The Variant Caller Learned to See
The Variant Caller Learned to See
Deep learning reframed variant calling as an image problem and beat the hand-tuned statistics at their own benchmark. Here is where the neural nets have already won, where they still stumble, and what the foundation-model era changes next.
For most of the short-read era, calling a variant meant writing down a probability model. The dominant pipelines — GATK foremost among them — encoded a human expert's beliefs about how sequencing errors arise, then estimated genotype likelihoods and applied a stack of hand-tuned filters to separate real variation from artifact.1 Every threshold in that stack was a judgment call, and the quality of a callset depended on how well those judgments matched the messy reality of a given sequencer, prep, and genome region.
Then, in 2018, DeepVariant did something that in retrospect looks inevitable and at the time looked heretical: it threw away the hand-written error model. It rendered the reads piled up around each candidate site as a multi-channel image, handed that image to a convolutional neural network trained on high-confidence truth sets, and let the network learn the relationship between pixel patterns and true genotypes directly from the data.2 No expert-authored likelihood function, no VQSR tranches — just a classifier that had seen enough labeled examples to recognize a real variant when it saw one. It generalized across genome builds, across species, and across sequencing technologies, and it won the PrecisionFDA Truth Challenges outright. The conceptual shift was total: from modeling the error process to learning it.
How Calling Became a Vision Problem
The classical caller is a pipeline of human decisions. Local realignment cleans up indel-adjacent noise; base-quality recalibration adjusts the sequencer's own confidence scores; a Bayesian model computes genotype likelihoods; and a final filtering stage — hard thresholds or a trained recalibration model — decides which calls survive.1 It works, and for a decade it was the standard. But it is only as good as the assumptions baked into each stage, and those assumptions were written by people reasoning about a specific kind of data.
DeepVariant's move was to replace that reasoning with representation learning. Because the network is presented with an image of every read relevant to a putative variant at once, its convolutional layers can capture the complex dependencies among reads that a hand-written model has to approximate explicitly.2 The tool ships trained models rather than tunable filters, and adapting it to a new technology means training on new labeled data, not rewriting statistics. That single design choice — learn the error model instead of authoring it — is the through-line connecting nearly every advance that followed.
The classical caller encodes what an expert believes about sequencing error. The neural caller learns it from the data — and stops needing the expert's thresholds.
Where Deep Learning Has Already Won
The paradigm has not advanced evenly. It is essentially finished in one domain, transformative in a second, and still genuinely open in a third.
Germline short-read: effectively solved
On a well-covered, non-repetitive human genome, deep-learning callers reach SNP F1 scores around 99.9% on both Illumina and PacBio HiFi data.8 For most of the genome this is, for practical purposes, a solved problem. The remaining contest is over the last fraction of a percent — segmental duplications, the MHC, long homopolymers, and other regions where even the truth sets are uncertain. When someone asks whether AI "works" for germline variant calling, the honest answer is that the question is settled; the interesting arguments have moved to the hard 2%.
Long-read: where AI was not optional but essential
Oxford Nanopore and PacBio reads are long enough to resolve repeats and structural variation that short reads cannot, but they historically carried systematic, context-dependent error profiles — exactly the pattern classical callers handle worst. This is where learned callers earned their place rather than merely improving on the status quo. Clair3 pairs a fast pileup network for the easy majority of candidates with a full-alignment network for the hard ones, delivering strong accuracy at low coverage and high speed.3 PEPPER-Margin-DeepVariant added haplotype-aware phasing that lifted nanopore calling to clinical-grade accuracy,4 and Clair3-Trio folded Mendelian inheritance across a mother–father–child family directly into the network to cut inheritance-violating errors.5 The direction of travel is still steep: recent Clair3 releases fold the raw sequencer signal — per-base dwell time — in as an additional input channel, and a dedicated model now extends the approach to long-read RNA. Deep learning is a large part of why an entire sequencing modality became usable in the clinic.
Somatic: the honest frontier
Cancer variant calling is where the paradigm is proven but the problem is not solved. Somatic mutations appear at low and variable allele fractions, tumor purity and ploidy distort the signal, and there is no clean germline truth to train against. NeuSomatic was the first convolutional approach built for this setting, summarizing tumor and matched-normal alignments into feature matrices a network could classify.6 VarNet pushed further with weak supervision, training on 4.6 million high-confidence variants drawn from 356 tumor genomes so that the model could learn without exhaustively hand-labeled truth.7 Both demonstrate that the approach transfers to cancer — but generalization remains fragile: a model trained on synthetic tumors can fall to an F1 near 0.43 on a real one.7 Most of the genuinely open research in AI variant calling lives in this gap.
Germline calling on a clean genome is solved. Somatic calling is not. Nearly all of the interesting research lives in that gap.
| Tool / class | Domain | Representation | What it changed |
|---|---|---|---|
| DeepVariant | Germline, all platforms | Pileup image → CNN | Reframed calling as image classification; swept PrecisionFDA |
| Clair3 | Germline long-read | Pileup + full-alignment | Speed and accuracy at low coverage for ONT and PacBio |
| PEPPER-Margin-DeepVariant | Germline nanopore | Haplotype-aware alignment | Lifted nanopore calling to clinical-grade accuracy |
| NeuSomatic | Somatic (cancer) | Feature matrix → CNN | First convolutional somatic caller |
| VarNet | Somatic (cancer) | Tumor/normal image → weakly supervised CNN | Scaled somatic learning without hand-labeled truth |
A representative slice, not a catalog. The unifying idea across every row is the same: encode the local evidence as a structured input and let a network learn the decision that a filter stack used to make by hand.
What Still Trips the Models Up
The accuracy numbers are real, but they come with conditions that a careful practitioner keeps in view — because each one is a place where a callset can quietly go wrong.
Training-data provenance is the ceiling. Almost every deep-learning caller is trained on Genome in a Bottle truth sets, built from a small number of exhaustively characterized samples. A model can only learn to call what its truth set labels correctly, so the regions those truth sets struggle to resolve are precisely the regions the models struggle with too. The training data is not a neutral input; it is the boundary of what the tool can know.
Distribution shift breaks trained models. A network tuned to one chemistry, coverage depth, or library prep can degrade on another. This is why every new sequencer chemistry — a new nanopore pore, a new PacBio release — triggers a retraining cycle rather than a free lunch. A model's accuracy is a claim about data that resembles its training set, not a universal guarantee.
Calibration and opacity matter in the clinic. A network emits a genotype and a quality score, but the reasoning behind them is not directly inspectable. For research that is tolerable; for a regulated clinical assay it is a real obstacle, because validation and error accountability depend on understanding why a call was made, not just how often the tool is right on a benchmark. Explainability is not a philosophical nicety here — it is a validation requirement.
Structural variants remain comparatively unsolved. The SNP-and-indel problem is largely conquered; deep-learning detection of large insertions, deletions, and rearrangements is far less mature, and remains an active area rather than a settled one.
The Future State
Four trajectories are visible from where the field stands in 2026, and they point toward callers that look quite different from a CNN staring at a pileup.
The backbone stops being frozen. DeepVariant's original Inception-based backbone is now years old, and proof-of-concept work swapping in modern architectures such as EfficientNet reports consistent gains in stability and accuracy on standard benchmarks. The practical shift is philosophical: the variant caller becomes a living model that inherits improvements from the broader deep-learning field, not a fixed artifact.
Transformers move into refinement and joint modeling. Attention-based frameworks are now being applied to the filtering-and-refinement stage — learning which raw calls to trust across the output of conventional callers.9 In parallel, new work trains a single network on long- and short-read data together, so that the complementary strengths of each technology are combined inside the model rather than reconciled afterward.
Genomic foundation models carry a prior over the genome itself. DNA language models — the Nucleotide Transformer, DNABERT-2, HyenaDNA, Evo, Caduceus — are pre-trained on raw sequence at scale, and already predict the functional effects of variants with meaningful accuracy, in some cases zero-shot.10 The trajectory points toward callers and interpreters that carry a learned prior over what a human genome is allowed to look like, rather than reasoning only from the reads in front of them.
Calling moves upstream to the signal and outward to multi-omics. Folding the raw sequencer signal directly into the model — as dwell-time-aware long-read calling now does — pushes intelligence closer to the instrument. Extending the same machinery to RNA, methylation, and integrated multi-omic context pushes it outward, toward a picture of a sample that no single-modality caller can produce.
The next caller may not look at a pileup at all. It may carry a pre-trained prior over what a human genome is allowed to be.
The Honest Summary
Artificial intelligence did not merely improve variant calling; it changed what a variant caller is — from a probability model a human authored into a function learned from truth data. For germline SNPs and indels on a well-covered genome, that transition is essentially complete and the accuracy ceiling is high enough that the tool choice rarely decides the result. For long reads, it was the difference between an unusable error profile and a clinically deployable one. For somatic mutations, structural variants, and the hard corners of the genome, the paradigm is proven but the work is unfinished — and that is where the field's real momentum sits.
What this means for practice is that the analyst's job has shifted, not disappeared. It is no longer about finding a magic tool; it is about knowing what each model was trained on, where its training distribution ends, and whether the sample in hand sits inside or outside that boundary. A learned caller is only as trustworthy as the truth set behind it and the validation you put in front of it. Understanding that — not running the tool, which anyone can do — is what a serious genomics engagement is actually for.
Lexington, KY · Genomics · Transcriptomics · Proteomics
Publications
- McKenna A., Hanna M., Banks E., et al. (2010). The Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation DNA sequencing data. Genome Research, 20(9), 1297–1303. doi:10.1101/gr.107524.110
- Poplin R., Chang P.-C., Alexander D., et al. (2018). A universal SNP and small-indel variant caller using deep neural networks. Nature Biotechnology, 36, 983–987. doi:10.1038/nbt.4235
- Zheng Z., Li S., Su J., Leung A.W.S., Lam T.-W. & Luo R. (2022). Symphonizing pileup and full-alignment for deep learning-based long-read variant calling (Clair3). Nature Computational Science, 2, 797–803. doi:10.1038/s43588-022-00387-x
- Shafin K., Pesout T., Chang P.-C., et al. (2021). Haplotype-aware variant calling with PEPPER-Margin-DeepVariant enables high accuracy in nanopore long-reads. Nature Methods, 18, 1322–1332. doi:10.1038/s41592-021-01299-w
- Su J., Zheng Z., Ahmed S.S., Lam T.-W. & Luo R. (2022). Clair3-Trio: high-performance Nanopore long-read variant calling in family trios with Trio-to-Trio deep neural networks. Briefings in Bioinformatics, 23(5), bbac301. doi:10.1093/bib/bbac301
- Sahraeian S.M.E., Liu R., Lau B., Podesta K., Mohiyuddin M. & Lam H.Y.K. (2019). Deep convolutional neural networks for accurate somatic mutation detection (NeuSomatic). Nature Communications, 10, 1041. doi:10.1038/s41467-019-09027-x
- Krishnamachari K., Lu D., Swift-Scott A., et al. (2022). Accurate somatic variant detection using weakly supervised deep learning (VarNet). Nature Communications, 13, 4248. doi:10.1038/s41467-022-31765-8
- Abdelwahab O. & Torkamaneh D. (2025). Artificial intelligence in variant calling: a review. Frontiers in Bioinformatics, 5, 1574359. doi:10.3389/fbinf.2025.1574359
- Abdelwahab O. & Torkamaneh D. (2026). A Transformers-based framework for refinement of genetic variants. Frontiers in Bioinformatics, 5, 1694924. doi:10.3389/fbinf.2025.1694924
- Dalla-Torre H., Gonzalez L., Mendoza-Revilla J., et al. (2025). Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nature Methods, 22, 287–297. doi:10.1038/s41592-024-02523-z