CMSV: Long-Read-Based Structural Variation Detection Through a CNN–Mamba Model
Genes · 30.05.2026 · 17 (6) · 633
DOI: 10.3390/genes17060633 ↗Kurz gefasst
Fachartikel in der Zeitschrift Genes, online erschienen am 30.05.2026. Verfasst von Song Cheng und 1 weiteren Autor:innen. Beteiligte Einrichtung: School of Computer Science and Technology, Xinjiang University, Urumqi 830046, China. Hauptthemen laut MEDLINE: Genomic Structural Variation, Software. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Science and Technology Department of Xinjiang Uygur Autonomous Region, Key Research and Development Project of Xinjiang Uygur Autonomous Region.
Abstract (Originalsprache)
Background/Objectives: Structural variations are important forms of genomic variation and are closely related to genomic diversity and many human diseases. Long-read sequencing has improved the ability to detect structural variations in complex genomic regions, but existing methods still mainly rely on manually designed heuristic rules and often have difficulty jointly modeling local SV signatures and cross-subsegment contextual modeling. To address this problem, we propose CMSV, a structural variation detection and genotyping method for long-read sequencing data.
Methods
CMSV extracts multi-channel position-level features from alignment results and combines a multi-scale convolutional encoder with stacked Mamba modules for window-level candidate region detection. Candidate variants are then integrated and optimized through DBSCAN-based density clustering and length-based clustering. Genotypes are inferred based on variant-supporting reads and reference-genome-supporting reads. CMSV is designed to support several major structural variation types, including DEL, INS, DUP, INV, and TRA/BND. In our real-data benchmarks, the strongest validation is provided for DEL and INS, while DUP, INV, and TRA/BND are further evaluated using simulated multi-type datasets.
Results
Experiments on real HG002 DEL/INS benchmarks, simulated multi-type datasets, and family-based datasets show that CMSV is competitive across PacBio CCS, PacBio CLR, and ONT platforms within the corresponding evaluation settings. Additional held-out chromosome evaluation on GRCh38 chr13–chr22 was further conducted to assess chromosome-level generalization beyond the full HG002 benchmark. CMSV shows stable performance in DEL/INS detection and genotyping on real-data benchmarks, while simulated multi-type evaluations further support its ability to detect DUP, INV, and TRA/BND. The results also show that CMSV can effectively model complex variant signals and maintain good family-level consistency in trio-based evaluation.
Conclusions
CMSV provides an effective deep learning framework for long-read structural variation detection and genotyping across sequencing platforms and coverage levels.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- School of Computer Science and Technology, Xinjiang University, Urumqi 830046, China.
- Department of Electronic Engineering, Tsinghua University, Beijing 100084, China.
ORCID-Profile: Hongbing Ma
Förderung
- Science and Technology Department of Xinjiang Uygur Autonomous Region
- Key Research and Development Project of Xinjiang Uygur Autonomous Region
Zitieren
APA
Cheng, S., & Ma, H. (2026). CMSV: Long-Read-Based Structural Variation Detection Through a CNN–Mamba Model. Genes, 17(6), 633. https://doi.org/10.3390/genes17060633
Vancouver
Cheng S, Ma H. CMSV: Long-Read-Based Structural Variation Detection Through a CNN–Mamba Model. Genes. 2026;17(6):633. doi:10.3390/genes17060633
BibTeX
@article{Cheng2026,
author = {Cheng, Song and Ma, Hongbing},
title = {CMSV: Long-Read-Based Structural Variation Detection Through a CNN–Mamba Model},
journal = {Genes},
year = {2026},
volume = {17},
number = {6},
pages = {633},
doi = {10.3390/genes17060633}
}
EUVIMED