What is a polygenic risk score, and can your raw DNA file produce one?
Somewhere in the upgrade tier of most DNA interpretation services sits a polygenic score. It is usually the feature that justifies the higher price, and it is the one least often explained.
The concept is genuinely useful and the limits are genuinely serious. Both parts are needed to read one sensibly.
What is a polygenic risk score?
A single number summarising many small genetic effects at once. Most common traits are not driven by one variant but by thousands, each contributing a little. A polygenic score adds those contributions up, weighting each by how strongly a study associated it with the trait, and reports where you sit in a distribution.
The output is almost always a relative position, not an absolute quantity. “Higher than average” and “in the upper part of the distribution” are the honest forms. A score does not say what will happen to you.
Contrast that with a single-gene finding. A pathogenic variant in a specific gene is a discrete fact about one position. A polygenic score is a statistical summary in which no individual marker matters much.
Where do polygenic scores come from?
From genome-wide association studies, published and catalogued publicly. The PGS Catalog, maintained by EMBL-EBI, holds thousands of published scores along with the study that produced each one, the variants it uses, and the populations it was developed and evaluated in. Any service can download them.
That last detail is the important one. Because the scores are open, the score itself is not where a service adds value. The value is in whether the score is appropriate for the person it is applied to, and in how the result is presented.
Two services can compute the same published score from the same file and present it very differently, which is one reason their reports disagree.
Why do polygenic scores work less well for some people?
Because most of the underlying studies were conducted in European-ancestry cohorts. A score’s weights are learned from a particular population, and they transfer imperfectly to people whose genomes differ in which variants are common and in how variants travel together. This is a well documented, unresolved limitation of the field.
The consequence is uncomfortable and rarely stated on product pages: the same score is measurably less informative for people of African, East Asian, South Asian and admixed ancestry. It is not that those genomes are less studied by accident — it is that the research base was built unevenly.
A service reporting a polygenic score without saying which ancestries it was validated in is omitting the single fact you would need to judge it. We do not report polygenic scores for that reason, among others.
Can a consumer DNA file support a polygenic score at all?
Partially. Scores often draw on tens or hundreds of thousands of variants, and a consumer array carries only some of them — so services impute the rest statistically from reference panels. That works reasonably for common variants and adds a layer of estimation on top of an estimate.
Imputation is a legitimate and widely used technique, not a shortcut. But it means a score computed from an array file is not the same object as a score computed from sequencing, even when both carry the same name.
If a report gives you a polygenic score, the questions worth asking are which published score it used, which population it was evaluated in, and how much of it came from imputed rather than measured data.
How should you read a polygenic score if you have one?
As a description of a group you resemble, not a statement about you. Genetics explains a minority of the variation in most common traits; environment, age, behaviour and chance carry the rest. A position in a distribution is one input among several, and rarely the largest.
It is also worth knowing that these scores change. As larger and more diverse studies are published, published scores are revised, and a result computed today may be recalculated differently in two years. That is science working correctly, and it is a poor foundation for a decision that feels permanent.
Anything with a health dimension belongs in a conversation with a clinician who knows your history. A line in a raw DNA file is not a diagnosis, and a number derived from thousands of lines is not one either.
How we turn published research into a report → · Why services disagree about your results →