Do Polygenic Risk Scores Improve Heart-Disease Prediction? Reading the Evidence and Its Ancestry Gap

A polygenic risk score (PRS) collapses the small effects of thousands of common DNA variants into a single number meant to rank a person's inherited risk of coronary artery disease. The published evidence points in two directions at once: these scores do track who develops heart disease and can sharpen prediction at the margins, yet the versions that dominate the literature were built almost entirely in people of European genetic ancestry and lose much of their accuracy in everyone else. No major cardiology guideline currently endorses PRS for routine risk stratification, and the ancestry gap is a central reason why.

What a polygenic risk score actually measures

Coronary artery disease is highly polygenic. Instead of one broken gene, risk is spread across hundreds of thousands of common variants, each nudging risk up or down by a tiny amount. A PRS weights each of those variants by the effect size estimated in a genome-wide association study (GWAS) and adds them up. The output is a relative ranking: it places you somewhere on a bell curve compared with the population the score was trained on, not a diagnosis and not a fixed probability.

In their 2021 review in Nature Reviews Cardiology, Klarin and Natarajan describe the case for optimism. Across multiple cohorts, CAD polygenic scores associate with incident disease, and post hoc analyses of statin and PCSK9-inhibitor trials suggest that people in the highest genetic-risk strata may show larger absolute benefit from preventive therapy. The authors also note a research use that is easy to overlook: enriching clinical trials for high-genetic-risk participants could shrink the sample size needed to detect a treatment effect, in some estimates by roughly threefold. Those are real signals of biological validity.

What "improves prediction" means, and what it does not

The honest version of the evidence is that PRS adds information on top of standard risk factors, but the added information is incremental rather than transformative. Age, blood pressure, lipids, smoking, and diabetes already capture most of the predictable risk in a middle-aged population. A genetic score earns its place mainly by reclassifying a subset of people, particularly younger adults whose conventional risk factors have not yet declared themselves, and by identifying the tails of the distribution.

The 2023 Nature Medicine study by Aragam and colleagues shows what a state-of-the-art score can do. Their model, GPSMult, was built from GWAS data spanning five ancestries, more than 269,000 CAD cases and over 1.1 million controls, plus ten CAD risk factors. In UK Biobank participants of European ancestry, the score carried an odds ratio of about 2.1 per standard deviation and flagged roughly 20 percent of the population with a threefold increased risk, alongside about 14 percent with a threefold decreased risk. That is a meaningful spread. It is also a population-level statistic. Being in a high-risk stratum raises the odds; it does not mean disease is certain, and a low score does not license ignoring blood pressure or tobacco.

The ancestry gap is the load-bearing limitation

Here the evidence turns cautionary. Because the underlying GWAS data have been overwhelmingly European, the resulting scores predict best in European-ancestry populations and substantially worse in African, East Asian, South Asian, and Hispanic or Latino populations. Klarin and Natarajan put it plainly: PRS predictions are poor in individuals of non-European ancestry and, in some cases, create the potential for racial disparities. Martin and colleagues, writing in Nature Genetics in 2019, quantified the problem and warned that deploying current scores clinically could widen rather than narrow health inequities, because the people who would benefit least are precisely those already underserved by research.

The mechanism is technical but important. Variant frequencies, the strength of statistical linkage between markers, and the true effect sizes all differ across ancestral populations, so a score calibrated in one group is systematically miscalibrated in another. This is not a flaw that more careful clinical use can patch over; it is baked into the training data.

Multi-ancestry scores are the field's answer, and they help. In the GPSMult analysis, the authors attributed their roughly 38 percent overall improvement over a prior score to three ingredients: about 26 percent from a larger CAD GWAS, about 9 percent from incorporating multi-ancestry summary statistics, and about 3 percent from borrowing signal across correlated risk factors. The multi-ancestry contribution is real but partial, and prediction generally remained strongest in the European-ancestry groups that still supply most of the data. Progress is measured in narrowing the gap, not closing it.

Why it is not yet routine

Put together, the evidence supports a measured conclusion. CAD polygenic scores are biologically valid and improving, but they add modest incremental accuracy over established risk factors, their performance is uneven across ancestries, and there is not yet trial evidence that acting on a score changes hard outcomes better than acting on conventional risk assessment. Questions of calibration, standardization across the many competing scores, and equitable performance remain open, which is why guideline bodies have stopped short of recommending PRS for general screening.

This article is educational and is not medical advice. Whether genetic risk information is useful for any individual is a decision to make with a qualified clinician who can weigh it against personal and family history.

For a curious reader, the reasonable stance is neither dismissal nor hype. A PRS is one more piece of information whose value depends heavily on your ancestry, your age, and what you already know about your conventional risk, and its greatest current contribution may be to research design and to exposing how unrepresentative genomic databases still are.