Genetic counselling · Perspective
Polygenic scores: what we predict, what we know
The promise of genetic prediction, its clinical value and the meaning of ‘predisposition’.
Polygenic risk scores (PRS) are statistical measures that combine information from many positions in our DNA to help estimate the likelihood of developing a particular disease. Scores have been developed for coronary artery disease, type 2 diabetes, several cancers and other conditions. Their purpose is to identify differences in disease likelihood that routine clinical assessment may not fully capture, early enough to inform prevention. [1]
But what does such an estimate add to clinical decisions, and what does it tell us about the biology of the person it concerns? The first question is about the usefulness of prediction; the second, about what we mean by ‘predisposition’.
How a polygenic score is calculated
In some inherited conditions, a disease-causing change in a single gene has a decisive role. In many common diseases, much of the inherited contribution is spread across numerous genetic differences, known as variants, each usually having a small effect. Both forms of genetic contribution can coexist in the same disease. [1]
A PRS combines these small effects into a single number. Genome-wide association studies (GWAS) compare large groups of people with and without a disease to estimate how genetic differences are associated with it. To calculate the score, the number of copies of each selected variant that a person carries—zero, one or two—is weighted according to its estimated association with the disease, and the weighted contributions are added together. The biological significance of every position is not necessarily known: some act as indirect markers of other DNA changes because they tend to be inherited together. Statistical association can therefore support prediction without fully explaining the underlying mechanism—a distinction to which we will return. [2]
The promise of more targeted prevention
PRS can identify increased genetic susceptibility even in people who do not carry a rare pathogenic variant. Khera and colleagues demonstrated this in 2018: around 8% of the population studied had scores associated with at least three times the risk of coronary artery disease seen in the remainder. This proportion was about twenty times the frequency of carriers of pathogenic variants causing familial hypercholesterolaemia with a comparable increase in risk. [3] The number of people for whom more intensive prevention might be considered was therefore much greater than the number identified by testing for rare variants.
This information can also be obtained early: the genetic positions used in a score can be examined at any age, before symptoms or abnormalities appear in routine tests. [4] Earlier recognition becomes useful, however, when it informs a specific choice: whether and when to start preventive treatment, when screening should begin, or how often it should be repeated. Two applications illustrate these decisions: treatment initiation in cardiovascular prevention, and the timing and frequency of breast cancer screening.
First application: cardiovascular prevention
Decisions about preventive treatment already draw on tools such as SCORE2, which considers age, sex, blood pressure, lipid measurements and smoking. A consensus statement published in the European Heart Journal in 2025 examines polygenic information as a possible addition to this assessment, without recommending routine use. It may be particularly useful when an existing estimate lies close to a treatment threshold. The question is therefore how much a score changes the overall assessment, and what follows from that change. [5]
In a 2024 study, adding polygenic information to SCORE2 moved 9.6% of participants in the intermediate-risk category, where treatment decisions are less clear-cut, into the high-risk category. Those reclassified experienced cardiovascular events at approximately twice the rate of those who remained in the intermediate category. [6] Better discrimination is only the first step, however: it must then be shown to lead to more appropriate treatment and fewer events.
The MI-GENES trial provided an early test of this possibility in people at intermediate risk who were not taking a statin. Participants received either a conventional estimate of coronary heart disease risk or one that also included a genetic score. Among the 203 participants who completed the trial, those given the genetic estimate were more likely to start statins and had lower LDL cholesterol at six months. [7] A follow-up analysis published in 2025 recorded two major cardiovascular events in the genetic-score group and nine in the comparison group over a median of 9.5 years. This was, however, a post hoc analysis of a small sample, with only eleven events in total. The finding is encouraging, but the evidence remains limited. [8]
Second application: breast cancer
For breast cancer, assessment combining several factors is already available as a clinical tool: healthcare professionals can access the BOADICEA model through CanRisk. [9] The version described in 2019 combined a PRS covering 313 genetic positions with family history, pathogenic variants in BRCA1, BRCA2, PALB2, CHEK2 and ATM that disrupt production of the full-length protein, mammographic density and other factors. [10,11] The score thus adds to clinical history and testing for single-gene causes rather than replacing them.
Gao and colleagues examined the additional information provided by a different PRS, covering 105 positions, in carriers of pathogenic variants. For a woman with a pathogenic CHEK2 variant and no first-degree relative with breast cancer, the estimated probability of developing the disease by age 80 was approximately 15% at the 10th PRS percentile and 37% at the 90th. These estimates fall on either side of 20%, the threshold the study used as a reference for enhanced surveillance with breast MRI. About three in ten CHEK2 carriers without an affected first-degree relative fell below it, whereas almost all BRCA1 or BRCA2 carriers remained above it, regardless of PRS. [12] Adding a PRS can therefore refine the estimate even for women carrying pathogenic changes in the same gene. This does not establish that less intensive surveillance is safe for CHEK2 carriers with a low score: the analysis did not test such an intervention.
The randomised WISDOM trial examined whether screening could be tailored safely in a broader population, rather than specifically in carriers. It assigned 28,372 women aged 40–74 to annual mammography or screening informed by clinical factors, breast density, PRS and testing of nine genes. In the latter group, mammography was recommended every two years for average risk, and annually or more frequently for elevated risk. Deferral until age 50, or until the estimated five-year probability reached 1.3%, applied only to women aged 40–49 with a low estimate. Safety was assessed through the rate of stage IIB or higher cancers. After a median follow-up of 5.1 years, tailored screening met the prespecified criterion for non-inferiority to annual mammography on this measure. Fewer mammograms were performed, but no reduction in biopsies was demonstrated. [13] The trial therefore evaluates the whole strategy; it neither isolates the contribution of PRS nor demonstrates a reduction in mortality.
WISDOM adjusted screening schedules without permanently excluding any group. A more radical approach would restrict screening to women with a high PRS. In Huntley and colleagues’ modelling study of UK screening, the 20% of women with the highest scores were estimated to account for approximately 37 in every 100 breast cancer cases—almost twice the share corresponding to the group’s size. Yet restricting screening to this group would leave out most women who would develop the disease. [14]
Four conditions for clinical use
In both settings, the usefulness of a score depends on its place within a broader assessment. Four conditions matter. The first is model reliability. Discrimination describes how well a model distinguishes people who will develop a disease from those who will not. Calibration asks whether predicted probabilities agree with what is observed: among people given an estimate of 20%, do roughly 20 in 100 develop the disease within the specified period? [15] A model can discriminate well while overestimating probabilities. When decisions depend on numerical thresholds, that discrepancy can lead to unnecessarily intensive surveillance or treatment.
The second is suitability for the population in which the score is used. Many PRS were developed predominantly in people of European genetic ancestry. Variant frequencies and patterns of co-inheritance differ between populations, so the associations on which a score relies may differ too, and its predictive performance may not transfer unchanged. Validation in diverse groups is therefore part of clinical development. [16] This also matters within Europe. In Greece, for example, European ancestry alone is insufficient: absolute estimates also depend on how common the disease is locally, and appropriate calibration in the Greek population cannot be assumed.
The third is that a result must be interpreted in context. A low PRS neither rules out disease nor removes the need to investigate a single-gene condition when family history warrants it. [17] Equally, a high score does not automatically determine the appropriate intervention. That decision depends on the expected benefits, possible harms and burden of the intervention, alongside the preferences of the person concerned.
The fourth concerns the organisation of care. Before testing, it should be clear who will explain the result, who will arrange follow-up and which interventions are actually available. Sharing a number is not enough: in the INFORM trial, providing cardiovascular risk information alongside online lifestyle advice did not improve objectively measured physical activity at twelve weeks. [18] This finding concerns a particular intervention and outcome. It does not negate the possibility that genetic information, used within organised care, can influence other decisions, such as starting a statin in MI-GENES.
What, exactly, does the score describe?
Using a number well requires understanding what it represents. Calling a PRS a ‘genetic predisposition’ can suggest that a test has identified a biological property and measured its magnitude. This returns us to our second question: what does the score tell us about the person’s biology? Rémy Furrer and Eric Turkheimer address it in Polygenic risk scores are not genetic predispositions (Nature Human Behaviour, 2026). They argue that the score should not be equated with a property of the organism: it is constructed from population associations, and its ability to predict does not mean that it describes the mechanism through which a particular person might develop a disease. [19]
In the authors’ account, a predisposition must be grounded in an underlying structure and manifest under identifiable conditions. Glass, for example, is fragile because of its structure, and that fragility becomes apparent when it is struck. The pathogenic repeat expansion in HTT that causes Huntington’s disease comes closer to this account: a specific DNA change has a clearer mechanistic connection to disease. A PRS is different. Its value can change when the included positions or their estimated contributions change, as with the 313- and 105-position breast cancer scores discussed above. Even with a fixed score, the probability assigned to it depends on the model, age, family history and disease frequency in the population. An estimate can therefore change without a change in the person. The authors do not deny the biological significance of genetic characteristics; they question whether the score itself is such a characteristic. [19]
This is a substantial objection, but it does not settle the matter. Revising an estimate does not negate biological differences in susceptibility; it shows that our account of them remains incomplete. Nor does the absence of a full mechanistic explanation remove predictive value, as illustrated by the SCORE2 reclassification and the differing estimates for CHEK2 carriers. Our two initial questions remain distinct: how well a score predicts, and what it explains about biology, must be judged separately.
The language of probability
That distinction also matters when discussing results. Furrer and Turkheimer warn that presenting a score as an unchanging internal trait may encourage a belief in a genetically predetermined future. [19] ‘You have a genetic predisposition’, when based solely on a PRS, may therefore suggest more certainty than the evidence allows. The word ‘risk’ deserves similar care. It is an established scientific term, but to the person receiving the result it may chiefly convey a threat. Nor do terms such as ‘rare’ or ‘common’ mean the same probability to everyone. Recommendations for communicating probabilities therefore favour numerical estimates, where available, over exclusively verbal descriptions. [20]
‘Estimated likelihood of developing the disease’ is often clearer, but the accompanying number also needs explaining. The 95th percentile means that a score is higher than those of 95% of a reference population, not that there is a 95% probability of disease. This confusion has been documented among patients and some primary care professionals interpreting example PRS reports. [21] Relative measures alone are also insufficient: ‘three times the risk’ says little without knowing three times what. An absolute probability, with a defined timeframe and an explanation of uncertainty, is clearer—for example, the number expected to develop the disease per 100 people with comparable characteristics. [20] Careful language allows serious information to carry precisely the weight the evidence supports.
This is a central task of genetic counselling: to explain a finding in relation to personal and family history and the available options, and to check what the person making the decision has understood. Predictive information becomes useful when the reasons for an intervention, its expected benefits and the remaining uncertainties are clear, and when statistical prediction is not presented as a complete biological explanation. Genetic prediction can widen the possibilities for prevention. To support a decision, however, it must also leave room for the judgement of the person whose life it concerns.
References
- Lewis CM, Vassos E. Polygenic risk scores: from research tools to clinical instruments. Genome Med. 2020;12:44.
- Choi SW, et al. Tutorial: a guide to performing polygenic risk score analyses. Nat Protoc. 2020;15:2759–2772.
- Khera AV, et al. Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. Nat Genet. 2018;50:1219–1224.
- O’Sullivan JW, et al. Polygenic Risk Scores for Cardiovascular Disease: A Scientific Statement From the American Heart Association. Circulation. 2022;146:e93–e118.
- Schunkert H, et al. Clinical utility and implementation of polygenic risk scores for predicting cardiovascular disease. Eur Heart J. 2025;46:1372–1383.
- Li L, et al. Integration of a polygenic score into guideline-recommended prediction of cardiovascular disease. Eur Heart J. 2024;45:1843–1852.
- Kullo IJ, et al. Incorporating a Genetic Risk Score Into Coronary Heart Disease Risk Estimates: Effect on Low-Density Lipoprotein Cholesterol Levels (the MI-GENES Clinical Trial). Circulation. 2016;133:1181–1188.
- Naderian M, et al. Effect of Disclosing a Polygenic Risk Score for Coronary Heart Disease on Adverse Cardiovascular Events. Circ Genom Precis Med. 2025;18:e004968.
- University of Cambridge. About CanRisk.
- Lee A, et al. BOADICEA: a comprehensive breast cancer risk prediction model incorporating genetic and nongenetic risk factors. Genet Med. 2019;21:1708–1718.
- Mavaddat N, et al. Polygenic Risk Scores for Prediction of Breast Cancer and Breast Cancer Subtypes. Am J Hum Genet. 2019;104:21–34.
- Gao C, et al. Risk of Breast Cancer Among Carriers of Pathogenic Variants in Breast Cancer Predisposition Genes Varies by Polygenic Risk Score. J Clin Oncol. 2021;39:2564–2573.
- Esserman LJ, et al. Risk-Based vs Annual Breast Cancer Screening: The WISDOM Randomized Clinical Trial. JAMA. 2026;335:763–774. Published online 12 December 2025.
- Huntley C, et al. Utility of polygenic risk scores in UK cancer screening: a modelling analysis. Lancet Oncol. 2023;24:658–668.
- Wand H, et al. Improving reporting standards for polygenic scores in risk prediction studies. Nature. 2021;591:211–219.
- Lennon NJ, et al. Selection, optimization and validation of ten chronic disease polygenic risk scores for clinical implementation in diverse US populations. Nat Med. 2024;30:480–487.
- Abu-El-Haija A, et al. The clinical application of polygenic risk scores: A points to consider statement of the ACMG. Genet Med. 2023;25:100803.
- Silarova B, et al. Effect of communicating phenotypic and genetic risk of coronary heart disease alongside web-based lifestyle advice: the INFORM Randomised Controlled Trial. Heart. 2019;105:982–989.
- Furrer RA, Turkheimer EN. Polygenic risk scores are not genetic predispositions. Nat Hum Behav. 2026;10:1177–1178.
- Bonner C, et al. Current Best Practice for Presenting Probabilities in Patient Decision Aids: Fundamental Principles. Med Decis Making. 2021;41:821–833.
- Lewis ACF, et al. Patient and provider perspectives on polygenic risk scores: implications for clinical reporting and utilization. Genome Med. 2022;14:114.