What Genotyping Costs, From $40 Arrays to $600 Genomes

A consumer SNP array costs $50 to $200 retail and measures roughly 600,000 to 900,000 pre-selected positions. A 30x whole genome costs $300 to $600 from direct-to-consumer sequencing shops and $1,500 to $3,000 from a CLIA/CAP lab that issues a signed clinical report. An exome sits between them, $300 to $700 research-grade. A single-locus PCR or TaqMan genotype, the kind a core facility runs, is about $0.30 to $1.00 per assay per sample plus probe cost. Those are four different products with the same word attached, and the price gap tracks a real difference in what you get back.

The price ladder, and what each rung measures

Assay	Typical price	What you get	Raw output
Single SNP (TaqMan/KASP)	$0.30–$1.00 per assay	One position	A call in a plate reader export
Consumer SNP array	$50–$200	~650k–900k genotyped sites	.txt (rsid, chr, pos, genotype)
Research array + imputation	$40–$80 at volume	Same sites, plus ~10–40M imputed	VCF/BGEN with dosages
Exome (WES), 100x	$300–$700	~1–2% of the genome, coding	FASTQ/CRAM, gVCF
WGS 30x, research	$300–$600	~4–5M variants genome-wide	FASTQ/CRAM, VCF
WGS 30x, clinical	$1,500–$3,000	Same data, plus interpretation and sign-off	PDF report, sometimes VCF

The array number is low because an array does not discover anything. The chip has oligo probes for positions someone chose in advance, mostly common variants with minor allele frequency above 1% in the reference panels available when the chip was designed. Manufacturing is cheap at scale and the readout is a fluorescence intensity cluster, not a sequence. The marginal cost of adding a person is a chip position and a technician’s time.

Sequencing is expensive because you pay per base, and 30x coverage of 3.1 Gb is roughly 90–100 Gb of base calls per person. The trajectory from the first genomes to today was mapped out two decades ago, when the $1,000 genome was a target rather than an invoice line ¹. Reagent cost on current high-output instruments is now the smaller part of the bill. Library prep, sample handling, compute, and QC are the rest.

What the array misses, and whether you care

An array genotypes what it was designed to genotype. Two consequences matter.

First, rare variants. If a variant sits below about 0.5% frequency, it is probably not on the chip, and if it is, the calls are less reliable because the intensity clusters are sparse. Deep sequencing of 1,070 Japanese individuals recovered millions of variants, the large majority rare and absent from earlier catalogs ². If your ancestry is underrepresented in the panels used to design the chip and to impute from, the gap is wider. This is also why “ancestry-specific” framing of chip coverage is shakier than it sounds: most human genetic variation is found within populations rather than between them, so a chip tuned on one cohort still captures a lot elsewhere, and still misses the rare tail everywhere ³.

Second, structural variation. Arrays give you log R ratio and B-allele frequency, which can flag large CNVs, but nothing at the resolution of a read-depth or split-read caller. Repeat expansions, inversions, and mitochondrial heteroplasmy are out of reach.

For common-variant work, polygenic scores, pharmacogenomic star alleles that are tagged by common SNPs, ancestry, an array plus imputation is close to a sequencing genome and costs a tenth as much.

Imputation: the cheapest coverage you can buy

Take the raw file, convert it, and send it to an imputation server. This is free and takes hours.

# 23andMe-style txt -> PLINK -> VCF
plink --23file genome.txt WOOLF ME 1 --snps-only just-acgt --make-bed --out me
plink --bfile me --recode vcf-iid bgz --output-chr chrM --out me

# check and fix REF/ALT against the reference before upload
bcftools +fixref me.vcf.gz -- -f GRCh38.fa -m top 2> fixref.log

Consumer files are usually GRCh37, sometimes GRCh38. Check the header, and lift over with CrossMap or picard LiftoverVcf if the server wants a different build. The TOPMed Imputation Server phases with Eagle and imputes with Minimac4 against a panel of roughly 97,000 deeply sequenced samples.

What comes back: tens of millions of sites with an imputation quality score (R2 in the INFO field). Filter hard. R2 >= 0.8 is a reasonable working threshold for anything you plan to act on analytically, and it will leave you with something in the low tens of millions of sites in a well-represented ancestry, fewer otherwise. Everything below that is a probability distribution wearing a genotype’s clothes. Imputed dosages are also the wrong input for any clinical question: a rare pathogenic variant imputed at R2 = 0.6 is a hypothesis, not a finding, and confirming it requires a clinical lab and a clinician to interpret.

The most common failure mode we see is silent strand mismatch at A/T and G/C sites. bcftools +fixref with -m top handles the flip, but read the log. If more than a percent or two of your sites need fixing, something upstream is wrong with the build or the manifest.

The costs that show up after the invoice

Storage. A 30x genome is roughly 50–60 GB as gzipped FASTQ, 60–90 GB as BAM, and 15–25 GB as CRAM aligned to a known reference. Keep CRAM, not BAM; the reference-based compression is lossless for your purposes and the savings are large. On S3 Standard at $0.023/GB/month, 25 GB is about $0.58 a month. Glacier Deep Archive is about $0.001/GB/month, roughly $0.03 a month, with retrieval latency measured in hours. Keep one archived copy of the CRAM and one hot copy of the VCF, which is under a gigabyte.

Compute. Calling variants from CRAM with DeepVariant takes a few hours on 32 vCPUs. On spot instances that is single-digit dollars per genome. GATK HaplotypeCaller in the standard best-practices path is slower and produces a gVCF that is more convenient if you ever joint-call with other samples. We use DeepVariant for single-sample work because the SNV precision and recall are better out of the box and there is no BQSR step to tune.

Annotation. Free, but opinionated. VEP with the --cache --offline --everything flags, plus gnomAD frequencies and ClinVar, is what we run. Budget an hour of wall time and a 60 GB cache download.

Interpretation. This is where the real money goes, and it is the part no pipeline removes. A clinical WGS report at $2,000 is mostly a board-certified lab director’s time and liability, not sequencing reagent.

What we would buy

If the question is “what common-variant information can I get for the least money,” buy an array and impute. Sixty dollars, two weeks, and a VCF with dosages.

If the question is “I want the substrate I will keep analyzing for the next twenty years,” buy 30x WGS and insist on the CRAM or FASTQ, not just a report. Sequencing once at 30x and re-analyzing as annotation databases improve is the better economics. The variant callers and the ClinVar classifications change every year. Your reads do not.

Do not buy an array first and sequence later if you already know you want sequence. The array will not reduce the sequencing cost.

Questions people also ask

How long does genotyping take? A consumer array is typically 3–6 weeks from kit mailing to results, most of which is shipping and batching, not lab time. Research WGS is 2–6 weeks. Clinical WGS with a signed report is 4–12 weeks, and the interpretation, not the sequencing, sets that clock. Imputation on a public server runs in a few hours once you upload.

Can I do a genotype test at home? You collect the sample at home (saliva or a cheek swab) and a lab does the work. Nothing about the chemistry happens in your kitchen. Saliva kits stabilize DNA at room temperature for weeks, which is why they dominate the consumer market.

What does a genotyping test test for? Whatever positions the assay targets. A chip targets a fixed, published list, which is why you can download the manifest and check whether a specific rsID is on it before you buy. A genome targets everything the reads cover, which is nearly all of the non-repetitive genome at 30x.

How much is “full” genetic testing? There is no complete test. Short-read 30x WGS leaves repeat expansions, some structural variants, and segmental duplications poorly resolved. Long-read WGS closes much of that gap and currently runs $1,000–$2,000 research-grade.

Is a $50 array result reliable enough to act on? For common variants and aggregate scores, the genotype concordance with sequencing is high. For a single rare variant flagged as pathogenic, it is not, and that finding needs orthogonal confirmation in a clinical laboratory and a conversation with a clinician before it means anything.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. Simon T. Bennett, C.L. Barnes, Anthony J. Cox, et al. Toward the $1000 Human Genome. Pharmacogenomics, 2005. https://doi.org/10.1517/14622416.6.4.373 ↩

2. Masao Nagasaki, Jun Yasuda, Fumiki Katsuoka, et al. Rare variant discovery by deep whole-genome sequencing of 1,070 Japanese individuals. Nature Communications, 2015. https://doi.org/10.1038/ncomms9018 ↩

3. Jeffrey C. Long, Jie Li, Meghan Elisabeth Healy. Human DNA sequences: More variation and less race. American Journal of Physical Anthropology, 2009. https://doi.org/10.1002/ajpa.21011 ↩