Learn Explainer
Heart disease polygenic scores: what the studies found, and what one number can carry
Coronary artery disease is a disease of many genes: hundreds of common variants each nudge the odds a little. Scores built from them separate people at the tails, and FinnGen showed the effect in Finns. Added to a doctor's usual risk equation, they have moved its accuracy by about two hundredths, or not at all.
Key takeaways
- A 6.6-million-variant score put 8 percent of UK Biobank participants at three or more times the risk of everyone else, about 20 times as many people as carry a familial hypercholesterolaemia mutation of similar effect[2].
- In 135,300 Finns in FinnGen, people in the top 2.5 percent of a score were diagnosed with coronary heart disease 4.35 years earlier on average than those in the middle, and an estimated 63.9 percent had it by age 80 against 37.2 percent[5].
- Added to the pooled cohort equations, a score raised the C statistic from 0.76 to 0.78 in 352,660 UK Biobank participants[3]. In a Finnish cohort the same kind of addition gave no gain at all[5].
- Among people in the top fifth of a genetic score, those with at least three of four healthy habits had a 46 percent lower relative risk of coronary events than those with one or none, in an observational analysis[13].
- Scores built mostly from European data predict less well in people of African ancestry: an odds ratio of 1.25 per standard deviation against 1.72 in Europeans for one of the newest[15].
Coronary artery disease is the narrowing of the arteries that feed the heart muscle. A 2022 genetic study of 1,165,690 participants, 181,522 of whom had the disease, found 241 separate places in the genome linked to it[1]. A polygenic score adds up a person's copies of the higher-risk versions at many such places. In 2018 one built from 6.6 million variants placed 8 percent of UK Biobank participants at three or more times the risk of the rest[2]. What that buys in practice is smaller than the headline: in one large study, adding a score to age, cholesterol, blood pressure and smoking raised the standard accuracy measure from 0.76 to 0.78[3].
Hundreds of small pushes towards one disease
Familial hypercholesterolaemia is the best-known single-gene route to heart disease: one faulty copy of a cholesterol gene and very high LDL cholesterol from childhood. It is not where most inherited risk comes from. Khera and colleagues took it as present in about 0.4 percent of people and as raising coronary risk up to three-fold[2]. Their polygenic score found a group twenty times larger at a comparable or greater risk, simply by adding up common variants.
8.0%
of 288,978 UK Biobank participants at three-fold or greater risk of coronary artery disease[2]
20×
as many people as carry a familial hypercholesterolaemia mutation of similar effect[2]
12%
of future cases caught by a typical coronary score when 5% of people who stay well are flagged[4]
The 8 percent and the 12 percent look at the same kind of score from opposite ends. The 12 percent asks how many of the people who go on to develop the disease a score would catch, if the cut-off were set so that only 1 in 20 people who stay well are flagged. Across 27 published coronary scores the median answer was 12 percent[4]. A typical score gave an odds ratio of about 13 between the top and bottom 1 percent, which sounds dramatic and still left 88 percent of cases outside the flagged group, because most cases come from the crowded middle of the curve. Our guide to what a percentile means walks through that arithmetic; familial hypercholesterolaemia has its own article.
Finland as a test bench
In 2020 Mars and colleagues took 135,300 FinnGen participants, roughly 3 percent of Finnish adults at the time, and linked them through their personal identity codes to the national hospital discharge register, which runs back to 1968, and the death register, which runs back to 1969[5]. That gave up to 46 years of follow-up, and 20,179 of the participants had coronary heart disease. FinnGen has since grown past 500,000 people, almost a tenth of the population[6].
Their coronary score had 6,412,950 variants, weighted from UK Biobank data and tuned with sequenced Finnish genomes as the reference for which variants travel together[5]. Against people between the 20th and 80th percentile, the top 2.5 percent had about twice the rate of coronary heart disease, a hazard ratio of 2.03, and the bottom 2.5 percent about 0.61 times the rate.
63.9%
with coronary heart disease by age 80 in the top 2.5% of the score, against 37.2% in the middle[5]
4.35 years
earlier diagnosis, on average, for the top 2.5% than for the middle of the distribution[5]
Those are group averages, and part of FinnGen was recruited through hospital biobanks and disease cohorts, which the authors note may overstate risk somewhat[5]. The paper also reports a less flattering result. In the population-based FINRISK cohort, which recorded blood pressure, cholesterol and smoking, the score did not improve the ten-year prediction of coronary heart disease over the clinical equation at all: a C-index of 0.823 for the clinical model and 0.820 with the score added. It did better at one task, reclassifying people whose disease began before 55, where it picked up 13 percent of early cases that the clinical equation had missed.
A later study by the same group pooled six biobanks and found that a coronary score built the same way, from a GWAS without UK Biobank data, worked almost identically in FinnGen and in other European biobanks, an odds ratio of 1.53 per standard deviation in Finland against 1.54 for the European groups combined. Within Finland, people from the early-settled south and west and the later-settled east and north showed essentially no difference[7].
The finding that scores transfer well to Finns does not mean Finns score the same. Several of the score's variants are noticeably more or less common in Finland, a product of the same population history that left the Finnish disease heritage. We computed the effect-allele frequencies for the score's variants from gnomAD's Finnish and non-Finnish European genomes.
| Variant and gene | Higher-risk letter | Weight in the score | Frequency, non-Finnish European | Frequency, Finnish |
|---|---|---|---|---|
| rs10455872, LPA | G | 0.304 | 6.9% | 3.5% |
| rs2891168, CDKN2B-AS1, 9p21 | G | 0.174 | 49.6% | 43.3% |
| rs429358, APOE, the ε4 letter | C | 0.091 | 13.8% | 19.5% |
| rs7177201, ADAMTS7 | T | 0.075 | 24.4% | 37.4% |
| rs11591147, PCSK9 | G | 0.251 | 98.4% | 95.6% |
Source: Our calculation from gnomAD v4 genome allele counts[8], with the weights and higher-risk letters of PGS003438 as the report holds them[9]. Gene names from Ensembl.
The differences pull in both directions. The LPA variant that tags high lipoprotein(a) is half as common in Finnish genomes, and the letter at PCSK9 that lowers the score, T, is nearly three times as common, 4.4 percent against 1.6. The APOE ε4 letter and the ADAMTS7 variant run the other way. Summed over the 223 variants placeable in every gnomAD group, the expected Finnish total sits about 0.13 standard deviations above the non-Finnish European one, by our estimate: a file at the 90th percentile of the European reference would rank about 88th among Finns.
What a score adds to a clinical risk equation
A doctor estimating ten-year heart risk usually works from an equation. The one most studies test against is the American pooled cohort equations, built from age, sex, cholesterol, blood pressure, diabetes and smoking. The usual yardstick is the C statistic: pick one person who went on to have a coronary event and one who did not, and it is the chance the model ranked the first one higher. 0.5 is a coin toss.
| Study | People and events | Score alone | Conventional model | Conventional plus score |
|---|---|---|---|---|
| Elliott 2020, UK Biobank | 352,660; 6,272 events | 0.61 | 0.76 (pooled cohort equations) | 0.78 |
| Inouye 2018, UK Biobank | 482,629; 22,242 cases | 0.623 | 0.670 (six factors, including self-reported high cholesterol) | 0.696 |
| Mars 2020, FINRISK | 20,165; 1,209 events in ten years | Not reported alone | 0.823 (pooled cohort equations) | 0.820 |
Source: Elliott et al. 2020[3]; Inouye et al. 2018[10]; Mars et al. 2020[5]. C statistic or C-index; each study used its own score and follow-up.
In Elliott's study the gain was 0.02, with a confidence interval of 0.01 to 0.03, and at a 7.5 percent ten-year risk threshold the score moved a net 4.4 percent of the people who later had an event into the higher group, and a net 0.4 percent of those who stayed well in the wrong direction[3]. The authors called the improvement statistically significant, yet modest. In 2022 the American Heart Association summarised the field in similar terms: gains of a few hundredths in the C statistic in large cohorts, larger in younger subgroups, and the judgement that such scores were beginning to enter clinical practice and might be considered in select scenarios[11].
Age is where the score earns more. In 330,201 UK Biobank participants, the 241-variant score in our report was associated with a 72 percent higher rate of heart attack per standard deviation in people under 50, against 42 percent over 60, and the age pattern was repeated in Biobank Japan[12].
Genetic risk and lifestyle in the same people
In 2016 Khera and colleagues studied 55,685 people in three prospective cohorts and one cross-sectional study[13]. They split a score of up to 50 variants into fifths and scored lifestyle on four factors: not smoking, not being obese, physical activity at least once a week and a healthier diet. People in the top fifth of the score had 1.91 times the coronary event rate of the bottom fifth. Within that top fifth, people with at least three of the four factors had a 46 percent lower relative risk of coronary events than people with one or none.
The same relative difference showed up at every level of genetic risk: 45 percent lower in the lowest fifth, 47 percent in the middle[13]. Lifestyle was observed, not assigned, and the authors write that its association with coronary events cannot be taken as causal. Its score was small by today's standards, and most participants were white.
Accuracy outside European ancestry
The study behind our score's weights was predominantly European[1]; the one behind the score tested across six biobanks below was 77 percent European[7]. When the eMERGE network tested Inouye's metaGRS in three groups, its hazard ratio per standard deviation was 1.53 in European-ancestry participants, 1.53 in Hispanic participants and 1.27 in African-ancestry participants[14]. Across six biobanks, Mars and colleagues found accuracy similar in European and Asian groups and much lower in the smaller groups of African ancestry, where one estimate, 1.10, could not be told apart from no association[7].
Patel and colleagues built GPSMult with genetic data from five ancestries specifically to close the gap, and it outperformed every earlier coronary score they compared it with, in every group[15]. The gap remained. They put part of it down to how few people of African ancestry the underlying genetic studies have included.
The score inside our report
The report's coronary card uses PGS003438: the 241 genome-wide significant variants of the 2022 study of 1,165,690 people, weighted by that study's effect sizes[1, 9]. Every weight is positive, from 0.027 to 0.517, so each copy of a higher-risk letter adds to the total. It is a far smaller score than the multi-million-variant ones above, and the report reads its positions directly from a file without imputing the rest.
Not every variant pulls equal weight: a rare variant barely changes the spread of totals because few people carry it. Measured as each variant's share of that spread in a European population, ten of the 241 carry about a quarter of it.
Two variants in the 9p21 region, inside CDKN2B-AS1, are among the top five. The LPA region on chromosome 6 is the other heavy block. Ten of the score's variants lie in the megabase of chromosome 6 around LPA, four of them inside the gene, and together they carry about 13 percent of the spread, by our estimate. One of them, rs10455872, is also one of the two letters the report's separate lipoprotein(a) card reads, so a file carrying it moves both cards. Two APOE letters, rs7412 and rs429358, are in the score too; our APOE article covers what else they are known for.
What the card shows, file by file
The card's heading is the percentile, written as an ordinal such as 63rd percentile, with the trait and the score's name: the 241-variant primary-prevention score of Marston and colleagues. Under it the card gives the percentile against the European reference, the number of the 241 variants used, and a line on which way the number runs: a higher percentile is a stronger genetic tendency, roughly that many in 100 of the reference group scored lower, and it is a rank inside a group, not a probability of the condition. The reference itself is modelled, from the frequencies of each higher-risk letter in gnomAD's non-Finnish European genomes, with an average total of 12.37[8]; on a chip it is rebuilt over the variants the chip read.
| File | What counts as covered | What the card shows |
|---|---|---|
| Chip export (23andMe, AncestryDNA, MyHeritage and others) | Score positions the array actually read, including two reference copies | 90% or more of the 241 read: the percentile to one decimal. 50 to 89%: an estimate as a whole number with a tilde. Under 50%: Examined, not reportable, with the count |
| Plain VCF from whole-genome sequencing | Cannot be counted: a missing line may be a match to the reference or an unread position | Always an estimate, with how many of the 241 were seen in the file and a note that the rest were treated as the common genotype |
| Plain VCF from an exome or gene panel | Recognised as a targeted capture | Not reportable: most of the score lies outside the sequenced regions |
| gVCF | Positions in a reference block or a call with at least 10 reads | The same three bands as a chip. At least 20 covered positions with no variant call at all is read as an unreadable file, with no percentile |
| BAM or CRAM (Deep Read) | Not used for the score | No change: the card comes from the report's VCF or gVCF |
Source: Aimosti's scoring engine and report card as of this page's review date; the depth floor of 10 reads is the one on the data page[16].
Chips are where the bands matter. The ten exports we measured read between 281 and 495 of the 1,457 positions behind all of the report's disease scores[16], and the report does not impute the rest, so on some chips the coronary card will say how many of the 241 were found instead of giving a number. We have not published the coronary score's own count per chip. Polygenic scores in plain words explains why a percentile from a third of a score, with the rest guessed, would look like a result without being one.
Some things the card leaves out by design. It gives no absolute or lifetime risk, no colour scale and no high-risk label at any percentile: every polygenic card carries the same neutral marker and the tag Tier 2, well-supported. It has no Finnish reference curve, so a Finnish file is ranked against non-Finnish Europeans, a difference of a few percentile points by the estimate above. It does not combine the score with the Lp(a) card, with a familial hypercholesterolaemia finding or with any measurement. The sample report shows the card on a synthetic genome.
What Aimosti would (and wouldn't) show you
With the reader's consent to sensitive findings, the report carries a coronary artery disease card built from PGS003438, the 241-variant score of Marston and colleagues. It gives a percentile against a modelled European-ancestry reference, says how many of the 241 variants the file covered, and shows no number when a chip read fewer than half of them.
What we won't claim
We won't turn the percentile into a chance of a heart attack, combine it with cholesterol, blood pressure or family history, or call anyone high-risk on the strength of it. We don't fill in variants the file never read, and we don't present a European-reference percentile as equally accurate for every ancestry.
Bottom line. Coronary artery disease scores pick out groups at the tails of the distribution whose risk does differ, more so in younger adults, while most people and most cases sit in the middle. Added to a clinical equation that already knows age, cholesterol and blood pressure, they have changed its accuracy by a couple of hundredths or not at all.
Questions people ask
Is a polygenic risk score for heart disease the same as a genetic test for familial hypercholesterolaemia?
No. A familial hypercholesterolaemia test looks for one rare variant with a large effect on LDL cholesterol. A polygenic score adds up hundreds or millions of common variants with small effects. In UK Biobank a score found about 20 times as many people at comparable risk as familial hypercholesterolaemia mutations do[2].
Does a high heart disease polygenic score mean a heart attack is likely?
Not by itself. In FinnGen, an estimated 63.9 percent of people in the top 2.5 percent of a score had coronary heart disease by age 80, against 37.2 percent in the middle of the distribution[5], so roughly a third of the top group had not. A typical coronary score catches 12 percent of future cases when 5 percent of people who stay well are flagged[4].
How much does a polygenic score add to a standard cardiovascular risk calculator?
A little. In 352,660 UK Biobank participants the C statistic rose from 0.76 for the pooled cohort equations to 0.78 with the score[3]. In the Finnish FINRISK cohort it did not rise at all, 0.823 against 0.820[5]. Its link with heart attack was stronger in people under 50 than over 60[12].
Which coronary artery disease score does Aimosti use?
PGS003438 from the PGS Catalog: the 241 variants a 2022 genetic study of 1,165,690 people linked to the disease, published by Marston and colleagues in 2023[9, 12]. The percentile is set against a modelled European-ancestry reference.
I am Finnish. Does the score work for me?
Coronary scores built mostly from other European data worked about as well in FinnGen as in other European biobanks, an odds ratio of 1.53 per standard deviation against 1.54[7]. The report ranks Finnish files against a non-Finnish European reference, which by our estimate sits about 0.13 standard deviations lower on this score.
References
- Aragam KG, Jiang T, Goel A, et al. Discovery and systematic characterization of risk variants and genes for coronary artery disease in over a million participants. Nature Genetics, 2022. doi:10.1038/s41588-022-01233-6 181,522 cases among 1,165,690 participants of predominantly European ancestry; 241 associations, including 30 new loci. The source GWAS of PGS003438 (GCST90132314).
- Khera AV, Chaffin M, Aragam KG, et al. Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. Nature Genetics, 2018. doi:10.1038/s41588-018-0183-z Read in the author manuscript, PMC6128408. CAD score of 6,630,150 variants; 8.0% of 288,978 testing participants (23,119) at three-fold or greater risk; 2.3% at four-fold, 0.5% at five-fold; 20-fold more people than familial hypercholesterolaemia mutations, taken as 0.4% of the population with up to three-fold risk. White British participants.
- Elliott J, Bodinier B, Bond TA, et al. Predictive accuracy of a polygenic risk score-enhanced prediction model vs a clinical risk score for coronary artery disease. JAMA, 2020. doi:10.1001/jama.2019.22241 Abstract and key points read. 352,660 UK Biobank participants, 6,272 incident CAD events over a median 8 years. C statistic 0.61 for the score, 0.76 for the pooled cohort equations, 0.78 combined; change 0.02 (0.01 to 0.03). NRI at 7.5%: 4.4% for cases, -0.4% for non-cases.
- Hingorani AD, Gratton J, Finan C, et al. Performance of polygenic risk scores in screening, prediction, and risk stratification: secondary analysis of data in the Polygenic Score Catalog. BMJ Medicine, 2023. doi:10.1136/bmjmed-2023-000554 Coronary artery disease: median detection rate at a 5% false positive rate 12% across 27 scores; top versus bottom 1% odds ratio about 13 for a typical score, still a detection rate of 12%.
- Mars N, Koskela JT, Ripatti P, et al. Polygenic and clinical risk scores and their impact on age at onset and prediction of cardiometabolic diseases and common cancers. Nature Medicine, 2020. doi:10.1038/s41591-020-0800-0 Published abstract read at nature.com. Coronary heart disease figures read in the authors' manuscript deposited in the University of Helsinki repository (hdl.handle.net/10138/319904): FinnGen n = 135,300 with 20,179 CHD cases; 6,412,950-variant score; HR 1.31 per SD; top 2.5% versus 20th to 80th percentile HR 2.03, bottom 2.5% HR 0.61; risk by age 80 37.2% versus 63.9%; onset 4.35 years earlier; FINRISK 10-year C-index 0.823 clinical, 0.820 clinical plus score; NRI 1.1 (-0.1 to 2.2) overall, 3.5 for early-onset CHD.
- FinnGen research project. FinnGen, University of Helsinki, 2026. Read 10 October 2026: genome and longitudinal health data of more than 500,000 Finns, almost 10% of the population; data freeze 13 of more than 500,000 individuals.
- Mars N, Kerminen S, Feng YA, et al. Genome-wide risk prediction of common diseases across ancestries in one million people. Cell Genomics, 2022. doi:10.1016/j.xgen.2022.100118 CAD odds ratio per SD: FinnGen 1.53 (25,706 cases); pooled European 1.54; UK Biobank South Asian 1.41; BioBank Japan 1.32; UK Biobank African/Caribbean 1.32; Mass General Brigham African 1.10 (0.96 to 1.26). Essentially no variability between Finland's early- and late-settlement regions. CAD score built from Nikpay et al. (European 77%, 13% South Asian), Table S4.
- gnomAD v4 genomes: allele counts, non-Finnish European and Finnish groups. Genome Aggregation Database, Broad Institute, 2024. Counts as committed in data/articles/what-a-percentile-means/gnomad_populations.tsv; rs10455872 (6-160589086-A-G) re-checked against the gnomAD API on 10 October 2026: 4,723 of 68,006 non-Finnish European and 376 of 10,618 Finnish alleles.
- PGS003438 (PRS241_CAD). PGS Catalog, 2023. Read 10 October 2026: 241 variants, method 'Genome-wide significant SNPs', GWAS of 1,165,690 participants (GCST90132314), evaluation in European-ancestry participants.
- Inouye M, Abraham G, Nelson CP, et al. Genomic risk prediction of coronary artery disease in 480,000 adults: implications for primary prevention. Journal of the American College of Cardiology, 2018. doi:10.1016/j.jacc.2018.07.079 metaGRS of 1.7 million variants. C-index 0.623 alone; 0.670 for six conventional factors; 0.696 combined. HR 1.71 per SD.
- O'Sullivan JW, Raghavan S, Marquez-Luna C, et al. Polygenic risk scores for cardiovascular disease: a scientific statement from the American Heart Association. Circulation, 2022. doi:10.1161/CIR.0000000000001077
- Marston NA, Pirruccello JP, Melloni GEM, et al. Predictive utility of a coronary artery disease polygenic risk score in primary prevention. JAMA Cardiology, 2023. doi:10.1001/jamacardio.2022.4466 Abstract and key points read. 241 genome-wide significant variants; 330,201 UK Biobank participants without CAD or lipid-lowering therapy, 4,454 heart attacks over 10 years. HR per SD 1.72 under 50, 1.46 at 50 to 60, 1.42 over 60; age interaction replicated in Biobank Japan.
- Khera AV, Emdin CA, Drake I, et al. Genetic risk, adherence to a healthy lifestyle, and coronary disease. New England Journal of Medicine, 2016. doi:10.1056/NEJMoa1605086 Read in the author manuscript, PMC5338864. 55,685 participants in three prospective cohorts and one cross-sectional study; up to 50 SNPs; top versus bottom quintile HR 1.91; favourable versus unfavourable lifestyle in the top quintile HR 0.54; standardised 10-year event rates 10.7% v 5.1% (ARIC), 4.6% v 2.0% (WGHS), 8.2% v 5.3% (MDCS).
- Dikilitas O, Schaid DJ, Kosel ML, et al. Predictive utility of polygenic risk scores for coronary heart disease in three major racial and ethnic groups. American Journal of Human Genetics, 2020. doi:10.1016/j.ajhg.2020.04.002 Abstract read. eMERGE: 45,645 European-ancestry, 7,597 African-ancestry and 2,493 Hispanic participants; metaGRS HR per SD 1.53, 1.53 and 1.27 (European, Hispanic, African ancestry).
- Patel AP, Wang M, Ruan Y, et al. A multi-ancestry polygenic risk score improves risk prediction for coronary artery disease. Nature Medicine, 2023. doi:10.1038/s41591-023-02429-x GPSMult. External validation odds ratio per SD: African 1.25 (33,096), European 1.72 (124,467), Hispanic 1.61 (16,433), Million Veteran Program; South Asian 1.83 (16,874), Genes & Health.
- What your DNA file can actually read: measured on 14 real files. Aimosti, 2026. Measured 6 October 2026. Ten chip exports read 281 to 495 of 1,457 polygenic-score positions.
Last reviewed . Every number on this page links to the source it comes from; if one of them has moved, tell us.