Polygenic risk scores (PRS) have advanced enormously in recent years and are now well on their way to becoming an established part of modern medicine. But there are still some issues such as accuracy and applicability that need to be solved before they can truly become mainstream.
With increasingly sophisticated genetic technology at their fingertips, scientists have developed correspondingly sophisticated genetic tests. Searching for single monogenic mutations used to be the norm, but it’s now possible to screen for multiple small mutations, often single nucleotide polymorphisms (SNPs), that together contribute to risk for specific diseases such as heart disease or cancer. Previously used primarily as research tools, these polygenic tests are now rolling out to clinics and are being developed as prognostic tools by companies and researchers working across a broad swath of diseases and health conditions.
“A lot of these studies have really started to show that polygenic risk score is not necessarily just an independent kind of factor, or another tool in our toolbox for clinical care, but it is also additive,” says Elissa Levin, senior director, clinical affairs and policy at population genomics company Helix.
With limited official regulation, due mostly to the speed the field has developed, there has been some controversy about the accuracy and validity of some of these tests. But this is slowly improving, and they have the potential to make a difference to people who are at risk, particularly as they have more of a focus on common diseases than other genetic tests.
Evolution of polygenic tests
“Polygenic scores allow us to identify individuals that are invisible to traditional risk assessment methods, but have increased risk compared to the rest of the population,” comments Giordano Bottà, Ph.D., CEO of Italian health tech company Allelica, which dubs itself “the polygenic risk score company.”
Single monogenic mutations, like the well-known BRCA1 or BRCA2 mutations, can significantly impact a woman’s risk of developing breast cancer, but only a small percentage of the population have these mutations.
If you take polygenic risk into account, this risk becomes more of a spectrum across the population. The WISDOM clinical study, which is ongoing in California, aims to recruit 100,000 women and randomly assign them to screening with a mammogram once a year or at an interval based on the results of a genetic test assessing their overall risk of breast cancer.
“On the whole, it actually reduces the number of mammograms that are done, because most women will fall into the every-other-year category instead of the every-year category. But there are going to be some women where we might actually say ‘No, you need to have a mammogram of every six months, because you’re at increased risk for breast cancer,’” explains Alicia Zhou, CSO at Color, a genomics company that is a partner of the WISDOM study.
While this study is still ongoing, polygenic risk scores estimating breast cancer risk have actually been used in the clinic for several years. In 2017, Myriad Genetics developed riskScore to help women assess their 5-year and lifetime risk of developing breast cancer. Importantly, this test takes into account other health factors as well as genetics.
“While existing clinical models for predicting someone’srisk of developing breast cancer are good, they’re not great. With riskScore, we posited that adding a limited amount of targeted genetic data would hone those models to provide more accurate results, particularly among women with a pronounced family history of breast cancer, but who do not have the most well-known genetic variants associated with high risk,” says Thomas Slavin, senior vice president of medical affairs, oncology at Myriad, which has a focus on developing genetic tests evaluating cancer risk.
Another area where PRS are coming into their own is for cardiovascular risk. Allelica conducts research in this area and has developed a PRS that can identify individuals with normal levels of cholesterol that actually have a threefold increased risk for cardiovascular disease based on their genetics.
“If you have an average level of LDL cholesterol, you think that you are fine, but if you have a high polygenic score, this average is not average for you anymore, it creates risk due to this interaction. LDL cholesterol is able to form plaques more easily in individuals with a high polygenic risk score,” says Bottà.
PRS are not yet used as widely as they could be, but in countries such as the U.S. they are becoming integrated into standard risk assessments. Zhou believes that advances in genomic sequencing over the last few years have helped improve polygenic risk scores.
“This area has been revolutionized recently. There’s been a set of investigators and researchers, such as Amit Kheera and Sek Kathiresan, who really looked at polygenic scores in a completely different way. They came up with this idea of a genome-wide polygenic score,” she explains.
This essentially means doing a low coverage whole genome sequence, rather than running a sample through a SNP chip containing around 750,000 possible variant binding sites. Color has now adopted this technique for several reasons, one of which is that it already has genomic sequencing capacity on site.
“It allows us to keep the price low, because we’re not sequencing very high depth. But actually, if we capture just a little bit of your genome across the entire genome, we have the same statistical power to be able to run these scores as you would get off of a genotyping array,” says Zhou.
Another advantage of low-coverage genome sequencing is that it is not restricted to a specific set of SNPs, as with a genotyping array. This can be an issue, as many currently available chips have been optimized for populations with European ethnicity so the included SNPs may not accurately reflect the population being tested.
Barriers to widespread rollout
Although polygenic tests have greatly improved over the last couple of years, there are still some significant barriers making greater rollout and use difficult.
A key problem with these tests is that people with European ethnicity make up around 80% of all published GWAS studies, which form many of the training sets for creating polygenic risk scores.
“The score itself has been trained to be biased towards European populations, unless they have the good representation of other ethnic populations within that training data. That’s actually something that is still a struggle for the field, it is not a solved problem,” says Zhou.
This means that if someone from a different ethnic background wants to use a polygenic risk score, depending on the score and the training set it uses, it may be less meaningful for them than someone from European ancestry.
That noted, this should improve in the not-too-distant future as there are several large biobank initiatives aiming to collect more data on populations of non-European origin. For example, the All of Us program in the U.S. will recruit one million individuals to form part of a biobankstyle cohort with a goal that 70% of the volunteers will be from underrepresented communities. Other biobanks such as the Million Veteran Program are also much more diverse than earlier initiatives, with around 30% non-European participants.
A number of researchers and companies have developed software tools to make better use of current data sets by more accurately accounting for differences in genetic variation seen in underrepresented groups—for example, the Tractor software developed by Elizabeth Atkinson, Ph.D., a researcher at the Broad Institute, and her colleagues.
“Thankfully, there is a growing field of researchers working on building more diverse cohorts and novel methodologies that can better handle underrepresented and admixed ancestries,” says Atkinson. (see sidebar “Tractor—Driving Towards More Diverse Genetic Tests” p.29)
“We effectively ‘paint’ everyone’s genome using a reference panel so we can tell the ancestry backbone on which each allele is falling at each spot in the genome, in each person. We use this local ancestry information in a novel GWAS model to get ancestry-specific association results. In other words, we estimate the effect of having a risk variant in each ancestry background, rather than just having a sort of aggregate measure.”
Another barrier to more widespread rollout has been a lack of consistent regulation. Like many areas in genetics, the speed with which the science has developed means regulation has struggled to keep up.
“Polygenic risk scores are a new area of discovery for the field of genetics. In order to reach maturity, we need to start being able to compare polygenic risk scores against each other, and be able to have some sort of standard rubric for how polygenic risk scores should be thought of,” says Zhou.
Most genetic tests used in the clinic are classified as Laboratory Developed Tests (LDTs), which means they do not undergo as much regulatory scrutiny as an FDA-approved device or test. This puts pressure on individual developers and companies to make sure the tests they produce are accurate and representative of the populations they are testing.
Many have stepped up to the mark and collaborated to create publicly available resources such as ClinVar, a database of genetic variants of clinical interest; and the Clinical Genome Resource, or ClinGen, a database containing information about genes and variants for use in precision medicine and research.
In 2018, the FDA formally recognized ClinGen as a resource of valid scientific evidence, meaning test developers can use data from this database to support applications for marketing approval without having to repeat the same data collection.
The direct-to-consumer genetic testing company 23andMe ran into regulatory problems with their health-related genetic risk scores in 2013 when the FDA blocked their rollout. But it has since achieved approval of a number of different genetic risk score-based tests for conditions such as Parkinson’s and Alzheimer’s diseases, among others, making it a leader in this area.
“23andMe was the first direct-to-consumer genetics service authorized by the FDA to provide health reports on disease carrier status, genetic health risks, and pharmacogenetic response,” explains Shirley Wu, director, health products at 23andMe.
“Over the years, we have put significant resources behind improving the methods used to develop these models, which lead to substantial improvements in model accuracy. These methods range from the ethnic diversity of the research participant cohorts used to train the models to the process used to select genetic variants for inclusion in the polygenic scores.”
More recently, Helix received FDA authorization for its Helix Laboratory Platform, the first time the FDA has authorized a sequencing platform.
“This is a game shifter because our entire platform has already been validated, and validated not on a SNP-bySNP basis, but across genes. And so, it really gives you the ability to validate a risk score on an FDA-approved platform,” says Levin. This means that if a test developer uses the Helix platform to help develop their product, it can help expedite the downstream FDA approval process.Levin believes that the regulation for these kinds of tests is going in the right direction, but can still be improved.
“This is an opportunity to begin to pivot and to bring the quality and the safety to all the patients that are involved, because while I absolutely believe in the LDT model, there’s also proven to be a lot of bad players out there. At some point there needs to be additional differentiation and this is an opportunity for that.”
Looking forward
It seems clear that PRS is here to stay. While some refinements are needed to make the scores suitable for broader rollout, they are already an important tool for modern medicine. Continuing decreases in the cost of sequencing and genotyping tools are also making these kinds of tests more accessible.
“We have clearly demonstrated across many different diseases that there is clinical validity and clinical utility. Over the next few years, we’re going to continue to see that get refined, we’re going to see it broaden out to additional populations that may not have been well represented from a genomic diversity standpoint,” says Levin.
Having appropriate oversight and ensuring a diversity of populations are represented in the training of these tests is of crucial importance to rolling them out to mainstream medicine and clinics around the world.
It’s also important to make sure tests are not being developed just for the sake of having a test, as polygenic risk scores are more appropriate for some types of disease prediction than others.
“One of the things that I think the field has to become really good about is to recognize what things do we actually think there’s also genetic component for, and what things are there not a genetic component for and we’re just trying to train a polygenic score, to try to find an association,” emphasizes Zhou.
Bottà agrees and also points out the importance of avoiding over-generalization. “The intended use of the polygenic risk score is very important. What is useful, in which population, and for what? The ‘which population’ is critical. The predictive power can decrease in a different ancestry group, so the polygenic score must have been assessed in the exact population where you claim it can provide any value.”
We are not quite there yet, but ultimately the aim of many companies and researchers developing these tests is to make them one routine factor that is tested and recorded when patients are assessed for disease risk.
“I think we are rapidly shifting into what I call a ‘genotype first’ world where instead of having people present with rare phenotypes we are going to be at a place where many people have had genotyping done and whole-genome sequencing done. And now we’re going to be using the genotype to try to predict your likelihood of experiencing as a phenotype,” says Zhou.
Tractor—Driving Towards More Diverse Genetic Tests
To make the best use of currently available datasets, a number of researchers are developing software to try and ensure that people of underrepresented ancestry are as accurately represented as possible.
Elizabeth Atkinson, Ph.D., based at the Broad Institute, is one such researcher. She has developed a software program called Tractor with her colleagues in the Analytic and Translational Genetics Unit at Massachusetts General Hospital.
“Admixed individuals, by which we mean people with recent ancestry from multiple continents, are often left out of large-scale genomics studies due to their complex genetic makeup. This means that we know more about the underlying genetic basis for traits and diseases in European-descent individuals than those of other ancestries,” she explains. “This is obviously not only not equitable, but feeds into the concerning health disparities that are currently observed across ancestries.”
This is important for the development of accurate polygenic risk scores, because most are trained on or based on large GWAS studies. To combat this disparity, Atkinson and colleagues wanted to develop a software tool that could allow admixed people to be easily included in large GWAS studies, which was how Tractor was born.
They are not the only ones using software-based approaches to try and improve the representation of different ancestries in GWAS studies. “Thankfully, there is a growing field of researchers working on building more diverse cohorts and novel methodologies that can better handle underrepresented and admixed ancestries,” says Atkinson.
“Different researchers have taken different strategies; here we focus on leveraging local ancestry—the ancestry of each specific genomic segment in each person in a cohort to control for population structure at an especially fine scale and boost our statistical power in GWAS studies on admixed participants.”
The Tractor program effectively ‘paints’ each GWAS participant’s genome to more accurately assess the impact of ancestry on any genetic associations that are picked up by the study.
“We estimate the effect of having a risk variant in each ancestry background, rather than just having a sort of aggregate measure. These ancestry-specific p values and effect size estimates are likely to be helpful in downstream efforts, such as the construction of better-performing polygenic risk scores for understudied and diverse populations,” concludes Atkinson, who is currently working to test these ideas with her colleagues. Tractor is open source and freely available for researchers who want to use it in their studies.