Is Pharmacogenetic Testing Legit? A Gene-by-Gene Answer
Pharmacogenetic testing is legitimate for a specific, short list of gene-drug pairs and considerably weaker for the thing most people buy it for. Genotype predicts the enzyme activity that clears many drugs, and for a few dozen pairs that prediction is strong enough that professional guidelines specify dose adjustments. Whether that prediction improves the outcome of choosing an antidepressant is a separate question, and the trial evidence there is thin and contested. The useful distinction is between pharmacokinetics (how fast your body clears a molecule) and pharmacodynamics (whether the molecule works on you). Genotype does the first job well and the second job poorly.
The gene-drug pairs that hold up
The strongest evidence sits where a single enzyme dominates clearance of a drug with a narrow therapeutic window. Thiopurine methyltransferase (TPMT) and NUDT15 variants predict severe myelosuppression from thiopurines. DPYD variants predict fluoropyrimidine toxicity. HLA-B*57:01 predicts abacavir hypersensitivity, and HLA-B*15:02 predicts carbamazepine-induced severe skin reactions. CYP2C19 loss-of-function alleles reduce the conversion of clopidogrel to its active metabolite. CYP2C9 and VKORC1 together explain a substantial fraction of warfarin dose variance.
These pairs share a structure. The phenotype is a hard clinical event or a measurable drug concentration, the effect size is large, and the mechanism is a single metabolic step. The Clinical Pharmacogenetics Implementation Consortium (CPIC) publishes graded guidelines for these pairs, and the FDA lists a subset in drug labeling. Regulatory interest in genotype-stratified pharmacokinetics goes back more than two decades and was formalized early through joint FDA and industry workshops on how genetic data should enter drug development.1 The field’s original framing was exactly this: inherited variation in drug-metabolizing enzymes as a predictable source of interindividual response differences.2
The counterexample that should calm anyone’s enthusiasm is antidepressant response. In the STAR*D sample, variants in the pharmacokinetic genes were not associated with response or tolerance to citalopram.3 The cohort is large and well characterized, and the negative result bears on precisely the claim that consumer psychiatric panels make. Enzyme genotype shapes plasma concentration. Plasma concentration is not the same as whether a serotonergic drug relieves depression.
Combinatorial psychiatric panels are a different product
Tests like GeneSight combine genotypes from a handful of pharmacokinetic and candidate pharmacodynamic genes into a proprietary algorithm that sorts drugs into colored bins. Two problems follow. First, the weighting is not public. You cannot reconstruct how the report moved a drug from green to yellow, and you cannot audit it against CPIC. Second, the pharmacodynamic inputs (serotonin transporter and receptor variants, for example) have small and inconsistently replicated effects. Candidate-gene work on psychiatric traits has a long record of associations that did not survive genome-wide scrutiny.4
A careful review of these decision-support tools reached a measured conclusion. Genotype-guided prescribing in psychiatry has plausible pharmacokinetic grounding, but the commercial algorithms outrun their evidence, and clinicians should treat the category recommendations as suggestions rather than directives.5 The practical downside is not just a wasted fee. A yellow bin next to a drug that would have worked can steer a prescriber away from it, and the report gives no confidence interval to argue with.
Why CYP2D6 breaks most pipelines
If you only fix one thing in your own analysis, fix CYP2D6. It metabolizes a large share of prescribed small molecules, and it is the hardest pharmacogene to genotype correctly. The locus sits next to two pseudogenes, CYP2D7 and CYP2D8, with high sequence similarity. Copy number varies from zero to more than five. Hybrid alleles formed by unequal crossover between CYP2D6 and CYP2D7 produce structures that a naive variant caller reports as ordinary heterozygous SNVs.
Short reads of 150 bp mismap between CYP2D6 and CYP2D7, which inflates or erases variant calls in the duplicated regions. Targeted sequencing of the full gene has repeatedly turned up novel alleles that panel genotyping cannot see, which is why allele discovery and clinical testing remain partly separate activities.6 Dedicated callers exist because generic pipelines fail here. Constellation was built specifically to assign CYP2D6 phenotype from whole-genome data at scale.7
Array-based tests have a related but worse problem. They interrogate a fixed set of positions, so any allele not on the chip is silently called as the reference. Estonian Biobank investigators translated array and imputed genotypes for 44,000 participants into CPIC recommendations. The difficulties were concentrated in exactly these places: structural variation, rare alleles absent from the panel, and diplotype-to-phenotype mapping that changes as guidelines are revised.8 Treat any report that does not state its allele definition table version and its copy-number method as incomplete.
Calling pharmacogenes from your own WGS
Assume you have a 30x whole-genome CRAM aligned to GRCh38 and a joint-called or GVCF-derived VCF. Two passes give you most of what a clinical panel would report, with the advantage that you can inspect every intermediate.
For the broad set of genes, use PharmCAT. It takes a VCF, matches diplotypes against the CPIC allele definition tables, and emits a JSON plus HTML report keyed to published guidelines.
pharmcat_pipeline sample.GRCh38.vcf.gz \
-o pharmcat_out/ \
-research cyp2d6 \
-reporterJson
The failure mode to watch is reference confidence. A variant-only VCF cannot distinguish a homozygous-reference position from a position with no coverage, and PharmCAT’s preprocessor will assume reference unless you tell it otherwise. If your caller emitted a GVCF, keep it and let the preprocessor read depth, or generate an all-sites VCF with bcftools mpileup restricted to the PharmCAT position list. Otherwise a dropout in CYP2C19 will read as a normal metabolizer diplotype with full confidence.
For CYP2D6, run a depth-aware caller on the alignment rather than the VCF. Cyrius is our default because it models copy number and the common CYP2D6-CYP2D7 hybrids directly from read depth and breakpoint evidence.
python star_caller.py --manifest manifest.txt \
--genome 38 --prefix sample --outDir cyrius_out/ --threads 4
Cross-check with Aldy, which solves diplotype and copy number as an optimization over the PharmVar allele definitions.
aldy genotype -p wgs -g CYP2D6 sample.GRCh38.cram
When Cyrius and Aldy disagree, do not average them. Open the region in IGV, look at coverage across exon 9 and the downstream CYP2D7 boundary, and check whether soft-clipped reads cluster at a hybrid breakpoint. Disagreements almost always land on structural alleles, which is the case where the activity score changes most.
Long reads change the calculus. PacBio HiFi or Oxford Nanopore data phase the whole locus in single reads and resolve hybrids without inference. If pharmacogenetics is your primary motivation and you are choosing a platform, that is the argument for long reads.
What a genotype gets you and what it does not
Convert diplotypes to activity scores using the current CPIC tables, and record the table version in your notes. Phenotype assignment is a moving target, and a report generated three years ago may map the same diplotype to a different category today. Then treat the result as one input among several. Enzyme activity is modified by inhibitors and inducers, by hepatic and renal function, and by age and adherence. Strong CYP2D6 inhibition by a co-prescribed drug can convert a genetic normal metabolizer into a functional poor metabolizer.
This is why genotype belongs alongside measurement. Plasma drug concentration, liver enzymes, and renal function describe the state you are in now. Genotype describes a constraint that does not change. Practical guidance for using pharmacogenomics in clinical research makes the same point: genotype is most informative as a covariate on measured pharmacokinetics. It is not a replacement for them.9 The early promise of genotype-directed prescribing was oversold in the popular press two decades ago, and the durable wins have been narrow and mechanistic.10
Do not start, stop, or change the dose of any medication based on a genotype you called yourself. Diplotype calls belong in a conversation with a prescribing clinician or a clinical pharmacist, who can weigh them against your current medications and labs. If a call would meaningfully change a decision, confirm it in a CLIA-certified laboratory before acting on it.
Questions people also ask
How accurate is genetic testing for antidepressants? The genotyping itself is accurate for common alleles in CYP2C19 and CYP2D6, with the caveats above about copy number and rare alleles. Accuracy at predicting which antidepressant will work for you is a different matter, and pharmacokinetic genes did not predict citalopram response or tolerance in the STAR*D sample.3
What are the downsides of using GeneSight testing? The combining algorithm is proprietary, so its recommendations cannot be audited against CPIC guidelines or reproduced from the underlying genotypes. The pharmacodynamic gene inputs carry weak evidence, and a colored bin can discourage a prescriber from a drug that would have helped.5
Is there a test to see which antidepressant works best? No test currently predicts which antidepressant will be effective for an individual. Genotype can flag unusually fast or slow clearance, which is information about likely concentration and side-effect risk rather than efficacy.
Which pharmacogenetic test is best? For the gene-drug pairs with strong evidence, choose sequencing-based testing that reports CYP2D6 copy number and hybrid alleles and names its allele definition table version. If you already have whole-genome data, PharmCAT plus Cyrius reproduces most panel content and shows you the evidence behind each call.7
Footnotes
-
Pharmacogenetics and Pharmacogenomics in Drug Development and Regulatory Decision-Making: Report of the First FDA–PWG– PhRMA–DruSafe Workshop. The Process of New Drug Discovery and Development, 2006. https://doi.org/10.1201/9781420004236-15 ↩
-
Daniel W Nebert. Pharmacogenetics and pharmacogenomics: why is this relevant to the clinical geneticist?. Clinical Genetics, 1999. https://doi.org/10.1034/j.1399-0004.1999.560401.x ↩
-
Eric J. Peters, Susan L. Slager, Jeffrey B. Kraft, et al. Pharmacokinetic Genes Do Not Influence Response or Tolerance to Citalopram in the STAR*D Sample. PLoS ONE, 2008. https://doi.org/10.1371/journal.pone.0001872 ↩ ↩2
-
S Shifman, A Bhomra, S Smiley, et al. A whole genome association study of neuroticism using DNA pooling. Molecular Psychiatry, 2007. https://doi.org/10.1038/sj.mp.4002048 ↩
-
Zane Zeier, Linda L. Carpenter, Ned H. Kalin, et al. Clinical Implementation of Pharmacogenetic Decision Support Tools for Antidepressant Drug Prescribing. American Journal of Psychiatry, 2018. https://doi.org/10.1176/appi.ajp.2018.17111282 ↩ ↩2
-
Yao Yang, Mariana R Botton, Erick R Scott, et al. Sequencing the CYP2D6 Gene: From Variant Allele Discovery to Clinical Pharmacogenetic Testing. Pharmacogenomics, 2017. https://doi.org/10.2217/pgs-2017-0033 ↩
-
Greyson P Twist, Andrea Gaedigk, Neil A Miller, et al. Constellation: a tool for rapid, automated phenotype assignment of a highly polymorphic pharmacogene, CYP2D6, from whole-genome sequences. npj Genomic Medicine, 2016. https://doi.org/10.1038/npjgenmed.2015.7 ↩ ↩2
-
Sulev Reisberg, Kristi Krebs, Maarja Lepamets, et al. Translating genotype data of 44,000 biobank participants into clinical pharmacogenetic recommendations: challenges and solutions. Genetics in Medicine, 2019. https://doi.org/10.1038/s41436-018-0337-5 ↩
-
Xiaolong Tom Zhang, Jacob Blacutt, Thomas Lloyd, et al. Enhancing clinical research with pharmacogenomics: a practical perspective. Bioanalysis, 2025. https://doi.org/10.1080/17576180.2025.2481019 ↩
-
Alison Abbott. With your genes? Take one of these, three times a day. Nature, 2003. https://doi.org/10.1038/425760a ↩
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.