Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
VarNet-T: Deep Learning Brings Precision to Tumor-Only Cancer Variant Detection
VarNet-T, a deep learning framework, substantially improves identification of cancer mutations from tumor-only sequencing data, achieving up to 33% better accuracy than existing methods and enabling more reliable patient stratification for immunotherapy.
Background
Accurate variant calling typically requires paired tumor and normal tissue samples for comparison. However, many clinical and archival samples lack matched normals, limiting tumor mutation burden (TMB) estimation—a critical biomarker for immunotherapy response. Existing tumor-only callers produce unreliable estimates, hindering precision oncology workflows.
Key Findings
- AUPRC of 0.773 on benchmark datasets, representing 20-33% improvement over the second-best tumor-only caller
- >3-fold improvement in TMB-high classification accuracy with F1 score of 88%, compared to 43-55% for competing methods
- Pearson correlation of 0.82 with tumor-normal TMB estimates, substantially outperforming DeepSomatic (0.57) and Mutect2 (0.21)
- A single pan-cancer model generalized effectively across cancer types and unseen tumor classes
Why It Matters
VarNet-T addresses a critical clinical bottleneck: many tumor samples from biobanks and archival sources lack matched normals. By enabling accurate TMB estimation from tumor-only data, it could improve immunotherapy patient selection and expand precision oncology to previously unassessable populations.
Limitations
The study focused on whole-genome and whole-exome sequencing. Generalization to panel-based or lower-depth sequencing, clinical validation in prospective cohorts, and integration into clinical workflows remain important next steps.
Original paper: Improved tumor-only variant calling and mutation burden estimation with VarNet-T. — Nature communications. 10.1038/s41467-026-71705-4
