How Accurate Is Genetic Testing? It Depends on the Variant More Than the Test

“Genetic testing accuracy” is not one number. A single genotype call has an accuracy. A test does not. On short-read whole-genome sequencing at 30x, a heterozygous SNV in a uniquely mappable, non-repetitive region is called correctly better than 99.9% of the time. Other calls are far less reliable, sometimes closer to a coin flip depending on the pipeline. They include a 40-bp insertion in a homopolymer, a CNV boundary, and a variant in PMS2 exon 13, which sits next to a pseudogene. Anything in the HLA region belongs with them. So the useful question asks for the expected error rate for this class of variant, in this region and on this platform, with this pipeline and at this depth. “Is this test accurate” is not that question. Then, separately, ask whether the interpretation of that variant is correct. The second question is usually where the real error lives.

Analytical accuracy: the part you can measure

Analytical validity means the pipeline reports the bases that are in the sample. It decomposes into a few measurable things.

Sequencing error rate. Illumina short reads run around Q30 to Q35 per base at the read level, meaning roughly 1 in 1,000 to 1 in 3,000 raw base errors. At 30x coverage, majority voting across independent reads drives per-genotype error far below that, and careful library and analysis choices push whole-genome SNV accuracy into the range where a genome can be treated as a general-purpose genetic test rather than a targeted one ¹.

Coverage uniformity. This is where most real failures come from. The reported “30x average” hides the distribution. What you want to look at is the fraction of the callable genome at ≥20x. GC-rich first exons, segmental duplications, and long homopolymers drop out. Run:

mosdepth --by 500 --fast-mode sample sample.cram
samtools coverage sample.cram

and look at the per-region output. The mean is not enough. If your 30x genome has 4% of coding bases below 10x, those bases carry no information, and a “negative” result there means nothing.

Mappability. Short reads cannot be placed uniquely in regions with near-identical paralogs. The limitation is structural and is not a depth problem. Adding coverage does not help if the read is 150 bp and the duplication is 20 kb at 99% identity. The paralog problem has been understood since the first careful computational annotations of the human genome, and it still governs which genes are reliably callable today ². Practical consequence: SMN1/SMN2, PMS2, CYP2D6, GBA, and the pericentromeric regions need targeted assays or long reads.

Variant class. Roughly, SNV calls are excellent and small indels are good but worse, especially in repeats. Structural variants from short reads are weak. GIAB benchmarking on HG002 with the v4.2.1 truth set typically shows SNV F1 above 0.995 and indel F1 in the 0.99 range inside high-confidence regions, and considerably worse outside them. That “inside high-confidence regions” clause is doing enormous work. About 10% of the genome is excluded from the benchmark entirely because no consensus truth exists.

Pipeline choice. Two labs can run the same FASTQs and disagree on thousands of calls. Bioinformatics is a first-order determinant of clinical genomic accuracy, not an implementation detail, and pipeline differences have produced discordant clinical results on identical raw data ³. If you have your own data, this is your lever.

What we would run on our own data

Start from the CRAM aligned to GRCh38, or T2T-CHM13 if the lab offers it. T2T fixes real alignment problems but breaks most annotation tooling, so we keep GRCh38 as primary and T2T as a second opinion.

Small variants: DeepVariant for SNVs and indels. It outperforms GATK HaplotypeCaller on indels in repetitive context in most published benchmarks, and it has no parameter tuning to get wrong.

run_deepvariant --model_type=WGS \
  --ref=GRCh38.fa --reads=sample.cram \
  --output_vcf=sample.dv.vcf.gz --num_shards=32

Then check yourself against truth:

hap.py HG002_truth.vcf.gz sample.dv.vcf.gz \
  -f HG002_high_confidence.bed -r GRCh38.fa \
  -o happy_out --engine=vcfeval

You cannot run hap.py on yourself (no truth set exists for you), but you can run it on a GIAB sample processed by the same lab and pipeline. If a provider will not share their hap.py output on a reference sample, treat their accuracy claims as marketing.

Structural variants: call with two tools and intersect. Manta plus Delly, or Manta plus GRIDSS, then require support from both for anything you take seriously. Expect 60–80% sensitivity for deletions above 1 kb and much worse for insertions. Short-read SV calls are hypotheses. They are not findings.

CNVs: on WGS use read-depth methods (CNVpytor, or GATK gCNV with a panel of normals). On exome data, do not trust CNV calls at all without orthogonal confirmation.

Genotype concordance as a cheap sanity check: if you also have an array from a consumer service, extract the overlapping sites and compare.

bcftools isec -n=2 -p isec_out array.norm.vcf.gz wgs.dv.vcf.gz

Concordance below about 99% at array sites means something is wrong: a sample swap, a strand-flip handling bug, or a reference mismatch. This one comparison catches the single most consequential error mode in all of genetics, which is your data not being your data.

Interpretive accuracy: the part that fails more often

A correct base call with a wrong interpretation is worse than no test. The main failure modes:

Classification drift. ClinVar classifications change. Variants of uncertain significance get reclassified in both directions, and pathogenic calls occasionally get downgraded when population frequency data from gnomAD shows they are too common to cause a rare severe disease. Your VCF is static. Reannotate yearly:

vep -i sample.vcf.gz --cache --assembly GRCh38 \
  --plugin CADD --plugin SpliceAI --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG \
  --af_gnomade --pick_allele_gene -o sample.vep.vcf --vcf

Penetrance in unselected people. Almost all penetrance estimates come from families that came to clinical attention because they had the disease. When the same variants are found in unselected sequenced adults, the risk is typically lower, sometimes much lower. This is the single largest source of misinterpretation for a healthy person sequencing themselves, and it is why “pathogenic variant found” and “will get the disease” are different statements. Sequencing healthy adults for predisposition has been studied prospectively, and the observed rate of actionable findings and downstream outcomes differs substantially from what family-based studies would predict ⁴.

Ancestry bias in reference data. Polygenic scores and allele frequency filters were built mostly on European-ancestry cohorts. A PRS trained on UK Biobank loses much of its discriminative power in African-ancestry individuals. Treat any PRS as a population-level statistic with a large ancestry-dependent error bar. It is not a personal number.

Secondary and incidental findings. Sequencing everything means finding things you did not ask about. Whether to look at all of them is a real decision with a real literature, and European guidance explicitly argues against unbounded opportunistic screening of genomes for unrelated conditions ⁵. Our view: decide your disclosure preferences before you look, write them down, and separate the analysis into tiers.

Anything with clinical weight needs confirmation in a clinical laboratory on a fresh sample and interpretation with a genetic counselor or physician. That covers ACMG secondary findings, cancer predisposition genes and pharmacogenomic star alleles. Research-grade sequencing is for hypothesis generation and self-education. It is not a diagnosis, and a variant found in your own pipeline should never change what you do medically until it has been confirmed and discussed with a clinician.

Where the field is, and where accuracy claims come from

Genome sequencing moved into ordinary clinical practice fast enough that interpretation capacity, not sequencing capacity, is now the constraint ⁶. The early technical challenges of personal genomes, coverage gaps, assembly ambiguity, and the difficulty of separating real variation from artifact, were mapped out over a decade ago and remain a good checklist for evaluating any provider’s output ⁷. What has changed is that the raw data is cheap enough that you can hold it, reprocess it, and audit the claims made about it.

Questions people also ask

Why do people opt out of genetic testing? The dominant reasons are not technical. Reviews of testing uptake find concerns about insurance and employment discrimination, uncertainty about what the result would change, cost, and anxiety about learning unactionable risk, alongside straightforward logistical barriers ⁸. Preference for not knowing is a legitimate position and a common one.

What are the negatives of genetic testing? The main ones are below.

• uncertain results (a VUS you cannot act on)
• false reassurance from a negative in a poorly covered region
• overestimated risk from family-based penetrance figures
• findings about relatives you did not consent for them
• downstream cost from follow-up imaging or specialist visits triggered by ambiguous findings

Decision-making research consistently shows people underestimate how much of a genome report is uncertain rather than binary ⁹.

Why are doctors pushing genetic testing? Usually because a specific result would change specific management. That can mean confirming a diagnosis or choosing a surveillance interval. It can also mean identifying a heritable condition that affects family members, or avoiding a drug metabolized by a variant enzyme. If a clinician orders a test, the right question is what decision the result will change.

What do people most want to know from sequencing? Studies of large-scale sequencing cohorts find participants prioritize treatable or preventable conditions and information relevant to children, and consistently want more than researchers plan to return ¹⁰.

Is whole-genome sequencing more accurate than a consumer array? For variant discovery, yes, by a wide margin. Arrays only interrogate the roughly 600,000 to 1 million sites they were designed for and impute the rest, so a rare variant not on the chip is invisible. Arrays also have higher false-positive rates specifically for rare pathogenic variants, because the clustering algorithms are tuned for common sites. At a given site included on the array, array genotyping is quite accurate.

Footnotes

1. Radoje Drmanac, Brock A Peters, George M Church, et al. Accurate Whole Genome Sequencing as the Ultimate Genetic Test. Clinical Chemistry, 2015. https://doi.org/10.1373/clinchem.2014.224907 ↩

2. Ru-Fang Yeh, Lee P. Lim, Christopher B. Burge. Computational Inference of Homologous Gene Structures in the Human Genome. Genome Research, 2001. https://doi.org/10.1101/gr.175701 ↩

3. W. Gregory Feero. Bioinformatics, Sequencing Accuracy, and the Credibility of Clinical Genomics. JAMA, 2020. https://doi.org/10.1001/jama.2020.19939 ↩

4. for the PeopleSeq Consortium, Emilie S. Zoltick, Michael D. Linderman, et al. Predispositional genome sequencing in healthy adults: design, participant characteristics, and early outcomes of the PeopleSeq Consortium. Genome Medicine, 2019. https://doi.org/10.1186/s13073-019-0619-9 ↩

5. On behalf of the European Society of Human Genetics, Guido de Wert, Wybo Dondorp, et al. Opportunistic genomic screening. Recommendations of the European Society of Human Genetics. European Journal of Human Genetics, 2020. https://doi.org/10.1038/s41431-020-00758-w ↩

6. Helen K Brittain, Richard Scott, Ellen Thomas. The rise of the genome and personalised medicine. Clinical Medicine, 2017. https://doi.org/10.7861/clinmedicine.17-6-545 ↩

7. Michael Snyder, Jiang Du, Mark Gerstein. Personal genome sequencing: current approaches and challenges. Genes & Development, 2010. https://doi.org/10.1101/gad.1864110 ↩

8. Megan Smith-Uffen, Nicci Bartley, Grace Davies, et al. Motivations and barriers to pursue cancer genomic testing: A systematic review. Patient Education and Counseling, 2021. https://doi.org/10.1016/j.pec.2020.12.024 ↩

9. Leland E Hull, Jason L Vassy. Toward Greater Understanding of Patient Decision-Making Around Genome Sequencing. Personalized Medicine, 2017. https://doi.org/10.2217/pme-2017-0037 ↩

10. Courtney L. Scherr, Sharon Aufox, Amy A. Ross, et al. What People Want to Know About Their Genes: A Critical Review of the Literature on Large-Scale Genome Sequencing Studies. Healthcare, 2018. https://doi.org/10.3390/healthcare6030096 ↩