Skip to content

How to Read Pharmacogenetic Test Results

Oak
A laboratory instrument with two glass columns of glowing helical filaments, one holding three copies, beside robotic fingers sorting notched tokens.

By the end of this guide you will be able to take a pharmacogenetic (PGx) report and translate each gene result into a star allele diplotype. You will also convert that diplotype into a standardized phenotype and check whether the lab’s call is supported by the underlying data. You will also know which genes a given test cannot call correctly and why. You need a PGx report from GeneSight, OneOme, Genomind, Invitae, or a hospital panel. Ideally you also have access to your own sequencing data as a VCF file. VCF is variant call format, the standard text format listing positions where you differ from the reference. If you have them, add the aligned reads as a CRAM or BAM file. Everything below runs on a laptop except the CYP2D6 copy number work, which wants the read-level data.

A note on scope. This guide is about reading and verifying measurements. Every decision about starting, stopping, or changing a medication belongs to a prescribing clinician who can see your full history. PGx results modify how a drug behaves in you. They do not tell you which drug you need.

1. Separate the three layers of the report

Every PGx report contains three logically distinct layers, and most reports blur them together. Reading the report well means pulling them apart.

The first layer is genotype: the actual variants observed, usually written as rsIDs (reference SNP identifiers, e.g. rs3892097) or as star allele diplotypes like CYP2C19 *1/*2. This layer is a measurement. It is either right or wrong, and it can be checked.

The second layer is phenotype: a category like “poor metabolizer” or “intermediate metabolizer” assigned by translating the diplotype through a published table. This layer is an inference from a consensus mapping. It changes when the consensus changes.

The third layer is the drug recommendation: green, yellow, and red columns, or text like “consider alternatives.” This is a clinical interpretation layered on top of the phenotype, and it is where reports differ most from each other. Some vendors use CPIC guidelines directly. Others use proprietary algorithms that combine genes in ways they do not disclose. When two reports disagree about you, they usually agree at layer one and diverge at layer three.

Write down the layer-one content for every gene before you read anything else. That is the only part of the report that is your data.

2. Decode star allele nomenclature

Star alleles are haplotype names. A haplotype is a specific combination of variants inherited together on one copy of a chromosome. The CYP2D6 *4 allele is a defined set, with rs3892097 (the 1846G>A splice defect) as its core defining change. The name does not refer to a single variant. You have two copies of most genes, so a result is a diplotype: two star alleles separated by a slash.

By convention *1 is the reference haplotype, meaning no known function-altering variant was found. This is the single most important thing to understand about PGx reports: *1 is a statement about what the assay looked for and did not find. The call is not a statement that the gene is normal. A targeted panel that genotypes eight CYP2D6 positions will call you *1/*1 even if you carry a rare loss-of-function variant at a ninth position. Sequencing the whole gene finds variants that targeted panels miss by construction, and the CYP2D6 locus in particular has a long tail of rare and novel alleles.1

Some star alleles are structural rather than point variants. CYP2D6*5 is a whole-gene deletion. CYP2D6*1xN or *2xN denotes a gene duplication or multiplication. These require copy number detection, not SNV genotyping, which is why a report that omits copy number for CYP2D6 is incomplete.

Suballeles appear as *2.001 or *4A. They differ at positions with no known functional consequence and collapse to the same functional call. You can generally ignore the decimal.

3. Convert diplotype to phenotype using the activity score

For CYP2D6, CPIC (the Clinical Pharmacogenetics Implementation Consortium) uses an activity score system. Each allele gets a value: 1.0 for normal function, 0.5 for decreased function, 0 for no function. Sum the two alleles, multiply duplicated alleles by their copy number, and the total maps to a phenotype band.

Worked example. Suppose your report says CYP2D6 *1/*4. *1 scores 1.0, *4 scores 0. Total is 1.0, which CPIC maps to intermediate metabolizer. Now suppose it says *4/*4xN. Both alleles are non-functional, so the duplication is irrelevant and the score is 0, a poor metabolizer. And *1/*2x2 gives 1.0 plus 2.0, a score of 3.0, an ultrarapid metabolizer.

CPIC defines the bands as follows. An activity score of 0 is poor metabolizer and 0.25 to 1.0 is intermediate. A score of 1.25 to 2.25 is normal, and above 2.25 is ultrarapid. If your report gives a phenotype without showing the activity score, compute it yourself from the diplotype and confirm the two agree. Disagreements happen, usually because the report follows an older mapping. The 2019 consensus reclassified activity score 1.0 from normal to intermediate, so older reports call *1/*4 normal and newer ones call it intermediate.

Other genes use different schemes. CYP2C19 and CYP2C9 have their own allele function tables and diplotype-to-phenotype lookups. TPMT and NUDT15 use a similar no-function/decreased-function framework. DPYD is scored by gene activity score over four well-characterized variants. HLA-B*57:01 and HLA-B*15:02 are presence-or-absence risk alleles for specific hypersensitivity reactions, and the report should say carrier or non-carrier. They are not metabolizer genes at all. Do not try to map them onto a metabolizer scale.

Pull the current tables from the CPIC site rather than trusting the report’s internal version:

# CPIC publishes allele definition, function, and diplotype-phenotype
# tables as downloadable files per gene.
curl -sL "https://api.cpicpgx.org/v1/allele?gene=eq.CYP2D6" \
  -H "Accept: application/json" | jq '.[] | {name, clinicalfunctionalstatus, activityscore}'

4. Check the lab call against your own data

This is the step most people skip, and it is where you learn the most. If you have a VCF from whole-genome sequencing, you can independently call most PGx genes and compare.

The tool we would use is PharmCAT. It takes a VCF, calls diplotypes for the CPIC genes, and emits both a machine-readable JSON and an HTML report. It is the reference implementation for genotype-to-phenotype translation, which makes it a fair arbiter of layer two.

# Normalize and restrict to PGx positions first.
# PharmCAT ships a positions VCF; use it as the template.
java -jar pharmcat-preprocessor.jar \
  -vcf mysample.vcf.gz \
  -refFna GRCh38.primary_assembly.genome.fa \
  -refVcf pharmcat_positions.vcf.bgz \
  -o ./prep

java -jar pharmcat.jar \
  -vcf ./prep/mysample.preprocessed.vcf.bgz \
  -reporterJson \
  -o ./out

Then compare gene by gene:

jq -r '.genes.CPIC[] | [.gene, (.recommendationDiplotypes[]?.label // "none"),
  (.recommendationDiplotypes[]?.phenotypes[]? // "")] | @tsv' \
  ./out/mysample.report.json | column -t

Two failure modes to expect. First, positions with no coverage in your VCF are treated as reference, which silently biases calls toward *1. PharmCAT’s preprocessor flags missing positions, so read that output before trusting the calls. Second, PharmCAT will not call CYP2D6 copy number from a VCF alone. You must supply an external CYP2D6 call.

Whole-genome sequencing gives you a far larger PGx variant inventory than any panel, including rare variants that appear in only one or two people in a cohort.2 That breadth is the reason to do the comparison. It is also why the comparison often shows the panel calling *1 where sequencing finds something.

5. Handle CYP2D6 separately

CYP2D6 is the hard one and deserves its own step. It sits next to two highly similar pseudogenes, CYP2D7 and CYP2D8. It has common whole-gene deletions, duplications up to at least five copies, and hybrid alleles where part of CYP2D6 is fused to part of CYP2D7. Short reads mismap in this region routinely, and a 150 bp read cannot phase variants across the 4.4 kb gene body.

For short-read data, run a dedicated caller:

# Cyrius is purpose-built for CYP2D6 from WGS CRAM/BAM.
python star_caller.py \
  --manifest manifest.txt \
  --genome 38 \
  --prefix mysample \
  --outDir ./cyrius \
  --threads 4

Cyrius reports a diplotype plus a filter field. Treat any result other than PASS as unresolved rather than negative. If Cyrius and your commercial report disagree on copy number, the copy number call is the more likely error on the panel side, because many panels do not assay it at all.

Long-read sequencing resolves this class of problem properly, because a single read spans the full gene and captures phase directly. The same advantage applies to other structurally messy PGx genes. Long-read work on NAT2, for instance, has been used to benchmark allele, diplotype, and phenotype call accuracy against other testing strategies, and it exposes the limits of methods that cannot phase.3 Targeted short-read panels can be validated to high accuracy for the variants they include, and a well-designed panel with full-gene coverage is a reasonable middle path.4 The general trend is clear: next-generation sequencing gives broader and more complete PGx characterization than fixed genotyping assays.56

If you only have array data or low-pass sequencing, be careful. Low-pass sequencing with imputation performs comparably to arrays for common PGx variants, but both depend on the imputation reference panel and both degrade for rare alleles.7 Neither handles CYP2D6 structural variation.

6. Read the drug-level section skeptically

Now go back to layer three. For each drug listed, find out which gene-phenotype pair drives the call and whether a published guideline supports it.

CPIC assigns evidence levels to gene-drug pairs. Level A means there is enough evidence to change prescribing. Level B means evidence is moderate. Levels C and D mean the pair is listed but no prescribing action is recommended. A commercial report that color-codes fifty drugs is usually drawing on far fewer level A pairs than the visual design implies. Ask, for any yellow or red cell, which of the roughly two dozen level A gene-drug pairs it corresponds to.

Psychiatric panels are where this matters most, because that is where combinatorial proprietary algorithms are common. A report may downgrade a drug based on a weighted combination of CYP2D6, CYP2C19, and pharmacodynamic genes such as SLC6A4 or HTR2A. The pharmacokinetic genes have solid mechanistic support. The pharmacodynamic ones are considerably weaker, and CPIC does not issue actionable recommendations for most of them. Reviews of PGx in psychiatry have consistently flagged the gap between the marketing of combinatorial tests and the evidence for their specific algorithms.

Also note what the report does not include. In a preemptive sequencing program of 10,000 patients, the large majority carried at least one actionable variant across a small set of genes.8 Actionable findings are the norm rather than the exception. The interesting question is therefore coverage rather than whether you have anything at all. The practical bar for adoption has been framed as noninferiority to standard prescribing, which is a lower bar than superiority and a reasonable one.9

7. Build a durable summary you can hand to a clinician

Write a one-page document you keep indefinitely. Genotype does not change, so this is a one-time cost with lifetime value. The permanence is the core argument for preemptive testing over ordering a test at the moment a prescription is needed.10

For each gene, include the diplotype and the activity score where applicable. Add the phenotype under the current CPIC mapping, the assay type, and the date. State explicitly which genes were sequenced fully and which were genotyped at fixed positions. A clinician reading CYP2D6 *1/*1 (targeted panel, 9 variants, no CNV assay) gets information that CYP2D6: normal metabolizer hides.

Gene      Diplotype     AS    Phenotype (CPIC 2024)   Assay
CYP2D6    *1/*4         1.0   Intermediate            WGS 30x + Cyrius (PASS, CN=2)
CYP2C19   *1/*17        n/a   Rapid                   WGS 30x, PharmCAT
CYP2C9    *1/*2         1.5   Intermediate            WGS 30x, PharmCAT
DPYD      Ref/Ref       2.0   Normal                  WGS 30x, 4 variants assessed
TPMT      *1/*1         n/a   Normal                  WGS 30x, PharmCAT
NUDT15    *1/*1         n/a   Normal                  WGS 30x, PharmCAT
SLCO1B1   *1/*5         n/a   Decreased function      WGS 30x, PharmCAT
HLA-B     *57:01 neg          Non-carrier             WGS 30x, imputed - confirm if needed

Flag imputed HLA calls as provisional. HLA typing from short-read WGS is good but not equivalent to a clinical-grade HLA assay, and the drug associations there are hypersensitivity risks where a false negative carries real weight. If an HLA result would inform a specific prescribing decision, that confirmation is a conversation to have with your clinician.

Common problems

Every gene reads *1/*1. Check whether the assay covered the gene at all and how many positions it interrogated. Uniform reference calls across a panel are more often an artifact of limited content than a real finding.

Two reports give different phenotypes from the same diplotype. Compute the activity score yourself and check which CPIC consensus version each report used. The 2019 CYP2D6 reclassification explains most CYP2D6 disagreements. For CYP2C19, check whether *17 was counted as increased function.

PharmCAT returns no call for a gene. This usually means required positions were missing from your VCF rather than absent from your genome. Read the preprocessor’s missing-positions report. Variant callers often drop positions in low-complexity or high-homology regions, and PGx genes are enriched for both.

Cyrius returns a non-PASS filter. The locus is ambiguous in your data, commonly due to a hybrid CYP2D6-CYP2D7 allele or unusual copy number. Do not substitute the commercial panel’s call as a resolution unless that panel measured copy number. Long-read sequencing of the locus is the clean answer.

The report lists a drug you take with no gene attached. Some reports include drugs with only pharmacodynamic or weak-evidence associations. Look up the gene-drug pair’s CPIC evidence level before giving it weight.

You are tempted to change something based on the report. Do not. A phenotype call tells you about metabolic capacity. It does not tell you what you should take. Dose and drug selection depend on indication and other medications. They also depend on kidney and liver function and much else. Bring the summary from step seven to a prescriber or a pharmacist and let them integrate it.

Footnotes

  1. 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 ↩

  2. Clint Mizzi, Brock Peters, Christina Mitropoulou, et al. Personalized Pharmacogenomics Profiling Using Whole-Genome Sequencing. Pharmacogenomics, 2014. https://doi.org/10.2217/pgs.14.102 ↩

  3. Shobana John, Erin C. Boone, Byunggil Yoo, et al. Characterization of NAT2 Using Long‐Read Sequencing: Allele, Diplotype, and Phenotype Call Accuracy Compared to Other Testing Strategies. Clinical Pharmacology & Therapeutics, 2026. https://doi.org/10.1002/cpt.70427 ↩

  4. Anna Gréen, Nils Johan Fredriksson, Michael Andresen, et al. Development and analytical validation of a targeted short‐read next generation sequencing‐based pharmacogenetic panel for comprehensive variant detection. British Journal of Pharmacology, 2026. https://doi.org/10.1111/bph.70480 ↩

  5. Ute I. Schwarz, Markus Gulilat, Richard B. Kim. The Role of Next-Generation Sequencing in Pharmacogenetics and Pharmacogenomics. Cold Spring Harbor Perspectives in Medicine, 2018. https://doi.org/10.1101/cshperspect.a033027 ↩

  6. Alireza Tafazoli, Henk-Jan Guchelaar, Wojciech Miltyk, et al. Applying Next-Generation Sequencing Platforms for Pharmacogenomic Testing in Clinical Practice. Frontiers in Pharmacology, 2021. https://doi.org/10.3389/fphar.2021.693453 ↩

  7. Kaja Wasik, Tomaz Berisa, Joseph K. Pickrell, et al. Comparing low-pass sequencing and genotyping for trait mapping in pharmacogenetics. BMC Genomics, 2021. https://doi.org/10.1186/s12864-021-07508-2 ↩

  8. SL Van Driest, Y Shi, EA Bowton, et al. Clinically Actionable Genotypes Among 10,000 Patients With Preemptive Pharmacogenomic Testing. Clinical Pharmacology & Therapeutics, 2013. https://doi.org/10.1038/clpt.2013.229 ↩

  9. R B Altman. Pharmacogenomics: “Noninferiority” Is Sufficient for Initial Implementation. Clinical Pharmacology & Therapeutics, 2011. https://doi.org/10.1038/clpt.2010.310 ↩

  10. Munir Pirmohamed. Pharmacogenetics: past, present and future. Drug Discovery Today, 2011. https://doi.org/10.1016/j.drudis.2011.08.006 ↩

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.