Neural network approach to somatic SNP calling in WGS samples without a matched control.

Somatic variants are usually called by analysing the DNA sequences of a tumor sample in conjunction with a matched normal. However, a matched normal is not always available for instance in diagnostic settings. To unlock such data for basic research single-sample somatic variant calling is required. Previous approaches can not easily be applied in the case of typical whole genome sequencing (WGS) samples.We present a neural network-based approach for calling somatic single nucleotide polymorphism (SNP) variants in tumor WGS samples without a matched normal. The method does not require any manual tuning of filtering parameters and can be applied under the conditions of a typical WGS experiment. We demonstrate the effectiveness of the proposed approach by reporting its performance on 5 SNP datasets corresponding to 5 different cancer types. The proposed method is implemented in Python 3.6 and available as a GitHub repository at https://github.com/heiniglab/deepSNP.