Which Genes Actually Change How You Respond to Cannabis, and Which Ones Are Being Oversold

The best-supported finding in this field concerns CYP2C9. In 43 healthy volunteers given oral THC, people homozygous for the reduced-function CYP2C9*3 variant had a median THC area under the curve about three times higher than people with two normal-function copies, and 70 percent lower exposure to the inactive metabolite. Those carriers also trended toward more sedation.

Almost everything else marketed as cannabis genetics is weaker than that. Variants in FAAH, AKT1 and COMT appear in the literature for defensible reasons, but the psychosis gene-environment findings in particular have a replication record that should make anyone cautious about selling a test built on them.

Two patients take the same 10 mg edible. One feels almost nothing and the other spends four hours on the floor. Clinicians see this constantly, and genetics is a real part of the explanation alongside tolerance, body composition, food, route and product accuracy.

The risk is that a plausible mechanism gets sold as a finished clinical tool. A patient who pays for a cannabis DNA report and then follows its product recommendations is being given precision language over an evidence base that does not currently support precision.

C. Sachse-Seeboth, Jurgen Brockmoller and colleagues in the Department of Clinical Pharmacology at University Medicine Gottingen gave oral THC to 43 healthy volunteers genotyped for CYP2C9 variants and published the results in Clinical Pharmacology and Therapeutics in 2009.

The CYP2C9*2 allele made no difference. The *3 allele did. Median area under the curve for THC was threefold higher in CYP2C9*3/*3 homozygotes than in *1/*1 homozygotes, and area under the curve for the inactive metabolite 11-nor-9-carboxy-THC was 70 percent lower, which is exactly the pattern you expect when the enzyme that converts one to the other is impaired. Carriers of *3 also showed a trend toward increased sedation after THC.

This matters clinically because it is a genuine exposure difference, not a subjective one. A person carrying two reduced-function copies takes the same milligram dose and achieves roughly triple the systemic THC exposure. It is the single most defensible pharmacogenetic statement anyone can make about cannabis. It also comes from one study of 43 people, containing a small number of *3/*3 homozygotes, and it has not been replicated at scale.

Rongrong Jiang and colleagues at Hokuriku University characterized cannabidiol metabolism in human liver microsomes, publishing in Life Sciences in 2011. They identified eight monohydroxylated metabolites and tested 14 recombinant human cytochrome P450 enzymes, seven of which metabolized cannabidiol. Correlation and inhibition studies pointed to CYP3A4 and CYP2C19 as the major contributors, with 7-hydroxylation predominantly catalyzed by CYP2C19.

That specific step matters because 7-hydroxy-cannabidiol is pharmacologically active and circulates alongside the parent compound. If CYP2C19 activity varies between people, then the ratio of parent cannabidiol to its active metabolite should vary too.

Here is where honesty is required. The mechanism is clear and the clinical quantification is not. The pharmaceutical cannabidiol interaction trials have genotyped CYP2C19 in participants, and CYP2C19 status is a recognized variable in that setting, but there is no published dose-adjustment rule for cannabidiol based on CYP2C19 genotype and no trial showing that adjusting by genotype improves outcomes. Anyone presenting one is extrapolating.

The FAAH C385A variant, formally rs324420, is the most interesting entry in this literature and the most frequently misapplied. Iva Dincheva, Francis Lee and colleagues at Weill Cornell built a knock-in mouse carrying the human mutation and compared it directly with human carriers, publishing in Nature Communications in 2015.

Both species showed the same thing. Reduced FAAH expression means less breakdown of anandamide, meaning higher endocannabinoid tone. In mouse and human, carriers showed enhanced fronto-amygdala connectivity, better fear extinction learning and decreased anxiety-like behavior. That is a well-executed cross-species result about the body’s own cannabinoid signaling.

It is not a result about response to cannabis. FAAH degrades anandamide; it does not metabolize THC or cannabidiol. A person with the variant has a differently tuned endogenous system, which is a reasonable hypothesis generator for how they might respond to an exogenous cannabinoid, and a hypothesis is what it remains. The closest direct evidence comes from a 2008 study in Addiction of 105 daily cannabis-using students at the University of Colorado, in which the FAAH variant interacted with five days of abstinence to affect craving, and a CNR1 variant interacted with abstinence to affect withdrawal. Small sample, specific population, no replication at scale.

This is the part of cannabis genetics that reaches the public, and it deserves the most careful reading of anything on this page.

Marta Di Forti and Robin Murray’s group at the Institute of Psychiatry in London studied 489 first-episode psychosis patients and 278 controls, publishing in Biological Psychiatry in 2012. The AKT1 rs2494732 locus by itself was not associated with psychotic disorder, with lifetime cannabis use, or with frequency of use. The interaction was significant: among people with a history of cannabis use, C/C carriers had roughly twice the odds of psychotic disorder compared with T/T carriers, odds ratio 2.18 with a confidence interval of 1.12 to 4.31. Among daily users the point estimate was 7.23, with a confidence interval running from 1.37 to 38.12. A confidence interval that wide is telling you the estimate rests on very few people.

The COMT story went further and then reversed. Avshalom Caspi and colleagues reported in 2005 that carriers of the COMT valine158 allele who used cannabis in adolescence were more likely to develop psychotic symptoms, with no such effect in methionine homozygotes. Stanley Zammit and colleagues at Cardiff found no such interaction in 2007 in the British Journal of Psychiatry. A 2018 meta-analysis in PLoS One pooled 13 studies and found a significant interaction only in the case-only studies, which have lower validity and tend to overestimate effects, and which identified the methionine genotype rather than the valine genotype as the risk allele. In studies with dichotomous or continuous psychosis outcomes there was no evidence of interaction at all. The authors concluded that evidence for the interaction remains unconvincing.

Candidate-gene studies pick one variant and look. Genome-wide studies look everywhere, and they are the correct test of whether single variants carry meaningful predictive weight.

Emma Johnson and colleagues published a genome-wide association meta-analysis of cannabis use disorder in The Lancet Psychiatry in 2020, combining 20,916 cases and 363,116 controls. Twin and family heritability of liability is estimated at 50 to 70 percent. The study found two genome-wide significant loci: one in FOXP2 with an odds ratio of 1.11, and a previously identified chromosome 8 locus near CHRNA2 and EPHX2 with an odds ratio of 0.89.

Daniel Levey and colleagues went larger still in Nature Genetics in 2023, meta-analyzing more than one million individuals including 64,314 cases across four ancestry groupings, and identified 22 loci in the European-ancestry analysis. The pattern is unambiguous: liability is substantially heritable and extremely polygenic, with individual variants carrying odds ratios close to one. A report that tells you a single polymorphism determines your cannabis experience is not consistent with what the genome-wide data show.

The market is well ahead of the science here, and the gap is not subtle. Reports that recommend specific strains, ratios or product categories based on a saliva sample are making claims that no published study supports.

The specific things that have not been demonstrated are worth listing plainly. No randomized trial has compared genotype-guided cannabis dosing with ordinary careful titration. No genotype-based dosing rule for THC or cannabidiol has entered standard clinical practice. No test has been shown to predict who will have a bad experience, who will benefit for pain, or which cannabinoid ratio suits a given person.

What is genuinely worth knowing is narrower and more useful. If you carry two reduced-function CYP2C9*3 alleles, oral THC will likely produce substantially higher exposure than the label implies, which is a reason to start lower rather than a reason to buy a different product. And beyond genetics entirely, the variables that reliably change outcomes remain dose, route, timing, food and product accuracy, all of which a patient can control without a DNA test.

This evidence base is layered and the layers differ enormously in quality. The CYP2C9 pharmacokinetic finding is a measured exposure difference with a clear mechanism and a plausible clinical consequence, from a controlled study with genotyped participants. Treat it as probably true and under-replicated.

The enzymology of cannabidiol metabolism is solid in vitro and undertested in people. The FAAH work is high quality and answers a question about endogenous signaling rather than about cannabis. The psychosis gene-environment findings are the weakest layer, resting on modest samples, wide confidence intervals, and a meta-analysis that found the effect only in the least reliable study design.

Candidate-gene research has a well-documented history in psychiatry of producing striking initial findings that do not survive replication. The COMT and cannabis literature is a textbook example: an influential 2005 result, a failed replication in 2007, and a 2018 meta-analysis that found the association only in case-only designs and with the opposite risk allele from the original report.

Sample sizes here are small by modern genetic standards. The CYP2C9 study enrolled 43 people. The FAAH and CNR1 withdrawal study enrolled 105 undergraduates aged 18 to 25 at a single university. These are useful studies, not definitive ones, and they are routinely cited in consumer materials as though they were.

No study shows that genetic testing improves cannabis outcomes. There is no published randomized comparison of genotype-guided dosing versus standard careful titration, and no evidence that any commercially available cannabis DNA report helps a patient choose a product, a ratio or a dose.

The genetics also do not show that anyone’s risk of cannabis-related psychosis can be read off a single variant. Even taking the AKT1 finding at face value, it describes an interaction in a population, not a prediction for an individual, and the genome-wide data indicate that liability is spread across many variants of very small individual effect.

Pharmacogenomics works well where the pharmacology is simple: a single drug, a single dominant metabolic route, a measurable concentration and a defined therapeutic window. Cannabis fails on all four counts. It delivers dozens of active compounds through routes with wildly different bioavailability, in products whose labeled content is frequently inaccurate, toward outcomes measured largely by self-report.

That is why the field has produced one clean pharmacokinetic finding and a long tail of associations. The obstacle is not that the genetics are uninteresting. It is that genetic variance is a small term in an equation dominated by dose, route, product accuracy and tolerance.

I see these reports in clinic. A patient brings in a printout that says their genotype means they should avoid certain products or that they metabolize cannabinoids slowly, and it reads with all the confidence of a lab result. When I ask what study supports the recommendation, there usually is not one.

The one piece of genetics I do find clinically useful is the CYP2C9 story, because it maps onto something I observe: a subset of people for whom an ordinary oral dose behaves like a much larger one. But I do not need a genetic test to manage that patient. I need to start them at 2.5 mg instead of 10 and watch what happens. Careful titration finds the same answer, costs nothing, and works for the many causes of variability that genetics does not explain.

CYP2C9*3 homozygosity substantially raises oral THC exposure and is the one well-documented pharmacogenetic effect in cannabis. CYP2C19 shapes cannabidiol metabolism with no validated dosing rule attached. FAAH, AKT1 and COMT findings are either about endogenous signaling rather than cannabis, or have replication problems serious enough to disqualify them from clinical use. Start low and titrate slowly, which handles genetic variation and everything else at once.

Sort any claim in this area into one of three bins before acting on it. Measured drug exposure differences, meaning CYP2C9 and THC. Plausible mechanism without clinical quantification, meaning CYP2C19 and cannabidiol. And population-level statistical associations with unresolved replication, meaning the psychosis genes. Only the first bin belongs anywhere near a dosing decision, and even there the practical response is the same one you would take without the test.