NEW YORK — As polygenic risk scores begin to be rolled out in the clinic, a few studies have suggested that they may not add as much to disease risk predictions as hoped.
By adding PRS to standard clinical assessments of disease risk, researchers have hoped to refine such predictions, particularly among patients who otherwise fall into a gray, intermediate level of risk, but who could potentially benefit from the stepped-up screening or precautionary treatments given to high disease-risk groups.
There has been, noted Alexander Charney, assistant professor of psychiatry and of genetics and genomic sciences at the Icahn School of Medicine at Mount Sinai, a lot of excitement and interest in PRS and how they could be applied clinically, which he suggested should be tempered. "There's this increase with hype and then the enthusiasm goes down, and the reality starts to set in," he said. "And then the reality is somewhere between zero and the peak of the hype."
His group and another team at Vanderbilt University have done research indicating that the addition of a PRS may not improve predictions as much as expected. However, others have suggested that large studies, which are increasingly being conducted, are needed to capture the effects of the addition of PRS. Further, those effects may be most prominent among younger individuals who have not yet accumulated some of the more traditional risk factors, they said.
"Like in any other aspect in medicine, you have to look at the totality of evidence, you can't focus on a single study," Carlos Iribarren, a researcher at Kaiser Permanente Northern California, said. "The overall evidence now is leaning toward that [PRS are] helpful."
Into the clinic
PRS scores are just starting out in the clinic. In late 2019, the Broad Institute's Amit Khera launched a new Preventive Genomics Clinic at Massachusetts General Hospital that, as part of its services, offers PRS testing alongside monogenic testing for heart disease.
"The traditional approach for human genetics is to focus on rare monogenic mutations, and they're super important for the 1 percent of the world who has one," Khera said. "So, the question is: What can you say about the remaining 99 percent of the world's population?"
Previously, he and his colleagues developed a handful of PRS for common diseases, including one for coronary artery disease that encompasses 6.6 million SNPs. This sweeping PRS, which they reported in Nature Genetics in 2018, forms the basis of the test they are now offering.
Similarly, Genomics plc and researchers at Stanford University are also beginning to use polygenic risk scores to predict cardiovascular disease.
A handful of studies, though, have suggested that PRS might not add too much to the determination of who is and who is not at increased risk of disease.
Researchers led by Vanderbilt University's Jonathan Mosley conducted an analysis of more than 7,000 individuals to examine whether adding a PRS to clinical gauges of coronary heart disease improved their ability to classify people's risk. As they reported in the Journal of the American Medical Association last year, they found that while the PRS they used — one developed by Khera and his colleagues — was associated with heart disease, it did not improve the accuracy of risk stratification. In all, they noted that about 10 percent of the individuals in their analysis, who were a mean of 62 years old and all of white, European ancestry, were reclassified to a higher or lower risk category.
Mosley noted that what they showed is similar to what other studies have found but that they differed in their interpretation. "And our conclusion was that this is a really small magnitude," Mosley said.
He added that PRS do not appear to perform better than other biomarkers for coronary artery disease, like C-reactive protein, that have not been widely adopted by the field to predict disease.
Similarly in schizophrenia, Mount Sinai's Charney and his colleagues analyzed whether a PRS could predict outcomes in adults with psychosis. That study, published in Nature Medicine in September, found that a schizophrenia PRS did not improve outcome predictions as compared to data gleaned from routine psychiatric assessments.
This finding suggested to Charney that PRS might not outperform data that's collected more easily clinically. "If I can get the same information by just talking to the patient and I'm already talking to the patient anyway, why would I order a genetic test that's going to cost money and then we have to wait for the result, then someone has to do the analysis of the polygenic risk score?" Charney said.
He added that PRS have been a boon for research, and that his skepticism is on the clinical side. He particularly is unsure about the common vision that PRS will be able to predict who will develop common conditions across medicine. "That, I think, is not in line with what the data show," Charney said.
Differing interpretations
Other researchers, though, view the field differently and said that the effects of PRS on risk prediction could be missed if studies don't take factors like age into account or if they aren't large enough
Age is a particularly important factor, Stanford University's Euan Ashley said in an email. "Age eventually trumps everything in every prediction model because everyone eventually dies of something and it is usually heart disease or cancer," he added.
Studies could miss the effect of PRS by not looking across a wide spectrum of ages. A recent paper from Genomics plc researchers and others, including Ashley, that appeared in Circulation: Genomic and Precision Medicine found their integrated risk tool for coronary artery disease — which included both a PRS and established risk assessment tools — led to improvements in reclassification for 5.9 percent of their cohort. They further reported that the improvement was greatest among men between the ages of 40 and 54 years.
PRS might also have more clinical utility in younger ages. "The unmet need in terms of clinical medicine is not necessarily predicting risk in the 70-year-old after the clinical risk factors have already set in; but actually, identifying the high-risk individuals who fly under the radar very early in life," the Broad's Khera said. "And so especially for preventable disease, that's where I think the value is going to be."
Mosley, though, had a different interpretation. He instead views studies that find PRS perform better among younger individuals as showing a modest performance among younger individuals and a poor performance among older individuals.
Large studies may also be needed to gauge whether PRS add statistically significant changes to risk prediction, Kaiser Permanente's Iribarren added. For instance, in a 2016 Circulation: Genomic and Precision Medicine study, he and his colleagues drew upon a cohort of more than 51,000 people of European ancestry to examine whether the inclusion of four different PRS for coronary heart disease could shift risk classification. The inclusion of those scores improved reclassification by 4 percent or 5 percent, depending on the score used, and improved the C statistic, they reported.
Mosley additionally pointed out that his team's study, though smaller, had similar results to other studies, including one from Khera's group, but that he, again, had a different interpretation of those results and their clinical relevance.
Isotta Landi, a postdoctoral fellow at Mount Sinai and first author of the Nature Medicine paper on which Charney was senior author, noted in an email that bigger samples sizes generally are better for machine learning-based analyses. She added, though, that their study was highly powered.
More to prove
Even with these differences in interpretation, studies into the best ways of applying PRS — which scores are the most useful and which interventions they may indicate — still need to be done, as does the development of scores based on genome-wide association studies of more diverse populations.
Some PRS include only a handful of SNPs, while others take millions of SNPs into consideration. Khera, for instance, is using a PRS of millions of SNPs. But Iribarren said that it is unclear to him whether just a few streamlined markers might suffice for making predictions or if a genome-wide tack would make better sense.
Khera, though, said that more is likely better. "If you compare like a one-million-variant score to one that's only the top 50 or the top 200 or top 20,000, you do definitely do better [with the larger one]," he said.
He added, though, that the effect has to do with how polygenic a trait is — for cystic fibrosis, for instance, as it is not polygenic, common variants likely matter less.
At the same time, interventions associated with a high PRS score need to be further evaluated. "Because it's never enough to just say you're at high risk," Khera said. A PRS has to be something "you can use to empower people," he added.
Eventually, the field needs clinical trials in which people determined to be at high risk by a PRS are randomized to an intervention or not, Khera said, as that will be the bar that needs to be cleared to make PRS part of the standard of care.
"I think we're at an exciting tipping point [with] the ability to look at someone's DNA and say something important about the risk for all of these diseases," Khera added. "And now the onus is on the scientific community to say, 'OK, well, how can we prove it is useful and figure out how it's useful and how it integrates into medicine?'"