What Genetic Testing for Personalized Medicine Measures, and What It Costs

Pharmacogenetic testing runs roughly $150 to $500 out of pocket for a targeted panel of 10 to 25 genes, and a clinical-grade whole genome runs $600 to $3,000 depending on depth, turnaround, and whether a report is included. Those two options answer different questions. A panel genotypes a fixed list of known variant positions and returns a star-allele diplotype per gene. A genome sequences everything and lets you re-analyze in five years when the evidence changes, but it calls some of the most clinically important pharmacogenes badly unless you use the right tools. If you only want to know your CYP2C19 and DPYD status before a specific prescribing decision, order the panel through a clinician. If you want data you own and can re-query, sequence the genome and do the star-allele calling yourself.

Everything below is about measurement and interpretation. Prescribing decisions belong to a clinician who has your full clinical picture.

What “personalized medicine genetic testing” usually means in practice

Three distinct products get sold under the same phrase.

Pharmacogenetic panels. A fixed set of positions, usually genotyped by TaqMan, MassARRAY, or a small amplicon NGS panel. A typical gene list is below.

• CYP2C19
• CYP2C9
• CYP2D6
• CYP3A5
• VKORC1
• DPYD
• TPMT
• NUDT15
• SLCO1B1
• UGT1A1
• HLA-B (*57:01, *15:02)
• CYP4F2
• G6PD

You get a diplotype like CYP2C19 *1/*17 and a phenotype label (rapid metabolizer). The evidence base for these gene-drug pairs is curated by CPIC and PharmGKB, and only a few dozen pairs have Level A guidelines.

Hereditary risk panels. BRCA1/2, Lynch syndrome genes, and cardiomyopathy and arrhythmia panels. Familial hypercholesterolemia as well (LDLR, APOB and PCSK9). These are diagnostic tests with real consequences and belong in a genetic counseling workflow. A self-serve one is not the place for them.

Whole-genome or whole-exome sequencing. 30x short-read WGS produces a BAM or CRAM plus a VCF with 4 to 5 million variants. Exome is cheaper but misses intronic and structural variants, including some that matter pharmacologically. The case for sequencing the whole genome once and interpreting it repeatedly has been made since the earliest personalized-sequencing efforts, and it remains the stronger long-run position ¹.

Why array-based consumer tests are the wrong substrate

Consumer genotyping arrays interrogate roughly 600,000 to 900,000 positions. They are cheap and reproducible at the sites they cover, and useless at the sites they do not. Two concrete failure modes.

CYP2D6. The gene sits next to a 93% identical pseudogene, CYP2D7, and comes in whole-gene deletions (*5), duplications and multiplications (*1xN, *2xN), and CYP2D6-CYP2D7 hybrid alleles (*13, *36, *68). Copy number and hybrid structure change the predicted activity score. Arrays do not measure copy number at this locus reliably, and neither does a naive short-read variant caller that only emits SNVs and small indels.

Rare variants. Arrays genotype known common positions. A novel loss-of-function variant in DPYD or TPMT will not appear on the array at all. A “normal metabolizer” call from an array is therefore a statement about the variants tested. It is not a statement about the gene. Next-generation sequencing was adopted for pharmacogenomics partly because it captures rare and novel functional variation that fixed-content genotyping misses ².

If you already have array data, treat it as a screen. It is fine for confirming a family variant you know about and unreliable as a negative.

Working with your own WGS for pharmacogenes

Assume you have a 30x CRAM aligned to GRCh38 and a joint or single-sample VCF. The pipeline we would run:

1. Check coverage at the pharmacogene loci before believing anything.

samtools depth -a -r chr22:42126499-42130881 \
  --reference GRCh38.fa sample.cram | \
  awk '{s+=$3; n++} END {print "mean", s/n}'

That interval is CYP2D6 on GRCh38. If mean depth there is well below your genome-wide mean, or if you see a depth cliff mid-gene, the caller output is suspect. A 30x genome typically gives 25 to 35x here with short reads, but mapping quality drops in the CYP2D7 homologous stretch.

2. Call star alleles with a tool built for the problem. Do not use a generic annotator.

PharmCAT takes a VCF and returns diplotypes and CPIC-mapped phenotypes:

java -jar pharmcat-pipeline.jar \
  -vcf sample.g38.vcf.gz \
  -reporterJson -matcherHtml -o out/

PharmCAT explicitly refuses to call CYP2D6 from a plain VCF. Feed it structural calls from a specialized caller. Two options:

# Aldy: read-depth + phasing based, handles CNV and hybrids
aldy genotype -p wgs -g CYP2D6 --profile illumina sample.cram

# Cyrius: CYP2D6-specific, designed for 30x WGS
python star_caller.py --manifest manifest.txt \
  --genome 38 --prefix sample --outDir out/ --threads 4

Then pass the result into PharmCAT with -po (outside calls). Cyrius and Aldy disagree on a small fraction of samples, usually involving *68 and *4 tandem arrangements. If they disagree and the gene matters for a decision in front of you, a clinical lab’s targeted CYP2D6 assay (long-range PCR plus copy number) is the arbiter. Order it through a clinician.

3. Check the phenotype-relevant genes short reads handle poorly.

HLA-B typing for *57:01 and *15:02 needs an HLA-aware caller (HLA-LA, Optitype for class I). Do not read HLA types out of a standard VCF. G6PD is X-linked and hemizygous calls in males look like homozygous calls, which is fine, but female heterozygotes have intermediate activity and a genotype does not give you an enzyme level.

4. Record what you could not call. A pharmacogenetic report that silently omits a no-call is worse than one that flags it. PharmCAT’s matcher HTML lists positions with insufficient coverage. Keep it.

Cost, and what the economics literature says

Sequencing cost estimates in the literature vary by an order of magnitude because studies count different things: reagents only, reagents plus labor, or the full cost including interpretation, genetic counseling, and confirmatory testing. Careful costing work puts interpretation and reporting, not the sequencer run, as a large share of the total ³. The same accounting problem appears in pharmacogenomics specifically, where micro-costing versus gross-costing methods produce materially different per-test figures ⁴.

On cost-effectiveness, the evidence is uneven and gene-specific rather than uniformly favorable. Systematic reviews of pharmacogenetic and pharmacogenomic screening tests find that a minority of evaluated gene-drug pairs show clear cost-effectiveness, with quality of the underlying economic models a recurring limitation ⁵. A separate review reached a similar conclusion: the strongest cases are drugs with severe, predictable, genotype-linked toxicity, and the general case remains unproven ⁶. Take that seriously. “My genome will optimize all my prescriptions” is not what the data supports. “Knowing my DPYD and TPMT status before a relevant prescribing decision has asymmetric value” is closer.

Implementation is the other cost nobody quotes. Getting a result into a workflow where it fires at the right moment requires structured storage, standardized phenotype terms, and clinical decision support, and this has been the persistent bottleneck in health systems that tried ⁷. As an individual, you are your own decision support layer. That means storing diplotypes in a form you can hand to a clinician, and re-running interpretation when CPIC updates guidelines, which happens a few times a year.

What sequencing does not tell you

A genome is a static baseline. It gives you allele-level metabolizer predictions, carrier status, and polygenic scores of contested clinical utility. It does not give you current enzyme induction from a drug you are taking, current renal or hepatic function, or your actual drug exposure. The field’s own framing has always been that genomic data is one input into a model that includes environment, phenotype, and clinical history ^{8 9}. The public expectation of what a personal genome delivers has consistently run ahead of the evidence, and that gap has been documented since the early post-Human Genome Project era ¹⁰.

The practical consequence: pair genotype with measurement. Baseline CBC, comprehensive metabolic panel, lipid panel with ApoB and Lp(a), HbA1c, hsCRP. Proteomic and transcriptomic panels move over weeks and months, and they tell you about state in a way sequence cannot. Interpretation of any of this in the context of a treatment decision requires a clinician.

Questions people also ask

How expensive is pharmacogenetic testing? Self-pay targeted panels run about $150 to $500 in 2026. Clinical single-gene assays (a standalone CYP2D6 with copy number, for example) are typically $200 to $400. Insurance coverage is inconsistent and usually tied to a specific prescribing indication. Published cost figures scatter widely because studies differ in whether they include labor, interpretation, and confirmatory testing ⁴.

Is whole-genome sequencing better than a pharmacogenetic panel? For breadth and reuse, yes. For the specific genes that need copy number and phasing (CYP2D6 above all), a targeted clinical assay is more reliable than short-read WGS unless you run Cyrius or Aldy and accept the residual discordance rate. Many people end up doing both.

Can I interpret my own results? You can compute diplotypes and phenotype labels yourself with PharmCAT plus Cyrius, and that is a reasonable engineering exercise. Translating a phenotype into a prescribing action is a clinical decision that needs a clinician with your history, current medications, and lab values.

Do I need to re-test as science advances? No. Re-analyze. A CRAM and VCF from a 30x genome stay valid, and re-running PharmCAT annually against updated CPIC and PharmGKB content costs nothing but compute. This is the main argument for sequencing once at depth rather than genotyping repeatedly.

Footnotes

1. Edward D Esplin, Ling Oei, Michael P Snyder. Personalized Sequencing and the Future of Medicine: Discovery, Diagnosis and Defeat of Disease. Pharmacogenomics, 2014. https://doi.org/10.2217/pgs.14.117 ↩

2. Bahareh Rabbani, Hirofumi Nakaoka, Shahin Akhondzadeh, et al. Next generation sequencing: implications in personalized medicine and pharmacogenomics. Molecular BioSystems, 2016. https://doi.org/10.1039/c6mb00115g ↩

3. Kurt Christensen, Dmitry Dukhovny, Uwe Siebert, et al. Assessing the Costs and Cost-Effectiveness of Genomic Sequencing. Journal of Personalized Medicine, 2015. https://doi.org/10.3390/jpm5040470 ↩

4. Stavroula Siamoglou, Kariofyllis Karamperis, Christina Mitropoulou, et al. Costing Methods as a Means to Measure the Costs of Pharmacogenomics Testing. The Journal of Applied Laboratory Medicine, 2020. https://doi.org/10.1093/jalm/jfaa113 ↩ ↩²

5. Elizabeth J. J. Berm, Margot de Looff, Bob Wilffert, et al. Economic Evaluations of Pharmacogenetic and Pharmacogenomic Screening Tests: A Systematic Review. Second Update of the Literature. PLOS ONE, 2016. https://doi.org/10.1371/journal.pone.0146262 ↩

6. M Verbelen, M E Weale, C M Lewis. Cost-effectiveness of pharmacogenetic-guided treatment: are we there yet?. The Pharmacogenomics Journal, 2017. https://doi.org/10.1038/tpj.2017.21 ↩

7. Ann M Moyer, Pedro J Caraballo. The challenges of implementing pharmacogenomic testing in the clinic. Expert Review of Pharmacoeconomics & Outcomes Research, 2017. https://doi.org/10.1080/14737167.2017.1385395 ↩

8. Geoffrey S. Ginsburg, Huntington F. Willard. Genomic and personalized medicine: foundations and applications. Translational Research, 2009. https://doi.org/10.1016/j.trsl.2009.09.005 ↩

9. Johanne Tremblay, Pavel Hamet. Role of genomics on the path to personalized medicine. Metabolism, 2013. https://doi.org/10.1016/j.metabol.2012.08.023 ↩

10. Brenda Wilson, Stuart G Nicholls. The Human Genome Project, and recent advances in personalized genomics. Risk Management and Healthcare Policy, 2015. https://doi.org/10.2147/rmhp.s58728 ↩