deepvariant

Why we included this project

DeepVariant does something unusual: it treats DNA sequencing data as images. Aligned reads from a BAM or CRAM file are turned into pileup tensors, and a convolutional neural network classifies each one, writing the results as standard VCF or gVCF. Google ships pretrained models for specific sequencing platforms, so labs working with Illumina whole-genome or exome, PacBio HiFi, or Oxford Nanopore reads can run the matching model instead of training their own. It targets germline variant calling in diploid organisms, and the same machinery has been extended to trio calling in DeepTrio and somatic calling in DeepSomatic. The README documents case studies for each supported platform, which makes reproducing the pipeline on your own data considerably easier.