Ancestry and polygenic risk

Dr. Mike Weale, Associate Director, Precision Health,?Genomics plc

In our recent preprint paper, we show that polygenic risk scores (PRSs) are powerful tools for predicting an individual’s genetic risk for a multitude of different diseases and conditions. However, our paper also shows that the predictive power varies depending on a person’s genetic ancestry. Genetic ancestry points to the geographic origins of a person's ancestors between 10 to 100 generations ago, and is typically estimated by comparing that person's genome to others sampled from a range of geographies.?

As illustrated by the figure below, PRS performance is higher for people with European ancestry, and lower for people with other ancestries, across both diseases and quantitative traits. For diseases, using the odds ratio per standard deviation of PRS as a measure of PRS performance, predictive performance was reduced on average by 9.1%, 14.3% and 27.6% in individuals of South Asian, East Asian and African ancestry, respectively. We are not the first investigators to see this - it appears to be a general pattern seen across a wide range of PRSs.

This variability is due to the data sources available to develop PRSs. We need data from studies that recorded the associations between different genetic variants and disease outcomes in huge numbers of individuals, but in most of the largest studies available, the majority of the people were of European ancestry. It’s not surprising then, that more data from people with European ancestry has translated into better predictions in people of European ancestry.

We must address this bias in data sources, and various initiatives such as H3Africa and GenomeAsia100K are underway to promote genetic association studies in currently under-represented ancestry groups. This problem is not unique to PRSs - other risk factors already in use in clinical risk assessment tools, also show variation in predictive performance across ancestries, and more data are needed to improve the usefulness of these for people of non-European ancestries as well (Vasan & van den Heuvel 2022).?

There has been some debate about whether PRSs should be excluded from clinical risk prediction tools until performance is equal across ancestry groups. However, it is important to recognise that adding a PRS to existing clinical risk assessment tools improves those tools’ performance for everyone, and results in better prediction for everyone, regardless of their ancestry. Omitting PRSs means everybody gets a less accurate estimate of their disease risk.?

Clinical utility also depends on the average level of risk among different ancestry groups. In our preprint, we illustrate this with the example of type 2 diabetes (see figure below). People of South Asian ancestry have a higher average risk of developing type 2 diabetes, compared to those with European ancestry. Although the relative PRS performance, measured as the odds ratio per standard deviation of PRS, is lower in people of South Asian ancestry (1.9 vs 2.3 in European ancestry), the changes in absolute risk are larger. The lifetime risk of type 2 diabetes is greater than 60% in people of South Asian ancestry in the top 3% of the PRS distribution, compared to around 40% in the equivalent risk group of people with European ancestry. This means, in this example of type 2 diabetes, the addition of PRS information is clinically powerful and important for individuals of South Asian ancestry.

We have looked at the effect of ancestry on clinical utility in more detail in a previous paper looking at cardiovascular disease risk. In that paper we found that adding a PRS to an established clinical risk tool generated improvements in risk prediction that were of similar size for people of African and European ancestries.

The need for more diverse data to improve PRS performance across all ancestries is clear, but even now, the benefits that PRS-powered risk prediction tools can bring to people of all ancestries are compelling.