ClairS-TO: a deep-learning method for long-read tumor-only somatic small variant calling.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate somatic variant detection is critical for cancer research and clinical applications.
• Existing methods often require matched normal samples, which are not always available.
• Current tumor-only callers struggle with long-read sequencing data.

Purpose of the Study:

• To develop a deep-learning-based method for long-read tumor-only somatic variant calling.
• To address the limitations of existing algorithms in scenarios lacking matched normal samples.
• To provide a reliable tool for somatic variant discovery in diverse cancer research settings.

Main Methods:

• Developed ClairS-TO, a deep learning model utilizing an ensemble of two neural networks.
• Trained networks on cancer cell line data (COLO829, HCC1395) for somatic variant identification.
• Evaluated performance on both long-read (ONT, PacBio) and short-read sequencing data.

Main Results:

• ClairS-TO demonstrated superior performance compared to existing methods (DeepSomatic, smrest) on long-read data.
• ClairS-TO also outperformed established short-read callers (Mutect2, Octopus, Pisces, DeepSomatic).
• The tool showed reliability across various sequencing coverages, variant allelic fractions, and tumor purities.

Conclusions:

• ClairS-TO is a robust and accurate tool for tumor-only somatic variant calling, particularly for long-read data.
• The method effectively distinguishes somatic variants from germline variants and artifacts without matched normal samples.
• ClairS-TO offers a valuable open-source solution for advancing cancer genomics.