There Is No Blood Test That Tells You Which Antidepressant Will Work

No blood or saliva test currently predicts which antidepressant will make you feel better. What the commercial tests measure is pharmacokinetics: how fast your liver enzymes clear a drug, mostly CYP2D6 and CYP2C19, plus a handful of pharmacodynamic variants of weak evidentiary standing. That information is real and occasionally useful, particularly for avoiding a dose that will produce either no exposure or unusually high exposure. It is not a matching algorithm for efficacy, and the trial data on clinical outcomes are modest at best. Separately, a plasma or serum drug level (a genuine blood test) tells you whether the drug is in you and roughly at what concentration, which answers a different and often more actionable question.

What the panels genotype

GeneSight, Genomind, NeuroIDgenetix, and their competitors are all variations on the same core: a targeted genotyping panel over pharmacokinetic genes, wrapped in a proprietary interpretation layer that bins drugs into colored columns (“use as directed” / “moderate gene-drug interaction” / “significant gene-drug interaction”). The genes that carry nearly all the defensible signal:

• CYP2C19: sertraline, citalopram, escitalopram and amitriptyline. Key star alleles are *2 (rs4244285, splice defect, loss of function), *3 (rs4986893), and *17 (rs12248560, increased expression). CPIC has published dosing guidance for SSRIs and TCAs keyed to these diplotypes.
• CYP2D6: paroxetine, fluoxetine, venlafaxine, nortriptyline and most TCAs. Hard to call from short-read data because the locus includes a pseudogene (CYP2D7), common structural variation (whole-gene deletion *5, duplications), and hybrid alleles.
• CYP2B6, CYP3A4/5, CYP1A2 and UGT: minor contributors for specific drugs (bupropion via CYP2B6 *6).

Then there is the second tier, which is where the marketing outruns the evidence. It covers SLC6A4 5-HTTLPR (serotonin transporter promoter length polymorphism), HTR2A rs7997012, COMT Val158Met and MTHFR C677T. These are association findings with small effect sizes and an inconsistent replication record. Reviews of the field have been blunt that the pharmacodynamic markers on these panels do not support drug-selection decisions, and that panel composition varies between vendors in ways that produce different recommendations for the same patient.¹² Older assessments made the same point about the panels being marketed ahead of the evidence.³ The structural criticism is that depression is heterogeneous at the level of diagnosis, so a genotype-to-drug mapping inherits that heterogeneity and cannot resolve it.⁴

What the outcome trials show

The trial you want to read is PRIME Care, a VA randomized trial of pharmacogenomic testing in veterans starting treatment for major depressive disorder. It had roughly 2,000 participants across 22 sites and randomized whether the clinician received the test result immediately or after 24 weeks.⁵ The design is the right one: it randomizes access to the information, not the drug, and measures remission over months rather than a proxy.

The general shape of results across this literature: pharmacogenomic-guided arms show a measurable reduction in prescribing of drugs flagged as having a predicted gene-drug interaction, and small, often short-lived differences in symptom scores. Prescribing behavior moves more than outcomes do. That is consistent with the mechanism being real but narrow. Avoiding a CYP2D6 poor-metabolizer dose of a drug cleared primarily by CYP2D6 prevents a specific failure mode, and that failure mode is a minority of treatment failures.

The reading, in our view: these panels are a modest safety and tolerability tool. They do not predict efficacy. Guidance on implementation says as much, framing the tools as decision support that must be interpreted alongside clinical history rather than as a selection algorithm.² If you want the direction the field is heading, it is multimodal measurement-based care, combining symptom trajectories, imaging, and molecular data rather than a single genotype lookup.⁶

Calling CYP2C19 and CYP2D6 from your own genome

If you have a 30x whole-genome BAM or CRAM, you already own the raw data these panels sell. The tooling is good now.

Star-allele calling from short reads:

# Aldy (CYP2D6, CYP2C19, CYP2B6, and ~30 other genes)
aldy genotype -p wgs -g CYP2D6 --genome hg38 -o cyp2d6.aldy sample.cram

# PyPGx (includes structural variant detection via depth profiles)
pypgx run-ngs-pipeline CYP2D6 CYP2D6-pipeline \
  --variants chr22.vcf.gz --depth-of-coverage depth.zip --control-statistics control.zip

# Then map diplotype -> phenotype with PharmCAT
pharmcat -vcf sample.preprocessed.vcf.gz -reporterJson

Notes from doing this:

• PharmCAT’s preprocessor is mandatory. It requires a VCF with the exact reference positions it expects, normalized, with <NON_REF> handled and multi-allelics split. Run pharmcat_vcf_preprocessor.py and read its warnings without skipping past them.
• CYP2D6 needs a caller that models copy number. Aldy and PyPGx both do it from read depth. If you only genotype SNVs, you will call *1/*1 on a sample that is *1/*5 (one gene deleted) and get the metabolizer phenotype wrong in the direction that matters.
• Coverage matters more than you would guess. Below roughly 20x, CYP2D6 copy-number calls get unstable, and the hybrid alleles (*13, *68, *4.013) are easy to miss. Long reads (PacBio HiFi or ONT) resolve the locus cleanly and are the right answer if you have them.
• Activity score to phenotype is a convention. It is not a measurement. CPIC assigns each allele an activity value (0, 0.25, 0.5, 1.0) and sums them: 0 is poor metabolizer, >2.25 is ultrarapid. The boundaries have been revised more than once.
• 5-HTTLPR is an indel in a repeat region of the SLC6A4 promoter and is not reliably genotyped from short-read WGS. If a report gives you an L/S call from an array or short reads, treat the call itself as uncertain, on top of the weak underlying evidence.

Outside depression, the same machinery is better validated for specific drug-gene pairs, for example CYP2D6 and antipsychotic exposure in schizophrenia treatment.⁷ The gene is the same. What differs is how tightly exposure is tied to the outcome you care about.

The other blood test: serum drug levels

Therapeutic drug monitoring is the test people are often reaching for when they ask this question. It measures the drug itself in plasma, by LC-MS/MS, reported in ng/mL. It is standard for nortriptyline and other TCAs, where the exposure-response relationship is established and toxicity is dose-limiting, and it is available for SSRIs and SNRIs through reference labs even though reference ranges there are looser.

What it answers that genotype cannot:

• Is the drug present at all (adherence, absorption, a supply-chain substitution).
• Is your measured concentration consistent with your predicted metabolizer phenotype. Genotype predicts and a level measures. Phenoconversion from co-medications is common enough that predicted and measured routinely disagree.
• For CYP2D6 substrates, the parent-to-metabolite ratio (venlafaxine/O-desmethylvenlafaxine, for instance) is an in vivo readout of enzyme activity.

Draw timing is the whole ballgame. Levels are conventionally trough, drawn 12 to 24 hours after the last dose depending on the drug’s half-life. They are also drawn at steady state, meaning after roughly five half-lives on an unchanged dose. A level drawn two hours post-dose or three days into a new dose tells you almost nothing. Interpreting the number, and changing anything based on it, is a clinician’s job.

How we would use this

Sequence once, keep the raw data, and run star-allele calling yourself. It is a few hours of work and it gives you the same pharmacokinetic information a $300 panel sells, with the diplotype calls visible and auditable rather than collapsed into a colored column. Ignore the pharmacodynamic markers. Treat a CYP2C19 or CYP2D6 poor- or ultrarapid-metabolizer result as one line item to hand a prescriber before starting a drug metabolized by that enzyme, alongside a level drawn at steady state if a drug is not producing the expected effect. Do not expect any of it to tell you which molecule will work. Selecting and adjusting a psychiatric medication is a clinical decision, and the sequencing-derived data is an input to that conversation. The data does not substitute for it.

Questions people also ask

Does the GeneSight test really work? It reliably genotypes CYP2C19, CYP2D6, and a few other pharmacokinetic genes, and those calls are generally accurate. The claim that its output improves depression outcomes is supported by small effects in randomized trials, with prescribing behavior shifting more than symptoms, and the pharmacodynamic genes on the panel do not support drug selection.⁵¹

Is there a blood test to determine which antidepressant will work best? No. Assays based on patient-derived neurons and other cell-based screens are in development, but none has randomized outcome data supporting clinical use.

Will antidepressants show up in a blood test? Not on a standard chemistry or CBC panel, and not on typical employment drug screens. They are detectable by targeted LC-MS/MS, which is how therapeutic drug monitoring works. Some antidepressants, bupropion and venlafaxine in particular, have produced false-positive amphetamine results on immunoassay screens. Confirmatory mass spectrometry then resolves them.

How long do antidepressants stay in your system? It scales with elimination half-life, which varies widely. Venlafaxine is around 5 hours (plus a longer-lived active metabolite) and sertraline about 26 hours. Fluoxetine is 1 to 4 days, with norfluoxetine at 7 to 15 days. Five half-lives is the usual rule for substantial clearance, so days for most and weeks for fluoxetine. Metabolizer phenotype shifts these numbers for specific drugs.

Will antidepressants affect my other blood tests? Some can. SSRIs are associated with hyponatremia via SIADH, mirtazapine and others can shift lipids and weight-related markers, and several agents affect platelet function. If you are tracking your own biomarkers longitudinally, note the start dates of any medication in the same timeline as the labs, and ask your prescriber what to monitor.

Footnotes

1. Hays Documentation Specialists, LLC, San Mateo, CA, Priya Hays, MS, PhD. Evidence Basis for Pharmacogenetic Testing in Psychiatry. Research Review, 2022. https://doi.org/10.52845/jmrhs/2022-5-3-6 ↩ ↩²

2. 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

3. Robert H. Howland. Pharmacogenetic Testing in Psychiatry: Not (Quite) Ready for Primetime. Journal of Psychosocial Nursing and Mental Health Services, 2014. https://doi.org/10.3928/02793695-20141021-09 ↩

4. Steven L. Dubovsky. The Limitations of Genetic Testing in Psychiatry. Psychotherapy and Psychosomatics, 2016. https://doi.org/10.1159/000443512 ↩

5. David W. Oslin, Sara Chapman, Scott L. Duvall, et al. Study design and implementation of the PRecision Medicine In MEntal health Care (PRIME Care) Trial. Contemporary Clinical Trials, 2021. https://doi.org/10.1016/j.cct.2020.106247 ↩ ↩²

6. Laura M. Hack. Precision Psychiatry: Using Neuroscience Insights to Inform Personally Tailored, Measurement-Based Care. 2021. https://doi.org/10.1176/appi.books.9781615379118 ↩

7. Hulunbuir Mental Health Center, Yakeshi, China, Xiong Zhang, Xiaoping Gu, et al. An Effective Method to Facilitate Personalized and Precise Medicine for Schizophrenia Treatment Based on Pharmacogenomics. Psychiatry and Clinical Psychopharmacology, 2021. https://doi.org/10.5152/pcp.2021.20176 ↩