on improving the detection of complex SVs. Unlike the most widely used SV detection tools that rely on prior assumptions about the types of SVs, gSV comprehensively captures a broad range of suspicious variant signals. This enables the detection of SVs that exhibit unexpected complex patterns that are often missed by conventional detection tools, especially those involving nested or multi-breakpoint events. Experiments on simulated and real datasets demonstrate that gSV achieves competitive performance on canonical SV classes and improved sensitivity for complex SV patterns. Furthermore, applications of gSV to four breast cancer cell lines and a Chinese breast cancer cohort identified SVs in breast cancer-associated genes, further illustrating the potential utility of gSV. In the following sections, we first detail the methodological framework of gSV and the experimental design for benchmarking in the Materials and methods section. In the Results and Discussion section, we present a quantitative comparison of gSV with existing tools using simulated and real data. We further validate the unique findings of gSV against the existing literature, illustrating the usefulness of our new tool in genomic and clinical studies. Materials and methods Detailed description of general structural variants The detailed workflow of gSV is shown in Fig. 1. We propose a novel encoding strategy that converts information from reads/reference into a matrix form, as illustrated in Fig. 1, which facilitates us to analyze one chromosome at a time. After encoding, the second step is detecting candidate regions with strong SV signals using the graph cut method [11]. In each candidate region, there may be multiple reads belonging to different subtypes, representing different SVs or genotypes, for which we carry out the clustering. Then we apply an assembling method (wtdbg2) to generate a long consensus [12] that covers the synthesized information from multiple reads within each cluster. After zooming in from the whole genome to candidate regions, our assembling is computationally much cheaper. Moreover, with a longer consensus, gSV has a chance to compare the differences between the reference and the underlying sequence comprehensively. With only one consensus, we cannot afford missing detailed/complex signals by using global alignment tools such as minimap2 [13] and NGMLR [3]. Therefore, we adopt the idea of MEM [14] between the consensus and the reference in our realigning step. In the finalizing step, we transform realignment results into SV callings. By properly incorporating the assembling and the MEM method, gSV provides a step (that did not exist before) to revisit the missed or misspecified signals for correcting the final calling. In the following, we shall explain these steps in detail. The workflow of gSV consists of six steps. Step 1. Encoding: converting reads/reference into a matrix representation to enable chromosome-scale analysis. Step 2. Detecting: using graph-cut method to detect candidate regions with strong SV signals. Step 3. Clustering: clustering reads within each complex candidate region into subtypes to resolve heterogeneous genotypes. Step 4. Assembling: generating long, accurate consensus sequences (via wtdbg2) for each cluster. Step 5. Realigning: aligning the consensus to the reference using MEMs to capture discrepancies missed by global aligners. Step 6. Finalizing: transforming realignment results into SV callings. Encoding Unlike existing alignment-based SV detection tools that rely on predefined SV models, our method directly captures all differences between the reference genome and sequencing reads without prespecifying SV types. These differences encompass sequencing errors and potential variant signatures. Here signatures are defined as consistent, region-level deviations from the reference genome across multiple reads, including split alignments, soft-clipping, and abnormal mapping distances or orientations that suggest the presence of SVs. To enhance the ability to capture complex alignment patterns that are often lost in traditional CIGAR-based parsing, we adopt a matrix-based encoding method to capture base-level inconsistencies, positional gaps, and strand discordance in a unified numerical form. For each chromosome and read, we encode the reference and aligned read as matrices representing nucleotides , , , and . Using the read’s reported genomic coordinates, we place the read matrix onto the reference coordinate system and compute their absolute difference. Because most bases match, the resulting matrices are sparse, enabling efficient memory usage and parallel processing. Read matrices are generated from standard Sequence Alignment Map (SAM) fields (FLAG, POS, CIGAR, and SEQ). Only essential alignment operations are retained in the main analysis: matches or mismatches (“M”/“=”/“X”), INSs (“I”), and DELs (“D”), while soft/hard-clipped bases are excluded. Supplementary alignments marking split reads are integrated with their primary alignments into a single matrix to preserve chimeric structure. Details of matrix construction, handling of CIGAR edge cases, and treatment of strand discordance are explained in Supplementary Note 1.3. In addition, when encoding each read, every clear, and easily recognizable signature will be recorded simultaneously. Taking the case of Step 1 in Fig. 1 as an example, when encoding this read, we record a DEL with 3 bps in positions 8–10. The user can choose whether or not to use the prior information for subsequent SV detection. Different from other SV callers, our matrix-based encoding method preserves all inconsistencies, increasing the possibility of detecting SVs. To facilitate the subsequent discussion, let 1 or 0 represents whether the th position corresponding to the reference is a variant or not, indicates the encoding result for the th position in the th read, and indicate the encoding result for the th position in the reference. Detecting After encoding the signals in a matrix form, we proceed to detect candidate regions with strong SV signals [11]. Using the above notations, we formulate the problem as follows: where indicates a neighboring set that satisfies . The objective function is designed to distinguish true SV signals from random sequencing error as well as the small variants such as single-nucleotide polymorphisms (SNPs), by exploiting their distinct alignment consistency patterns with respect to the reference genome. Specifically, in regions without SVs, reads are expected to align consistently to the reference, resulting in small mismatch loss . In true SV regions, reads from the variant allele systematically deviate from the reference, leading to substantially larger and more coherent discrepancies across multiple reads. This creates a collective increase in the mismatch loss when , making such positions unlikely to be classified as non-variant under the first term. In contrast, sequencing errors tend to be random and sparse across reads and positions, and, therefore, do not generate consistent increases in mismatch loss across a genomic region. As a result, they are less likely to produce a strong aggregate signal that justifies labeling a region as variant, especially when balanced against the penalty terms. The second term introduces a sparsity penalty controlled by . Each time a position is labeled as variant (), the model pays a fixed cost . Therefore, a locus will be called variant only when the reduction in the first-term mismatch loss is large enough to justify that cost. This helps prevent weak or isolated discrepancies (more likely to reflect sequencing error) from being over-interpreted as SV signals. The third term, controlled by , penalizes changes in the indicator between neighboring positions. In this way, it discourages fragmented solutions with many isolated variant labels and instead favors continuous candidate regions. This is consistent with the biological expectation that SVs usually affect continuous intervals rather than single and scattered bases like small variants. This formulation allows us to detect candidate regions with strong signals for subsequent analysis. Clustering In each candidate region, there may be multiple reads belonging to different subtypes, representing different SVs or genotypes within the region. Users may optionally employ prior information (i.e. genomic locations, lengths, and types of potential SVs derived from variant signatures in the alignment results during the encoding step) to assist detection. When this option is enabled (default setting), candidate regions are classified into two categories: regions with clear signatures and regions with complex or ambiguous signatures (usually containing complex SVs and some simple SVs with complex alignment patterns), where the distinction is defined based on the consistency of variant signatures extracted during the encoding step. If the option is disabled, all candidate regions are treated as regions with complex signatures, ensuring comprehensive detection of potential SVs. Regions with clear signatures typically correspond to cases where variant signals are consistent and well aligned, usually contain only one or two consistent and unambiguous signatures. For example, within such a region, most reads exhibit the same signature type (e.g. DEL signal), with highly similar breakpoint positions and comparable estimated lengths across reads. So we utilize frequency-based clustering to determine the subtype or genotype of SVs within these regions. Specifically, signatures with the same type, similar length and position are grouped together. By analyzing the number of supporting reads, we determine their genotypes and then make a final call of their corresponding SV. For reads in regions with complex and ambiguous signatures, the alignment results are often too chaotic to discern what is happening in the regions. For example, different reads may support different signature types (e.g. DEL, INS, or inversion signals), and even within the same signature type, the inferred breakpoint positions and lengths may vary substantially across reads, indicating a lack of consistent dominant pattern. Therefore, we use assembly and alignment techniques to help figure out the sequence structure. But directly assembling reads corresponding to different subtypes together may lead to problematic consensus that does not reflect the features of any subtype. To solve this problem, we cluster the reads in such regions first. Given that SVs usually cover many bases, if multiple subtypes of SVs exist, the reads corresponding to different subtypes exhibit significant deviations within the region. However, if only one subtype exists, the differences among reads are mainly due to other factors, such as sequencing or mapping errors. Importantly, we prefer only one subtype in one region, which is usually the case. This task differs somewhat from a classical clustering problem but fits within the framework of a statistical hypothesis test. The null hypothesis is that a pair of reads are from the same subtype, versus the alternative hypothesis that they are not from the same subtype. To construct a test statistic, we directly calculate the differences between each pair of reads. The difference is based on the Levenshtein distance [15] within the candidate region. Specifically, we can apply a user-defined threshold to give the critical region. Then we select all the reads that appear in that region and arrange them in order of their lengths, from the longest read to the shortest read. We use the longest read as a representative sequence and test the similarity between each read and the representative. If the similarity score between a read and the representative sequence of a subtype falls below a user-defined threshold for the first time, we assume that the null hypothesis is invalid and that this read belongs to a different subtype. Then we create a new subtype and use this read as a representative sequence. Subsequent reads will be compared for similarity with representative sequences of different subtypes and assigned to the category with the highest similarity score. This algorithm runs efficiently with a time complexity of O(), where is the number of reads in the candidate region. The Levenshtein distance, also known as the edit distance, is used to measure the difference between two sequences. Since sequences of different lengths will inevitably result in lower similarity scores, we use the left endpoint of the candidate region as the starting point to extract sequences of the same length for comparison. If there are reads that do not cover the left endpoint of the region, we will use the start position in the reference corresponding to the later read as the starting point to extract the sequence. Assembling For reads within a cluster, we utilize the assembly tool wtdbg2 [12] to generate a consensus. This assembling step enables us to obtain a longer contiguous sequence that covers the combined information from the reads within the cluster. wtdbg2 is also used in DeBreak [6] to detect long insertions by assembling the reads around them. Please note that assembly is usually computationally expensive, which makes these assembly-type methods unpopular in SV detection. But assembly indeed provides an option to recover the sequence information for comprehensive checking, especially when the alignment information is complex. Thus how to economically use it is the key issue here. Rather than using the assembly on raw reads, we apply wtdbg2 only in the candidate region. In this way, we take advantage of the assembly-based methods and save unnecessary computation at the same time. Moreover, most SVs may be in a simple format, such as DELs, INSs, INVs, and DUPs. We also provide the option to skip the regions with simple SV signatures and only carry out the assembly on the candidate regions with complex SV signatures, which further reduces the computational burden. Realigning We align the assembled consensus back to the reference genome to determine the variants in the cluster. In this step, we note that SVDSS [8] also utilizes a similar procedure, but the purpose is to detect DELs and INSs due to the use of an improper alignment tool. It is important to mention that, we only have one consensus for each cluster. Common alignment tools like minimap2 [13] and NGMLR [3] aim to minimize the overall alignment loss by mapping the sequence to a specific part of the reference. Due to their mechanism, these tools may not achieve the minimum loss. In our case with only one consensus, we cannot afford such an error that may completely disrupt detection. Thus, we adopt the idea of MEM for the realigning step. MEM identifies and reports the MEMs between the consensus and the reference while disregarding some gaps. This approach, despite yielding some gaps and multiple possible matches, keeps as much information as possible and allows us to infer the differences between the consensus and the reference more accurately. Moreover, MEM also enables us to capture information about potential complex SVs. In this paper, we utilize the copMEM2 [14] tool to carry out the MEM alignment. Finalizing In the finalizing step, the alignment results are transformed into the final SV callings. This step is similar to the encoding step, but with additional considerations for multiple possible matches and gaps therein. When using MEM, both forward matches and inversion matches are reported, along with their positions in the consensus and reference sequences, as well as the match lengths. Due to various factors such as sequencing errors, mapping errors, SNPs, and small indels, gaps may be present in alignment results, apart from the SVs. These gaps are typically short, covering only a few bps. We will skip the gap (<50 bps) not introduced by SVs, and convert inversion matches to the forward sequence. Subsequently, final SVs are determined based on the aforementioned encoding step. It is important to note that the final SVs may not necessarily fall into common SV types, such as DELs and INSs. Therefore, we infer complex SV types directly from the MEM results. Output format The detected SVs are reported in Variant Call Format (VCF). Two output files are generated: gSV_complex.vcf, which reports only complex SVs as unified events, and gSV.vcf, which reports both simple SVs and the decomposed components of complex SVs for compatibility with downstream analyses. For complex SVs, multiple events are encoded within a single record using a composite SVTYPE field, where each component specifies the variant type, genomic region, and length, and different components are concatenated using the “+” delimiter. For example: #CHROM POS ID REF ALT INFO chr1 17251 SV1 T END=18509;SVTYPE=COMPLEX:DUP:17251-17759-508+INV:17251-18509-1259;SVLEN=1259 In addition to the unified representation, each component event of a complex SV can also be reported separately as simple SV records in gSV.vcf. For example: #CHROM POS ID REF ALT INFO chr1 17251 SV1 T END=17759;SVTYPE=DUP;SVLEN=508 chr1 17251 SV2 T END=18509;SVTYPE=INV;SVLEN=1259 The above example only shows partial output information (lacking QUAL, FILTER, SUPPORT, etc.). The complete VCF output results are shown in Supplementary Note 1.4. Performance evaluation We use both simulated and real datasets to comprehensively assess SV detection performance. We compare gSV with five state-of-the-art SV callers: PBSV [2], Sniffles [16], cuteSV [5], DeBreak [6], and SVision-pro [10]. First, we conduct benchmark comparisons using simulated data, which provides controlled scenarios to quantify recall and precision across different SV types and complexities. Next, we perform benchmark comparisons using a real HG002 dataset, whose ground truth includes high-confidence INSs and DELs. To extend the evaluation across diverse SVs and real-world genomic complexity, we further test our method on additional datasets, including family trios for Mendelian inheritance validation [17], whole-genome sequencing of breast cancer cell lines [18], and population-scale breast cancer cohorts for clinical utility analysis. However, as these data lack a comprehensive ground truth, we perform manual inspection using the Integrative Genomics Viewer (IGV) to confirm alignment patterns, breakpoint support, and concordance with other evidence (e.g. read-depth signals). Simple structural variants and complex structural variants benchmark We use simulated data and HG002 as benchmarks for simple SVs to evaluate the performance of different methods. For simulated data, we utilize the simulator VarSim [19] to randomly insert variants from the Database of Genomic Variants (DGV) (2675 INSs, 4972 DELs, 1409 DUPs, and 362 INVs) into the reference genome (GRCh37). Then we utilize PBSIM2 [20] to simulate Continuous Long Reads (CLRs) at sequencing depths of 20, 30, and 40. For HG002 (30), we use the BAM file obtained by aligning SMRT sequencing data to the hs37d5 reference using pbmm2 (the minimap2 wrapper), and we utilize the Tier 1 benchmark set (including highly confident INSs and DELs) as the ground truth. The size distributions of SVs in simulated data and HG002 are shown in Supplementary Fig. 1. We use simulated complex SVs as the benchmark to evaluate the performance of different methods. Based on previous studies on complex SVs [1, 21], we classified complex SVs into five major categories with 12 subclasses according to their structural composition and simulated them (shown in Fig. 2). The first three combinations (ID 1–3) belong to one category, formed by linking adjacent simple subsidiary SVs together. The last two combinations (ID 4 and 5) belong to the second category, composed of overlapping and nested sub-SVs. Following the idea of creating complex SVs used in Sniffles and SVision, we utilize VISOR [22] to construct complex SVs and PBSIM2 to simulate reads at a depth of 20. Diagrams of simulated complex SVs. IDs 1–3 are constructed by linking adjacent simple sub-SVs. IDs 4–5 are composed of overlapping or nested sub-SVs. We use Truvari [23] to evaluate the results of gSV, PBSV [2], Sniffles [3], cuteSV [5], DeBreak [6], and SVision-pro [10] on simple and complex SVs. For complex SVs, we evaluate the different tools from both the regional and type perspectives. We first focus on evaluating whether the tools can identify the regions correctly. Since SVision-pro is the only tool capable of reporting complex SVs, we compare our gSV exclusively with it. We extract the start and end positions of each complex SV and use Truvari to calculate recall, precision, and F1 scores. Furthermore, we evaluate the detection results from the perspective of SV subtypes. Since all tools, other than SVision-pro, can only detect the parts of complex SVs, we divide each complex SV into its constituent simple subtype SVs, which form a new ground truth dataset. In other words, we examine each subcomponent of complex SVs for position, size, and type. For PBSV, Sniffles, cuteSV, and DeBreak, we directly compare the detection results with the split ground truth set. For gSV and SVision-pro, we split the detected complex SVs and merge them with the detected simple SVs to perform the evaluation. Mendelian consistency analysis For real data lacking ground truth, traditional metrics (e.g. recall, precision, and F1-score) cannot be directly applied to evaluate the detection results. Here we leveraged Mendelian consistency in family trios as a surrogate validation framework, where fewer observed Mendelian inheritance violations directly correlate with reduced false-positive rates. Mendelian consistency is a statistical measure used to evaluate whether observed SVs in a family trio (e.g. parents and offspring) adhere to Mendelian inheritance laws. It is typically calculated as the proportion of SVs in offspring that are genetically compatible with parental genotypes: (Total SVs−Mendelian violations SVs)/ Total SVs. We performed Mendelian consistency analysis on HiFi sequencing data from four previously published trios: the Chinese Trio (HG005/ HG006/ HG007; 50.6/43.1/40.2), the Yoruba Trio (NA19240/ NA19239/ NA19238; 30.4/ 27.5/ 26.8 ), the Southern Han Chinese Trio (HG00514/ HG00512/ HG00513; 26.2/30.8/45.3), and the Puerto Rican Trio (HG00733/ HG00731/ HG00732; 35.5/ 35/ 21.4), where sequencing depths are reported in the order of child/ father/ mother. We use minimap2 to align the sequencing reads to the GRCh38 reference genome. Five variant callers are employed: gSV, cuteSV, DeBreak, SVision-pro, and Sniffles. For the three callers without built-in merging capabilities (gSV, cuteSV, and DeBreak), we first performed independent variant calling for each sample, and then compared the results of two merging methods (Jasmine [24] and SURVIVOR [25]). Since Jasmine and SURVIVOR cannot effectively handle complex SVs, the complex SVs detected by gSV were converted into their constituent simple subtype SVs and merged together with other simple SVs. To eliminate the impact of de novo SVs, we retained only SVs detected in at least two samples within each trio (e.g. parent–parent or parent–child pairs). Finally, we used the Mendel plugin in BCFtools to assess Mendelian consistency. Tumor-normal paired cell line structural variants analysis To explore novel SV loci and complex structures, we examine four open-source breast cancer tumor-normal paired cell line datasets: HCC1395 (62.2/43.7), HCC1937 (60.6/56.4), HCC1954 (63.8/63.2), and HG008 (83.8/37.5), where sequencing depths are reported in the format of tumor/normal. Sequencing reads are aligned to the GRCh38 reference sequence using minimap2. And we use gSV, cuteSV, DeBreak, SVision-pro, and Sniffles for SV detection. Since most tools, except gSV and SVision-pro, do not report complex SVs, the complex SVs detected by gSV and SVision-pro were divided into their constituent simple subtype SVs to enable consistent and accurate comparisons across all callers. Due to the lack of ground truth, we performed manual inspection using IGV to visually inspect our unique findings. Given the impracticality of manually inspecting >20 000 SVs typically detected per sample, we implement a pipeline to prioritize candidate SVs. First, we analyze SV callsets across multiple tools using Jasmine to compare the overlap of detected SVs and identify the subset of tools supporting each SV. Next, we annotate all SVs with ANNOVAR [26] to prioritize functionally relevant variants (e.g. genes with potential clinical implications, exonic regions). Finally, we focus validation efforts on SVs uniquely identified by gSV to assess its ability to detect novel or challenging SVs missed by existing methods. In-house breast cancer data analysis We further analyze an in-house long-read sequencing dataset acquired at the People’s Hospital at Peking University from 234 samples (188 from breast cancer patients and 46 from individuals without a current breast cancer diagnosis), with sequencing depths ranging from 60 to 200. Focusing on breast cancer, we develop a 28-related-gene (more details, see Supplementary Table 1) panel for Pacific Biosciences platform-based sequencing. We apply gSV together with PBSV, Sniffles, cuteSV, DeBreak, and SVision-pro to detect SVs from each sample. And the complex SVs detected by gSV and SVision-pro were divided into their constituent simple sub-type SVs to enable consistent comparisons across all callers. We first filter out the detected SVs shorter than 50 bps. For comparison, an important step is to determine the intersections of the SVs detected by the tools. Given that the SVs detected by different tools may be slightly different in terms of position and/or length, we give some tolerance as follows: where and represent the start positions for the detected SVs, and and represent their lengths. For each type of SV, we go through each pair of candidate SVs from the different tools and determine whether they are identical. Then, across samples, we follow a similar procedure to determine whether an identical SV exists in different samples. After that, we performed manual inspection using IGV to confirm our unique findings. Results and discussion In this section, we present comparative results in different data sets and discuss the possible biological and clinical significance of the unique findings of gSV in breast cancer data. Performance evaluation of simple and complex structural variants in datasets with known ground truth For simple SVs, as shown in Fig. 3a, gSV outperforms the other tools for all read depths in terms of recall, precision, and F1-score. Figure 3b compares the performance of different tools across different types of simple SVs. For INSs, the alignment pattern is relatively simple, so most tools show comparable performance. For DELs, gSV shows a 0.7% increase in F1-score compared to the best-performing existing method, while larger differences are observed for DUPs and INVs (3.1% and 4.0%, respectively). By using the bootstrap method with details shown in Supplementary Note 1.5, we show that the three 95% confidence intervals of F1-score differences do not cover 0, indicating that the improvement is statistically significant. One possible explanation is that the alignment patterns of these SV types exhibit more complex scenarios, as illustrated in Fig. 3d and e, where both IGV screenshots and realignment results based on the assembled sequences are provided. Specifically, panel (d) shows a DEL where the corresponding region does not exhibit a clean DEL pattern but displays a discontinuous alignment with numerous mismatches. It is worth noting that the MEM results reveal a clear DEL signal spanning the affected region. Panel (e) shows a large-scale DUP where the variant signal is not directly discernible in most reads. Only a few of the split-read alignments exhibit partial truncation on the right flank, indicative of the DUP breakpoint. In contrast, the assembled sequence alignment clearly supports the DUP structure. Additionally, there is a pronounced elevation in read depth across the duplicated region. These patterns are different from those expected for simple DELs or DUPs in alignment models, rendering them undetectable by existing tools. This further demonstrates the unique capability of our method to resolve SVs with complex alignment patterns. In addition, we apply gSV to the HG002 dataset (only containing INSs and DELs) from the Genome in a Bottle (GIAB) consortium [27], and find that gSV performs similarly well to the benchmarking tools for INSs and DELs (Fig. 3c), consistent with the results shown in panel (b). In addition, we note that Oxford Nanopore Technologies (ONT) data are known to exhibit elevated indel error rates, particularly in homopolymer regions (e.g. consecutive identical bases such as “AAAA”). To evaluate the robustness of our method under such conditions, we further assessed its performance on the ONT HG002 dataset. Despite the presence of homopolymer-associated errors, gSV achieves a precision of and a recall of . We observe that, compared with PacBio HiFi data, slightly larger values of hyperparameters and are required to effectively suppress noise introduced by such sequencing errors. This adjustment helps reduce spurious candidate regions caused by indel-rich signals while preserving true SV signals. Panel (a) shows the performance of different methods in detecting simple SVs in terms of recall, precision, and F1-score. gSV uniformly performs better in all depths. Panel (b) compares the F1-scores of different methods for detecting specific types of simple SVs. gSV demonstrates a more pronounced advantage in detecting DUPs and INVs compared with existing methods. Panel (c) evaluates the performance of different methods in detecting simple SVs (INS and DEL) on the HG002 benchmark dataset. gSV performs similarly well to the benchmarking tools in INSs and DELs, consistent with the results shown in Panel (b). Panels (d) and (e) show the IGV screenshot of a DEL and a DUP uniquely detected by gSV, together with the corresponding alignment plots of assembled sequences. Both variants exhibit complex alignment patterns, which prevented their detection by other tools. These MEM-based alignments of the assembled sequences clearly reveal the underlying structural patterns, thereby explaining why these variants can be detected by gSV. (The contents in parentheses indicate their identifiers in the DGV). For complex SVs, the simulations were conducted at a sequencing depth of 20, which provides moderate coverage for complex SV detection. The left panel of Fig. 4 shows that gSV achieves substantially higher accuracy than existing tools in both breakpoint localization and SV subtype determination. More results for other types of simulated complex SVs are shown in Supplementary Fig. 2. The right panel shows the numbers of complex SVs detected by SVision-pro and gSV across different types, with red representing gSV and blue representing SVision-pro. For the DUP+DEL (ID2), as well as the highly complex DUP+INV (ID4) and DUP+INV+DEL (ID5) types that involve overlapping or nested sub-SVs, SVision-pro detects markedly fewer complex SVs than gSV. Among the detected complex SVs, SVision-pro, and gSV show comparable positional precision, as shown in Supplementary Fig. 2. In absolute terms, however, gSV identifies more true complex SV loci, and when complex SV subtype accuracy is considered, gSV exhibits a clear advantage. For the INV+DEL (ID1) and DUP+DEL+INV (ID3) categories, the numbers of complex SVs detected by SVision-pro and gSV differ (striped bars), but only to a limited extent. Notably, for ID2 and ID3, gSV shows a higher rate of false positives. This is likely due to the specific structure of this SV type, where a DUP and a DEL occur adjacently. The DUP introduces an additional sequence into the read, while the DEL removes a segment from the reference. During alignment, the extra sequence introduced by the DUP may be incorrectly aligned—often with low alignment quality—to the region corresponding to the DEL, particularly when the lengths of the DUP and DEL are similar. This mismatch between inserted and missing segments can lead to ambiguous alignments and generate spurious SV signals, thereby increasing the false positive rate. Performance evaluation of simulated complex SVs (each simulated at 20 depth). The left panel shows the performance of different methods in detecting one example type (ID4: DUP–INV) of complex SV by evaluating the region and type, respectively. gSV demonstrates better performance resolving these complex cases, underscoring its capability to accurately detect and characterize multiple breakpoints or nested SVs, which are often missed by conventional detection tools. The right panel shows the numbers of complex SVs detected by SVision-pro and gSV across different types (striped bars), and the corresponding numbers of true positive SVs (solid bars). To mitigate false positives, users can apply stricter detection parameters. For example, requiring a larger minimum number of supporting reads or imposing higher mapping quality thresholds can help reduce spurious calls. Moreover, integrating external annotation information can further improve precision. For instance, SVs located in structurally challenging regions, highly repetitive sequences, homopolymers or microsatellite tracts, nested or overlapping SVs, and regions near telomeres or centromeres, are more prone to alignment ambiguity and can be optionally removed to reduce false positives. In the future, improvements could focus on addressing alignment ambiguities arising from highly repetitive regions or from adjacent and compensatory SV patterns. For example, refining realignment strategies to better distinguish between inserted sequences and alignment gaps may help reduce spurious signals. In addition, incorporating long-range or haplotype-aware information could further improve the accurate characterization of such variants. Mendelian consistency assessment in family trios Figure 5a shows the Mendelian consistency of different tools in four family trio data. We can see that gSV demonstrates greater Mendelian consistency compared with other tools, reflecting its higher precision in SV detection. Figure 5b illustrates a Venn diagram that further reveals the overlap of Mendelian-consistent SVs in the Chinese trio across different tools. Only 10.43% of SVs are shared by all methods, representing a set of high-confidence variants. Approximately 70% of SVs are detected by at least two tools, indicating strong concordance across methods. Notably, gSV identifies 1660 unique Mendelian-consistent SVs (6.18%), which may represent additional true variants missed by other tools. The partial overlap among tools also highlights their complementarity in SV detection. Detailed results for other trios are provided in the Supplementary Fig. 3. Mendelian consistency analysis. (a) Overall Mendelian consistency of different SV detection tools across four family trios. gSV demonstrates higher Mendelian consistency compared with other tools. (b) Venn diagram showing the overlap of Mendelian-consistent SVs detected by different tools in the Chinese trio(HG005/HG006/HG007, 50.6/43.1/40.2). Structural variants analysis in breast cancer tumor-normal cell lines Following the analytical framework in the Performance Evaluation subsection of the Methods section, we obtain the gSV-unique-detected SVs and then perform literature validation to corroborate their biological relevance. Figure 6a demonstrates a somatic DEL in HCC1954-tumor only detected by gSV, located within the 5-Hydroxytryptamine Receptor 1A (HTR1A) exonic region. Published studies [28] indicate that HTR1A significantly inhibits triple-negative breast cancer (TNBC) cell development in vivo and in vitro through downregulation of both canonical and noncanonical TGF- pathways. Notably, the expression level of HTR1A was significantly lower in TNBC tissues compared with that in paracancerous tissues, and knockdown of HTR1A significantly enhanced the migration and invasion of MDA-MB-231 and Hs578T breast cancer cell lines. Figure 6b demonstrates a germline DUP in the HCC1937 cell line exclusively identified by gSV. This SV resides within the FLG exonic region, which encodes filaggrin. The partial exonic DUP disrupts the filaggrin architecture, potentially leading to skin barrier function, which may lead to higher cancer susceptibility [29]. Another studies show that FLG were found amplified in of breast tumors [30]. Additional gSV-unique-detected SVs localized to exonic regions of breast cancer-associated genes are provided in Supplementary Figs 5–12. Furthermore, Supplementary Figs 13–15 show exon-proximal upstream SVs in breast cancer-associated genes. Although these non-coding alterations do not directly disrupt core coding sequences or protein domains, they may alter the binding sites of transcription factors. Such regulatory perturbations could drive oncogenic progression through dynamic upregulation or downregulation of critical gene expression programs. Supplementary Figs 16–18 show gSV-unique-detected SVs in exonic regions of genes not currently linked to breast cancer. These findings provide potential candidates for investigating SVs in cancer genomics, even in the absence of established oncogenic associations. Panels (a) and (b) show IGV screenshots of SVs uniquely detected by our gSV in HCC1954 and HCC1937, with sequencing depths of 63.8/63.2 (tumor/normal) for HCC1954 and 60.6/56.4 for HCC1937. Panel (c) shows a DUP (Chr7:DUP at EGFR in chromosome 7: 55167022-55167776) uniquely detected by gSV in six breast cancer patients, which shows an American College of Medical Genetics and Genomics (ACMG) class of 3 (unclear breast cancer risk). Structural variants analysis of in-house breast cancer data In this in-house dataset, we analyzed targeted sequencing data of 28 breast cancer-related genes from 234 samples, including 188 breast cancer patients and 46 individuals without a breast cancer diagnosis. Under this setting, gSV detects eight unique SVs that are not detected by the other tools. We show one such example of a DUP in EGFR in Fig. 6c. It shows an American College of Medical Genetics and Genomics (ACMG) class of 3, which means an unclear risk related to breast cancer [31]. The same SV is detected in five other patients but not in healthy people, indicating a possible association with breast cancer. Details of the remaining seven gSV-unique-detected SVs are provided in Supplementary Table 2. Although these variants are located in non-exonic regions, their translational implications warrant further investigation, as they may reside in critical regulatory elements influencing gene expression, splicing mechanisms, or epigenetic modifications that could underlie phenotypic variability or disease susceptibility. Effect of and on structural variant detection performance To investigate the impact of hyperparameters and on the graph-cut step and overall SV detection performance, we systematically evaluated nine parameter combinations (, ) on both simple and complex simulated datasets (chromosome 1, 20 coverage). Performance was assessed using precision, recall, F1 score, and the number of SV candidate regions generated by the graph-cut procedure. Notably, the number of candidate regions is positively correlated with runtime. The results are summarized as heatmaps in Fig. 7. controls the threshold for selecting candidate regions; larger values substantially reduce the number of candidate regions, resulting in improved precision but decreased recall, consistent with stricter filtering of candidate SV signals. The parameter balances the relative contributions of the data term and the smoothness term in the graph-cut model. Increasing enforces stronger smoothness, leading to fewer candidate regions. This reduces false positives but may increase false negatives. In practice, for noisier sequencing data (e.g. nanopore data, which typically has a higher error rate than PacBio), larger values of and are preferred to suppress spurious signals and improve robustness. Effects of and on SV detection performance. Heatmaps showing precision, recall, F1-score, and the number of candidate regions under different combinations of and on simulated datasets (chromosome 1, 20 coverage). Results are presented for both simple (first row) and complex SV scenarios (second row). Increasing leads to fewer candidate regions and a clear trade-off between precision and recall, with higher precision but lower recall. In contrast, increasing enforces stronger smoothness in the graph-cut model, reducing candidate regions and false positives at the cost of increased false negatives. In addition, the number of candidate regions decreases substantially with increasing and , which reduces the computational burden and consequently shortens the runtime. Comparison of maximum exact match and conventional aligners for complex structural variants To evaluate the effectiveness of MEM-based realignment in resolving complex SVs, we performed a comparison on simulated datasets containing five different types of complex SV events. For each genomic region containing complex SVs, reads were first assembled into contigs using wtdbg2, followed by realignment using MEM, minimap2, and NGMLR, respectively. Breakpoint accuracy was assessed by comparing the predicted breakpoints from each method to the ground truth annotations within the same genomic region. As shown in Fig. 8, MEM-based realignment achieved higher breakpoint accuracy compared with conventional long-read aligners. A representative example (DUP+DEL) is shown in Fig. 8b and c, where minimap2 and NGMLR produce a series of consecutive mismatches rather than correctly resolving the underlying SV. In contrast, MEM-based realignment clearly reveals the SV structure: the DUP is indicated by the orange segment sharing identical reference coordinates with the blue segment, while the DEL is evidenced by a gap on the reference between the green and orange segments, despite their adjacency on the read. Comparison of realignment tools. (a) Breakpoint accuracy of MEM, minimap2, and NGMLR across simulated genomic regions containing five types of complex SVs (ID1–ID5). Panels(b) and (c) show the realignment results of the three methods on assembled reads within a genomic region containing a DUP+DEL event. Both minimap2 and NGMLR only exhibit consecutive mismatches without clear SV signals. In contrast, the MEM realignment reveals distinct SV patterns: the leftmost and second orange and blue segments share identical reference coordinates, indicating a DUP event, while the second and third segments are adjacent on the read but separated by a gap on the reference, indicating a DEL event. Computational efficiency To assess the computational efficiency of our pipeline under different configurations, we evaluated the runtime and memory usage across modes. Figure 9 illustrates the chromosome-level assembly runtime and peak memory usage in HCC1937, comparing the default (normal) mode and the complex mode, where all candidate regions are forced into the complex branch. The runtime (s) and peak memory usage (GB) for chromosome-level assembly in HCC1937, comparing the default mode and the complex mode, where all candidate regions are forced into the complex branch. Each chromosome was processed independently, enabling parallel execution. Overall, the total assembly runtime in the default mode is 36% of that in the complex mode, while the peak memory usage for both modes is roughly same across all chromosomes. The runtime is primarily driven by the number of candidate regions entering the assembly step, indicating that computational cost is highly dependent on complex region selection. On average, the assembly time per region is 1.6 s (as shown in Supplementary Fig. 19). For most chromosomes, runtime scales approximately proportionally with chromosome length, suggesting a stable computational behavior. However, an exception is observed for chromosome 16 in HCC1937, where the runtime is significantly higher than expected. This is due to a small number of highly complex regions, where the assembly of a single region can exceed 30 min, thereby substantially increasing the total runtime. Notably, this phenomenon appears to be sample-specific, as it is only observed in chromosome 16 of HCC1937 but not observed in HCC1954 (as shown in Supplementary Fig. 19). Experiments were conducted on a Supermicro server (88 CPU cores, 256 GB RAM), with parallel execution configured to process up to six chromosomes simultaneously. The total runtime of the normal mode is 11.7 h, comprising encoding ( 2.6 h, 22.4%), detection (graph-cut, 0.78 h, 6.7%), and the complex branch (clustering, assembly, realignment, and finalization; 8.29 h, 70.9%). In contrast, the complex mode requires 23.5 h, with encoding ( 2.6 h, 11.2%), detection ( 0.75 h, 3.2%), and the complex branch accounting for the majority of the runtime ( 20.13 h, 85.6%). Based on these observations, we recommend using the default normal mode for most regions to maximize computational efficiency. For specific regions of interest that are expected to be highly complex, the complex mode can be applied to enhance sensitivity and avoid missing complex SVs. Additionally, users can adjust the level of parallelization according to the available memory resources on their system. Conclusion Our proposed gSV inherits high sensitivity from alignment-based methods and high specificity from assembly-based methods. The inclusion of MEM further improves its capability in capturing detailed and complex signals. This hybrid design enables gSV to detect both simple SVs and complex SVs formed by non-prespecified combinations of basic SV types. Benchmark evaluations on simulated data with ground truth show its superior performance in recall and precision compared with other existing popular tools. When we apply the new tool to detect SVs in public datasets and in-house breast cancer dataset, we obtain promising findings not detected by existing tools. Overall, gSV provides a general and easy-to-interpret pipeline for detecting SVs. Key Points Existing structural variant (SV) detection tools are limited in resolving SVs with complex alignment patterns due to their reliance on predefined variant models. Existing structural variant (SV) detection tools are limited in resolving SVs with complex alignment patterns due to their reliance on predefined variant models. General structural variants (gSV) integrates alignment-based and assembly-based evidence using a maximum exact match strategy, enabling capture of diverse and complex SV signals. General structural variants (gSV) integrates alignment-based and assembly-based evidence using a maximum exact match strategy, enabling capture of diverse and complex SV signals. Benchmarking on simulated and real long-read sequencing datasets demonstrates that gSV achieves competitive performance on canonical SV classes and improved sensitivity for complex SV patterns. Benchmarking on simulated and real long-read sequencing datasets demonstrates that gSV achieves competitive performance on canonical SV classes and improved sensitivity for complex SV patterns. Application of gSV to breast cancer cell lines and a population-scale breast cancer cohort reveals previously unresolved SVs in cancer-associated genes, highlighting its utility in genomic and clinical studies. Application of gSV to breast cancer cell lines and a population-scale breast cancer cohort reveals previously unresolved SVs in cancer-associated genes, highlighting its utility in genomic and clinical studies. Supplementary Material Contributor Information Jingyu Hao, Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, 999077, China. Jiandong Shi, Department of Statistics and Data Science, The Chinese University of Hong Kong, Shatin, New Territories, Hong Kong SAR, 999077, China. Sheng Lian, Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, 999077, China. Zhen Zhang, Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, 999077, China. Yongyi Luo, Department of Statistics and Data Science, The Chinese University of Hong Kong, Shatin, New Territories, Hong Kong SAR, 999077, China. Taobo Hu, Department of Breast Surgery, Peking University People’s Hospital, No. 11, Xizhimen South Street, Xicheng District, Beijing, 100044, China. Toyotaka Ishibashi, Division of Life Science, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, 999077, China. Depeng Wang, GrandOmics Inc, Zhongguancun Life Science Park, Changping District, Beijing, 102206, China. Shu Wang, Department of Breast Surgery, Peking University People’s Hospital, No. 11, Xizhimen South Street, Xicheng District, Beijing, 100044, China. Xiaodan Fan, Department of Statistics and Data Science, The Chinese University of Hong Kong, Shatin, New Territories, Hong Kong SAR, 999077, China. Weichuan Yu, Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, 999077, China. Author contributions J.H. and J.S. developed and implemented the analysis algorithms, carried out data analysis, and drafted the paper. S.L., Z.Z., and Y.L. helped improve the analysis methods. S.W., X.F., T.I., and W.Y. conceived the idea, designed, and supervised the study. T.H. and D.W. coordinated sample collection and data acquisition. All authors helped to proofread and revise the manuscript. Conflicts of interest None declared. Funding The work was supported by grants 7015-23G, T12-101/23-N, R4012-18, MHP/033/20, and ITS/043/23 from the HKSAR government, internal grants 3030_009, BGF.001.2023, CSSET24SC01, OKT26EG06, and Z1056 from HKUST, and the National Key Research and Development Program of China (grant no. 2021YFE0203200). Data and Code availability Sample HG002, HG008, HCC1395, HCC1937, and HCC1954, have been consented for commercial use and for public posting of Personally Identifying Genetic Information (PIGI), which allows open-access, public posting of extensive genetic data. The studies involving human participants related to breast cancer data were reviewed and approved by Peking University People’s Hospital ethics committee. The patients/participants provided their written informed consent to participate in this study. HG002 data were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/AshkenazimTrio/HG002_NA24385_son/PacBio_CCS_15kb/alignment/. The Tier1 benchmark SV callset and high-confidence HG002 region were obtained from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/AshkenazimTrio/analysis/NIST_SVs_Integration_v0.6/. HG008 normal-tumor paired cell data were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/PacBio_Revio_20240125/. HCC1395 normal-tumor paired cell data were downloaded from https://downloads.pacbcloud.com/public/revio/2023Q2/HCC1395/. HCC1937 normal-tumor paired cell data were downloaded from https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305185&display=metadata and https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305182&display=metadata. HCC1954 normal-tumor paired cell data were downloaded from https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305163&display=metadata and https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305160&display=metadata. The Chinese Trio were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/ChineseTrio/. The Yoruba Trio, the Southern Han Chinese Trio, and the Puerto Rican Trio were downloaded from http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/HGSVC2/working/. The human reference genome GRCh37 was downloaded from http://ftp-trace.ncbi.nih.gov/1000genomes/ftp/technical/reference/phase2_reference:assembly_sequence/. The human reference genome GRCh38 was downloaded from https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/release/references/GRCh38/. Sequencing data for in-house breast cancer samples have been deposited in the Genome Sequence Archive for Human (GSA-Human) at the National Genomics Data Center (NGDC) under accession codes HRA003120, HRA003172, and HRA008465. gSV is publicly available at https://bioinformatics.hkust.edu.hk/Software.html or https://github.com/jhaoae/gSV. References Associated Data Supplementary Materials Data Availability Statement Sample HG002, HG008, HCC1395, HCC1937, and HCC1954, have been consented for commercial use and for public posting of Personally Identifying Genetic Information (PIGI), which allows open-access, public posting of extensive genetic data. The studies involving human participants related to breast cancer data were reviewed and approved by Peking University People’s Hospital ethics committee. The patients/participants provided their written informed consent to participate in this study. HG002 data were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/AshkenazimTrio/HG002_NA24385_son/PacBio_CCS_15kb/alignment/. The Tier1 benchmark SV callset and high-confidence HG002 region were obtained from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/AshkenazimTrio/analysis/NIST_SVs_Integration_v0.6/. HG008 normal-tumor paired cell data were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/PacBio_Revio_20240125/. HCC1395 normal-tumor paired cell data were downloaded from https://downloads.pacbcloud.com/public/revio/2023Q2/HCC1395/. HCC1937 normal-tumor paired cell data were downloaded from https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305185&display=metadata and https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305182&display=metadata. HCC1954 normal-tumor paired cell data were downloaded from https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305163&display=metadata and https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=SRR28305160&display=metadata. The Chinese Trio were downloaded from https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/data/ChineseTrio/. The Yoruba Trio, the Southern Han Chinese Trio, and the Puerto Rican Trio were downloaded from http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/HGSVC2/working/. The human reference genome GRCh37 was downloaded from http://ftp-trace.ncbi.nih.gov/1000genomes/ftp/technical/reference/phase2_reference:assembly_sequence/. The human reference genome GRCh38 was downloaded from https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/release/references/GRCh38/. Sequencing data for in-house breast cancer samples have been deposited in the Genome Sequence Archive for Human (GSA-Human) at the National Genomics Data Center (NGDC) under accession codes HRA003120, HRA003172, and HRA008465. gSV is publicly available at https://bioinformatics.hkust.edu.hk/Software.html or https://github.com/jhaoae/gSV.