What does a risk score measure?
A polygenic risk score (PRS) is a statistical calculation combining information from multiple genetic variants. In NHGRI’s definition, the score does not itself incorporate environmental or other factors. A genetic score should therefore not be read as a summary of a person’s entire health risk. This distinction is a starting point for understanding the question a biobank-derived model actually answers.
Who a biobank includes matters alongside its size
The All of Us team’s 2024 Nature paper introduced a research resource containing 245,388 whole-genome sequences. In that release, 77% of participants came from communities historically underrepresented in biomedical research. Genomic information could be studied alongside health records. These figures describe the data release used in the publication; they are neither the programme’s current participant count nor a measure of diagnostic performance.
New research: a larger dataset is not always the best option
Published on 14 September 2026, Tsuo and colleagues’ study evaluated polygenic scores for 32 traits and diseases using All of Us and UK Biobank data. Diversity improved some predictions, particularly in underrepresented groups. However, combining the two resources was not always optimal: for less polygenic traits, training on All of Us alone performed best for participants of African genetic ancestry. This account is limited to model comparisons reported in the study’s public abstract.
An editorial lesson for AI projects: look beneath the average
PRS is a statistical score; we are not labelling every PRS method as generative AI. The lesson we draw for AI and machine-learning projects is to avoid relying on a single average performance figure. We recommend specifying the target population, data version and test cohort, then examining error and uncertainty by age, sex and relevant genetic-similarity groups. This is our editorial evaluation recommendation, not a claim that the study tested all AI models.
From research performance to clinical use
Better prediction in research data does not by itself show that decisions based on a score improve patient outcomes. Our editorial approach separates model evaluation, clinical-benefit research and product-specific regulatory authorisation. This article does not announce a current product authorisation. Before use in a particular population, we recommend assessing performance in that population and evaluating the intended workflow separately.
A question for the next project meeting
Place “Who was this model evaluated for?” alongside “How many people trained it?” Our recommendation is to record data provenance, an external test cohort, subgroup uncertainty and the model version together. Missing evidence can then remain visible instead of disappearing behind a large total participant count. This educational article provides no individual genetic risk calculation, personal treatment recommendation or evidence of clinical effectiveness for RAI products.
Sources
- NHGRI — Polygenic Risk Score (PRS) ↗
Accessed: 2026-10-27 - All of Us Research Program Genomics Investigators — Genomic data in the All of Us Research Program (Nature, 19 February 2024) ↗
Accessed: 2026-10-27 - Tsuo et al. — All of Us diversity and scale yield context-dependent improvements in polygenic prediction (Nature Genetics, 14 September 2026) ↗
Accessed: 2026-10-27
This article is for education and information. Referencing an organization does not imply its endorsement of RAI Health Pro. This is not a claim of clinical effectiveness for a product or model.