Clair3 Variant Calling: How to Run It on Your Own Long-Read Genome
Clair3 is the germline small-variant caller we would use on any long-read whole genome. For most people the entire job is one command. Point run_clair3.sh at a sorted BAM, a reference FASTA, and a model directory that matches the basecaller that produced the reads. Everything else is tuning. The parts worth understanding are why Clair3 runs two different neural networks over the same data, how to pick the right model, and where the output will be wrong in ways the quality score does not warn you about.
This page covers the mechanics of running Clair3 and the flags that change results rather than runtime. It also covers how to measure your own accuracy against a truth set and the failure modes to plan around.
What the two-stage architecture buys you
Clair3 calls variants twice. The first pass builds a pileup tensor, a summarized count of bases and indels stacked at each position, and feeds it to a small network. This is cheap and resolves the large majority of sites, because most of the genome is either homozygous reference or an obvious heterozygous SNP. Sites where the pileup model is uncertain get promoted to the second pass.
The second pass builds a full-alignment tensor, which encodes individual reads with their base identities, quality scores, strand, and mapping quality across a window around the candidate. That representation preserves haplotype structure the pileup throws away, so it recovers indels in homopolymers and variants in repetitive context that pileup counts alone cannot resolve. The cost is roughly two orders of magnitude more compute per site, which is exactly why Clair3 only spends it where it helps.
The fraction promoted is controlled by --var_pct_full and --ref_pct_full. The default for --var_pct_full is 0.3, meaning the lowest-quality 30% of pileup variant calls. The default for --ref_pct_full is 0.3 for ONT and 0.1 for HiFi, applied to reference calls. Raising these improves recall slightly and costs a lot of time. We leave them alone. A later engineering pass on Clair3 reworked the candidate selection and tensor generation to cut whole-genome runtime substantially without changing the calling model.1
The command we run
For an ONT genome aligned with minimap2 against GRCh38:
run_clair3.sh \
--bam_fn=sample.sorted.bam \
--ref_fn=GRCh38_no_alt.fa \
--threads=32 \
--platform=ont \
--model_path=/opt/models/r1041_e82_400bps_sup_v500 \
--output=clair3_out \
--sample_name=SAMPLE \
--gvcf \
--use_longphase_for_intermediate_phasing \
--remove_intermediate_dir
A few notes on the non-obvious parts. --gvcf produces merge_output.gvcf.gz alongside the VCF, which records confident reference blocks. Without it you cannot distinguish “reference at this position” from “no data at this position”, and that distinction matters if you ever want to say something about a site where a variant was not called. The cost is disk and a modest runtime increase.
--remove_intermediate_dir deletes the per-chunk tensors and logs. Keep them the first time you run, because they are the only way to diagnose a chunk that failed silently. --include_all_ctgs forces calling on decoys and unplaced contigs, which we leave off for GRCh38. If you are working on a non-human or a custom reference, you need it.
For PacBio HiFi, set --platform=hifi and use the matching model. For short reads, Clair3 has an ilmn platform, but on Illumina data we would use DeepVariant or GATK instead. Benchmarks of coding-region variant discovery on short reads show the choice of caller interacts strongly with the aligner and the variant type, and the established short-read pipelines are better characterized there.2
Model selection is the decision that matters most
Clair3 models are trained against a specific chemistry and basecaller configuration. An R10.4.1 flow cell basecalled with a super-accuracy (sup) model has a different error profile from the same flow cell basecalled with hac, and the model name encodes this. The name r1041_e82_400bps_sup_v500 means R10.4.1, E8.2 pore, 400 bases per second, sup basecalling, and Dorado v5.0.0 models. Using a model trained for a different basecaller version is the single most common source of quietly degraded accuracy. It lowers your F1 by a percentage point or two, mostly in indels. It does not crash.
ONT distributes newer models through the Rerio repository rather than bundling them with Clair3 releases. Check what basecaller version produced your reads (dorado writes it into the BAM @PG line) and pick the closest available model. If your exact version has no model, prefer the nearest earlier one from the same chemistry over a newer one from different chemistry.
Several Clair3 variants exist for situations the base model does not cover. Clair3-Trio uses a trio-to-trio network across a child and both parents jointly, which improves calls by enforcing Mendelian consistency during inference rather than filtering afterwards.3 Clair3-MP takes aligned data from two platforms at once and improves on either alone, which is relevant if you have both ONT and HiFi for the same person.4 There is also work incorporating the ONT move table, the signal-to-base mapping emitted by the basecaller, to give the caller access to information lost in the base sequence.5
Coverage, and what accuracy to expect
Clair3 on ONT reaches high SNP accuracy at moderate depth and keeps improving to roughly 40x, after which returns flatten. Indels are the harder problem and remain the dominant error class on nanopore data, concentrated in homopolymers and short tandem repeats. Comparative evaluations of nanopore callers consistently find Clair3 among the strongest performers on both SNVs and indels.6
Below about 15x we would not trust heterozygous calls without inspection, because allele sampling noise makes a true het look homozygous often enough to matter. For haploid regions, chrX and chrY in a male sample, use --haploid_precise or --haploid_sensitive on those contigs in a separate run. Calling them as diploid produces spurious heterozygous calls.
For any use of these results in a medical context, a clinical laboratory has to perform its own validation and a clinician has to interpret the findings. Validation studies of long-read platforms for diagnostic use define the confirmation and reporting workflow explicitly, and a research-grade Clair3 VCF is not a substitute for it.7
Measuring your own accuracy
Do not accept a caller’s reputation on your own data. Run a Genome in a Bottle sample through the exact pipeline you plan to use and compare with hap.py:
hap.py HG002_GRCh38_v4.2.1_benchmark.vcf.gz clair3_out/merge_output.vcf.gz \
-f HG002_GRCh38_v4.2.1_benchmark_noinconsistent.bed \
-r GRCh38_no_alt.fa \
-o happy_out \
--engine=vcfeval \
--threads 16
The --engine=vcfeval flag matters. The default xcmp engine does not handle complex representational differences between an indel written one way in the truth set and another way by the caller. Use vcfeval and you get haplotype-aware comparison.
Read the stratified output. The headline F1 is not enough. Break performance down by variant type and by genomic context. Your real question is whether the specific regions you care about are called reliably, and a genome-wide number hides a 10x difference between unique sequence and segmental duplications.
Phasing
Clair3 phases internally as part of full-alignment calling, and --enable_phasing writes phased genotypes to the final VCF. We use LongPhase for the intermediate phasing step because it is considerably faster than WhatsHap at whole-genome scale and gives comparable phase blocks on long reads.
Phasing is one of the concrete advantages of long reads over short reads. With 20 kb reads you get phase blocks spanning hundreds of kilobases, which lets you determine whether two variants in the same gene sit on the same chromosome or on opposite copies. That distinction changes the interpretation of compound heterozygous findings entirely.
Where Clair3 will let you down
Clair3 calls germline SNVs and small indels, roughly under 50 bp. It does not call structural variants, copy number changes, or repeat expansions. For those you need Sniffles2 or cuteSV on the same BAM, plus a dedicated repeat genotyper. It is also not a somatic caller. Low-frequency mosaic variants will either be missed or called as low-quality heterozygotes.
Segmental duplications and the pseudoautosomal regions produce mapping-quality artifacts that no caller fixes. Filter on MQ or use the GIAB difficult-regions BED to flag calls rather than trusting them. Long reads help here relative to short reads, but they do not eliminate the problem.
Finally, RNA sequencing data needs a different tool. The splice-aware alignment and the allele-specific expression in RNA break assumptions the DNA model makes, and Clair3-RNA exists for that reason. Broader reviews of deep-learning callers make the same general point: these models encode the error profile of the data they were trained on, and applying them off-distribution degrades results in ways that are hard to see from the output alone.8
Questions people also ask
How long does Clair3 take on a 30x human genome? On 32 CPU threads, expect several hours end to end, dominated by the full-alignment stage. GPU acceleration helps the inference but not the tensor generation, which is CPU-bound. Recent engineering work on Clair3 targeted exactly this bottleneck and reduced whole-genome runtime considerably.1
Do I need a GPU? No. Clair3 runs on CPU and the released Docker and Singularity images are CPU-only by default. A GPU shortens the full-alignment inference step but the pipeline is usually limited by tensor construction and I/O.
Can I run Clair3 on amplicon or targeted data?
Yes, with --bed_fn restricting to your regions. Amplicon data has very high depth and non-uniform coverage, so set --chunk_size smaller and check that the pileup model is not saturating. Nanopore amplicon whole-genome workflows for viral targets use this pattern successfully.9
How does Clair3 compare with lighter-weight nanopore callers? Faster alignment-based callers exist and can approach deep-learning accuracy on SNVs with far less compute.10 They generally give up ground on indels. If runtime is not your constraint, Clair3 is the better default.
What reference should I use? GRCh38 without ALT contigs for compatibility with existing annotation resources. T2T-CHM13 gives better mapping in previously unresolvable regions, but annotation and truth-set support is still thinner. Running both is reasonable if you have the compute.
Footnotes
-
Zhenxian Zheng, Minggao He, Xian Yu, et al. Accelerated long-read variant calling with Clair3 for whole-genome sequencing. Bioinformatics, 2026. https://doi.org/10.1093/bioinformatics/btag181 ↩ ↩2
-
Yury A. Barbitoff, Ruslan Abasov, Varvara E. Tvorogova, et al. Systematic benchmark of state-of-the-art variant calling pipelines identifies major factors affecting accuracy of coding sequence variant discovery. BMC Genomics, 2022. https://doi.org/10.1186/s12864-022-08365-3 ↩
-
Junhao Su, Zhenxian Zheng, Syed Shakeel Ahmed, et al. Clair3-trio: high-performance Nanopore long-read variant calling in family trios with trio-to-trio deep neural networks. Briefings in Bioinformatics, 2022. https://doi.org/10.1093/bib/bbac301 ↩
-
Huijing Yu, Zhenxian Zheng, Junhao Su, et al. Boosting variant-calling performance with multi-platform sequencing data using Clair3-MP. BMC Bioinformatics, 2023. https://doi.org/10.1186/s12859-023-05434-6 ↩
-
Xian Yu, Zhenxian Zheng, Lei Chen, et al. Leveraging ONT move table values for signal aware variant calling. Bioinformatics, 2026. https://doi.org/10.1093/bioinformatics/btag376 ↩
-
Asmaa A. Helal, Bishoy T. Saad, Mina T. Saad, et al. Evaluation of the Available Variant Calling Tools for Oxford Nanopore Sequencing in Breast Cancer. Genes, 2022. https://doi.org/10.3390/genes13091583 ↩
-
Siddhartha Sen, Hillary P. Handler, Alec Victorsen, et al. Validation of a comprehensive long-read sequencing platform for broad clinical genetic diagnosis. Frontiers in Genetics, 2025. https://doi.org/10.3389/fgene.2025.1499456 ↩
-
Omar Abdelwahab, Davoud Torkamaneh. Artificial intelligence in variant calling: a review. Frontiers in Bioinformatics, 2025. https://doi.org/10.3389/fbinf.2025.1574359 ↩
-
Maina K. Titus, David Giesbrecht, Cliff I. Oduor, et al. Scalable long-read nanopore HPV16 amplicon-based whole-genome sequencing. Scientific Reports, 2025. https://doi.org/10.1038/s41598-025-18664-w ↩
-
Miao Cui, Yadong Liu, Xian Yu, et al. miniSNV: accurate and fast single nucleotide variant calling from nanopore sequencing data. Briefings in Bioinformatics, 2024. https://doi.org/10.1093/bib/bbae473 ↩
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.