The Neanderthal Gene That Explained Nothing¶
In August a paper landed that was, for me personally, almost too convenient to be believed.
Two changes in the growth hormone receptor gene, inherited from Neanderthals some forty-seven thousand years ago, turn out to be common in South Asian populations and vanishingly rare in European ones. Carrying them is associated with more lean muscle. The finding was published in Current Biology by a team including Svante Pääbo, and it went around the world in a week (Kanis et al., 2026).
I am South Asian. I have been visibly muscular my entire adult life without ever training for it, to the point that I was used as an anatomical model for neck musculature in medical school. And I have had my own raw genotype file sitting in a folder since 2022, downloaded from a consumer testing service and never seriously looked at.
So the question assembled itself in about four seconds, and so did the mistake.
Why the Answer Was Going to Be No Before I Looked¶
Here is the trap, and it is worth naming before any results, because naming it afterwards would be too late to be honest.
Why the paper made the news
A large difference in how common the variant is, which says nothing yet about what it does to any one person.
A genome-wide association study establishes that a statistical relationship exists across a population. It does not, and cannot, establish that the relationship explains anything about the particular human being reading the headline. Those are different claims that happen to sound identical when you are the human being in question and the finding flatters you.
Everything that follows is an attempt to hold that line against a story I badly wanted to be true.
The File, and What I Did with It¶
The raw download from a consumer genetics service is far less intimidating than its reputation. It is a text file of about sixteen megabytes, one variant per line, four columns:
# rsid chromosome position genotype
rs4477212 1 82154 AA
rs3094315 1 752566 AG
rs3131972 1 752721 GG
rs12124819 1 776546 AG
Six hundred thousand rows or so. Completely unmanageable for a person scrolling, and utterly trivial for a targeted lookup, which is the entire reason a project like this is possible at all for someone who is not a geneticist.
I built a panel of around 250 variants and ran it in two directions. Backward from traits I already knew I had, looking for the genetics underneath them. And forward from the genotype, for things that would never announce themselves in a mirror. Of the sixty-one variants that made the final report, fifty-five were recoverable from my chip (the genotyping array the testing service used, which reads a fixed set of positions rather than the whole genome). The other six simply were not on it, and are reported as absent rather than quietly guessed at.
The Rules I Wrote Before I Was Allowed to Look¶
This is the part I would keep if I could keep only one, and every rule was fixed before a single lookup ran.
The deflating fact went first. Before searching for anything, I established the finding that argues against the whole enterprise: archaic ancestry is depleted for the heritability of body composition traits, not enriched. Two independent analyses agree (McArthur et al., 2021; Wei et al., 2023). Putting that on the table first meant any positive result had to survive a prior pointing firmly the other way. It is remarkably hard to do this afterwards, and remarkably easy to skip.
Every claim got a grade. Statistical significance is necessary and nowhere near sufficient, so each line in the report carried one of four tiers:
| Tier | What it means | Examples |
|---|---|---|
| A | Deterministic or guideline-grade. True of this body. | Mendelian genotypes, pharmacogenomics with clinical guidelines |
| B | Robust and material. Replicated across ancestries, effect large enough to notice. | Alcohol metabolism, lactase persistence |
| C | Real but individually near-meaningless. Effect sizes stated in embarrassing plainness. | Most cognition and personality single variants |
| D | Folklore, labeled as such. | The "warrior gene" and its relatives |
Tier C is the one doing the work, and it is the tier most consumer genetics quietly omits. A finding can be entirely real, replicated, statistically unimpeachable, and still tell you absolutely nothing useful about yourself. Giving that its own category, instead of letting it sit next to Tier A looking equally important, is most of the discipline right there.
Unconfirmed stays unconfirmed. Rare pathogenic-looking calls in consumer array data are wrong roughly forty percent of the time (Tandy-Connor et al., 2018), because the probes misbehave precisely where variants are rare (Weedon et al., 2021). So the standing rule was that any alarming rare result gets flagged as a probable artifact until clinically retested, and never reported as a finding. If you take one thing from this article into your own life, take that one.
And some things were refused outright. I wanted to look at cognitive and behavioral traits, because of course I did. Those live in thousands of tiny-effect variants and can only be read through a polygenic score, and published scores are trained overwhelmingly on European-ancestry cohorts (Martin et al., 2019) with well-documented miscalibration when applied to a South Asian genome (Martin et al., 2017; Privé et al., 2022). The answer I got, which I kept because it was better than the question:
A number I could technically compute for you would be pseudo-quantitative garbage, and those websites you used years ago handed people exactly that. I won't.
That refusal was settled before the analysis began, not negotiated in the moment when a tempting number was already on the table. Refusals decided in advance are the only kind that hold.
Why the File Never Left My Machine¶
Years ago I uploaded this same data to third-party interpretation sites, stripping the identifying headers off first, feeling rather clever about it.
That instinct was right and it could not possibly have worked. Genotype data is the identifier. There is no header you can remove that de-identifies six hundred thousand of your own variants, because the variants are the thing that identifies you. I had performed a ritual, not a precaution.
Parsing locally narrows the problem rather than removing it. The file stayed in a folder on my machine, and only the lines each question needed were read. But the assistant doing the reasoning was a frontier model running in the cloud, so those lines did go to its provider. I made that call knowingly: the analysis needed more capability than anything I could run locally, and the data was mine to risk. It is not a choice I would make with anyone else's data, and it is not the only reasonable choice with your own. A model on your own machine keeps every line at home at the cost of weaker reasoning; a frontier model reasons better and sees whatever you show it. If you are weighing whether to upload your own data somewhere, that trade is the whole lesson.
What Was Actually in There¶
The variant the paper used as its marker was not on my chip. The other archaic change on the same inherited segment was, and I carry one copy of it. Because the two travel together on the inherited Neanderthal segment (the introgressed haplotype), one copy of the second is a strong indicator that I carry one copy of the whole thing. Confirming it properly would require sequencing rather than an array.
So: probably a carrier. Which felt, briefly, wonderful.
Then the arithmetic.
The variant, drawn to scale against the thing it was supposed to explain
One copy: 285 g, what I actually carry
The difference from population average this was invoked to explain: several kilograms.
Two hundred and eighty-five grams. Per copy. That is the effect on body weight, almost all of it muscle, drawn from over 1.1 million people, and it is a real and replicated finding.
It is also about the weight of a large apple, set against a physique that differs from the population mean by several kilograms. My one copy accounts for single-digit percent of the thing I was trying to explain. The variant is very probably in my genome and it is very nearly irrelevant to the question I asked it.
The same pattern holds for the other trait it touches, incidentally. It associates with roughly three millimeters of additional height per copy. The paleoanthropologist John Hawks made exactly this point in his commentary on the paper: three millimeters is not much for one person, and folk reasoning about archaic ancestry is a poor way to explain individual traits.
The Correction, Which Was the Most Useful Thing That Happened¶
Somewhere in the middle of all this, discussing an association between this variant and jaw morphology reported in earlier work, the assistant framed things in a way that implied Neanderthal ancestry meant a more retruded jaw.
I pushed back, because that is backwards. Neanderthal mandibles were large, robust, and positioned further forward than ours. What they lacked was a bony chin, which is a modern human novelty, and the popular imagination confuses the chin with the jaw entirely.
It conceded immediately, which is not the interesting part. The interesting part is what the correction then dragged into the light.
The variant's association in living people runs in the opposite direction from actual archaic anatomy. Which is only a paradox if you have been quietly assuming that an inherited archaic allele reproduces the archaic phenotype. It does not. An allele operating on a modern human developmental background produces modern human variation, full stop.
That assumption had been sitting underneath the entire line of reasoning, unstated, doing damage. Challenging one anatomical detail did not merely fix a fact. It exposed a category error that would have propagated through everything downstream, unnoticed, for as long as I cared to keep asking questions.
And I could only catch it because mandibular anatomy happens to be inside my training. One step outside that, and I would have nodded along.
In fairness to the paper, its authors read these findings the other way: they count the shorter mandibular ramus, the higher risk of overbite and the tendency toward shorter tooth roots seen in carriers among Neanderthal characteristics. My point is narrower, and concerns where the jaw sits.
What Transfers¶
- Put the deflating context on the table before you search, not after you find something. Afterwards is too late to be honest with yourself.
- Grade every claim, and make sure your weakest grade says out loud that the finding is individually meaningless. Most consumer genetics omits that tier entirely, which is precisely how it sells.
- Decide what you will refuse to compute before a tempting number exists.
- Genotype data cannot be anonymized by deleting a header. Parse it where it sits.
- Press hard on something inside your own expertise. It is the only place you can referee, and what you learn there tells you how much to trust everything else.
Why Only One Result Appears Here¶
The analysis returned around sixty results, including pharmacogenomic findings and disease risk loci. Those are my medical record and they stay private.
The single variant discussed here is ancestry-informative rather than disease-predictive, it is common (about one in five copies of the gene in South Asia is the Neanderthal version), and nothing adverse can be inferred from it by an insurer, an employer, or anybody else. A genome is also partly information about relatives who consented to none of this, which is the other reason the published slice is one common variant rather than a panel.
Deciding which single result was safe to publish took considerably longer than running the analysis did.
What This Does Not Show¶
One person, one array, no clinical outcome, nothing measured downstream. The variant call is indirect, inferred from a linked marker rather than the one the paper itself used, and unconfirmed by sequencing. This was a physician poking at his own data out of curiosity on a weekend. It is not a service, not a study, and not advice.
Content last reviewed September 3, 2026. Review dates are maintained as described on the About page.