# A polygenic score needs evidence from the people using it

*Draft • Research checked 4 October 2026 • AI-assisted DomDNA editorial content. No independent clinical review or publication is claimed. General education, not individual medical advice.*

A provider describes a polygenic score as “validated in thousands of people”. You still need to know which people, how the researchers selected them and whether they tested the score outside the group that helped build it. A large total can hide a small evaluation group for the population relevant to a proposed use.

Your reader question is: “Show me the validation for this score, this version and the population you intend to use it in.” That request concerns prediction quality. It does not require posting your ancestry percentages or genetic file in a public comment.

## Follow the score into a new group

Researchers combine genetic associations to construct a polygenic score. The [PGS Catalog](https://www.pgscatalog.org/about/) records published scores, their development samples and evaluations of predictive performance. A catalogue record can help you locate a paper and its methods. Inclusion in the catalogue does not certify that a provider’s product suits you.

In their [2019 analysis](https://www.nature.com/articles/s41467-019-11112-0), Duncan and colleagues examined the first decade of polygenic-scoring studies and compared performance across populations. They found that European-derived scores performed worse in several other ancestry groups. Treat this as historical research about the scores and datasets they examined, rather than a verdict on every model available in 2026.

That distinction matters when you read a newer product claim. Ask for current evidence about the named model, not a percentage copied from an older review. An updated score may differ in its training data or method. The provider needs to show the evaluation that supports its claim.

## Ask about more than the headline sample size

Consider a fictional provider that develops a score with 100,000 participants, then tests it in 8,000 new participants. The evaluation includes 7,800 people in one reported group and 200 in another. “Validated in 8,000 people” tells you the total. You need the group-specific results and their uncertainty to assess the second group’s evidence.

These invented counts do not describe a real product. They show a question you can put to a methods table: how many participants contributed to each performance estimate? Avoid treating “included” and “validated with useful precision” as interchangeable descriptions.

Also ask whether the evaluation measures the same outcome the product promises. A score that separates groups with and without a diagnosis may still require further evidence before a provider uses it to assign ten-year probabilities. Save the exact performance metric and its explanation instead of accepting “high accuracy” without a definition.

## Keep ancestry terms attached to their method

[NHGRI defines genetic ancestry](https://www.genome.gov/genetics-glossary/Genetic-Ancestry) through inherited genetic relationships. Researchers use particular methods and labels to describe their samples. Read how the study assigned those labels; do not treat nationality, identity and a research ancestry grouping as identical variables.

A broad label can include people with different genetic backgrounds and lives. In your comparison notes, record the authors’ method, the group sizes and any limits they describe. You can ask a provider how it handles people who do not fit the categories in its evaluation, without trying to assign yourself to a category from appearance.

NHGRI’s [polygenic-risk explainer](https://www.genome.gov/Health/Genomics-and-Medicine/Polygenic-risk-scores) highlights the historic lack of diversity in genomic research. The page dates to 2020. We use it for that established limitation and the probabilistic nature of scores, not as a current inventory of services or clinical adoption.

## Make the provider’s answer reviewable

Request a score identifier, version and evaluation citation in one message. Ask the provider to point to population-specific performance and explain any uncertainty or calibration checks. Keep the reply with the report. “Our science team has considered diversity” leaves you without a result you can inspect.

Do not infer that weak evidence for a group means low health risk for someone in that group. It means the prediction evidence needs attention. A qualified clinician or genetic counsellor can discuss the clinical relevance of a personal report.

Your next step is to request the validation table for one score you are considering. [DomDNA’s educational quiz](https://domdna.com/quiz) offers general lifestyle topics; it does not generate or validate polygenic scores.