Deep convolutional neural networks for accurate somatic variant calling

We present DeepSomatic, the first convolutional neural network approach for somatic variant calling, which significantly outperforms previous techniques on different sequencing platforms, sequencing strategies, and tumor purities. DeepSomatic summarizes sequence alignments into small matrices and can incorporate more than a hundred features to capture variant signals effectively. It can be used universally as a stand-alone somatic variant caller or with an ensemble of existing callers to achieve the highest accuracy.