Improved tumor-only variant calling and mutation burden estimation with VarNet-T

2026-04-09 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Life Sciences & Biology, Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, medium

Summary

VarNet-T, an end-to-end weakly supervised deep learning framework, has been introduced for accurate somatic variant calling and tumor mutation burden (TMB) estimation from tumor-only sequencing data. This framework addresses the challenge of distinguishing somatic mutations from germline mutations or sequencing artifacts when matched normal samples are unavailable, a common issue in clinical diagnostics and retrospective analyses. VarNet-T was trained using millions of high-confidence variants and demonstrated a 20-33% performance improvement over existing methods on public datasets. Furthermore, it achieved over 3x higher accuracy in TMB-high status classification across 1000 tumor samples spanning 10 solid cancer types, indicating significant potential to enhance patient selection for immunotherapy. The framework is publicly available at https://github.com/skandlab/VarNet under a PolyForm Noncommercial License 1.0.0.

Key takeaway

For AI Scientists and Research Scientists developing cancer diagnostics, VarNet-T offers a robust solution for tumor-only somatic variant calling and TMB estimation. Its demonstrated 20-33% performance improvement and >3x higher accuracy in TMB-high classification suggest that integrating this deep learning framework could significantly enhance diagnostic precision and patient stratification for immunotherapy. Consider evaluating VarNet-T for your tumor-only sequencing pipelines to improve accuracy and clinical utility.

Key insights

VarNet-T improves tumor-only somatic variant calling and TMB estimation using weakly supervised deep learning.

Principles

• Weakly supervised deep learning enhances tumor-only variant calling.
• Tumor-only sequencing can accurately estimate TMB for immunotherapy selection.

Method

VarNet-T is an end-to-end weakly supervised deep learning framework trained on millions of high-confidence variants to identify somatic variants from aligned tumor reads without a matched normal sample.

In practice

• Utilize VarNet-T for tumor-only sequencing in cancer research.
• Apply VarNet-T for TMB-high status classification in clinical diagnostics.

Topics

• VarNet-T
• Tumor-only Variant Calling
• Weakly Supervised Deep Learning
• Tumor Mutation Burden
• Immunotherapy Patient Selection

Code references

• skandlab/VarNet

Best for: AI Scientist, Research Scientist

Related on AIssential

• Multi-Class Brain Tumor Classification Using Advanced Deep Learning Models: A Comparative Study — EfficientNetB0 significantly improves multi-class brain tumor classification from MRI images, especially for subtle meningiomas, outperforming other CNNs.
• DUET: Dual-Paradigm Adaptive Expert Triage with Single-cell Inductive Prior for Spatial Transcriptomics Prediction — DUET unifies regression and retrieval with single-cell priors for robust spatial gene expression prediction from histology.
• 3 Questions: Building predictive models to characterize tumor progression — AI and ML tools can decode tumor evolution, especially ecDNA amplifications, to predict and counter treatment resistance.
• Immunotherapy drug target identification using machine learning and patient-derived tumour explant validation — MIDAS, a multimodal GNN, effectively identifies novel immuno-oncology targets by integrating diverse biological data.
• Mitosis Detection in the Wild: Multi-Tumor and Context-Aware Generalization in the MIDOG 2025 Challenge — Automated mitosis detection models exhibit significant generalization gaps "in the wild" beyond curated hotspot regions.
• A deep learning–based digital biopsy for predicting early recurrence in gastric cancer — The RSA model combines deep learning with clinical data to predict gastric cancer recurrence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.