Polygenic Risk Scores

A polygenic risk score is a single number that summarizes the inherited component of a person's risk for a common disease by adding up the small effects of many genetic variants. It belongs to the study of genetic variation, sitting alongside the single-variant concepts of the genetic variants and inheritance pattern pages, but it answers a different question: not which variant a person carries, but where their whole genome places them on a population distribution of risk. The concept matters because most common diseases, including the leading causes of age-related death such as heart disease, type 2 diabetes, and several cancers, are highly polygenic, shaped by thousands of variants each contributing a little rather than by one decisive gene. For longevity, this is the genetic architecture that actually governs most late-life disease risk, and it is probabilistic and partly modifiable rather than fixed. The scale of the enterprise is now large: the Polygenic Score Catalog curates thousands of published scores spanning hundreds of traits, and national research programs have begun returning polygenic results to participants. A polygenic score is therefore the practical bridge between genome-wide association discovery and an individual estimate of inherited risk.

Mechanically, a polygenic score is a weighted sum: for each included variant, the number of risk-raising alleles a person carries is multiplied by an effect size estimated in a genome-wide association study, and the products are added across all variants. The effect sizes come from discovery studies that scan hundreds of thousands of people and estimate how strongly each common variant associates with the trait. Because each common variant typically shifts risk only slightly, the predictive power of a score comes from aggregating thousands to millions of them. A central technical challenge is linkage disequilibrium, the correlation between physically close variants, which must be handled so that the same underlying signal is not counted several times. The simplest approach, clumping and thresholding, keeps the strongest variant in each correlated block and includes variants below a chosen significance threshold, while Bayesian methods such as LDpred and PRS-CS instead reweight every variant according to the local correlation structure and the expected distribution of effects. The resulting score is developed in one sample, tuned in a second, and evaluated in an independent third to avoid overfitting, and its performance is expressed as variance explained, as the area under a prediction curve, or as the risk carried by the top percentiles relative to everyone else. The output is a position in a distribution, which becomes a usable risk only after calibration against age, sex, and ancestry.

The single most influential body of evidence for the clinical promise of polygenic scoring is the work of Khera and colleagues published in Nature Genetics in 2018. Validating a 6.6-million-variant genome-wide polygenic score for coronary artery disease in 288,978 participants of the UK Biobank, they showed that 8 percent of the population carried more than threefold the average lifetime risk, a fraction in the high-risk tail comparable in effect size to the rare familial hypercholesterolemia genotype yet affecting roughly 20 times as many people. They repeated the demonstration for atrial fibrillation, type 2 diabetes, inflammatory bowel disease, and breast cancer, each yielding a high-risk tail of similar magnitude. This study reframed polygenic risk from an academic measure of heritability into a potential clinical instrument, because it showed that common-variant risk could reach magnitudes previously associated only with monogenic disease. Independent work reinforced the picture: Inouye and colleagues built a coronary score in roughly 480,000 adults the same year and found a fourfold risk gradient across the score that added information beyond conventional risk factors, while Mavaddat and colleagues engineered a 313-variant breast cancer score in 2019 that placed the top 1 percent at roughly fourfold risk. Together these studies established the size and the shape of the effect that polygenic scores can deliver.

Translating polygenic scores into clinical and longevity practice means treating them as an additional, independent risk factor that reshapes a prior probability rather than as a verdict. The clearest use is reclassifying people whose conventional risk sits near a decision threshold, where a high score can tip the balance toward starting a statin, intensifying screening, or beginning earlier lifestyle change, and a low score can support a more conservative path. Mars and colleagues showed in the FinnGen cohort in 2020 that combining polygenic and clinical risk predicts better than either alone and can move the estimated age of disease onset, which is exactly the kind of decision-relevant shift a score should provide. The dominant failure mode of translation is the ancestry gap: because roughly 80 percent of discovery participants have been of European ancestry, scores predict substantially less well elsewhere, and Martin and colleagues quantified a roughly 4.5-fold loss of accuracy in African-ancestry individuals, so deploying current scores uniformly would widen health disparities. The second failure mode is deterministic interpretation, treating a high score as a diagnosis rather than a probability, which a polygenic score never warrants. Used correctly, with calibration, ancestry awareness, and integration with clinical data, a polygenic score is a tool for stratifying risk and timing prevention across the long arc of healthspan.

Core Concepts

From Single Genes to the Polygenic Model

Most of the genetics covered elsewhere on this site concerns variants with large, often decisive effects: a single change in CFTR causes cystic fibrosis, an expanded repeat in HTT causes Huntington disease. Common diseases behave differently. Heart disease, type 2 diabetes, obesity, and most cancers are not caused by one gene but by the combined action of thousands of common variants, each shifting risk only slightly, layered on top of environment and chance. This is the polygenic model, and its statistical foundation reaches back to Ronald Fisher’s infinitesimal model of 1918, which reconciled the discrete inheritance Mendel described with the smooth, continuous variation seen in traits like height by proposing that such traits arise from many genes of small effect. A polygenic risk score is the modern, data-driven realization of that idea: rather than asking whether a person carries a single causal mutation, it asks where the sum of their many small genetic effects places them in the distribution of inherited risk. The shift from a single-gene to a polygenic frame is the conceptual core of this page, and it changes the kind of answer genetics can give from a yes-or-no diagnosis to a position on a probability gradient. It is why the same person can carry no high-penetrance mutation yet still sit at the high end of inherited risk for a common disease.

GWAS: The Raw Material of a Score

The raw material of a polygenic score is the genome-wide association study, which compares the genomes of large numbers of people with and without a trait to estimate how strongly each common variant associates with it. A typical association study tests millions of single-nucleotide variants across hundreds of thousands of people and reports, for each variant, an effect size and a measure of statistical confidence. Early studies were underpowered and found only a handful of variants, which is why Manolio and colleagues described the missing-heritability problem in 2009: the significant hits explained only a small fraction of the heritability that twin and family studies implied. The resolution, demonstrated by Yang and colleagues in 2010, was that most heritability is carried by many variants whose individual effects are too small to clear the stringent genome-wide significance threshold, but which collectively explain a large share of variance, roughly 45 percent for height in their analysis. This realization is what made polygenic scoring possible, because it justified including thousands of sub-significant variants rather than only the handful that reached significance. The quality of a polygenic score is therefore bounded by the size and the design of the discovery association study behind it, and scores improve as those studies grow. The GWAS Catalog and large biobanks such as the UK Biobank are the public infrastructure that supplies this raw material.

Building the Score: Clumping, Thresholding, and Bayesian Shrinkage

Turning association statistics into a score is not a matter of simply adding up every significant variant, because nearby variants are correlated through linkage disequilibrium and would otherwise count the same underlying signal several times. The oldest and simplest approach, clumping and thresholding, addresses this by keeping the most strongly associated variant within each correlated block and discarding its neighbors, then including variants whose association p-value falls below a chosen threshold that is tuned to maximize prediction. This method is transparent and still widely used, but it discards information by hard-thresholding and by pruning. Bayesian methods improved on it by keeping all variants and instead shrinking their weights according to the local correlation structure and an assumed distribution of true effect sizes. Vilhjálmsson and colleagues introduced LDpred in 2015, which models linkage disequilibrium explicitly and substantially improved accuracy, and Ge and colleagues introduced PRS-CS in 2019, a continuous-shrinkage method that adaptively reweights variants using an external reference panel. These methods matter because the choice of construction method can change a score’s accuracy as much as a moderate increase in discovery sample size. The common thread is that every credible method has to model linkage disequilibrium, because ignoring the correlation between variants is the fastest way to build a misleading score.

Heritability, Variance Explained, and the Missing Heritability Problem

Understanding what a polygenic score can and cannot do requires distinguishing several kinds of heritability. Narrow-sense heritability is the proportion of trait variance attributable to additive genetic effects, estimated historically from twins and families. SNP-based heritability is the proportion attributable specifically to the common variants measured in a genome-wide study, and it is generally lower than the family-based figure. A polygenic score captures only part of even the SNP-based heritability, because finite discovery samples estimate effect sizes imperfectly, so there is a hierarchy: a score explains less than the SNP heritability, which explains less than the total heritability. The gap between SNP heritability and total heritability is part of what Manolio and colleagues called missing heritability in 2009, and it reflects rare variants, structural variants, gene-by-gene and gene-by-environment interactions, and imperfect tagging that common-variant scores do not capture. This hierarchy explains why polygenic scores predict imperfectly even for highly heritable traits, and why they improve steadily as discovery cohorts grow and as rare-variant and multi-ancestry data are added. It also sets a realistic ceiling: a polygenic score will never explain more than the heritability of the trait, and for most diseases it currently explains a good deal less.

From Score to Risk: Percentiles, Liability, and Calibration

A raw polygenic score is just a weighted sum, a number whose units are arbitrary, so it becomes interpretable only by placing it in the distribution of scores across a reference population. This is why polygenic risk is usually reported as a percentile: a person in the 95th percentile has a higher genetic burden than 95 percent of the reference group. Converting that relative position into an absolute risk, the actual probability of developing the disease, requires a further step that combines the percentile with the disease’s overall frequency and with the person’s age, sex, and ancestry, often through a liability-threshold model in which disease appears once combined liability crosses a threshold. Calibration is the discipline of making sure that the predicted absolute risk matches the observed frequency, and a score can rank people correctly yet still be poorly calibrated, overstating or understating everyone’s absolute risk. This distinction matters clinically, because a decision to start a statin or schedule earlier screening depends on absolute risk, not on a percentile alone. A percentile that is not calibrated against the right reference population, particularly the right ancestry, can be badly misleading even when the underlying ranking is sound.

The Polygenic Score Catalog and Reproducibility

As polygenic scores proliferated, the field faced a reproducibility problem: scores were published with incomplete descriptions of which variants and weights they used, making them hard to reproduce or compare. The Polygenic Score Catalog, described by Lambert and colleagues in 2021, addressed this by providing an open repository in which scores are deposited with standardized metadata, including the variants, the weights, the discovery study, and the populations in which the score was evaluated. The catalog now curates thousands of published scores spanning hundreds of traits, and it is paired with reporting standards intended to make a score’s provenance and performance transparent. This infrastructure is what allows a clinician or researcher to ask not just whether a score exists for a disease but how it was built, in whom it was validated, and how well it performs across ancestries. Reproducible, well-documented scores are a precondition for safe clinical use, because a score whose construction and validation cannot be inspected cannot be trusted with a medical decision. The catalog is the practical starting point for anyone evaluating a polygenic score today.

How a Polygenic Score Is Built

Step One: Discovery GWAS and Summary Statistics

Every polygenic score begins with summary statistics from a discovery genome-wide association study: a table listing, for each common variant, the estimated effect on the trait and the confidence in that estimate. The size and composition of this discovery study set the upper limit on how well the resulting score can perform, because effect sizes estimated in small samples are noisy and biased by the winner’s curse, the tendency for the variants that happen to reach significance to have overstated effects. Larger discovery studies produce more accurate effect-size estimates and therefore better scores, which is the main reason polygenic prediction has improved steadily as biobanks have grown into the hundreds of thousands and beyond. Crucially, the ancestry composition of the discovery study is also fixed at this stage, and because roughly 80 percent of association-study participants have been of European ancestry, most available scores inherit a European-centric calibration from the very first step. The discovery study is thus the source of both the predictive power and the central limitation of a polygenic score.

Step Two: Variant Selection and Weighting

The second step converts raw summary statistics into a defined set of variants and weights. This is where linkage disequilibrium is handled, either by clumping and thresholding, which prunes correlated variants and keeps those below a tuned significance cutoff, or by a Bayesian method such as LDpred or PRS-CS, which retains all variants and shrinks their weights according to the correlation structure and the expected distribution of effects. Tuning parameters, such as the p-value threshold or the assumed fraction of variants that are truly associated, are chosen in a separate training sample to maximize prediction without overfitting. The product of this step is the score itself: a specific list of variants, each with a numeric weight, that can be applied to any genotyped individual. Two scores built from the same discovery data but with different construction methods can differ noticeably in accuracy, which is why the method and its parameters are part of a score’s identity and are recorded in the Polygenic Score Catalog. This step is where most of the methodological art of polygenic scoring lives.

Step Three: Computing and Validating the Score

Computing a score for an individual is arithmetic: at each variant in the score, the number of risk-raising alleles the person carries is multiplied by that variant’s weight, and the products are summed across the whole list. The resulting number is then placed in the distribution of scores from a reference population to yield a percentile, and calibrated against age, sex, and ancestry to yield an absolute risk. Validation is the essential check that this score actually predicts the trait, and it must be done in an independent target sample that was not used to build or tune the score, because evaluating a score in its own training data gives a falsely optimistic result. Performance is reported as variance explained on the liability scale, as the area under a receiver-operating curve, or as the risk carried by the top percentiles relative to the rest, the framing Khera and colleagues used when they reported that 8 percent of the population sat above threefold average coronary risk. A score that performs well in one independent cohort of the same ancestry may still perform poorly in a different ancestry, which is why validation across populations is now expected rather than optional.

Why Scores Fail to Transfer Across Ancestries

The most important operational fact about polygenic scores is that they do not transfer cleanly between ancestries, and understanding why is essential to using them responsibly. A score encodes the statistical relationships between measured variants and the trait as they appear in the discovery population, but several of those relationships differ across populations. Linkage-disequilibrium patterns, the correlations between nearby variants, differ, so a variant that tags a causal signal well in one population may tag it poorly in another. Allele frequencies differ, so a variant that is common and informative in one group may be rare and uninformative in another. Effect sizes can themselves differ because of gene-by-environment and gene-by-background interactions. Martin and colleagues quantified the combined result in 2019, finding that European-trained scores predict roughly 4.5-fold less accurately in African-ancestry individuals, with intermediate losses for other groups, and Duncan and colleagues confirmed the same systematic degradation across many traits. Because roughly 80 percent of discovery participants have been of European ancestry, the practical effect is that current scores work best for the already best-studied populations. This is a structural problem with a structural solution: diversifying the discovery cohorts that feed the scores.

Clinical & Longevity Relevance

Coronary Artery Disease and Lipid Management

Coronary artery disease is where polygenic scoring is closest to clinical use, because the disease is common, polygenic, and highly modifiable once high risk is recognized. The work of Khera and colleagues in 2018 and Inouye and colleagues the same year established that a polygenic score can identify a high-risk tail of the population whose inherited coronary risk rivals that of a monogenic disorder, and that this information is largely independent of the conventional risk factors clinicians already measure. The practical value is concentrated in reclassification: a person of intermediate conventional risk who also has a high polygenic score may be moved into a range where guidelines favor starting a statin or controlling blood pressure more aggressively, and a low score can support a more conservative approach. Because high polygenic coronary risk is substantially offset by lipid-lowering therapy and by favorable lifestyle, identifying it early creates a long runway for prevention. The score does not diagnose heart disease; it shifts the probability that the disease will develop, and it does so in a way that prevention can act on.

Breast Cancer and Cancer Screening

Cancer screening is a second domain where polygenic scores are maturing toward clinical use, with breast cancer the most developed example. The 313-variant score of Mavaddat and colleagues placed women in the top 1 percent at roughly fourfold the risk of the middle of the distribution and performed consistently across independent European-ancestry cohorts, a level of stratification that can inform the age at which mammographic screening begins and how often it is repeated. Combined with clinical risk factors and family history, such a score can identify women who might benefit from earlier or more intensive screening or from a risk-reducing discussion, well before any clinical sign appears. The continuous nature of the score means it adds resolution across the whole population rather than only flagging an extreme tail. The major caveat is that the score performs less well in non-European-ancestry women, so applying it equitably is not yet possible without ancestry-appropriate validation, a limitation that applies to cancer screening as forcefully as to any other use.

Type 2 Diabetes and Obesity

Type 2 diabetes and obesity illustrate how polygenic risk connects directly to prevention, because both are highly preventable and both carry strong polygenic signals anchored by well-known variants. The TCF7L2 variant rs7903146 is the strongest common signal for diabetes and the FTO variant rs9939609 the strongest for body mass index, but in each case the single variant is modest and prediction depends on aggregating many. Khera and colleagues showed in 2019 that a body mass index score marks a weight trajectory detectable from early childhood, with the top decile carrying roughly 25-fold higher odds of severe obesity, and Mars and colleagues showed in 2020 that a diabetes score identifies people who cross the diagnostic threshold years earlier. Because the trajectory is established early and the diseases are modifiable, a high score marks a long window in which diet, physical activity, and clinical care interact with inherited predisposition. This is the clearest illustration that polygenic risk and prevention reinforce each other rather than competing.

Integrating Polygenic and Clinical Risk

Across every disease studied, polygenic scores deliver the most value when combined with the clinical information a physician already uses rather than when read alone. Mars and colleagues demonstrated in the FinnGen cohort in 2020 that polygenic and clinical risk are largely independent across coronary disease, type 2 diabetes, atrial fibrillation, breast cancer, and prostate cancer, so combining them predicts better than either alone and can shift the estimated age of onset. A polygenic score therefore behaves like an additional, independent risk factor, comparable in effect to a major conventional one for several diseases, and its independence from clinical factors is precisely what gives it incremental value. The integration is most decision-relevant for people whose conventional risk sits near an action threshold, where the extra information can change management. This additive, probabilistic role, layered onto rather than replacing clinical judgment, is the realistic clinical identity of a polygenic score.

Longevity-Specific Considerations

For a longevity-oriented reader, polygenic risk is the genetic architecture that actually governs most age-related disease, because the leading causes of late-life morbidity and death, including coronary disease, type 2 diabetes, several cancers, and obesity-related conditions, are polygenic rather than monogenic. This matters for two reasons. First, polygenic risk is probabilistic and continuous, so it raises or lowers the odds of late-life disease rather than determining them, and the gap between a high score and the eventual outcome is exactly the space in which decades of prevention can act. Second, the evidence consistently shows that favorable lifestyle and appropriate clinical care substantially offset high polygenic risk for these conditions, so a high score is best read as an early and durable motivation rather than a fate. A polygenic score can also identify high inherited risk long before any clinical sign, lengthening the window for preventive action across the healthspan. The longevity reading of a polygenic score is therefore that inherited risk sets a starting position on a long road, not a destination, and that the road remains open to modification.

Equity and Ancestry Considerations

No application of polygenic scoring can be discussed honestly without confronting its ancestry problem, because the scores in current use were built mostly in European-ancestry data and predict substantially worse elsewhere. Martin and colleagues quantified in 2019 that European-trained scores lose roughly 4.5-fold accuracy in African-ancestry individuals, with intermediate losses for East Asian, South Asian, Hispanic, and admixed populations, and Duncan and colleagues confirmed in 2019 that this degradation is systematic and scales with genetic distance from the discovery cohort. The root cause is that roughly 80 percent of genome-wide association study participants have been of European ancestry, so the linkage-disequilibrium patterns, allele frequencies, and effect sizes the scores encode are calibrated to that population. The consequence is stark: deploying current scores uniformly would deliver the most benefit to the already best-studied populations and the least to under-represented ones, widening rather than narrowing health disparities. This is the single most important caveat to attach to any polygenic result outside European-ancestry populations, and the only durable fix is to recruit the full range of human ancestry into the discovery studies that feed the scores. Until that happens, confidence in any polygenic result should scale with how well the relevant ancestry has been studied.

Limitations and Open Questions

Several limitations constrain how far polygenic scores can be pushed today. Ancestry transferability is the largest, and it is structural rather than incidental. Beyond ancestry, scores capture only the heritability tagged by common variants in the discovery study, so they miss rare high-penetrance variants entirely and explain a fraction of total genetic risk. Individual-level discrimination is modest for most traits, so a score stratifies populations well but predicts any one person’s outcome only weakly. Calibrating a percentile into an accurate absolute risk is technically demanding and population-specific, and a poorly calibrated score can mislead even when its ranking is correct. Population structure, assortative mating, and indirect parental genetic-nurture effects can inflate or confound the associations a score is built on, so part of a measured effect may not be a direct causal genetic effect. And prospective evidence that returning polygenic risk to patients actually changes behavior and improves outcomes is still emerging, so most current data describe predictive accuracy rather than demonstrated clinical benefit. None of these limitations negates the value of polygenic scoring, but each marks a place where the science is still maturing and where honest interpretation must acknowledge uncertainty.

Practical Application

Reading a Polygenic Risk Report

The single most useful skill built on this page is reading a polygenic risk report correctly. A report typically gives a percentile or a relative risk, and the right first questions are what disease the score predicts, what relative risk the reported percentile corresponds to, in what ancestry the score was trained, and whether the person being assessed matches that ancestry. A high percentile should be read as a raised probability, often a twofold to several-fold increase in relative risk, not as a diagnosis, and a low percentile as a reduced but not absent risk, remembering that most disease cases still arise from the large middle of the distribution. A polygenic result is most actionable when it sits near a clinical decision threshold, where it can tip the balance toward earlier screening or preventive therapy, and least useful when it merely confirms an already-clear assessment. Direct-to-consumer polygenic estimates deserve particular caution, because they are often not calibrated to individual absolute risk and rarely convey the ancestry caveat. The recurring error to avoid is collapsing a probability into a certainty in either the alarming or the reassuring direction.

Tools, Databases, and When to Escalate

A small set of public resources operationalizes polygenic risk. The Polygenic Score Catalog is the repository for published scores, recording the variants, weights, discovery study, and validation populations for each, and it is the practical place to see how a given score was built and in whom it was tested. The GWAS Catalog records the underlying variant-trait associations that scores are built from, which is useful for judging the strength and the ancestry composition of the evidence. Large biobanks such as the UK Biobank, FinnGen, and the United States All of Us research program are the cohorts in which most scores are developed and validated, and All of Us has begun returning polygenic results to participants, making questions of calibration and equity concrete. Because all of these resources describe scores and populations rather than the particulars of one person, any clinical, screening, or reproductive decision based on a polygenic score should be escalated to a physician or genetic counselor, who can place the score alongside clinical risk factors, family history, ancestry, and validated guidelines, and who can explain what the score does and does not predict. A polygenic score is a tool for shifting a probability and timing a decision, and it is most valuable in the hands of someone who reads it as exactly that.