Fig. 1.

NeuSomatic overview. a Toy example of input matrix preparation for a given candidate somatic SNV. Sequence alignment information in a window of seven bases around the candidate somatic mutation is extracted. The reference sequence is then augmented by adding gaps to account for insertions in the reads. The augmented alignment is then summarized into the reference matrix, the tumor count matrix, and the normal count matrix. The count matrices record the number of A/C/G/T and gap (‘-‘) characters in each column of the alignment, while the reference matrix records the reference bases in each column. The count matrices are then normalized by coverage to reflect base frequencies in each column. Separate channels are reserved to record the tumor and normal coverages. b The input 3-dimensional matrix and the proposed NeuSomatic network architecture. The input matrix consists of reference channel, tumor and normal-frequency channels, coverage and position channels, followed by several channels summarizing the alignment features. When used in ensemble mode, NeuSomatic also includes additional channels for other individual methods features. NeuSomatic network architecture consists of nine convolutional layers structured in four blocks with shortcut identity connections. We use two softmax classifiers and one regressor on the final layer to predict the mutation type, size, and position