What a percentile means on a genetic report: a rank inside one reference group

A rank inside a group

Take a polygenic score for coronary artery disease, the subject of its own guide. Our report uses one with 241 variants, published in 2023 and listed in the PGS Catalog as PGS003438^{[2, 3]}. For each variant the engine counts how many copies of the higher-risk allele a file carries, zero, one or two, multiplies by the variant's weight and adds the 241 products. The total for one person might be 12.9. That number means nothing until it is set beside the totals other people get.

So the total is compared with a reference distribution: the spread of totals expected in a reference group. If 90 in 100 of that group would score lower, the file is at the 90th percentile. The percentile is a place in a queue of people sorted by their score. A finishing place in a race tells you how many runners crossed the line behind you, and nothing about how fit any of them are.

Why the 90th percentile is not a 90 percent risk

A rank is spread evenly by construction: every percentile holds one in a hundred people. Disease is not. What differs between percentiles is how many of those people develop the condition, and in large studies the difference is real but modest.

The first figure comes from 288,978 UK Biobank participants, where a score built from 6.6 million variants placed 8.0 percent of people at more than three times the risk of coronary artery disease^[4]. Even so, about 89 in 100 of the people in its top percentile did not have the disease, by our subtraction. The second figure is a model estimate from a study of breast cancer in women of European ancestry, using UK incidence and mortality rates^[5]. At the 99th percentile the estimated lifetime risk was about one in three, enough for the authors to note that these women would meet the UK's clinical definition of high risk. Their predicted risk was about four times that of women in the middle fifth of the distribution^[5]. For comparison, Cancer Research UK puts the lifetime risk for all UK women born in 1961 at 1 in 7^[6].

A worked example with published numbers

Screening researchers make the point with detection rates. Applied to coronary artery disease, the genome-wide score from that UK Biobank study would flag 5 percent of people who never develop the disease and catch only 15 percent of those who do, missing the other 85 percent; Wald and Old called that little better than identifying people at random^[7]. In the queue a percentile sorts, the affected and the unaffected overlap almost everywhere.

How the report builds its reference group

Studies usually rank a score within their own cohort. Our report has no cohort, so it builds the reference group from public allele frequencies. For every variant in a score it takes the frequency of the higher-risk allele in non-Finnish European genomes from gnomAD v4, a public database of variation from 76,215 sequenced genomes^[8]. With those frequencies the expected average and spread of the total follow directly, assuming each variant is inherited independently of the others. A file's percentile is its total's position on a normal curve with that average and spread.

Two corrections keep the curve honest. Variants that sit close together on a chromosome are inherited together more often than independence assumes, which widens the real spread; we measured how much for each score in 525 unrelated European genomes from the 1000 Genomes Project^[9] and widen the curve by that ratio: 2.64 times the variance for the venous thromboembolism score, 1.63 for kidney cancer, and less than 1.2 for the other nine. And on a chip, which reads only some of a score's variants, the curve is rebuilt over exactly the variants the chip read, so the question becomes where the file ranks among people scored on the same markers.

Until 6 October 2026 the frequencies came from the PGS Catalog's files, and checking them against gnomAD showed that for several scores the catalogue column carried the other allele's frequency or a non-European one. A wrong average moves every file's percentile alike, invisibly. The reference now comes from one source for every score, and the frequencies we used for the table below match it at all 1,466 variants both could place.

Change the group and the rank moves

Allele frequencies differ between populations, and so does a score's expected average. gnomAD keeps Finnish genomes apart from other Europeans, because Finland's population went through a bottleneck that left it with its own pattern of variants^{[10, 8]}. We recomputed each score's reference from four other gnomAD groups and asked where a file at exactly the European 90th percentile would land against each. The table is our arithmetic on gnomAD's counts: an estimate, under the report's own assumptions, with the European linkage correction applied to every group.

For Finnish frequencies the move is small. Across the eleven scores in the table, a file at the European 90th percentile lands between the 86th and the 94th against a Finnish reference, because no score's Finnish average sits more than about a quarter of a standard deviation from the European one. The other groups move it much further: against African frequencies the same file lands anywhere from the 48th percentile (breast cancer) to the 99th (LDL cholesterol), and against East Asian ones from the 58th (atrial fibrillation) to the 98th (venous thromboembolism). Nothing about the file changed in any of these rows. Ancestry percentages move the same way when the reference panel changes, as our guide to them shows.

A shifted average is only one of the problems. About 79 percent of all participants in genome-wide studies have been of European descent, so most published weights were estimated in Europeans, and they predict less well elsewhere: in UK Biobank, prediction accuracy for 17 traits was 2.5 times lower in East Asian and 4.9 times lower in African participants than in European ones^[11]. Re-centring the curve moves the rank; it cannot repair weights never measured in that group. This is why a reporting standard for polygenic scores asks studies to state the ancestry of every sample a score was built and tested in^[12].

What the report prints beside each percentile

The rank never stands alone on the card. Under it the report prints the percentile with its reference (63.4 · European reference), a markers-used line, and a sentence headed Which way the number runs, which reads the rank back as roughly 63 in 100 of that group scoring lower and ends: It is a rank inside a group, not a probability of the condition. Then come the score's statement that it is a tendency and not a diagnosis, the ancestry caveat and the PGS Catalog identifier. The breast cancer card adds a banner saying its reference is built in women. How precise the number is depends on the file.

Chips are the common case for coverage trouble. Across the ten chip exports on our measured data page, the arrays read between 281 and 495 of the 1,457 polygenic-score positions the page counts^[13], so on a chip many scores land in the estimate band or are not reportable. A plain VCF never reaches the precise state at all, because it cannot show that an absent variant was read; the card counts only the variants seen in the file. The sample report shows the cards from a full-coverage gVCF.