What a DNA Test for Medication Can and Cannot Tell You

Oak
pharmacogenomics cyp2d6 whole-genome-sequencing pharmcat cpic
A laboratory instrument rotates a carousel of colored drug ampoules while glowing amber fluid drains at different speeds through tall enzyme columns.

A DNA test for medication, properly called pharmacogenetic or PGx testing, genotypes a small set of genes that encode drug-metabolizing enzymes and transporters and reports whether you carry variants that change how fast you clear certain drugs. For roughly 20 gene-drug pairs the evidence is strong enough that the Clinical Pharmacogenetics Implementation Consortium (CPIC) publishes prescribing guidance with an A-level rating: thiopurines and TPMT/NUDT15, fluoropyrimidines and DPYD, clopidogrel and CYP2C19, codeine and CYP2D6, abacavir and HLA-B*57:01, allopurinol and HLA-B*58:01, warfarin dosing and CYP2C9/VKORC1. For choosing an antidepressant or an antipsychotic, the genotype tells you about clearance and exposure, not about whether the drug will help. Any test marketed as identifying which psychiatric medication “works for you” is selling a combinatorial algorithm on top of the same pharmacokinetic genes, and the evidence for those algorithms improving outcomes is thin. Prescribing decisions belong with the clinician who is treating you, and a PGx result is one input among many.

The genes that carry real evidence, and what they change

Pharmacogenetics is old enough to be unglamorous. The core observation, which predates cheap sequencing by decades, is that common loss-of-function and gain-of-function variants in a handful of cytochrome P450 enzymes produce order-of-magnitude differences in plasma exposure at a fixed dose 1. The variants are organized as haplotypes, or star alleles, curated by PharmVar: CYP2C19*2 (rs4244285, a splice-disrupting variant) and *3 are non-functional, *17 increases transcription, and the diplotype maps to a metabolizer phenotype that CPIC then maps to guidance. CYP2D6 is the messiest and most consequential, with over 150 named alleles, whole-gene deletions (*5), duplications, and hybrid genes formed by recombination with the neighboring pseudogene CYP2D7.

Three categories are where genotype changes practice most sharply. DPYD variants (*2A rs3918290, c.2846A>T rs67376798, the intronic HapB3 variant rs56038477, and *13) reduce dihydropyrimidine dehydrogenase activity and raise the risk of severe fluoropyrimidine toxicity, which is why pre-treatment DPYD genotyping is now standard in much of Europe. TPMT and NUDT15 do the same for thiopurines. The HLA alleles are different in kind: HLA-B*57:01 for abacavir hypersensitivity and HLA-B*15:02 for carbamazepine-induced Stevens-Johnson syndrome are immunological risk markers, not metabolism, and they behave like near-deterministic contraindications rather than dose adjustments. Those are conversations to have with a prescriber before starting a drug, not after.

Why we would call pharmacogenes from whole-genome sequencing rather than buy a panel

Commercial PGx tests genotype a targeted set of positions, typically on an array or a small amplicon panel, usually a few dozen to a few hundred variants. That design has a specific failure mode: anything not on the panel is silently assumed to be reference. If you carry a rare loss-of-function variant in CYP2C9 that the array does not probe, you are reported as a normal metabolizer with no indication of uncertainty. Whole-genome sequencing at 30x covers the full coding and regulatory extent of these genes and lets you interrogate rare variants alongside the common star alleles, which is the main argument for WGS-based pharmacogenetic profiling over fixed panels 2. Short-read WGS also recovers structural variation in pharmacogenes, and analysis of the 1000 Genomes whole genomes found pharmacogene structural variants to be common and population-specific, including CYP2D6 hybrids and gene deletions that arrays cannot see 3.

The counterargument is honest engineering: sequencing gives you more raw information and more ways to be wrong. Star allele calling from short reads requires dedicated software with its own assumptions, and next-generation sequencing for pharmacogenetics has known limits around phasing, homologous regions, and CNV breakpoints 4. Depth matters, since the variant-calling sensitivity you get at 15x is not the sensitivity you get at 30x, especially for heterozygous calls in GC-rich exons 5. If you are ordering sequencing with pharmacogenetics in mind, ask for 30x or better and insist on receiving the BAM or CRAM, not only a VCF.

The pipeline we would run

Start from an aligned CRAM against GRCh38 and a joint or single-sample VCF from GATK HaplotypeCaller or DeepVariant. The workhorse is PharmCAT, which implements CPIC and DPWG allele definitions and produces a per-gene diplotype, a phenotype, and the guideline text.

# normalize and fill in reference calls at PGx positions
pharmcat_vcf_preprocessor -vcf sample.vcf.gz \
  -refFna GRCh38_full_analysis_set.fna -bf sample

java -jar pharmcat.jar -vcf sample.preprocessed.vcf.bgz \
  -reporterJson -o pharmcat_out

The preprocessor exists because of a subtlety that trips up almost everyone the first time. A position absent from your VCF is ambiguous: it could be homozygous reference or it could be a no-call in a region with zero coverage. PharmCAT will not assume, so you must supply either a gVCF or a VCF that carries explicit reference calls at the defining positions. Run bcftools norm -m-both -f ref.fa beforehand if your caller emits multiallelic records, and check the PharmCAT warnings section for positions flagged as missing. A diplotype called from two-thirds of its defining positions is a guess wearing a label.

PharmCAT deliberately does not call CYP2D6 copy number from a VCF. For that, run Cyrius, which models read depth across CYP2D6/CYP2D7 to resolve deletions, duplications, and the common hybrid alleles:

python star_caller.py --manifest cram_manifest.txt \
  --genome 38 --prefix sample --outDir cyrius_out --threads 4

Feed the Cyrius diplotype back into PharmCAT with -matcherHtml and the outside-call file. Aldy and StellarPGx are reasonable cross-checks, and we would run at least two callers on CYP2D6 and treat disagreement as an unresolved result rather than picking the answer you prefer. For HLA-B*57:01 and *15:02, standard germline variant calling is not enough. Use a dedicated typer such as HLA*LA on the CRAM, since the MHC region’s divergence from the reference means reads there are frequently mismapped or filtered.

Interpretation is the part that does not automate. The framework question, posed clearly a decade ago for whole-genome interpretation in adults, is how to move from sequence to a defensible statement about phenotype when most variants have weak or conflicting annotation 6. For pharmacogenetics the answer is narrower than for disease genetics: restrict yourself to CPIC level A and B pairs, use PharmVar haplotype definitions rather than single rsIDs, and ignore the long tail of variants of uncertain significance in these genes. WGS in healthy people generates plenty of findings that look actionable and are not, and pharmacogenetics is one of the few domains where the evidence base is tight enough to be useful without a clinical geneticist reading over your shoulder 7.

Mental health, specifically

The searches that bring people here are mostly about psychiatry, so here is the position plainly. Genotype affects psychotropic exposure in ways worth knowing. A CYP2D6 poor metabolizer on a drug cleared primarily by CYP2D6 will reach higher plasma concentrations than a normal metabolizer at the same dose, and a CYP2C19 ultrarapid metabolizer may reach lower ones. CPIC has guidelines for SSRIs and tricyclics on exactly that basis, framed as dose and drug-selection considerations for a prescriber.

What genotype does not do is predict response. The combinatorial tests that assign drugs to green, yellow, and red categories combine the same pharmacokinetic genes with pharmacodynamic markers of much weaker evidence, and their proprietary weightings are not published in a form you can audit. Separately, no DNA test diagnoses depression, bipolar disorder, ADHD, or schizophrenia. Polygenic scores for psychiatric traits explain a small fraction of variance, have poor transferability across ancestries, and are research instruments. If you are unwell, the path runs through a psychiatrist.

Questions people also ask

Is genetic testing for medication worth it? For the CPIC level A pairs, yes, particularly before starting a fluoropyrimidine, a thiopurine, abacavir, or carbamazepine, and it is worth doing before you need it rather than during a crisis. For picking an antidepressant, the value is modest and specific: it tells your prescriber about exposure, not efficacy.

How much does a pharmacogenetic test typically cost? Standalone PGx panels generally run a few hundred dollars out of pocket, with insurance coverage varying by indication. If you already have 30x whole-genome data, the marginal cost is compute time, since PharmCAT, Cyrius, and Aldy are free and open source.

How accurate is pharmacogenetic testing? Genotyping accuracy for well-characterized SNVs is very high, above 99% concordance for the common star alleles. Accuracy at the phenotype level is lower, and CYP2D6 is where errors concentrate, because hybrid alleles and copy-number states are hard to resolve from short reads 3. Predicting clinical outcome is a separate and much weaker link, since dose, adherence, drug interactions, liver and kidney function, and smoking all move exposure as much as genotype does.

Can you get a DNA test to see what meds work for you? You can get a test that predicts how you process certain drugs. No test predicts which drug will work, and claims otherwise overstate what the underlying genetics support.

How do I get genetic testing for medication? Either through a clinician who orders a CPIC-aligned panel from a CLIA-certified laboratory and puts the result in your chart, or by sequencing your genome and running the open-source callers yourself. Results you generate at home are for your own understanding, and a clinician will typically want a clinical-grade confirmation of any allele before prescribing on it.

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. Howard L McLeod, William E Evans. Pharmacogenomics: Unlocking the Human Genome for Better Drug Therapy. Annual Review of Pharmacology and Toxicology, 2001. https://doi.org/10.1146/annurev.pharmtox.41.1.101 ↩

  2. Sylvan Manuel Caspar, Timo Schneider, Patricia Stoll, et al. Potential of Whole-Genome Sequencing-Based Pharmacogenetic Profiling. Pharmacogenomics, 2021. https://doi.org/10.2217/pgs-2020-0155 ↩

  3. Carissa A. Sherman, Katrina G. Claw, Seung-been Lee. Pharmacogenetic analysis of structural variation in the 1000 genomes project using whole genome sequences. Scientific Reports, 2024. https://doi.org/10.1038/s41598-024-73748-3 ↩ ↩2

  4. Yuan Ji, Sherin Shaaban. Interrogating Pharmacogenetics Using Next-Generation Sequencing. The Journal of Applied Laboratory Medicine, 2024. https://doi.org/10.1093/jalm/jfad097 ↩

  5. Rachel L Goldfeder, Dennis P Wall, Muin J Khoury, et al. Human Genome Sequencing at the Population Scale: A Primer on High-Throughput DNA Sequencing and Analysis. American Journal of Epidemiology, 2017. https://doi.org/10.1093/aje/kww224 ↩

  6. Frederick E. Dewey, Megan E. Grove, James R. Priest, et al. Sequence to Medical Phenotypes: A Framework for Interpretation of Human Whole Genome DNA Sequence Data. PLOS Genetics, 2015. https://doi.org/10.1371/journal.pgen.1005496 ↩

  7. Noralane M. Lindor, Stephen N. Thibodeau, Wylie Burke. Whole-Genome Sequencing in Healthy People. Mayo Clinic Proceedings, 2017. https://doi.org/10.1016/j.mayocp.2016.10.019 ↩