Germline SNP and small-indel variant caller that classifies pileup images of aligned reads with a CNN, with pretrained models per sequencing platform.
No providers recorded yet. Browse all providers
Every short read is a noisy witness. Calling a genetic variant means deciding, at each site where reads disagree with the reference, whether the disagreement reflects the genome or the instrument — and the errors that mislead a caller are not independent. Classical callers such as GATK model each error source with its own statistical component, assume reads err independently, and must be re-engineered for every new sequencing chemistry.
DeepVariant, developed by Google with Verily Life Sciences, replaces that assembly of hand-built models with a single learned one. Around each candidate site it renders the reference and all overlapping reads as an image-like tensor, and a convolutional neural network classifies it into one of three diploid genotypes: homozygous reference, heterozygous or homozygous alternate. Because the network sees every read together, it can learn the joint, correlated error structure that per-read likelihood models ignore. The method appeared as a bioRxiv preprint in December 2016, was released as open-source software with trained models in December 2017, and was published in Nature Biotechnology in 2018.
The paper's claim to universality is about transfer. A model trained on reads aligned to GRCh37 called variants on GRCh38 alignments at an overall F1 of 99.45%, against 99.53% for a model trained on GRCh38 directly. A model trained on human data reached F1 98.29% on a mouse dataset sequenced on a different instrument at different depth, outperforming a model trained on the mouse data itself (97.84%). In practice, users run a fixed, pretrained checkpoint chosen by --model_type on their own alignments; support for each new sequencing platform arrives as a newly released checkpoint.
The 2018 model used the Inception-v2 architecture on three-channel RGB pileup images; later releases moved to Inception-v3 on the multi-channel tensor layout. The v1.10 WGS checkpoint is about 22 million parameters. Models are trained on Genome in a Bottle reference samples with benchmarked truth sets: the v1.10 WGS model on 943 million labelled examples from replicates of HG001, HG002 and HG004–HG007, the PacBio model on 965 million and the ONT model on just over one billion. Current releases exclude HG003 from training and use it as the standard evaluation sample. On HG003 the v1.10 WGS model reaches SNP F1 0.9961 and indel F1 0.9960; the PacBio HiFi model reaches 0.9985 and 0.9942, and the hybrid PacBio–Illumina model 0.9992 and 0.9968.
In the original evaluation, a model trained only on NA12878 won the SNP "highest performance" award in the 2016 precisionFDA Truth Challenge on the blinded sample NA24385, and made 4,652 errors per genome against 9,531 for the next-best method tested. The pangenome-aware extension reduced errors by up to 25.5% relative to linear-reference DeepVariant across short-read platforms and mappers, and gave 23.6% more accurate calls on Element reads than existing methods.
DeepVariant is a standard germline calling step for research whole-genome and exome sequencing, population cohorts, and long-read projects. Multi-sample cohorts are joint-genotyped by merging per-sample gVCFs with GLnexus. It runs from a Docker image on CPUs, GPUs or TPUs. The shipped models are human-trained and support diploid calling only, with haploid handling for sex chromosomes; non-human projects can apply them directly or retrain on species-specific truth data using the published training tutorial.
DeepVariant demonstrated that variant calling could move from expert-tuned, technology-specific statistics to a general model optimized against truth data, and it won the Nanopore, PacBio and multi-technology categories of the 2020 precisionFDA Truth Challenge V2. Its pileup-tensor design became the template for Google's later genomics models — DeepTrio for families, DeepSomatic for tumour samples and DeepPolisher for assembly polishing. Its limits are explicit: calling is restricted to diploid SNPs and small indels, structural variants are out of scope, and accuracy depends on the upstream candidate-generation step, which bounds sensitivity.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.