Benchmark and Evaluation for Somatic Structural Variants Detection with Long-read Sequencing Data

Somatic structural variations (somatic SVs) are hallmarks of tumors, but their comprehensive detection remains technically challenging. Long-read sequencing (LRS) technology, which generates reads spanning large-scale SVs and their flanking sequences, enables a wide range of prospects for somatic SV detection. However, existing LRS-based somatic SV detection algorithms and pipelines exhibit variable performance that has not been systematically characterized. In this study, we conducted a rigorous evaluation of 51 LRS-based somatic SV detection strategies, integrating 3 reference genomes, 2 aligners, 5 SV callers, and 5 processing methods tailored for SV callers. We use both simulated datasets and empirical data from HCC1395/HCC1395BL cell lines sequenced on Oxford Nanopore (ONT) and Pacific Biosciences (PacBio) platforms for technical assessment. Our findings highlight the need for further refinement of specialized somatic SV detection tools, as no single strategy consistently outperforms across all scenarios. Workflows based on germline SV callers exhibit a high false-positive rate, which cannot be mitigated by increasing sequencing depth or tumor purity. Furthermore, challenges persist in detecting insertions, genomic tandem repeat regions, and ultra-long SVs. We delineate technical bottlenecks in current somatic SV detection approaches and provide recommendations for their further advancement. Additionally, we offer suggestions for selecting specific tools in different application scenarios. This work offers a comprehensive benchmark for somatic SV detection and valuable insights for future LRS-based tools development and methodological improvements.