Best Genetic Testing for Psychiatric Medications, and How Accurate It Is

The best test is a sequencing-based genotype of a small number of genes with real prescribing evidence. Those genes are CYP2D6, CYP2C19 and CYP2B6, plus HLA-B and HLA-A for specific anticonvulsants. Everything else on a commercial psychiatric panel is either redundant with those genes or supported by evidence too weak to change a prescription. If you have 30x whole-genome sequencing, you already have the data, and you can produce a report equivalent to GeneSight’s genotype section with open tools in about an hour of compute. What no test does is tell you which antidepressant will work. Genotype speaks to exposure, meaning how much drug ends up in your blood at a given dose. It says almost nothing about response. The prescribing decision belongs to a clinician who can see your history, your other medications, and your symptoms.

The genes with actual prescribing evidence

Two enzymes do most of the work in psychiatry. CYP2C19 is the main clearance route for escitalopram, citalopram and sertraline. It also contributes to amitriptyline and clomipramine. CYP2D6 handles paroxetine, fluvoxamine, venlafaxine and vortioxetine. It also handles atomoxetine, aripiprazole, risperidone and the tricyclics. The Clinical Pharmacogenetics Implementation Consortium (CPIC) publishes graded guidelines for these pairs, and the FDA lists a subset in its Table of Pharmacogenetic Associations. Those two documents are the entire defensible scope of psychiatric pharmacogenomics for metabolism.

Three more genes matter in narrower cases. CYP2B6 poor metabolizers have substantially higher efavirenz and bupropion-related exposures, and CPIC covers the former. HLA-B*15:02 carriage is associated with carbamazepine-induced Stevens-Johnson syndrome, with by far the highest prevalence in Southeast Asian ancestries. HLA-A*31:01 is associated with carbamazepine hypersensitivity in European ancestries. These HLA results describe risk of serious skin reactions and must be interpreted by a prescribing clinician before any anticonvulsant decision is made.

Note what is missing. There is no validated genetic predictor of which antidepressant will relieve depression. Depression is highly polygenic with a large environmental contribution, and effect sizes per variant are tiny ¹. Genome-wide studies of schizophrenia and bipolar disorder now find hundreds of loci, and they replicate across diverse cohorts, but they index disease liability rather than drug response ². The gap between a statistically solid association and a clinically usable test is the central difficulty in this field, and it has not closed ³.

What the trials of combinatorial panels showed

GeneSight is the most widely sold product, and it has the largest trial. GUIDED randomized about 1,167 patients with treatment-resistant depression to guided versus usual care. The primary endpoint, symptom improvement on the HAM-D17 at week 8, did not reach significance. Secondary endpoints for response and remission favored the guided arm by roughly five to six percentage points. The trial was not blinded to clinicians in a way that eliminates expectancy effects, and the comparison is between a test report and no report rather than between two genotype interpretations.

The PRIME Care trial in the Veterans Affairs system is the stronger design and the more sobering result. Giving prescribers pharmacogenomic reports reduced prescriptions of drugs with predicted gene-drug interactions, which is what you would expect. It did not produce a durable improvement in remission over the following months. The reasonable conclusion is that these panels change prescribing behavior more reliably than they change outcomes.

The proprietary part of these products is the interpretive algorithm that collapses many genes into a green, yellow, or red column. That algorithm is not published in a form you can audit, and it mixes genes with CPIC-level evidence together with genes that have none. We would not pay for that layer. The genotype underneath is commodity information.

How to get this from your own sequencing data

If you have a 30x whole-genome BAM or CRAM and a GATK-style VCF on GRCh38, the pipeline is short. Start with PharmCAT, which implements CPIC allele definitions and produces a per-gene diplotype and phenotype plus guideline text.

pharmcat_pipeline sample.vcf.gz -o pharmcat_out/ --reference-pgx-vcf

PharmCAT handles several genes well, including CYP2C19, CYP2B6 and CYP3A5. It also covers SLCO1B1, DPYD and others. It deliberately does not call CYP2D6 from a VCF, because that locus has whole-gene deletions, duplications and CYP2D6-CYP2D7 hybrid alleles that a variant caller cannot see. Call it separately with Cyrius, which uses read depth and variant profiles across the region.

python star_caller.py --manifest bams.txt --genome 38 \
  --prefix sample --outDir cyrius_out --threads 8

Then feed the Cyrius diplotype back to PharmCAT as an outside call, a tab-separated file with the gene name and diplotype, passed with -po. PharmCAT will merge it and emit the CYP2D6 activity score and phenotype in the same report.

For HLA, short-read genotyping needs a graph-aware caller. HLA-LA works on the WGS alignment directly.

HLA-LA.pl --BAM sample.cram --graph PRG_MHC_GRCh38_withIMGT \
  --sampleID sample --maxThreads 8

Check that the reported alleles reach four-field or at least two-field resolution with high quality, since HLA-B*15:02 must be distinguished from other B*15 alleles to be meaningful. If the call is ambiguous at two fields, order a targeted clinical HLA typing instead of trusting the WGS call.

Failure modes worth knowing about

Arrays are the main source of wrong answers. A genotyping chip interrogates a fixed set of positions, so a rare loss-of-function variant in CYP2C19 that is not on the chip returns as the reference allele. Sequencing avoids that class of error by observing the whole gene, and the design logic for why wider sequencing finds variants that targeted assays cannot has been laid out clearly in the methods literature ⁴. Rare and structural variation contributes disproportionately in neuropsychiatric genetics generally, which is the argument for whole-genome data over exomes or panels ^{5 6}.

Coverage is the second failure mode. CYP2D6 sits in a segmental duplication, and mapping quality there is poor. Cyrius is calibrated for roughly 30x uniform coverage. Below about 20x, its no-call rate rises and duplication detection degrades. Check mean depth over chr22:42,120,000-42,150,000 before trusting the result.

Third, phenoconversion can override genotype entirely. A CYP2D6 normal metabolizer taking a strong CYP2D6 inhibitor such as fluoxetine, paroxetine, or bupropion behaves like a poor metabolizer. Any report that lists your genotype without asking what else you take is describing a hypothetical you. This is also why measured drug concentrations, where an assay exists, beat genotype for answering whether your exposure is unusual.

What this testing cannot do

It cannot diagnose a psychiatric condition, and it cannot stratify you into a subtype that predicts drug class. Psychiatric genomics has produced real biology at the population level, including specific loci and pathway signals across disorders ⁷. Translating that into individual clinical prediction remains unsolved, and the methylation and expression layers that might eventually help are still research-grade ⁸. Treat any product that claims to pick an antidepressant from your DNA as claiming more than the evidence supports.

The practical value of a clean CYP2C19 and CYP2D6 result is narrower and real. It tells a prescriber whether standard dosing of a specific drug is likely to give you unusually high or unusually low exposure. That is a useful input to a conversation, and it is a conversation you should have with a clinician rather than resolve from a report.

Questions people also ask

How do I get genetic testing for psychiatric medications? Two routes. Order a CPIC-aligned clinical pharmacogenomics panel through a prescriber, or use existing 30x whole-genome data with PharmCAT plus Cyrius and bring the diplotypes to your clinician. The second route costs compute time instead of a test fee and covers more rare variants.

Does GeneSight testing work? It reliably reports your genotype. In GUIDED, the primary symptom endpoint did not reach significance, and secondary response and remission endpoints improved by a few percentage points. PRIME Care found changed prescribing without a durable remission benefit.

Is GeneSight or Genomind better? Neither has a demonstrated advantage over simply knowing your CYP2C19 and CYP2D6 phenotypes and applying published CPIC guidance. Genomind includes more genes with weaker evidence. We would not choose between them on the strength of the interpretive algorithm, because neither algorithm is auditable.

How accurate is it? Allele calling from good sequencing data is highly accurate for CYP2C19 and reaches high concordance for CYP2D6 with Cyrius. The prediction of clinical response from those calls is weak, and that is where the accuracy question really lands.

What about MTHFR, SLC6A4, COMT and HTR2A? They appear on panels without CPIC prescribing guidance for psychiatric drugs. Association findings for these candidate genes have largely not held up at genome-wide scale ¹.

Footnotes

1. Jonathan Flint, Kenneth S. Kendler. The Genetics of Major Depression. Neuron, 2014. https://doi.org/10.1016/j.neuron.2014.02.033 ↩ ↩²

2. Tim Bigdeli, Ayman Fanous, Yuli Li, et al. Genome-Wide Association Studies of Schizophrenia and Bipolar Disorder in a Diverse Cohort of US Veterans. Biological Psychiatry, 2020. https://doi.org/10.1016/j.biopsych.2020.02.468 ↩

3. U Demkow, T Wolańczyk. Genetic tests in major psychiatric disorders—integrating molecular medicine with clinical psychiatry—why is it so difficult?. Translational Psychiatry, 2017. https://doi.org/10.1038/tp.2017.106 ↩

4. 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 ↩

5. Whole Genome Sequencing for Psychiatric Disorders (WGSPD), Stephan J. Sanders, Benjamin M. Neale, et al. Whole genome sequencing in psychiatric disorders: the WGSPD consortium. Nature Neuroscience, 2017. https://doi.org/10.1038/s41593-017-0017-9 ↩

6. Erin L. Heinzen, Benjamin M. Neale, Stephen F. Traynelis, et al. The Genetics of Neuropsychiatric Diseases: Looking In and Beyond the Exome. Annual Review of Neuroscience, 2015. https://doi.org/10.1146/annurev-neuro-071714-034136 ↩

7. Patrick F. Sullivan, Arpana Agrawal, Cynthia M. Bulik, et al. Psychiatric Genomics: An Update and an Agenda. American Journal of Psychiatry, 2018. https://doi.org/10.1176/appi.ajp.2017.17030283 ↩

8. Jonathan Mill, Thomas Tang, Zachary Kaminsky, et al. Epigenomic Profiling Reveals DNA-Methylation Changes Associated with Major Psychosis. The American Journal of Human Genetics, 2008. https://doi.org/10.1016/j.ajhg.2008.01.008 ↩