Polygenic scores, in plain words: what the number adds up and what it leaves out

A sum of many small weights

Start with a genome-wide association study. It compares hundreds of thousands of people one common variant at a time and estimates how much each copy of a variant's effect allele goes with a trait. That estimate is the variant's weight. A polygenic score takes the weights from one such study, counts how many copies of each effect allele a person carries, zero, one or two, and adds up copies times weight across every variant in the score^[1].

No single variant does much. In that coronary artery disease score the median weight is about 0.04, and the biggest common one, rs10455872, carries 0.30. Leaving out any one variant barely moves the total; leaving out half of them moves it a lot, which matters later for chip files. Sports gene tests read one variant at a time; we checked the best-known, ACTN3.

Published scores are collected in the PGS Catalog, which describes itself as an open database of polygenic scores and the metadata needed to apply and evaluate them^[2]. It exists so that a published score can be reproduced exactly and compared with others^[5]. It often holds many scores for one disease: a 2023 analysis found performance figures for 108 breast cancer scores alone^[6]. Two services can give the same person different percentiles for one disease without either making an error, because they are not computing the same score.

From a total to a percentile

A total such as 12.90 means nothing on its own. It becomes readable once it is set against the totals of other people: the report places it on a reference distribution and reads off the share of that group who score lower, which is the percentile.

Our reference distributions are modelled, not measured in a cohort. For each variant we take how common its effect allele is in the non-Finnish European genomes of gnomAD v4^[7] and work out the average total and spread a European-ancestry population would show, widened where neighbouring variants tend to be inherited together. Every percentile on the report is labelled with this European reference.

A sum of hundreds of small numbers falls into a bell curve whether or not it predicts anything. The tidy curve under a percentile says nothing about accuracy; that has to be measured in people whose outcomes are known.

How much the report's scores explain

For a measurable trait such as cholesterol, accuracy is usually reported as the share of the differences between people that the score accounts for, written R². For a disease it is often the AUC: the chance that a randomly picked person who develops the disease scores higher than one who does not, where 0.5 is a coin toss and 1 is perfect. Here is what the sources of six of the report's scores measured.

Height is the exception, and even there the report's score is not the best available: a score of 12,111 variants reached 40 percent of the variation in European-ancestry samples^[13]. For BMI, the study that tested our score also built an index from six lifestyle factors, among them alcohol, exercise, sleep and smoking, and it explained 6.6 percent of BMI against the score's 2.9^[10]. Eye colour is a rare exception: for blue against brown, one variant does most of the work (how that switch works).

A rank, not a diagnosis

An AUC of 0.66 translates into weak separation between the people who will and will not develop a disease. In 2023 Hingorani and colleagues took 3,915 performance figures for 926 scores and 310 diseases from the PGS Catalog and asked a screening question: set the cut-off so that only 5 percent of people who stay well are flagged, and what share of those who become ill are caught? The median was 11 percent^[6].

The same paper translated a typical coronary artery disease score into ten-year odds for a 50-year-old. At the 2.5th centile the odds of developing the disease were 1:54; at the 97.5th they were 1:8, against background odds of 1:19 for everyone that age^[6]. Put as plain fractions, our conversion, that is about 1 in 55 at the bottom and 1 in 9 at the top.

A high percentile moves the odds; it does not decide the outcome. Most people who develop the disease come from the middle of the distribution, because that is where most people are: the top 2.5 percent of the coronary score contributed 7 percent of cases^[6].

Accuracy falls away from the study population

Weights are estimated in particular people, and most of those people have been of European ancestry. In 2019 Martin and colleagues, counting from the GWAS Catalog, put about 79 percent of GWAS participants as European, from a group that makes up 16 percent of the world's population^[14]. By 2021 the European share of GWAS samples had risen to 86 percent^[15].

The cost shows up as lost accuracy. Using European-derived weights for 17 measurable traits in UK Biobank, Martin's group found prediction accuracy 1.6-fold lower in Hispanic or Latino participants, 1.7-fold lower in South Asian, 2.5-fold lower in East Asian and 4.9-fold lower in African-ancestry participants, on average^[14]. Height, studied in 5.4 million people, follows the same pattern: 40 percent of the variation explained in European-ancestry samples, about 10 to 20 percent in others^[13]. One of the report's own scores shows it directly.

The loss does not wait for a continental boundary. Privé and colleagues applied scores for 245 traits, trained on north-western Europeans in UK Biobank, to nine ancestry groups of the same cohort and found that prediction already drops within European ancestries and falls further with genetic distance^[16]. The report's reference frequencies are non-Finnish European. We have not measured how much accuracy the disease scores lose in Finns; the chronotype score is the one whose published check was run in Finnish adults^[11].

Every percentile on the report therefore carries the same sentence, whatever the reader's ancestry: scores travel poorly across ancestries, and a percentile is materially less accurate for someone whose ancestry differs from the reference.

Chips, missing positions and imputation

The GWAS behind most weights used imputed genotypes. UK Biobank genotyped its participants on an array, then used imputation, statistical inference from a reference panel of sequenced genomes, to raise the number of testable variants to about 96 million^[17]. A score's variants are chosen from sets like that, so many were never on any consumer chip.

A consumer chip export is a list of the positions one array measured: between 563,320 and 955,958 rows in the ten exports we measured. Of the 1,457 positions the report's disease scores use, those ten chips read between 281 and 495, roughly a fifth to a third^[18].

Imputation could fill those gaps statistically, the way UK Biobank filled its own. Aimosti does not impute; what we measured before deciding has the numbers. A position the chip did not read stays unread. On a chip the report scores only the variants the array measured and rebuilds the reference distribution over exactly those variants, so the percentile answers a narrower question: among people scored on the same markers, where does this file rank. Then it applies a coverage rule to every score separately.

• 90 percent or more of the score's variants read: the percentile is shown to one decimal place.
• 50 to 89 percent read: the percentile is shown as an estimate, a whole number with a tilde, and the card says it rests on the variants covered and is enough for a tendency, not a precise rank.
• Under 50 percent read: no percentile. The card says the score was examined and is not reportable, gives the count, and calls it a coverage gap, not a result.

With a fifth to a third of the positions overall, no chip we measured can clear the halfway mark on every score, so on any of them some disease scores come back as not reportable. That is intended: a percentile computed from a third of a score, with the rest guessed, would look like a result and would not be one. Files that a provider imputed from low-coverage sequencing, and say so in their header, are read like chip exports, with a line added to the chip caveat saying the genotypes were imputed by the provider.

What the cards show, file by file

The disease scores sit in their own section, headed as a tendency and not a diagnosis, each card tagged Tier 2, well-supported. The Emerging scores, for traits anyone can measure, sit apart in the Emerging-research section because they are newer and less tested. There is no separate opt-in and no per-score reveal: every report includes both sets.

How a card treats coverage depends on the kind of file, because files prove different things about the positions they do not list.

On a plain VCF, silence is read as the reference allele, which is usually right but cannot be shown, so the card never gives the one-decimal percentile and says the true value could sit on either side of the estimate. A gVCF from the same run settles that; our guide to VCF, gVCF, BAM and CRAM explains why. On a gVCF that shows at least 20 of a score's positions as covered but not a single variant call among them, the card treats the file as unreadable and shows no percentile.

The Emerging cards add what the trait is and how much of it the score explained in its source study, the figures in the table above. Their direction line is written per score, because a higher eGFR score points towards better kidney filtering, the opposite way to the disease scores. A banner above them says a measurement of the trait says more than its score does.

What the cards leave out

The cards carry no absolute risk. None turns a percentile into a chance of disease, which would need age, sex, family history, measured risk factors and a model validated in people like the reader. There is no colour scale: every disease-score card has the same neutral marker whatever the percentile, and the report labels no one high-risk on the strength of a score.

The report does not infer sex from the file, so the breast and prostate cancer scores are shown to everyone with a banner saying they are not applicable, rather than reassuring, to people they were not built for. It has no reference curves for other ancestries. The disease cards do not yet say how much of each disease their score explains, a figure the Emerging cards carry. The sample report shows every card on a synthetic genome.