NEW YORK — Polygenic risk scores for the same disease can provide different risk estimates for the same person, an issue that could affect their use in the clinic.
Genome-wide association studies have linked millions of common genetic variants to the risk of developing a range of common human diseases or phenotypes and they are often bundled into polygenic risk scores (PRSs) to gauge genetic risk of developing that disease or phenotype. There has been increasing interest in applying those scores clinically to stratify people into risk categories to guide, for instance, how intensively they are monitored for a certain disease, or to recommend them for preventive treatments.
In addition to scores for a range of common conditions, there are now also multiple scores for the same disease. In a recent paper appearing in Scientific Reports, David Hunter, a professor of epidemiology and medicine at the University of Oxford, and his colleagues examined whether different PRS scores affected how people were classified.
"We were curious to know whether if you applied different statistical techniques to the same data to produce a polygenic risk score — because there's a range of ways of constructing them — whether this would have any material influence or whether everybody was essentially coming up with the same score," said Hunter, who also directs the Harvard-Oxford program in epidemiology.
For three different conditions — breast cancer, hypertension, and dementia — they compared how recently published PRSs classified individuals from the UK Biobank. They noted that, after accounting for linkage disequilibrium, 65 percent to 79 percent of SNPs contained within one PRS for a condition were also represented in the second. Each PRS for the same condition also had similar performance metrics as their counterpart, though the newer PRSs tended to have a slightly higher odds ratio. For instance, PRS-A for hypertension had an area under the curve (AUC) of 0.69 and an odds ratio of 1.83, while the second, PRS-B, had an AUC of 0.70 and an odds ratio of 2.18.
Despite this, Hunter and his colleagues only found modest correlations between the PRS pairs on risk prediction. That is, the scores predicted different levels of risk for many UK Biobank participants for the same condition.
Of the women that the PRS-A score placed in the top 1 percent of risk for breast cancer, only slightly more than 23 percent were also placed in that top risk category by PRS-B. The researchers reported similar findings for hypertension and dementia in this analysis and have since examined other diseases, indicating the issue is widespread.
To Hunter and his colleagues, their findings present a problem for using PRSs in the clinic. Depending on what score is used, people could be classified into different levels of risk, which could have an effect on what advice or treatments they are offered.
"This has obvious implications if you're going to start telling people about their risk," Hunter said. "At a minimum, we have to be careful to explain that this is still a science that's in some evolution."
This finding, said Greg Gibson, the director of the Center for Integrative Genomics at the Georgia Institute of Technology, is not unexpected.
For instance, he noted that in a Nature Genetics study from December 2021, researchers from the University of California, Los Angeles also examined the accuracy of PRSs at the individual level with UK Biobank data. They, too, found variations that would affect patients' risk stratification.
Gibson added that it is also a difference in perspective and focus. "Science papers focus on sensitivity and specificity, which is what goes into the AUC curve," he said. "But from a patient perspective, what you care about is positive predictive value or precision, or given a positive test, what's my likelihood of getting it?"
According to Hunter — who also noted that they were not the first to highlight this issue of variation between PRS — the differences in classifications his team uncovered do not appear to be attributable to the number of SNPs included in the risk score but instead are more likely due to differences in the statistical approaches used to generate the scores. He suggested that a meta-analysis approach proposed by Michael Inouye and colleagues in the Journal of the American College of Cardiology in 2018 — constructing an overarching genetic risk score by averaging all the existing scores — might "give you a more robust estimate than any one of the individual scores."
As Gibson noted, though, prediction is inherently difficult. The UCLA researchers in their Nature Genetics analysis found that their observed variance in PRS scores could be calculated and suggested that that individual-level uncertainty be included into PRS score discussions. But this type of approach discussing uncertainty, Gibson said, may not work in the clinic, as people are often not good at gauging risk.
"If it's 50 percent or even two- or three-fold higher risk, but if the baseline risk is 1 percent, you're still not likely to get it right," he added. "And so people might struggle with that idea."
Nongenetic clinical factors could also be added alongside PRSs to generate integrated risk scores. "There are not many common diseases for which you would not want to put the polygenic risk score into the context of the other known lifestyle and clinical risk factors," Hunter noted.
For instance in diabetes, while a PRS does add to risk discrimination, it is only part of the clinical picture to consider, he said. "It wouldn't make any sense for you to give me my diabetes PRS and tell me that that's meaningful if you were ignoring my age, sex, family history of diabetes, body mass index, and waist circumference," he added.
In such scenarios with added clinical data, Hunter suggested that the differences between PRSs might be less important.
Additionally, when PRSs are used could affect their predictive value, as nongenetic factors may have an increasing influence on disease risk over time. In Genome Medicine in 2021, Gibson and his colleagues found that the predictive value of PRSs for heart attack varied by age. By middle age, they found that PRSs added little to the risk information provided by environmental or other nongenetic risk factors. Genetic factors were better at predicting risk in younger individuals, while clinical and lifestyle factors had more sway in older adults.
Part of the issue with conflicting PRS risk estimates may also come down to how PRSs are discussed. Both Gibson and Hunter noted that the community often presents scores as immutable. "These scores from a personalized perspective were never meant to be predictive in a deterministic sense," Gibson said.
Hunter noted even if someone's SNPs are fixed, what SNPs are included in a PRS and how those SNPs are weighted are not. "That's a matter of the statistical algorithm and they don't correlate as highly as I had expected," he said.
"We're not the first people to notice this," Hunter added. "But again, I think it's still not sufficiently talked about in the risk score community."