Gene tests for antidepressants: what the trials show, and what your DNA file already holds

Two kinds of gene on one panel

A drug has two jobs before it can help. It has to reach the brain at a useful concentration, and it has to act on its target once it arrives. Genes can change either step. The first is pharmacokinetics, what the body does to the drug, and for many antidepressants it runs through a few liver enzymes of the cytochrome P450 family. The second is pharmacodynamics, what the drug does to the body, and for these drugs it centres on the serotonin transporter, the protein they block^[1].

The Clinical Pharmacogenetics Implementation Consortium (CPIC), the expert group whose guidelines turn genotypes into prescribing recommendations, reviewed both kinds in 2023 for thirteen antidepressants: six SSRIs, five SNRIs, vilazodone and vortioxetine^[1]. It made recommendations for three enzyme genes. CYP2D6 guides paroxetine, fluvoxamine, venlafaxine and vortioxetine; CYP2C19 guides citalopram, escitalopram and sertraline; CYP2B6 guides sertraline. For fluoxetine and duloxetine it found no basis for a gene-based recommendation^[1]. The older tricyclic antidepressants have a separate CPIC guideline, built on CYP2D6 and CYP2C19^[4].

The brain-target genes went the other way. Variants in SLC6A4 and HTR2A have been studied against response and side effects for years, and some meta-analyses found statistical associations, but CPIC declined to use them.

Its reasons are specific. There is no agreed way to turn an SLC6A4 or HTR2A genotype into a category, laboratories test different variants and so report different results, and the size and even the direction of the associations differed between ancestry groups^[1]. Many commercial panels nonetheless score these genes alongside the enzymes, inside a proprietary algorithm. CPIC describes trials of such combined tests as showing 'collectively positive but modest results', and says their opaque algorithms put them outside what its own process can evaluate^[1].

How large the enzyme effect is: escitalopram in 2,087 patients

The clearest measurement comes from a hospital laboratory in Oslo that monitors antidepressant levels. Researchers took 4,228 escitalopram blood measurements from 2,087 patients whose CYP2C19 had been genotyped, and compared each genotype group with people carrying two typical copies, *1/*1^[6].

The two inactive copies of a poor metabolizer meant 3.3 times the drug level of the typical group, and one inactive copy meant 1.6 times. At the other end, two copies of the increased-function *17 allele meant 20 percent less. Both ends also switched to a different antidepressant more often within a year: 3.3 times as often for the poor metabolizers and 3.0 times for the *17/*17 group^[6]. The study counted switches, not their reasons, but the pattern fits too much drug at one end and too little at the other.

That U shape is why CPIC's table for citalopram and escitalopram has rows at both ends. For poor metabolizers it recommends considering an antidepressant not mainly cleared by CYP2C19, or, if one of these two is used, a lower starting dose, slower titration and half the usual maintenance dose. For ultrarapid metabolizers it recommends considering an alternative^[1]. Our escitalopram and citalopram page restates every row, and the CYP2C19 page covers the gene's other drugs.

What the FDA's table says, and what it leaves out

The US Food and Drug Administration keeps a Table of Pharmacogenetic Associations, last updated on 10 September 2026, in three sections. Section 1 lists associations whose data support therapeutic management recommendations, section 2 those with a potential effect on safety or response, and section 3 those shown to affect a drug's pharmacokinetics only^[7].

Every entry is about how much drug reaches the blood and the risks that follow from it. None involves SLC6A4 or HTR2A, and the FDA attaches a caveat to the whole table: a listing 'does not necessarily mean the FDA advocates using a pharmacogenetic test before prescribing the corresponding medication', unless the test is a companion diagnostic^[7].

Response is a different matter, and the FDA has said so directly. In November 2018 it issued a safety communication about genetic tests that claimed to predict which medicines would work, naming antidepressants as its example.

Most companies it contacted afterwards removed specific drug names from their labelling and patient reports^[5]. Read together, the FDA's two documents draw one line: some gene effects on drug levels are accepted, and an effect on whether the drug works is not.

What the trials found

Three large trials and one meta-analysis carry most of the evidence. Each asks whether handing clinicians a test report improves outcomes for a group of patients, compared with usual care.

GUIDED enrolled 1,167 outpatients whose depression had not lifted on at least one antidepressant, and gave half of their clinicians a test report^[8]. The trial's sponsor was Assurex Health, and the report was GeneSight Psychotropic, which the registration describes as testing 'multiple pharmacokinetic and pharmacodynamic genes'^[9]. Its primary outcome, symptom improvement at week 8, did not differ significantly: 27.2 against 24.4 percent (p = 0.107). Two secondary outcomes did: response, 26.0 against 19.9 percent, and remission, 15.3 against 10.1 percent^[8].

PRIME Care, run by the US Department of Veterans Affairs at 22 medical centres, randomised 1,944 patients who were starting or switching an antidepressant^[2]. The VA funded it. The testing company supplied the genotyping, at no cost, and the report, and had no role in the design, analysis or write-up. Among patients prescribed an antidepressant in the first 30 days, the tested group got one with no predicted drug-gene interaction 59.3 percent of the time, against 25.7 percent in usual care^[2].

Over the whole 24 weeks, remission was higher with the test: odds ratio 1.28, a difference of 2.8 percentage points. The gap peaked at week 12 and was not significant at week 24, which the authors summarised as 'small nonpersistent effects on symptom remission'^[2]. They also note that the test's proprietary algorithm 'may not align' with CPIC's recommendations^[2].

PREPARE asked a different question, in Europe and across all drug classes. Hospitals, clinics and pharmacies in seven countries genotyped 6,944 patients receiving a first prescription for a drug with Dutch pharmacogenetic guidance, testing 50 variants in 12 genes, and treated those with an actionable result by that guidance^[10]. Clinically relevant side effects within 12 weeks fell from 28.6 to 21.5 percent, an odds ratio of 0.70. The trial was open-label, it measured side effects rather than response, and 97.7 percent of its patients reported European, Mediterranean or Middle Eastern ethnicity^[10].

A 2022 meta-analysis pooled 13 trials and 4,767 patients and found remission 1.41 times as likely with test-guided treatment^[3]. Its own risk-of-bias review is the reason to read that number with care. Prescribing clinicians knew who had been tested in every trial, all but two trials were sponsored by industry, and every trial tested CYP2C19 and CYP2D6 together with other genes, such as SLC6A4 and HTR2A, so the analysis could not say which genes drove the effect. Two of its authors declared commercial interests in pharmacogenomic testing^[3].

None of these trials tested whether a report can pick the antidepressant that will work for one person. They tested whether patients whose clinicians had reports did better on average. They did, modestly and not always for long, and the clearest fall in side effects came in the trial that followed published guidance rather than a proprietary score.

How common these results are in Finland

The largest Finnish figures come from SUPER-Finland, a study of people with psychotic disorders, which genotyped 9,262 unrelated participants and inferred CYP2D6 copy number with a reference panel built for the purpose^[11]. For CYP2C19, 3.5 percent were poor metabolizers, 28.7 percent intermediate, 39.7 percent normal, 24.3 percent rapid and 3.8 percent ultrarapid. For CYP2D6, 3.2 percent were poor and 6.6 percent ultrarapid, and 8.5 percent carried a duplicated copy of the gene^[11]. The authors caution that a cohort recruited through psychosis care may not mirror the whole population^[11].

We checked the CYP2C19 figures against gnomAD, the largest public collection of sequenced genomes. In its Finnish genomes the no-function *2 allele has a frequency of 18.65 percent and the increased-function *17 18.49 percent, and *3, a no-function allele common in East Asia, did not appear once in 10,550 Finnish allele copies^{[12, 13, 14]}. If alleles pair at random, those frequencies predict 3.5 percent poor and 3.4 percent ultrarapid metabolizers, within half a point of the cohort. That is our arithmetic and an estimate: it ignores rarer alleles.

Finland sits close to the rest of Europe here. Non-Finnish Europeans in gnomAD carry *2 at 14.80 percent and *17 at 22.00 percent, which by the same arithmetic gives about 2.2 percent poor metabolizers. East Asian genomes carry *2 at 31.19 percent and *3 at 6.10 percent, which puts the estimate near 14 percent^{[12, 13, 14]}.

Across all the genes such panels cover, nearly everyone has something to report: 99.5 percent of 487,409 UK Biobank participants^[15], and 98.8 percent of 930 genotyped patients at a Finnish university hospital, where 23.3 percent had an actionable gene-drug pair among their medicines^[16].

CYP2C19 from a chip export: four of ten could read it

CYP2C19 is the antidepressant gene a consumer chip has the best chance of answering, because its three common alleles each come down to one position: *2 and *3 switch the enzyme off, and *17 increases its function^[1]. Our report calls the gene from a chip only when the array settled all three. If one is missing, the card names what it did read and gives no metabolizer group, unless both copies already carry a variant, which settles the call on its own.

We ran the report's own chip reader over the ten public chip exports measured for our data page^[17].

The three 23andMe exports and the 2024 AncestryDNA export read all three positions. The other six all miss *17, the allele that is about as common in Finland as *2^{[12, 13]}. Two exports read only *3, the one allele no Finnish genome in gnomAD carried^[14]. A chip that cannot see *17 cannot tell a normal metabolizer from a rapid or ultrarapid one, so the report leaves the group open. The measured data page has the rest of what these ten files can read.

Which file reads which gene

Beyond the chip, the kind of file decides what the report can read; our guide to genome file types explains why a VCF, a gVCF and a BAM or CRAM differ.

CYP2D6 is the gene a chip cannot settle. Its variation includes whole-gene deletions, duplications and hybrids with the neighbouring pseudogene CYP2D7^[1], and 8.5 percent of the SUPER-Finland cohort carried a duplication^[11]. A chip reads single positions. In about 50,000 UK Biobank participants, CYP2D6 groups inferred from chip data matched calls that also used exome sequencing for 64.9 percent of people, against 99.4 percent for CYP2C19, with copy number left out altogether^[15]. Researchers have inferred CYP2D6 copy number from Finnish chip data with a purpose-built panel^[11]; our report does not attempt it.

From a BAM or CRAM, the Deep Read panel calls CYP2D6 from the reads, copy number included; the CYP2D6 page explains how. Its CYP2D6 card quotes CPIC's guidance for codeine, tramadol and tamoxifen, and the 2023 guideline's antidepressant rows for paroxetine, fluvoxamine, venlafaxine and vortioxetine^[1]. The Deep Read CYP2C19 card quotes CPIC's rows for citalopram, escitalopram, sertraline and five tricyclics, and the CYP2B6 card, explained on the CYP2B6 page, its sertraline row.

The report uses neither SLC6A4 nor HTR2A for any medicine. The single SLC6A4 position it shows sits among the Frontier findings, presented as a famous depression result that failed to replicate.

What a genotype cannot tell you

A genotype also predicts the enzyme only on average. CPIC lists what can move a person's real metabolism away from it: other medicines, age, diet, illness, smoking, pregnancy and epigenetic differences^[1]. Paroxetine and fluoxetine themselves inhibit CYP2D6, and with paroxetine a normal or intermediate metabolizer can come to behave like an intermediate or poor one. Rare variants that a test does not look for are missed outright, so the real phenotype can differ from the predicted one^[1].

CPIC also states the limit of its own tables. People 'on stable and effective antidepressant medication doses without significant tolerability concerns may not benefit from dose modifications' based on these genes, and a result is 'one of many pieces of clinical information'^[1]. A metabolizer group is a fact about an enzyme. What it means for one person's treatment is a judgement for the clinician who knows the rest of the picture.