DNA diet tests: what happened when trials checked the genotype

What the tests sell

The best picture of this market comes from people who went looking. A 2026 scoping review searched the web as a consumer would, between 2023 and 2025, and found 104 companies selling 204 genetic testing panels with nutrition advice attached^[1]. The mean price was US$234. Only 56 of the companies disclosed which genes they tested, and across those 56 the reviewers counted 3,309 different genes said to be used for diet advice. Weight loss was one of the three most common advice categories, after micronutrients and cardiovascular health.

The five genes reported most often across the panels were FTO, HLA-DQ, MTHFR, PPARG and TCF7L2^[1]. An earlier study looked at what 38 company websites accessible from Finland told buyers before they paid. Eight of the 38 websites had a clearly identifiable section explaining genetics, two offered any consultation before the test, and the authors concluded that the sites did not support a well-informed decision to buy^[3].

Two of these deserve their own pages, and have them. The common MTHFR variants are widespread and we refuse to report them as a clinical finding. HLA-DQ is the exception, and the review itself calls its research base for coeliac disease more substantial than the rest^[1]; our coeliac genes guide covers what a DNA file can and cannot rule out. The rest of this piece is about weight, because that is where the claim was tested head on.

DIETFITS: a trial built around the genotype

DIETFITS was designed to answer a question diet trials kept raising. In earlier trials comparing two diets, the average difference in weight lost was small, while people on the same diet ranged from losing about 25 kg to gaining about 5 kg^[4]. The trial's stated goal was to find out which diet works for which person, and its lead hypothesis was genetic. Three variants, in FABP2 (rs1799883), PPARG (rs1801282) and ADRB2 (rs1042714), were combined into a multi-locus pattern that sorted people into a low-fat genotype, a low-carbohydrate genotype or neither. The preliminary data the design paper cites for that pattern is a 2012 patent application on genetic markers for weight management^[4].

Between 2013 and 2015 the trial enrolled 609 adults aged 18 to 50 with a body mass index of 28 to 40 and no diabetes^[2]. Of them, 244 (40 percent) had the low-fat genotype and 180 (30 percent) the low-carbohydrate genotype. Both groups attended 22 small-group sessions over the year, aimed at the lowest fat or carbohydrate intake each person could keep up. The diets did separate: at 12 months carbohydrates made up 48 percent of energy in the low-fat group and 30 percent in the low-carb group, and fat 29 and 45 percent. 481 people, 79 percent, completed the trial.

The difference between the diets, 0.7 kg, had a 95 percent confidence interval from −0.2 to 1.6 kg, so the trial could not tell the two apart. The genotype result is the one that matters here. If the pattern did what it was designed to do, people with the low-fat genotype would have lost more on the low-fat diet and people with the low-carb genotype more on the low-carb one. The authors found no significant diet-genotype interaction, and concluded that neither of the two factors they tested, the genotype pattern or insulin secretion, helped identify which diet was better for whom^[2].

Food4Me: advice with and without the genotype

DIETFITS asked whether a genotype picks the right diet. Food4Me asked a different question: does genetic information make dietary advice work better? Adults in seven European countries took part online and were randomised to one of four arms^[5]. The control arm got conventional, population-wide dietary advice. The other three got personalised advice built from their own baseline diet; from their diet plus phenotype, meaning body measurements and blood markers; or from diet plus phenotype plus genotype for five variants chosen as diet-responsive.

Of the people enrolled, 1,269 completed the six months. Compared with the control arm, those given personalised advice ate less red meat, less salt and less saturated fat, took in more folate and scored higher on the Healthy Eating Index. Salt intake, for instance, was 0.65 g a day lower than in the control arm^[5]. The authors then compared the personalised arms with one another and found no evidence that adding phenotypic information, or phenotypic plus genotypic information, made the advice more effective.

That result cuts against the marketing. A buyer of a DNA diet plan may well eat better afterwards, since personalisation helped. The trial separated the ingredients, and the improvement came from advice built around what the person already ate; the genotype did not add to it.

FTO: real, small, and no guide to which diet

FTO is one of the five genes diet tests report most often^[1], and its link to weight is not an invention. In 2007 a genome-wide study found a common variant in FTO, rs9939609, associated with body mass index, and replicated it in 13 cohorts with 38,759 participants. The 16 percent of adults with two copies of the risk allele weighed about 3 kg more, and had 1.67 times the odds of obesity, compared with people with no copy^[6]. The allele is about as common in Finland. In gnomAD's genomes, 39.9 percent of Finnish chromosomes carry it, and 821 of 5,272 Finnish genomes have two copies, about one person in six (our arithmetic on gnomAD's counts)^[7].

The question a diet test needs answered is different: do carriers respond differently to a diet? A 2016 meta-analysis pooled individual data from eight randomised weight-loss trials, 9,563 people in all, with diet, exercise or drug interventions^[8]. Carriers of the FTO allele were heavier when the trials began, by 0.89 kg per copy. Once the interventions started, they lost the same amount as everyone else: the extra change per copy was −0.04 kg, with a confidence interval from −0.34 to 0.26 kg. The result held across intervention type, length, ethnicity, sex, age and starting BMI.

Even as a predictor of weight, one variant carries little. FTO had the strongest effect of the 97 variants in a common BMI score, 0.58 kg/m² per allele, and in 33,511 patients of the Mass General Brigham Biobank the whole 97-variant score accounted for 2.9 percent of the variation in BMI^[9]. The remaining 97 percent lies outside the score.

APOA2 and saturated fat

Gene-diet interactions are not imaginary, and the one in APOA2 has been reported in American, Mediterranean and Asian populations^[10]. People with two copies of the C allele of rs5082 (written −265T>C) tend to have a higher BMI, but in the studies that found it, only when they eat a lot of saturated fat. Among 907 older Mediterranean adults at high cardiovascular risk, the CC genotype went with a 6.8 percent higher BMI among those with a high saturated-fat intake and no significant difference among those with a low one; the same pattern appeared in Chinese and Asian Indian participants in Singapore^[10]. The CC genotype was carried by 1 to 15 percent of people, depending on the population.

Those were observational studies, which can show an association but not that changing the diet changes the outcome. The closest thing to a trial test came in 2025, from a secondary analysis of DIETFITS. Researchers kept the participants whose saturated-fat intake met a set criterion at all three time points, 22 g a day or more on the low-carb diet and less on the low-fat one, which left 309 people by our count of the paper's table^[11]. In that subgroup, TT carriers lost more on the low-carb diet at 3, 6 and 12 months, while C carriers did so only at 3 months, and the genotype-by-saturated-fat interaction appeared only at 12 months. Among the 264 participants who did not consistently meet the criterion, genotype made no significant difference.

That is a lead worth a trial of its own. It is a post-randomisation subgroup of one trial, chosen by how people ate after they were assigned, with an effect that faded by 6 months in one genotype group. It does not reverse the trial's main genotype result, and it is a long way from a consumer report that sorts people by APOA2 genotype.

Knowing a genotype changes little about eating

A last defence of the tests is motivational: even if the genotype does not pick the diet, seeing one's own DNA might make people stick to it. A review using Cochrane methods pooled 18 randomised and quasi-randomised studies to test that for several behaviours^[12]. For diet, seven studies with 1,784 people, the effect of receiving a DNA-based risk estimate was a standardised mean difference of 0.12, with a confidence interval from −0.00 to 0.24. The authors concluded that the evidence does not support the expectation that DNA-based risk estimates change behaviour.

There are smaller positive results. In a trial of 138 adults aged 20 to 35, those told they carried a risk version of the ACE gene and advised to limit sodium cut their intake by about 287 mg a day over a year, while the control group's rose by about 130 mg^[13]. The trial also targeted caffeine, vitamin C and added sugars, and its abstract reports a significant 12-month change only for sodium in that group. One nutrient in one subgroup of a small trial is the size of the evidence on that side.

What Aimosti's report reads, file by file

We read some of the same genes the diet tests read, and we label them differently. The report has eight traits filed under metabolism and diet, an FTO card in the Frontier section, which holds exploratory associations, and a 97-variant BMI score among the emerging research scores. None of them comes with a diet, a macronutrient split or a food list. What each file type can support differs, because a chip reads a fixed list of positions, a plain VCF says nothing about positions without a variant, and a gVCF records where the reads matched.

The trait cards are written as tendencies: the lactase card says symptoms, not genotype, are what matter clinically, and the caffeine card, for a slower metaboliser, calls the result a modest tendency in which habit and dose matter more. The FTO card describes the effect as small, and the BMI section says a measurement of the trait says more than its score. Deep Read, from a BAM or CRAM, is a pharmacogenomic panel; it reports CYP1A2 and ALDH2 genotypes for transparency, with no guidance attached.

For background on the two single-gene diet questions with a clear mechanism, see lactase persistence and the coeliac genes guide. For how a polygenic score is built and why it ranks rather than predicts, see polygenic scores in plain words.