The Best Genetic Testing for Medication Response

Oak
pharmacogenomics cyp2d6 pharmcat whole-genome-sequencing cpic
Glowing vine stalks in a dark forest, each bearing a different number of luminous pods, above a long ribbon of blue light on the ground.

If you want the most pharmacogenetic information per dollar, get a 30x whole-genome sequence aligned to GRCh38 and run it through PharmCAT for star-allele calling plus Cyrius for CYP2D6 copy number, then read the results against CPIC guidelines. That covers every gene a branded panel covers and roughly 20 more, and the BAM/CRAM stays yours for the next question you have. The tradeoff is real: a research-grade WGS call is not a CLIA-certified PGx report, and a prescriber who wants a signed lab document will ask for one. If a specific prescribing decision is imminent (fluoropyrimidine chemotherapy, thiopurines, abacavir), order a clinical PGx test from a CLIA lab and bring the result to the clinician. Everything below is about measurement and interpretation. Dosing and drug selection are a clinician’s job, and nothing here substitutes for that conversation.

Which genes carry the evidence

Pharmacogenetics is not uniformly strong. A small set of gene-drug pairs has large, replicated effect sizes and clinical guidelines behind them, and a much larger set has modest, inconsistent associations. The idea that individual variation in drug-metabolizing enzymes should drive dosing is old and well-supported in principle 1. What varies is how much a given variant moves an outcome you care about.

The pairs with CPIC Level A guidelines and the clearest mechanism:

  • DPYD and fluoropyrimidines (5-FU, capecitabine). Loss-of-function variants like c.1905+1G>A (*2A), c.1679T>G (*13), and c.2846A>T predict severe, occasionally fatal toxicity. This is the single highest-stakes result on most reports.
  • TPMT and NUDT15 with thiopurines. Same shape: reduced enzyme activity, myelosuppression risk.
  • CYP2C19 with clopidogrel, proton pump inhibitors, and several SSRIs. *2 (rs4244285) and *3 (rs4986893) are loss of function, *17 (rs12248560) increases expression.
  • CYP2D6 with tamoxifen, codeine/tramadol, many antidepressants and antipsychotics.
  • SLCO1B1 c.521T>C (rs4149056) and simvastatin-associated myopathy.
  • CYP2C9 plus VKORC1 (rs9923231) for warfarin.
  • HLA-B*57:01 (abacavir), HLA-B*15:02 and HLA-A*31:01 (carbamazepine). These are immunological, not metabolic, and they are close to binary in their predictive value.
  • CYP3A5 *3/*6/*7 for tacrolimus.
  • G6PD, X-linked, relevant to several oxidant drugs.

Note what is missing from that list: SLC6A4, COMT, MTHFR, and the other pharmacodynamic markers that fill out commercial psychiatric panels. The pharmacodynamic side of antidepressant response has been studied for two decades and the replicated effects are small 2.

Why CYP2D6 breaks most tests

CYP2D6 is the gene that separates a real pharmacogenetic pipeline from a genotyping chip. It sits next to two pseudogenes, CYP2D7 and CYP2D8, with high sequence identity. Real people carry whole-gene deletions (*5), duplications and multiplications (*1xN, *2xN, *4xN), and CYP2D6-CYP2D7 hybrid alleles (*13, *68, *36) formed by gene conversion. PharmVar catalogs well over a hundred star alleles.

A fixed SNP array or a small targeted panel genotypes a handful of positions and calls a diplotype from them. If you carry a duplication of a non-functional allele, or a *68+*4 tandem, or a deletion, that design will either miss it or call the wrong activity score. This is the main reason two commercial PGx reports on the same person can disagree on CYP2D6 phenotype. Consumer array raw data (23andMe, AncestryDNA) is worse still: no copy-number information, no phasing, and per-site error rates that are fine for ancestry and unacceptable for a metabolizer call.

Short-read WGS at 30x handles this if you use a caller built for the locus. Cyrius (Illumina) genotypes CYP2D6 from WGS by modeling read depth across CYP2D6/CYP2D7 and resolving the hybrid structures. Aldy and StellarPGx are reasonable alternatives, and Aldy works on targeted panels too. Sequencing study design matters here in the ordinary way: depth, uniformity, and alignment choices set what you can call 3.

Running it yourself from a WGS file

Start from a GRCh38 CRAM with chr-prefixed contigs. If your provider gives you GRCh37, realign rather than lifting over. The PGx region definitions are GRCh38-based and liftover of indel-heavy loci silently corrupts calls.

Call CYP2D6 first:

echo /data/sample.cram > cram_manifest.txt
python3 star_caller.py \
  --manifest cram_manifest.txt \
  --genome 38 \
  --prefix sample \
  --outDir cyrius_out \
  --threads 8

Cyrius emits a TSV with a Genotype column (for example *1/*4) and a Filter column. Treat anything other than PASS as unresolved rather than as a negative.

Then run the PharmCAT preprocessor and the caller:

bcftools norm -m-any -f GRCh38.fna -Oz -o sample.norm.vcf.gz sample.vcf.gz
python3 pharmcat_vcf_preprocessor.py \
  -vcf sample.norm.vcf.gz \
  -refFna GRCh38.fna \
  -o pcat_in/

printf "CYP2D6\t*1/*4\n" > outside_calls.tsv

java -jar pharmcat.jar \
  -vcf pcat_in/sample.preprocessed.vcf.bgz \
  -po outside_calls.tsv \
  -reporterJson \
  -o reports/

The failure mode that bites everyone: PharmCAT does not assume that an absent position is homozygous reference. If your VCF came from a variant-only pipeline, every uncalled PGx site becomes “unknown” and genes silently drop to Indeterminate. You need a gVCF or a VCF with reference calls emitted at the PharmCAT positions. Check reports/sample.report.json for the count of genes with no call before you read anything else. The second failure mode is phasing. Unphased data forces PharmCAT to report multiple possible diplotypes when two variants could be cis or trans, which matters most for CYP2C9 and CYP2B6. Long reads or read-backed phasing (WhatsHap) resolve this.

For HLA-B*57:01 and *15:02, short-read WGS is workable with HLA-LA or Optitype, but for a decision with real consequences, a lab typing result is the right input.

What the psychiatric panels do and do not show

GeneSight and Genomind both combine metabolizer genotypes with pharmacodynamic markers into a proprietary “combinatorial” score and a red/yellow/green medication sort. The metabolizer part is standard pharmacogenetics. The combinatorial algorithm is the product, and it is not public, which means you cannot audit how a category assignment was reached 4.

The trial evidence for these tools is mixed. Meta-analytic work has found effects on depression rating scales that are statistically detectable but modest, with most of the underlying trials industry-sponsored and not fully blinded 5. A randomized single-blind study of combinatorial testing in depression reported improvement in response measures relative to standard care 6. A group of academic psychiatrists reviewing clinical implementation concluded that the available tools were not yet supported well enough for routine antidepressant prescribing, while noting that specific CYP2C19 and CYP2D6 results do carry actionable information 7. Decision-support systems built around explicit, published pharmacogenetic rules have also been tested and show measurable effects on dosing and side-effect burden 8.

Our read: pay for the genotype, treat the color-coded medication list as one clinician-facing input among several, and do not let a green category substitute for clinical judgment or a yellow one rule out a drug that is otherwise indicated.

Sequencing once, using it many times

The argument for WGS over a panel is not only CYP2D6. The same CRAM answers carrier-status questions, gives you ACMG secondary-findings genes, and supports reanalysis as variant databases grow. Genome sequencing has become the first-line diagnostic test for rare disease precisely because one assay replaces a sequence of targeted ones 9. The cost of population-scale sequencing has fallen far enough that per-individual genomes are now routine infrastructure rather than a research event 10.

The cost of that flexibility is interpretive burden. A panel hands you eight genes and a report. A genome hands you a file and the obligation to know which tools to trust. If you are the kind of person who will run the commands above, the genome is the better purchase. If you will not, order a CLIA PGx panel that explicitly states it performs CYP2D6 copy-number analysis, and check that DPYD and TPMT/NUDT15 are on the gene list.

Questions people also ask

Is GeneSight or Genomind better? They test overlapping gene sets and differ mainly in the proprietary algorithm that sorts medications into categories. Neither algorithm is published in enough detail to audit, and neither has evidence strong enough to prefer one over the other on outcomes 4. If you are choosing between them, compare the gene lists and whether CYP2D6 copy number is measured.

Which pharmacogenetic test is best? For information per dollar and long-term use, 30x WGS on GRCh38 processed with PharmCAT plus Cyrius. For a result a prescriber will act on this month, a CLIA-certified panel that covers DPYD, TPMT, NUDT15, CYP2C19, CYP2C9, CYP2D6 with CNV, SLCO1B1, and HLA-B*57:01.

Can you get genetic testing to see what meds will work for you? You can get testing that predicts how fast you clear certain drugs and that flags a small number of high-risk immunological reactions. Prediction of whether a psychiatric medication will relieve symptoms is much weaker, because the pharmacodynamic associations behind it are small and inconsistently replicated 2.

Can I use my 23andMe raw data? No. Array data lacks copy-number resolution at CYP2D6, is unphased, and has per-site error rates unsuited to a clinical genotype. It is useful for orientation and nothing else.

Do I need a clinician? Yes, for any prescribing decision. A metabolizer phenotype is one input into a choice that also involves your history, other medications, organ function, and symptoms. Bring the report to the person who writes the prescription.

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. Chiara Fabbri, Concetta Crisafulli, Marco Calabrò, et al. Progress and prospects in pharmacogenetics of antidepressant drugs. Expert Opinion on Drug Metabolism & Toxicology, 2016. https://doi.org/10.1080/17425255.2016.1202237 ↩ ↩2

  3. David B. Goldstein, Andrew Allen, Jonathan Keebler, et al. Sequencing studies in human genetics: design and interpretation. Nature Reviews Genetics, 2013. https://doi.org/10.1038/nrg3455 ↩

  4. Chad A Bousman, Malcolm Hopwood. Commercial pharmacogenetic-based decision-support tools in psychiatry. The Lancet Psychiatry, 2016. https://doi.org/10.1016/s2215-0366(16)00017-1 ↩ ↩2

  5. Joshua D. Rosenblat, Yena Lee, Roger S. McIntyre. Does Pharmacogenomic Testing Improve Clinical Outcomes for Major Depressive Disorder?. The Journal of Clinical Psychiatry, 2017. https://doi.org/10.4088/jcp.15r10583 ↩

  6. Xiaoxiao Shan, Wenli Zhao, Yan Qiu, et al. Preliminary Clinical Investigation of Combinatorial Pharmacogenomic Testing for the Optimized Treatment of Depression: A Randomized Single-Blind Study. Frontiers in Neuroscience, 2019. https://doi.org/10.3389/fnins.2019.00960 ↩

  7. 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 ↩

  8. Michael S. Zastrozhin, Aleksandr S. Sorokin, Tatyana V. Agibalova, et al. Using a personalized clinical decision support system for bromdihydrochlorphenylbenzodiazepine dosing in patients with anxiety disorders based on the pharmacogenomic markers. Human Psychopharmacology: Clinical and Experimental, 2018. https://doi.org/10.1002/hup.2677 ↩

  9. Kristen M. Wigby, Deanna Brockman, Gregory Costain, et al. Evidence review and considerations for use of first line genome sequencing to diagnose rare genetic disorders. npj Genomic Medicine, 2024. https://doi.org/10.1038/s41525-024-00396-x ↩

  10. F Carrasco-Ramiro, R Peiró-Pastor, B Aguado. Human genomics projects and precision medicine. Gene Therapy, 2017. https://doi.org/10.1038/gt.2017.77 ↩