MambaSV: Accurate germline and somatic structural variant calling from long-reads with deep sequence model

Somatic structural variants (SVs) are difficult to detect from long-read sequencing due to the inherent complexity of tumor genomes characterized by low variant allele fractions (VAFs), variable tumor ploidy/purity, and highly complex rearrangement patterns. Deep learning models are underexplored in this problem due to the lack of labeled data. To leverage it, we introduce a novel data simulator that generates paired normal-tumor BAM files with accurate haplotype-aware SV labels to facilitate model training and benchmarking. Built on this, we present MambaSV, a GPU-native deep sequence model for haplotype-resolved somatic SV calling. The scalable, end-to-end, and haplotype-aware data simulator for long read sequencing (PacBio and ONT) is designed to capture real-world somatic SV complexities. It simulates SVs across five SV classes (INS, DEL, DUP, INV, BND) following empirically derived length distributions, partition them by haplotype and sample, and inject them together with SNPs and INDELs into CHM13 to create synthetic genomes. Longreads are simulated from the synthetic genomes with real statistics (e.g., length, error rate, etc.). For Fusions we generate artificial sequence constructs in FASTA format by concatenating relevant sequence fragments to reproduce true breakpoint signatures. To model tumor heterogeneity, we mix reads from SV and SNP-only genomes to achieve per-event VAFs from 0-100%, simulate copy-neutral loss of heterozygosity (cnLOH) in tumor, and haplotag both BAMs with third-party tools to mirror practical workflows. MambaSV is trained to discover SVs from matched tumor and normal alignments. MambaSV formulates SV discovery as a per-base multi-class classification task. At each base, a 30-feature vector (mapping qualities, supplementary evidence, etc.) is computed across three channels (hap1, hap2, unphased) and both samples. Haplotype-specific inputs combine each haplotype with the unphased channel. Leveraging a Siamese network with shared weights, MambaSV jointly processes tumor and normal samples with a bi-directional Mamba-2 backbone that preserves megabase-scale context without loss of single-base resolution. A lightweight decoder outputs haplotype-resolved somatic predictions, which are post-processed to merge contiguous signals and emit VCFs. Trained solely on HiFi BAMs and tested on held-out samples (~23 k SVs), MambaSV achieves F1 = 92.27 (HiFi) and 89.32 (ONT), outperforming Severus (69.18/70.31), Svision-pro (53.92/54.24), SAVANA (56.37/53.93), and Nanomonsv (42.42/42.20). To conclude, our comprehensive simulation framework fuels MambaSV's development by enabling rigorous training and broad generalizability in challenging real-world scenarios, while MambaSV unifies base-level features, haplotype structure, and long-range genomic context in a single deep model to set a new standard for somatic SV calling from long-read data.