Skip to content

What DeepVariant Is and How to Run It on Your Own Genome

Oak
A feathered dinosaur-like creature on a dark plinth under studio light, its layered iridescent plumage broken by a few mismatched bright feathers.

DeepVariant is an open-source germline variant caller from Google. It takes aligned sequencing reads (BAM or CRAM) plus the reference genome they were aligned to, and produces a VCF listing the single-nucleotide variants and small insertions and deletions where your genome differs from that reference. The distinguishing design choice is that it renders each candidate site as a stack of image-like tensors and classifies the genotype with a convolutional neural network. It does not use a hand-written statistical model of sequencing error. If you have a 30x whole-genome BAM and want a variant list you can trust for common analyses, DeepVariant is the caller we would reach for first.

Why the image-classification approach matters

Traditional callers such as GATK HaplotypeCaller compute genotype likelihoods from an explicit error model, then apply filters tuned by hand or by a separate recalibration step. DeepVariant replaces that with a supervised classifier. For each candidate position it encodes the features below as channels of a pileup tensor.

  • read bases
  • base quality
  • mapping quality
  • strand
  • whether each base supports the candidate allele
  • several other per-read features

The network outputs probabilities over three genotype classes: homozygous reference, heterozygous, and homozygous alternate.

The practical consequence is that error behavior particular to a platform or chemistry gets learned from training data rather than modeled in advance. That is why DeepVariant ships separate weights per data type, and why new sequencing chemistries are supported by retraining rather than by rewriting the caller. Element Biosciences avidity sequencing, for instance, was brought to high accuracy largely by retraining and by adjusting how reads are represented in the tensor 1. Independent comparisons of SNP-calling tools consistently place deep learning callers ahead of older probabilistic pipelines on Illumina short reads, particularly for indels 2.

The cost is opacity. When DeepVariant makes an odd call you cannot inspect a likelihood term and reason about it. You inspect the pileup and the training domain instead.

Running it

DeepVariant is distributed as a Docker image. Pin the version tag explicitly, because model weights and default behavior change between releases and you want your calls reproducible.

docker run -v "$PWD":/data google/deepvariant:1.8.0 \
  /opt/deepvariant/bin/run_deepvariant \
  --model_type=WGS \
  --ref=/data/GRCh38_no_alt_analysis_set.fasta \
  --reads=/data/sample.cram \
  --output_vcf=/data/sample.vcf.gz \
  --output_gvcf=/data/sample.g.vcf.gz \
  --num_shards=$(nproc) \
  --intermediate_results_dir=/data/dv_tmp \
  --logging_dir=/data/dv_logs

A few flags carry most of the weight. --model_type selects the weights and must match your data: WGS for short-read whole genome, WES for exome or panel, PACBIO for HiFi, ONT_R104 for recent Oxford Nanopore chemistry, and HYBRID_PACBIO_ILLUMINA when you have both. Using the WGS model on exome data inflates false positives at capture boundaries, so do not skip this.

--num_shards parallelizes the make_examples stage, which is CPU-bound and the longest part of the run. On a 64-vCPU machine a 30x human genome finishes in roughly one to two hours. The middle stage, call_variants, is neural network inference and benefits from a GPU or from NVIDIA Parabricks if you have one available. --intermediate_results_dir keeps the sharded TFRecord examples on disk. That is what you want the first time you run this, because it lets you restart a failed postprocess_variants without redoing the pileups.

Always request --output_gvcf. The gVCF records confidence at reference-matching positions, which is how you later distinguish “this position matches the reference” from “this position had no usable coverage”. Without it you cannot tell an absent variant from an absent measurement.

Two more flags matter for a personal genome. Pass --haploid_contigs=chrX,chrY and --par_regions_bed=... for a male sample so that non-PAR sex chromosome sites are called as haploid rather than forced into diploid genotypes. And run against the analysis set reference with alternate contigs excluded. Aligning to a reference with ALT contigs present and then calling naively scatters reads across paralogous sequence and depresses coverage in the regions you care about most.

Reading the VCF

DeepVariant writes a FORMAT field of GT:GQ:DP:AD:VAF:PL. The two you will use constantly are GQ and VAF. GQ is the phred-scaled genotype quality, and VAF is the fraction of reads supporting the alternate allele. The FILTER column holds either PASS, RefCall (the network concluded homozygous reference at a candidate site), or LowQual. For downstream analysis start by taking PASS records only.

bcftools view -f PASS -i 'GQ>=20' sample.vcf.gz -Oz -o sample.pass.vcf.gz
bcftools index sample.pass.vcf.gz
bcftools stats sample.pass.vcf.gz | head -40

Sanity-check the totals before you interpret anything. A 30x short-read human genome against GRCh38 should yield roughly 4.5 to 5 million variants, with a SNV-to-indel ratio near 8:1 and a transition-to-transversion ratio close to 2.0 to 2.1. A het/hom ratio far from about 1.5 to 1.6, or a Ti/Tv below 1.9, usually signals a contamination or alignment problem rather than an interesting genome.

For annotation, pass the filtered VCF to Ensembl VEP or bcftools csq with a matching GFF. Keep the coordinate build stamped in every filename. Mixing GRCh37 and GRCh38 positions is the single most common error we see in personal genome analysis.

Measuring accuracy on your own data

Benchmark numbers from papers describe reference samples under reference conditions. Your library, your coverage, and your alignment are different. If accuracy matters to you, measure it.

The standard method is to sequence a Genome in a Bottle reference sample alongside your own, then compare calls against the truth set with hap.py using the vcfeval engine, restricted to the high-confidence BED.

hap.py truth.vcf.gz sample.pass.vcf.gz \
  -f HG002_GRCh38_1_22_v4.2.1_benchmark.bed \
  -r GRCh38_no_alt_analysis_set.fasta \
  -o happy_out --engine=vcfeval --threads 16

Read the stratified output. The headline F1 is not enough. Tools that look near-identical genome-wide diverge sharply in segmental duplications, homopolymers, and other low-complexity sequence. The GA4GH stratification BEDs let you see that directly. Also note that older truth sets are themselves incomplete in hard regions, which is why newer long-read-derived benchmarks built on multi-generation pedigrees expand the evaluable fraction of the genome and expose errors the old sets could not score 3.

Two directions are worth watching. Pangenome-aware DeepVariant conditions the pileup representation on a graph reference rather than one linear sequence, which reduces reference bias and improves calling in regions where your haplotypes diverge from GRCh38 4. Separately, work on lighter network architectures shows the classifier can be made substantially cheaper without a large accuracy cost, which matters if you are running many samples 5.

Where it is weak

DeepVariant calls SNVs and small indels. It does not call structural variants, large repeat expansions, or copy number changes. Nor does it phase variants into haplotypes on its own, and it will not resolve the regions where short reads simply cannot map uniquely. Recovering both short and long variant classes across a whole genome takes a combination of read lengths and callers rather than one tool 6.

It also assumes a germline diploid sample. Somatic and mosaic variation at low allele fraction falls outside the training distribution, and the genotype classifier will tend to call those sites reference. Detecting mosaicism reliably generally requires a different assay design and often long reads 7. Applying variant calling to RNA-based data is an active area with its own error structure and its own caveats 8.

Finally, a variant call is a measurement. It is not a finding about your health. Even a high-confidence PASS call in a well-known gene is a hypothesis from a research-grade pipeline. Clinical interpretation requires confirmatory testing in an accredited laboratory and a conversation with a clinician or genetic counselor. Do not act on a research VCF.

Questions people also ask

What is DeepVariant used for? Germline variant calling: converting aligned reads into a list of SNVs and small indels with genotypes and quality scores. It is used for human whole genomes and exomes, for long-read data from PacBio and Oxford Nanopore, and widely in non-human genomics as well.

What is Google DeepVariant? It is the same tool. DeepVariant was developed at Google and released as open source in 2017, with the method published the following year and covered at the time as a notable reduction in genome error rates 9. Development continues in the public repository under a BSD-3 license.

What does it mean if an NGS test is positive? In the narrow sense it means the pipeline called a variant that matches something in a database or a filter you applied. It does not by itself mean the variant is real, nor that it explains any symptom. Confirmation and clinical interpretation belong with a clinician.

Is Deep Genomics a startup? Deep Genomics is a separate company working on RNA therapeutics and machine learning for drug discovery. It is unrelated to DeepVariant despite the similar name.

Footnotes

  1. Andrew Carroll, Alexey Kolesnikov, Daniel E. Cook, et al. Accurate human genome analysis with element avidity sequencing. BMC Bioinformatics, 2025. https://doi.org/10.1186/s12859-025-06191-4 ↩

  2. Shikhi Baruri, Sunita Khanal. SNP Detection Strategies in Genomic Research: A Comparative Review of Major Tools, Algorithms, Challenges and Applications. Nepal Journal of Biotechnology, 2026. https://doi.org/10.54796/njb.v14i1.481 ↩

  3. Zev Kronenberg, Cillian Nolan, David Porubsky, et al. The Platinum Pedigree: a long-read benchmark for genetic variants. Nature Methods, 2025. https://doi.org/10.1038/s41592-025-02750-y ↩

  4. Mobin Asri, Pi-Chuan Chang, Juan Carlos Mier, et al. Pangenome-aware DeepVariant. 2025. https://doi.org/10.1101/2025.06.05.657102 ↩

  5. Anastasiia Gurianova, Anastasiia Pestruilova, Aleksandra Beliaeva, et al. Rethinking DeepVariant: Efficient Neural Architectures for Intelligent Variant Calling. International Journal of Molecular Sciences, 2026. https://doi.org/10.3390/ijms27010513 ↩

  6. Robert J. M. Eveleigh, Sarah J. Reiling, J. Hector Galvez, et al. Finding an optimal sequencing strategy to detect short and long genetic variants in a human genome. 2025. https://doi.org/10.1101/2025.05.30.656631 ↩

  7. Carolina Montano, Luke Morina, Elizabeth Wohler, et al. P224: Using long-read sequencing for genomic and epigenomic analysis in patient-derived samples with somatic IDH1 mosaicism. Genetics in Medicine Open, 2024. https://doi.org/10.1016/j.gimo.2024.101121 ↩

  8. Satoshi Oota. AI/ML-Driven Gene Analysis: New Perspectives on Variant Calling in Normal Human Tissue Using scRNA-seq Data. Methods in Molecular Biology, 2026. https://doi.org/10.1007/978-1-0716-5284-8_12 ↩

  9. Chris Anderson. Google’s AI Tool DeepVariant Promises Significantly Fewer Genome Errors. Clinical OMICs, 2018. https://doi.org/10.1089/clinomi.05.01.21 ↩

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.