FocalSV enables target region–based structural variant assembly and refinement using single-molecule long-read sequencing data

Evaluation of SV calls in FocalSV and comparisons with existing SV callers

To assess the performance of insertion and deletion SV detection, we evaluated five assembly-based tools (FocalSV, PAV, SVIM-asm, Dipcall, and sawfish) and five alignment-based tools (cuteSV, SVIM, pbsv, Sniffles2, and SKSV) using nine long-read sequencing libraries of the HG002 sample. For FocalSV, we evaluated two operational modes: FocalSV(target), which uses predefined regions, and FocalSV(auto), which automatically identifies and expands regions with potential SVs. SV calls were benchmarked against the Genome in a Bottle (GIAB) SV gold standard (Zook et al. 2014) using Truvari (v4.0.0) (English et al. 2022), a widely adopted structural variant evaluation tool. Truvari compares SV calls from any tool's call set against a gold standard call set, both provided in Variant Call Format (VCF) files, by analyzing key metrics: reference distance, reciprocal overlap, size similarity, and sequence similarity.

In this study, we applied a moderate-tolerance parameter set in Truvari for SV comparisons, with the following settings: p = 0.5, P = 0.5, r = 500, and O = 0.01. Specifically, the parameter p (pctstim), which ranges from 0 to 1.0, controls the minimum required sequence similarity for two SVs to be considered identical. P (pctsize), also ranging from 0 to 1.0, defines the minimum allowable allele size similarity. O (pctovl), ranging from 0 to 1.0, establishes the minimum reciprocal overlap ratio, a key measure for comparing deletions and evaluating their breakpoint alignment. Lastly, r (refdist), which can range from 0 to 1000 bp, sets the maximum allowable difference between the reference positions of two SVs, helping assess breakpoint shifts in insertion events.

We first evaluated the average performance across different PacBio Hifi, CLR, and ONT data sets (Tables 2–4). In the HiFi data sets (Table 2), FocalSV(target) demonstrated the third-highest average F1 score for deletions (93.99%) and the highest for insertions (91.93%). FocalSV(auto) attained the highest average F1 score for deletions (94.29%) and the second-highest for insertions (91.03%). In the CLR data sets (Table 3), FocalSV(target) outperformed the other tools with the highest average F1 scores for deletions (92.95%) and insertions (90.71%). FocalSV(auto) ranked second, with comparable performance for deletions (92.92%) and slightly lower performance for insertions (89.16%). In the ONT data sets (Table 4), FocalSV(target) achieved the highest average F1 score for deletions (92.94%), and ranked third for insertions (89.69%). FocalSV(auto) ranked second in average F1 score for deletions (92.50%), and achieved the highest F1 score for insertions (90.01%). These results highlight the robustness of both FocalSV modes across different sequencing platforms, with either FocalSV(target) or FocalSV(auto) consistently ranking first or second in average F1 scores for deletions and insertions across nearly all data sets. In terms of overall genotyping accuracy, cuteSV and FocalSV consistently performed best, with accuracies frequently in the 98%–99% range.

When examining each data set individually (Tables 2–4; Supplemental Tables S1–S3), FocalSV outperformed all other tools, achieving the highest F1 scores for both deletions and insertions across all HiFi, CLR, and ONT libraries. In the context of three HiFi data sets (Table 2; Supplemental Table S1), FocalSV emerged as the top overall performer. For deletions, FocalSV(auto) outperformed all other tools on Hifi_L2 and Hifi_L3, exceeding the F1 score of the second-best tool, sawfish, by an average of 0.12%. On Hifi_L1, FocalSV(target) achieved the highest performance, closely followed by sawfish and FocalSV(auto). For insertions, both FocalSV(target) and FocalSV(auto) consistently outperformed all other tools across all three libraries, with FocalSV(target) achieving F1 score that surpassed the next-best tools by an average of 1.80%.

In the three CLR data sets (Table 3; Supplemental Table S2), FocalSV emerged as the top-performing tool in terms of F1 score and recall across nearly all libraries, demonstrating clear advantages. For deletions, FocalSV(auto) outperformed the second-ranked tools in F1 score on CLR_L2 and CLR_L3 by an average of 0.39%. On CLR_L1, FocalSV(target) exhibited the highest F1 score, closely followed by pbsv and FocalSV(auto). For deletion recall, FocalSV(target) ranked first, outperforming the second-ranked tools across all three libraries by an average of 1.48%. For insertions, both FocalSV(target) and FocalSV(auto) excelled across all libraries, with FocalSV(target) achieving an average F1 score 3.45% higher than the second-ranked tools. For insertion recall, FocalSV(target) exceeded the second-ranked tools on CLR_L1 and CLR_L2 by an average of 0.93%. On CLR_L3, FocalSV(auto) exhibited the second-highest recall, trailing the top by 0.23%.

In the three ONT data sets (Table 4; Supplemental Table S4), both FocalSV(target) and FocalSV(auto) maintained substantial leads. FocalSV(target) outperformed the second-ranked tool across ONT_L1, ONT_L2, and ONT_L3 by an average of 0.58% in deletion F1 score. For deletion recall, FocalSV(target) outperformed the second-ranked tools on ONT_L1 and ONT_L2 by an average of 0.58%. On ONT_L3, FocalSV(target) achieved the second-highest recall, trailing the top by 0.10%. For insertion F1, FocalSV(auto) achieved the highest F1 scores on ONT_L1, ONT_L2, and ONT_L3, outperforming the second-ranked tool, cuteSV, by an average of 0.30%. In terms of insertion recall, FocalSV(target) outperformed the second-ranked tool on ONT_L2 and ONT_L3 by an average of 0.51%. On ONT_L1, FocalSV(target) achieved the second-highest recall, trailing the top recall by 0.49%.

For these benchmarks, FocalSV(target) used regions of interest defined by SVs identified by GIAB. To avoid biased evaluation, we also measured false positive (FP) counts within regions lacking known SVs. Recognizing that some tools may report fewer SVs—leading to lower recall but artificially high precision—we further normalized FP counts by penalizing them according to each tool's average F1 score (https://www.researchgate.net/publication/268185911_The_truth_of_the_F-measure; Powers 2011; Christen et al. 2023), using the formula: normalized FP count = raw FP count × (1 − F1). To evaluate performance in SV-negative regions, we randomly selected 1000 regions without known SVs using the GIAB gold standard. All tools produced very few false positives. FocalSV(target) achieved the lowest FP counts across both raw (CLR_L1: 5; ONT_L1: 3) and normalized values (CLR_L1: 0.4; ONT_L1: 0.26). For HiFi data, it showed the second-lowest raw (Hifi_L1: 6) and lowest normalized FP value (Hifi_L1: 0.4) (Supplemental Fig. S1). These results demonstrate FocalSV(target)’s consistently high precision across both SV-rich and SV-negative regions.

In summary, FocalSV consistently outperformed other tools across nine HG002 data sets (HiFi, CLR, and ONT) based on average F1 scores and library-specific performance, demonstrating strong detection capabilities for both deletions and insertions. Although it occasionally ranked second or third in recall or precision, these variations were minor and did not diminish its overall effectiveness. In terms of genotyping accuracy, FocalSV consistently ranked first or second, with accuracy rates frequently reaching 98%–99%. These results highlight FocalSV as a reliable and high-performing tool for structural variant detection.

FocalSV demonstrates excellent resilience to different SV size ranges

So far, we evaluated SVs by averaging across all size ranges. However, we found that the tools we assessed showed varying levels of accuracy in detecting SVs of different sizes. To demonstrate the effect of SV size and compare the resilience of each tool, we plotted the F1 score against the range of SV sizes (Fig. 3; Supplemental Fig. S2). We utilized the same modest tolerance parameters in Truvari for benchmarking. FocalSV(auto) and FocalSV(target), along with all other tools, are visualized in Figure 3 and Supplemental Figure S2, respectively. Both modes of FocalSV exhibited comparable performance across all variant size ranges and libraries.

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 3.

F1 accuracy of SV detection across various size ranges on nine long-read data sets. (A–C) F1 accuracy plot for three HiFi data sets. Negative ranges denote deletions and the positive ranges denote insertions. The bar plot illustrates the benchmark SV distribution across these size ranges. The line plot displays the F1 scores for four distinct detection methods. Dashed lines indicate alignment-based, whereas solid lines represent assembly-based methods. (D–F) F1 accuracy plot for three CLR data sets. (G–I) F1 accuracy plot for three ONT data sets.

Overall, assembly-based tools such as FocalSV(auto), FocalSV(target), PAV, SVIM-asm, Dipcall, and sawfish showed better resilience than alignment tools such as SKSV, Sniffles2, pbsv, SVIM, and cuteSV when processing HiFi data (Fig. 3A–C; Supplemental Fig. S2A–C). Most alignment-based tools had a significant drop in performance with large INSs in the range of 2 kb–50 kb, whereas most assembly-based tools typically encountered performance challenges only with large INSs in the range of 9 kb–50 kb. Among all assembly-based tools, FocalSV, in both modes, continued to distinguish itself with exceptional performance across nearly all size ranges, except for 8 kb- to 10 kb-deletions and 10 kb- to 50 kb-insertions.

On CLR data (Fig. 3D–F; Supplemental Fig. S2D–F), assembly-based tools except Dipcall still outperformed alignment-based tools in terms of resilience, especially for INSs. Across all size ranges, FocalSV, in both modes, maintained a top-tier performance. Notably, FocalSV's performance is much better than other tools on CLR_L3, where it had the highest F1 score across all size ranges except for INSs in the range of 9 kb–10 kb and showed the least fluctuation.

On ONT data (Fig. 3G–I; Supplemental Fig. S2G–I), the difference between alignment-based and assembly-based tools was less pronounced. Two alignment-based tools, cuteSV and Sniffles2, demonstrated resilience comparable to assembly-based tools. FocalSV, in both modes, emerged as a top-performing tool for small- to large-sized deletions (50 bp–8 kb) and insertions (50 bp–2 kb). For large deletions (8 kb–50 kb) and insertions (2 kb–50 kb), FocalSV, in both modes, remained within the range of the best tools.

FocalSV achieves robust SV detection across evaluation parameters

While assessing different SV callers against a gold standard, we recognized the potential impact of breakpoint shifts and sequence similarity issues on the evaluation. This acknowledgment arises from the fact that SVs often span substantial genomic regions. In previous evaluations, we selected a set of moderate-tolerance but fixed parameters to demonstrate the overall performance for each tool. However, the choice of parameters such as breakpoint shift tolerance and sequence similarity, for identifying a call as a true positive, varies subjectively. Therefore, to comprehensively evaluate the effectiveness and stability of SV callers, we adjusted key evaluation parameters (p, P, r, and O) using Truvari for benchmarking (English et al. 2022). Specifically, we systematically evaluated the impact of four key parameters by conducting two complementary experiments: (1) a univariate sensitivity analysis, where each parameter was varied individually to assess its effect on F1 scores while holding others constant; and (2) a grid-based benchmarking experiment, in which all pairwise parameter combinations were explored to evaluate tool performance across the full parameter space. Parameters p, P, and O were adjusted in increments of 0.1 from 0 to 1, and r was adjusted in 100-bp increments from 0 to 1000 bp. Our goal was to examine how different levels of stringency or leniency in these parameters might impact SV detection across the nine data sets. We expected FocalSV to demonstrate relatively consistent and stable performance, even under more stringent conditions.

We first systematically varied each parameter and plotted curves to illustrate the changes in F1 scores across all libraries. When varying one parameter, the remaining parameters were held at their default moderate values. For deletion and insertion detection on Hifi_L1 (Fig. 4A,B), increasing the stringency of the matching threshold by reducing r from 1000 bp to 0 bp led to a consistent decline in F1 scores for all tools, with insertions showing greater sensitivity to r than deletions. When varying O, P, or p from 0 to 1.0—corresponding to increasing stringency (Fig. 4C–H)—F1 scores of all tools decreased. Notably, for P and p, F1 scores for both deletions and insertions dropped sharply when the thresholds exceeded 0.8. Changes in O had an even stronger impact: both deletions and insertions exhibited a substantial decline in F1 performance when O increased, with insertions again being more sensitive than deletions. Across all parameter settings, FocalSV, in both modes, consistently achieved the highest F1 scores. Similar trends were observed across other libraries, with FocalSV maintaining top performance in nearly all scenarios (Supplemental Figs. S3–S10).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 4.

F1 accuracy by changing four different evaluation parameters on Hifi_L2. (A,B) F1 score curves for deletions (DEL) and insertions (INS) across all tools as r is varied. r is the maximum reference location distance between SV call and gold standard SV. r varies from 0 to 1000 bp with a 100-bp interval. (C,D) F1 score curves for deletions and insertions across all tools as O is varied. O is the minimum reciprocal overlap between SV call and gold standard SV. (E,F) F1 score curves for deletions and insertions across all tools as P is varied. P is the minimum allowable allele size similarity between SV call and gold standard SV. (G,H) F1 score curves for deletions and insertions across all tools as p is varied. p is the minimum percentage of allele sequence similarity between SV call and gold standard SV. O, P, and p vary from 0 to 1 with a 0.1 interval.

We next varied each parameter pair and plotted F1 heat maps to illustrate performance changes across all libraries. For deletion calls on Hifi_L1, the F1 heat maps for FocalSV(auto) and FocalSV(target) are shown in Figure 5, A through I and Supplemental Figure S11, respectively. Both modes demonstrated comparable performance across all libraries under varying evaluation parameters. Based on the comprehensive benchmarking study by Liu et al. (2024), which thoroughly examined the impact of various parameter combinations, such as p-O, P-r, O-r, p-P, p-r, and P-O, on deletion evaluation, we selected p and O as the representative parameter pair, as they had the most substantial influence on deletion performance. As shown in Figure 5,A through I and Supplemental Figure S11, increasing the values of p and O imposed stricter correspondence requirements between the SV call and the gold standard, leading to a decline in F1 scores. The heat map and gradient analysis of Liu et al. (2024) demonstrated distinct performance patterns among tools as more stringent thresholds were applied. They showed that all read alignment-based SV callers, except for pbsv, exhibited significant performance drops when p or O exceeded 0.7, with F1 scores falling below 5% when either parameter reached 1.0. In contrast, FocalSV, in both modes, and other assembly-based SV callers from the study of Liu et al. (2024), namely Dipcall, SVIM-asm, and PAV, along with pbsv, maintained stable performance across the parameter grid. Even under the strictest conditions (p = 1.0, O = 1.0), these tools achieved F1 scores >69%. Notably, under this strict exact-match criterion in terms of sequence similarity and reciprocal overlap ratio, FocalSV, in both modes, excelled with an exceptional F1 score above 74% (Fig. 5A; Supplemental Fig. S11A), outperforming other stable tools and demonstrating its robustness. Across nearly all parameter settings, FocalSV consistently achieved higher F1 scores compared to Dipcall, SVIM-asm, PAV, pbsv, and sawfish (Supplemental Figs. S12–S15A), demonstrating superior performance across the grid. The overall performance trends for these tools were consistent with those previously reported by Liu et al. (2024) (Supplemental Figs. S12–S20A).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 5.

Deletion F1 accuracy of FocalSV(auto) by tuning different evaluation parameters (p and O). O is the minimum reciprocal overlap between SV call and gold standard SV. p is the minimum percentage of allele sequence similarity between SV call and gold standard SV. O and p vary from 0 to 1 with a 0.1 interval. (A–C) The F1 heat map for deletions by FocalSV(auto) on three HiFi data sets. Every cell in the heat map represents the F1 score under a specific pair of p and O evaluation. (D–F) The F1 heat map for deletions by FocalSV(auto) on three CLR data sets. (G–I) The F1 heat map for deletions by FocalSV(auto) on three ONT data sets.

For insertion evaluations on Hifi_L1 (Fig. 6A; Supplemental Figs. S21–S30A), the pair p and r were similarly selected as the most representative parameters based on the findings of Liu et al. (2024). Their study revealed that insertion detection was generally more sensitive to parameter variations compared to deletions. Combining their results and our experiment, we found that most tools, including FocalSV, experienced a decline in F1 scores when p exceeded 0.8 (i.e., when the required allele sequence similarity between the called SV and the benchmark exceeded 80%) (Fig. 6A; Supplemental Figs. S21–S30A). Additionally, F1 scores decreased when r was reduced. Once r dropped to 200 bp or less, indicating a more stringent reference distance threshold, performance deteriorated significantly across all tools. Similar to the trends observed for deletions, FocalSV, other assembly-based tools, and pbsv displayed greater robustness to stringent parameters compared to read alignment-based tools. Among all robust performing tools, FocalSV, in both modes, consistently achieved higher F1 scores than Dipcall, SVIM-asm, PAV, and pbsv, showing outstanding performance over the grid (Fig. 6A; Supplemental Figs. S21–S25A). We have also observed similar patterns on grid searches for FocalSV on other libraries (Figs. 5B–I, 6B–I: Supplemental Figs. S12–S30).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 6.

Insertion F1 accuracy of FocalSV(auto) by tuning different evaluation parameters (p and r). r is the maximum reference location distance between SV call and gold standard SV. p is the minimum percentage of allele sequence similarity between SV call and gold standard SV. p vary from 0 to 1 with a 0.1 interval. r varies from 0 to 1000 bp with a 100 bp interval. (A–C) The F1 heat map for insertions by FocalSV(auto) on three HiFi data sets. Every cell in the heat map represents the F1 score under a specific pair of p and r evaluation. (D–F) The F1 heat map for insertions by FocalSV(auto) on three CLR data sets. (G–I) The F1 heat map for insertions by FocalSV(auto) on three ONT data sets.

The robustness of assembly-based tools to evaluation parameters may indicate their capacity to accurately detect SV breakpoints and alternative allele sequences, as highlighted by Liu et al. (2024). To address this, we analyzed the distribution of SV breakpoint shift and alternate allele sequence similarity for FocalSV. Breakpoint shifts were defined as the maximum reference location difference between two compared SV calls, calculated as the greatest start/end location difference between true positive SVs and their corresponding benchmark SVs. Sequence similarity was quantified using the edit distance between compared SV calls, directly extracted from Truvari. Our results revealed that FocalSV achieved a near-zero breakpoint shift and near 100% SV sequence similarity with the benchmark callset (Supplemental Figs. S31–S34). In line with previous findings, our results confirm that FocalSV, like other assembly-based tools, demonstrates strong resilience across various parameter settings, underscoring its capacity to accurately capture both SV breakpoints and alternative alleles. Notably, FocalSV consistently outperformed other assembly-based tools, including Dipcall, SVIM-asm, and PAV, across all parameter configurations, further emphasizing its superior accuracy and reliability.

FocalSV exhibits reliable SV detection on HG005, validated by multitool consensus benchmarking

To further assess the robustness and generalizability of FocalSV, we extended our evaluation to the Chinese male sample HG005. As the GIAB consortium has not provided a gold standard SV callset for this sample, we developed a consensus-based evaluation framework leveraging agreement across multiple established SV callers. We first applied all benchmarked tools to the HG005 HiFi data set (Table 1) and collected their respective callsets based on the GRCh38 reference genome. From each, we extracted indel SVs (≥50 bp) marked as “PASS” in the VCF filter field and located within the high-confidence regions defined by GIAB. Although originally curated for HG002, this BED file encompasses regions with consistent support across sequencing platforms and excludes problematic areas such as segmental duplications and low-mappability areas. Given its stringent design, we applied it to HG005 to enable fair and broadly comparable evaluations. Following filtering, we merged and sorted the resulting VCF files, then applied distance-based clustering to group SVs of the same type within a 500-bp window. These clusters served as proxy gold standard sets for HG005, with each cluster annotated by the number of supporting tools. By adjusting the minimum support threshold, we generated benchmark sets at varying confidence levels. SV callers were evaluated by comparing their predictions to these clusters: a call was considered a true positive if it overlapped a benchmark cluster and matched the majority genotype. SVs without matching clusters were treated as false positives. We computed recall, precision, and F1 scores across multiple confidence thresholds to quantify performance.

Overall, FocalSV(auto) demonstrated the leading performance in deletion detection across a wide range of confidence thresholds. As shown in Supplemental Figure S35, FocalSV consistently achieved the highest F1 scores when the minimum number of supporting tools ranged from two to seven. For instance, at the stringent threshold of seven supporting tools, FocalSV reached an F1 score of 97.66% (Supplemental Table S4), outperforming the second-best tool, Sniffles2, by 0.19%. At thresholds above seven, cuteSV ranked highest, with FocalSV following closely in second place. The number of benchmark deletions ranged from 3985 to 4710 (Supplemental Fig. S35D; Supplemental Table S4), closely matching the 4116 Tier 1 deletions reported for HG002 in the high-confidence BED file. This consistency supports the reliability of our filtering and evaluation framework.

For insertion calls, FocalSV maintained strong performance, ranking first or second in F1 score across most confidence thresholds. As shown in Supplemental Figure S36 and Supplemental Table S5, FocalSV achieved the highest F1 scores when the number of supporting tools was two or three. At thresholds between four and eight, FocalSV consistently ranked second, closely following PAV, with an average F1 difference of 0.3%. When the threshold increased to nine supporting tools, FocalSV ranked third, trailing the top performer by 0.7%. At the 10-tool threshold, F1 scores dropped substantially for most tools, except pbsv, which reported significantly fewer insertions than others. This suggests the benchmark set at this threshold may be incomplete. As a result, recall becomes a more informative metric under such stringent criteria, and FocalSV achieved the highest recall among all tools. The number of benchmark insertions ranged from 4231 to 6238 as the support threshold increased from two to 10, closely aligning with the 5281 Tier 1 insertions reported in the HG002 high-confidence BED file, further validating the soundness of our evaluation framework.

In summary, FocalSV achieved the highest overall performance in deletion detection and consistently ranked among the top performers for insertion detection on the HG005 data set, highlighting its robustness, reliability, and broad applicability across diverse genomic contexts.

FocalSV maintains strong SV detection performance regardless of variations in phasing quality

FocalSV utilizes Longshot for phasing and read partitioning into two haplotypes prior to indel SV assembly; therefore, phasing quality can directly influence downstream SV detection accuracy. We assessed phasing accuracy using SNP-based switch error metrics (https://github.com/ywchoi/phasing) and PhaseQ scores derived from Strand-seq data, as detailed in the Supplemental Methods and Supplemental Figure S37. Across HiFi, CLR, and ONT data sets, we observed low switch error rates (median: 0%; mean: 2.2%–2.4%) and high proportions of correctly phased heterozygous variants (median: 97.8%–100%; mean: 93.3%–95.6%), with 68% of regions achieving 0% SER (Supplemental Fig. S38). Strand-seq analysis further confirmed robust phasing, with over 80% of phase blocks showing PhaseQ values >0.8 (Supplemental Fig. S39). Additional results are provided in the Supplemental Results.

To evaluate phasing quality, we further examined the distribution of phased read percentages across all target regions. Both HiFi and ONT data sets exhibited high phasing rates (Hifi_L1: mean 83%–87%, median 97%–98% across both modes; ONT_L1: mean 82%–88%, median 84%–90%) (Supplemental Fig. S40), whereas the CLR data set displayed lower phasing rates (mean 49%–52%, median 57%–58%) (Supplemental Fig. S40). To assess the effect of phasing on SV detection, we ranked all target regions by phased read percentage and divided them into high and low phasing groups. SV detection performance was then compared between these groups. Regions with higher phasing percentages showed moderately improved F1 scores across HiFi, CLR, and ONT data sets, particularly for deletions. For instance, deletion F1 scores for high versus low phasing regions were 95.1% versus 93.4% (Hifi_L1), 93.3% versus 91.8% (CLR_L1), and 94.5% versus 91.5% (ONT_L1) (Supplemental Table S6). In contrast, insertion detection was less affected, with F1 scores of 91.2% versus 91.2% (Hifi_L1), 90.1% versus 89.1% (CLR_L1), and 92.3% versus 90.2% (ONT_L1) (Supplemental Table S6). Overall, FocalSV demonstrated high phasing accuracy for both SNP phasing and local assemblies, while maintaining robust structural variant detection performance even in regions with lower proportions of phasing-informative reads.

FocalSV achieves superior performance in complex somatic SV detection in cancer data

To extend the evaluation to include SV detection beyond deletions and insertions, we analyzed additional SV types involving complex DNA rearrangements, such as translocations, inversions, and duplications. FocalSV and other relevant tools were applied to detect somatic SVs in two publicly available cancer libraries (Fang et al. 2021). Dipcall was excluded from this analysis, as it was not designed to detect these three types of SVs. Although PAV is designed to detect inversions, it failed to detect any inversions in the cancer data and was therefore also excluded from this analysis.

We conducted a comparative analysis by applying the selected tools to two publicly available tumor-normal paired libraries (PacBio CLR and ONT), as provided by Talsania et al. (2022). These libraries, along with the high-confidence HCC1395 somatic SV callset serving as the benchmark gold standard, formed the basis for evaluating the detection of three classes of somatic SVs. Each benchmarked tool was first applied independently to each library, generating VCF files. These VCFs were then processed using SURVIVOR (Jeffares et al. 2017) to identify somatic variants by comparing paired normal-tumor VCFs. The detailed method to detect somatic SVs is provided in the Supplemental Methods section. The identified somatic SVs were subsequently compared to the gold standard, which contained 137 translocations, 133 inversions, and 230 duplications.

FocalSV(auto) outperformed other tools (SVIM-asm, cuteSV, SVIM, pbsv, and Sniffles2) in terms of F1 score across nearly all sequencing libraries and SV types, except for inversions in ONT data (Table 5). Specifically, on PacBio CLR data, FocalSV(auto) outperformed the second-best tools in F1 score by 4.7%, 1.2%, and 7.6% for translocations, inversions, and duplications, respectively. On ONT data, FocalSV(auto) exceeded the second-best tools by 29.6% and 6.3% for translocations and duplications, respectively. For inversions, Sniffle2 achieved the highest F1 score (30.4%), followed by FocalSV(auto) (20.2%). We also applied FocalSV(target) to directly detect the high-confidence somatic SV callset based on predefined target regions. As expected, it achieved the highest F1 scores across all scenarios, followed by FocalSV(auto).

Computation cost of FocalSV

Finally, we evaluated FocalSV's memory requirement and runtime performance. For a relatively large SV—an insertion of 12.6 kb on Chromosome 21—FocalSV finished execution in 32 sec of elapsed time and 5 min and 20 sec of CPU time, using a maximum 5.6 MB of memory when run with eight threads. In comparison, detecting the smallest SV—a deletion of 50 bp on Chromosome 21—took a similar elapsed time of 33 sec and required 5 min and 30 sec ofCPU time, with a higher memory usage of 0.76 GB. On average, the CPU time was ∼5.5 min for one target region, reflecting consistent computational effort regardless of SV size.

Based on this estimate, detecting 20,000 SVs would require approximately 1800 total CPU hours—comparable to the runtime for FocalSV(auto) to assemble a single library. As such, assembling all SVs across an entire genome using FocalSV can still be more computationally demanding than alignment-based methods. Importantly, we did not compare FocalSV's runtime to existing SV callers, as its time efficiency is specifically optimized for targeted analyses involving a limited number of regions rather than whole-genome discovery. Whereas assembly-based SV detection is generally resource-intensive, FocalSV's region-based design enables efficient and rapid analysis when users focus on selected loci of interest.

FocalSV is designed to support multilevel parallelization, allowing tasks to be efficiently distributed across multiple servers, with each server assigned to distinct genomic regions. Within each region, FocalSV utilizes multithreading to manage assembly and variant detection tasks, maximizing CPU core utilization.