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

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Somatic variant calling typically requires matched normal samples, which are often unavailable in clinical settings.
• Lack of normal samples hinders accurate distinction between somatic mutations, germline variants, and sequencing artifacts.
• This limitation compromises cancer genome analysis in diagnostics and biobank studies.

Purpose of the Study:

• To introduce VarNet-T, a deep learning framework for somatic variant identification from tumor-only sequencing data.
• To address the challenge of analyzing cancer genomes without matched normal samples.
• To improve the accuracy of somatic mutation detection and tumor mutation burden estimation.

Main Methods:

• Developed VarNet-T, an end-to-end weakly supervised deep learning framework.
• Trained the model on millions of high-confidence variants.
• Benchmarked performance against existing methods using public datasets.

Main Results:

• VarNet-T demonstrated a 20-33% performance improvement over existing somatic variant calling methods.
• Accurate tumor mutation burden (TMB) estimation was achieved on 1000 tumor samples across 10 cancer types.
• VarNet-T showed >3x higher accuracy in TMB-high status classification compared to current approaches.

Conclusions:

• VarNet-T enhances the accuracy of somatic variant calling using tumor-only sequencing.
• The framework shows significant potential for improving patient selection for immunotherapy through precise TMB assessment.
• VarNet-T can increase the utility of tumor-only sequencing in cancer research and clinical diagnostics.