DeepSom: a CNN-based approach to somatic variant calling in WGS samples without a matched normal

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Somatic mutation detection typically requires paired normal and tumor DNA samples.
• The absence of matched normal samples, common in retrospective studies and diagnostics, poses a challenge for variant calling.
• Existing tumor-only methods show limitations, particularly with whole-genome sequencing (WGS) data.

Purpose of the Study:

• To develop and validate a novel computational tool for accurate somatic variant calling from tumor-only WGS data.
• To address the limitations of current tumor-only approaches in detecting single nucleotide polymorphisms (SNPs) and short insertions/deletions (indels).

Main Methods:

• Development of DeepSom, a convolutional neural network (CNN)-based approach for somatic variant detection.
• Utilizing tumor whole-genome sequencing data exclusively, without requiring matched normal samples.
• Performance evaluation across five diverse cancer datasets.

Main Results:

• DeepSom effectively identifies somatic single nucleotide polymorphisms and short insertion/deletion variants.
• The tool demonstrates superior performance compared to previous tumor-only variant calling methods on WGS samples.
• Validation across multiple cancer datasets confirms the robustness and accuracy of DeepSom.

Conclusions:

• DeepSom offers a robust solution for somatic mutation detection in tumor WGS data when matched normal samples are unavailable.
• This method enhances the utility of tumor-only sequencing for cancer research and diagnostics.
• The developed tool is publicly available, facilitating its adoption in the scientific community.