Teaching neural networks to spot typos in DNA

What it does DeepVariant is a germline variant caller for diploid organisms. It ingests aligned sequencing reads in BAM or CRAM format, encodes each candidate site as a pileup image tensor, and runs it through a convolutional neural network. The output is a standard VCF or gVCF file. The tool supports a wide range of sequencing technologies—Illumina, PacBio HiFi, Oxford Nanopore, Complete Genomics, Roche SBX, and RNA-seq—and includes DeepTrio for joint calling on parent-child trios or duos.

The interesting bit The core insight is treating variant calling as computer vision rather than statistical inference: each genomic position becomes an image that the model inspects for patterns. Under the hood it relies on Nucleus, a Google-built library that glues genomics file formats to TensorFlow. Deployment is Docker-first, with optional GPU and TPU acceleration, and Google quotes roughly $2.84 to $11.80 to call a 30x human whole genome on cloud instances depending on preemptible pricing.

Key highlights

• Broad sequencer support: NGS (Illumina/Element), PacBio HiFi, Oxford Nanopore R10.4.1, Complete Genomics, Roche SBX, and both Illumina and PacBio RNA-seq.
• DeepTrio extension for trio/duo joint calling, using GLnexus to merge output VCFs.
• All bundled models are trained exclusively on human data; other species require retraining or caution.
• Strictly diploid—only hom-alt, het, and hom-ref genotypes are supported.
• Docker-native with GPU/TPU support; an experimental pipeline can overlap CPU preprocessing with GPU inference.

Caveats

• Every model ships trained on human data, so non-human variant calling is effectively bring-your-own-model.
• Hard-coded diploid ploidy means polyploid or aneuploid organisms are out of scope.
• Unix-like only; Windows is explicitly unsupported.

Verdict Human genomics pipelines that need germline calling from diverse sequencers should consider this a default option. If you are studying polyploid crops, non-model organisms without retraining budget, or somatic tumor evolution, this is not your tool—point somatic work to the separate DeepSomatic repo instead.