Direct comparison
DeepVariant vs GATK HaplotypeCaller
Compares DeepVariant's deep-learning accuracy and GPU needs against GATK HaplotypeCaller's mature, CPU-based pipeline, and when switching pays off.
Written and maintained by CASRAI Editorial Board
Last updated
Ask CASRAI · included with Regulatory Radar
Ask about DeepVariant vs GATK HaplotypeCaller
Ask CASRAI answers research-administration questions and cites the passages behind every claim — and says so when the corpus does not cover something, instead of guessing. It comes with a Regulatory Radar subscription at $29 a month, alongside the daily digest of regulatory changes and the dashboard of what changed.
150 questions a day, on this site, over the API, or inside your own tools through the CASRAI MCP server.
Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.
How do DeepVariant, GATK HaplotypeCaller compare side by side?
The table below compares DeepVariant, GATK HaplotypeCaller across 13 procurement-relevant dimensions, from core method through best fit.
Side-by-side comparison
| Dimension | DeepVariant | GATK HaplotypeCaller |
|---|---|---|
| Core method | Converts aligned read pileups into image-like tensors and classifies each candidate site with a trained convolutional neural network (CNN) — variant calling treated as an image-classification problem rather than a statistical inference problem. | Determines 'active regions' with likely variation, performs local de novo reassembly of reads into candidate haplotypes, then scores each haplotype against the reads using a pair hidden Markov model (PairHMM) to compute genotype likelihoods. |
| Developer / origin | Google (Google Health / Google Brain), first released 2017, published in Nature Biotechnology (Poplin et al., 2018). | Broad Institute, part of GATK since GATK3 (with a substantial rewrite in GATK4) — the default germline caller in GATK Best Practices for over a decade. |
| Reported accuracy in independent benchmarks | Won best-SNP performance in the 2016 PrecisionFDA Truth Challenge and won the ONT, PacBio, and Multiple-Technologies categories of PrecisionFDA Truth Challenge V2 (2020), evaluated against Genome in a Bottle (GIAB) truth sets. Independent published comparisons have generally found DeepVariant producing fewer indel-calling errors than HaplotypeCaller, though the margin varies by dataset, coverage, and GATK/DeepVariant version — always re-check current-version numbers before citing a specific figure. | Consistently strong, well-characterized SNP accuracy and mature indel handling via local reassembly; it was the accuracy baseline several PrecisionFDA and GIAB-based benchmarking rounds were measured against. The gap to DeepVariant has generally narrowed release over release as both tools have been refined. |
| Compute requirements | Three-stage pipeline (make_examples → call_variants → postprocess_variants). Runs CPU-only but is compute-heavy; the call_variants CNN-inference stage benefits substantially from GPU acceleration, and Google documents an experimental GPU-accelerated mode for exactly this stage. | CPU-only by design — no GPU support needed or offered. Scales with standard multi-core parallelization (scatter-gather across genomic intervals), which is why it remains the default choice for institutions without GPU infrastructure. |
| Typical cost / runtime (30x WGS) | Google's own published reference figures for a 30x whole genome on Google Cloud: roughly $11.80 on non-preemptible instances, or about $2.84 using preemptible instances (cloud pricing changes over time and varies by region/instance type — treat as an order-of-magnitude reference, not a current quote). | No official standard cost benchmark comparable to DeepVariant’s published cloud figures, since it is far more often run on existing on-premises HPC/CPU infrastructure than provisioned per-job on cloud GPU instances; per-genome cost is typically dominated by ordinary CPU-hour pricing rather than a specialised accelerator. |
| Multi-sample / cohort (joint) genotyping | Produces per-sample gVCF output, then joint-genotypes a cohort with GLnexus (a separate, purpose-built merging/joint-calling tool), not with GATK's own CombineGVCFs/GenomicsDBImport/GenotypeGVCFs chain. | Run per-sample in -ERC GVCF mode, then joint-genotyped with GATK's own CombineGVCFs (or GenomicsDBImport for larger cohorts) followed by GenotypeGVCFs — the standard GATK cohort workflow. |
| Sequencing technology support | Ships separate pre-trained models tuned for Illumina short reads, PacBio HiFi long reads, and Oxford Nanopore long reads (plus hybrid/multi-technology modes) — its PrecisionFDA wins spanned all three. | Designed and most extensively validated for Illumina short-read data; usable on long-read platforms but without the technology-specific tuning DeepVariant's separate long-read models provide. |
| Integration with existing GATK pipelines | Drops into a GATK-adjacent workflow at the calling step — its gVCF output is format-compatible — but the joint-genotyping and several downstream QC conventions built around GATK's own tool suite need adjusting (e.g. swapping in GLnexus). | Native to GATK Best Practices end-to-end: base quality score recalibration, variant quality score recalibration (VQSR), and the whole downstream annotation/filtering toolchain assume HaplotypeCaller-shaped output. |
| Tooling maturity / operational track record | Actively maintained by Google, with a growing production track record (e.g. large consortium pipelines using DeepVariant+GLnexus), but a shorter operational history than HaplotypeCaller and a smaller base of accumulated pipeline-specific troubleshooting knowledge. | Over a decade of production use across clinical, translational, and population-scale genomics; extensive documentation, community troubleshooting history, and validated deployment patterns in regulated (e.g. clinical diagnostic) settings. |
| Retraining / customization | Because it is a trained model, DeepVariant can in principle be retrained or fine-tuned on lab-specific data (Google publishes training scripts and case-study models for non-human and non-standard-coverage data), but doing so credibly requires ML expertise and curated truth data most wet-lab or clinical genomics groups do not have in-house. | Behavior is controlled through documented, well-understood parameters (e.g. active-region thresholds, ploidy, ERC mode) rather than retraining — tuning it does not require machine-learning expertise, which is one reason it remains approachable for teams without a dedicated bioinformatics-ML function. |
| Regulatory / clinical validation history | Adopted in some clinical and large-scale research pipelines and increasingly cited in validation literature, but with a shorter track record of published clinical-validation studies than HaplotypeCaller, since it is the newer tool of the two. | The far longer incumbent in clinical and diagnostic germline pipelines, with a correspondingly larger body of published validation studies and regulatory-submission precedent to draw on — a real practical advantage where validation burden, not raw accuracy, is the binding constraint. |
| Explainability of individual calls | A CNN classification score is harder to manually inspect and explain than a probabilistic model's intermediate values — useful for aggregate accuracy, less convenient when a specific call needs to be manually adjudicated and justified. | The PairHMM/local-reassembly model produces intermediate likelihoods and haplotype evidence that are more directly interpretable, which can matter when a lab needs to manually justify a specific borderline call (e.g. in a clinical report). |
| Best fit | Labs with GPU access (or cloud budget for GPU instances) running at meaningful scale, working with PacBio/ONT long reads, or where the accuracy gain over HaplotypeCaller is worth added pipeline complexity and validation effort — particularly for indel-sensitive applications. | Labs on CPU-only infrastructure, pipelines already built around GATK Best Practices (VQSR, joint genotyping, established QC), regulated/clinical settings valuing a long validation track record, or workloads where DeepVariant’s accuracy edge is not the limiting factor. |
Common questions
Common questions about DeepVariant vs GATK HaplotypeCaller
Is DeepVariant actually more accurate than GATK HaplotypeCaller?
+
In successive PrecisionFDA Truth Challenges (2016, and Truth Challenge V2 in 2020) DeepVariant placed at or near the top on SNP and indel accuracy against GIAB truth sets, and several independent published comparisons have found it making fewer indel errors than HaplotypeCaller. The exact margin depends on the dataset, coverage, and the specific versions of each tool being compared, so treat 'more accurate' as a real, benchmark-supported pattern rather than a fixed, universal percentage.
Does DeepVariant require a GPU to run?
+
No — it runs CPU-only. A GPU is not required but is recommended for production-scale use, because it substantially speeds up the call_variants (CNN inference) stage. For a lab running occasional single-genome analyses, CPU-only DeepVariant is workable; for routine, high-volume calling, the GPU-accelerated path is where most of the runtime benefit shows up.
Can DeepVariant slot into an existing GATK-based pipeline?
+
Partially. DeepVariant's gVCF output is format-compatible with GATK's, so it can replace HaplotypeCaller at the per-sample calling step. But GATK's own joint-genotyping chain (CombineGVCFs/GenomicsDBImport → GenotypeGVCFs) is not the tool DeepVariant's authors validated for cohort merging — that role is filled by GLnexus — so a full swap means adjusting the cohort-calling and some downstream QC steps, not just the caller itself.
Which tool is better for PacBio or Oxford Nanopore long-read data?
+
DeepVariant, generally. It ships separate pre-trained models tuned specifically for PacBio HiFi and Oxford Nanopore reads, and won the ONT and PacBio categories of PrecisionFDA Truth Challenge V2. HaplotypeCaller was designed and is most extensively validated for Illumina short reads; it can be run on long-read data but without equivalent platform-specific tuning.
Is switching from HaplotypeCaller to DeepVariant worth it for a small lab?
+
Often not, if the lab lacks GPU infrastructure and runs CPU-only Illumina short-read pipelines already built around GATK Best Practices. The accuracy gain is real but incremental for many use cases, and it has to be weighed against the cost of GPU provisioning, revalidating a pipeline (especially in regulated settings), and adopting GLnexus for joint genotyping. It is more clearly worth it for labs already on GPU infrastructure, working with long-read data, or running indel-sensitive analyses where the accuracy difference has downstream consequences.
