The Genetics Are Real, and Small

A genome-wide study of 27,885 people who had taken these drugs found genuine genetic predictors of both how much weight they lost and how sick the drugs made them.^[1] The associations are real, statistically overwhelming, and located exactly where biology says they should be. The efficacy effect is about three-quarters of a kilogram per copy of the variant — and that gap, between a real finding and a useful test, is the entire subject of this page.

What was found

Researchers surveyed 23andMe customers who reported using a GLP-1 medication, and ran a genome-wide association study on 15,237 people of European ancestry with complete data — a median age of 52 and 82.4% women. Four variants reached genome-wide significance.

These are not marginal findings. A P value of 10⁻²⁸ is not a fluke, and the variants sit in the receptor genes the drugs actually bind — GLP1R for both drugs, GIPR for the one that also targets GIP. The GIPR variant affected only tirzepatide users, which is exactly what you would predict if the biology is what it looks like. Internal coherence of that kind is strong evidence.

Now look at the size of the efficacy effect

Three-quarters of a kilogram, per copy of the variant.

The efficacy association is −0.641% of BMI per allele, which the authors put at roughly 0.76 kg with a confidence interval from 0.34 to 1.27 kg. Someone carrying two copies is looking at something under two kilograms of expected difference, against total losses in these trials that run to fifteen or twenty.

So the honest framing of a test built on this variant is: it can tell you, with real statistical confidence, about somewhat less than a kilogram. That is a discovery about biology. It is not information that changes which drug you take or whether you take one.

The side-effect predictions are closer to useful

Predicting who will be badly nauseated is arguably the more valuable application, because it bears on a decision people actually face — which drug to start, how fast to titrate, whether to persist.

The models reach an area under the curve of 65.4% for nausea and 68.0% for vomiting. That measure runs from 50%, which is a coin flip, to 100%. A rough convention treats 70% as the floor for acceptable discrimination in a clinical tool. Both of these fall below it — better than guessing, not yet good enough to act on.

The GIPR finding is the most interesting of the four for a different reason: it is drug-specific. A variant that protects against vomiting only on tirzepatide is the first thing here that could in principle inform a choice between drugs, which is what genetic testing in this area has always been sold as doing.

What has and has not replicated

The efficacy variant was tested in two independent cohorts. In All of Us, with 4,855 participants and electronic health record data, it held — P = 0.001, and directionally consistent, at −0.47%. In UK Biobank it did not, although the authors note that analysis had low statistical power.

One replication and one failure is a genuinely mixed result, and the low-power caveat is a reasonable explanation rather than an excuse — an underpowered study failing to replicate is weak evidence against. It is still a mixed result, and reporting only the half that worked would be the easy version of this article.

Who ran it, and who it was run in

• The data are self-reported. Participants told a survey how much weight they lost and whether they felt sick. The All of Us replication used medical records, which is a stronger source.
• The GWAS was restricted to European ancestry, as most are for statistical reasons. Whether these variants behave the same in other populations is unaddressed.
• 82.4% of participants were women.
• The study comes from 23andMe, a company that sells genetic testing. That is a material disclosure in an article about whether to buy a genetic test, and it is not a reason to discount a Nature paper with these statistics behind it.

What this means if someone offers to test you

The position has genuinely moved, and it has moved to a more interesting place than either side of the old argument. There are real variants in the right genes with overwhelming statistical support. And the amount of your outcome they explain is small enough that a test would not change a decision.

So the questions to put to anyone selling one are now sharper, not softer: what effect size does this variant predict, in kilograms? What is the discrimination of your side-effect model, as a number? Has the association replicated? If a seller quotes the P value and not the effect size, they are quoting the half of this paper that sounds impressive.

The best-supported predictor of your response remains your response — the first weeks on the drug carry more information than any panel currently on offer, as covered in losing fast early.