In August 2020, the first American child was born after using a new kind of genetic test in hopes of reducing the risk of heart disease. In vitro fertilization offered her father, a physician, the opportunity to use pre-implantation testing to select the embryo that, according to this test from Genomic Prediction, had the lowest risk of developing heart disease and some other conditions. Heart disease doesn’t have a single genetic cause, and it also has environmental factors. But this new test uses polygenic risk score, or PRS, models to predict an embryo’s lifetime risk of diseases and conditions (like heart disease and schizophrenia) that are only partly caused by genetics.
Embryo testing is not new for parents going through IVF, and embryos are often screened for genetic diseases, but parents can now get this kind of genetic test offered by other companies like Genomic Prediction, such as Orchid and MyOme. However, genetics professionals (including me) have concerns about this high-tech solution to “have healthy babies,” as Orchid’s tagline puts it.
For many couples, the IVF journey can be long, fraught, and expensive, averaging about $12,000 to $17,000 per cycle (and a pregnancy may take multiple cycles). After fertilization and before implantation, couples can opt to have their less-than-week-old embryo screened with a genetic test. This typically identifies disorders caused by a single gene (like Huntington’s disease) or a whole chromosome (Down syndrome). Now, for approximately $1,100, they can add on a newer and much less understood genetic test: a PRS test.
At its core, a PRS model is fancy statistics applied to basic genetics. Researchers produce the model by comparing genomic data between people with and without some condition (or trait). But they aren’t comparing genes. Instead, they compare common genetic changes called single-nucleotide polymorphisms—SNPs for short, pronounced “snips.” These SNPs are single-letter differences in the DNA, whereas a gene is between several hundred to more than 2 million letters long. A gene codes for a protein, but SNPs may slightly alter how a gene is translated into a protein, or how the resulting protein functions. When studied in large groups, SNPs show small but significant correlations that suggest they influence inherited conditions. Using advanced statistics, researchers create a model using anywhere from dozens to millions of SNPs that influence the risk of a condition, adding them together for a composite score—the PRS.
For example, I might use a colorectal cancer PRS test and learn I have a lifetime risk of 5.1 percent, very close to the average American’s lifetime risk of 4.1 percent. But my friend might learn his lifetime risk is 0.4 percent. My genetic risk would be essentially “average,” while my friend’s risk would be significantly below average. However, most people will have near-average risks when using a PRS test, perhaps a little higher or lower but not significantly different, as few people have a set of SNPs conferring very high or very low risk. In theory, this risk information can then help make health care decisions—more frequent colonoscopies, for example—or screen embryos.
Hype for using PRS models in health care has been building for years. In 2018, researchers created a model using more than 6 million SNPs to stratify risk for coronary artery disease. Although this was an achievement, a follow-up study adding this PRS model to conventional risks (such as age) found that the model failed to predict heart disease any better than conventional risks did alone. This doesn’t mean the coronary artery disease PRS didn’t work; it’s more likely that the PRS model didn’t capture enough of the genetic risk to make a meaningful difference in the total risk score. Researchers expect to discover more SNPs that influence coronary artery disease in the future, which they can use to improve PRS models. But insufficient knowledge is a problem for essentially all current PRS models.
Another problem? “We’re not sure why they work,” said Peter Kraft, genetic epidemiologist at Harvard’s T.H. Chan School of Public Health. “We don’t know what these [SNPs] are doing biologically. Many [SNPs] are pleiotropic, so they have more than one effect. Pick the allele that reduces your risk of breast cancer and maybe that’s also increasing your risk of Type 2 diabetes.”
Indeed, at least two studies have found higher scores on measures of creativity to be correlated with higher schizophrenia PRS. Though this link isn’t definitive, it does raise the question as to whether selecting an embryo for low schizophrenia PRS might limit creative potential. And definitively proving any link between SNPs and a condition is going to be daunting, as the classic genetics approach simply doesn’t work at scale.
The typical approach to proving a genetic variation causes a condition is to create an organism or cellular model with the variant in an experiment, then take careful measurements to assess the relationship between the condition and the engineered and original versions of the cell or organism. But how do you experimentally measure a minute risk conferred by dozens, or millions, of SNPs? And what if the risk occurs only in combination with other SNPs, or in the context of certain environmental factors?
With all of these unanswered questions, the usefulness of current PRS models is dubious. Kraft used schizophrenia as an example: “The back-of-the-envelope calculations, if you take your average couple and you compare the average risk of their child getting schizophrenia versus screen five embryos and pick the embryo with the lowest risk, you’re talking about going from a 1 percent lifetime risk of schizophrenia to a 0.6 percent life risk of schizophrenia. So, rounding up, that’s a 1 percent risk of schizophrenia. Even if a parent has schizophrenia, you’re talking about going from 2 percent risk to 1.6 percent risk.” Put another way, that embryo’s risk of not getting schizophrenia went from 98 percent to 98.4 percent. “It’s unclear if, for many parents, that’s a huge game changer,” he said.
That these tests are already in use worries Hannah Wand, a genetic counselor and director of the preventive genetics program at Stanford University. “We have a tendency in genetics to put out tests when they’re technically feasible.” She thinks PRS testing is going to lead to a paradigm shift in health care. “We have not yet figured out how genetics integrates with routine life and public health. And insurance companies—if everyone has a PRS, is this going to be misinterpreted?” In theory, a high PRS could be misinterpreted as a preexisting condition.
And then there are ethical concerns. These tests teeter on the edge of eugenics. Lowering risks for diabetes is good, but without regulatory safeguards, future tests could, in theory, report on traits like predicted intelligence. And there are ethical issues that already exist for today’s PRS models. “This is an intervention that is never going to be available to everyone,” said Leila Jamal, bioethicist and associate director for cancer genomics at the Johns Hopkins Bloomberg School of Public Health Genetic Counseling Training Program. She meant cost, but also social equity due to a quirk of current PRS models: They are most accurate in people from whom the original model was created. “The models mostly account for European populations,” she added. So that is who will have the most accurate polygenic risk scores.
A recent report analyzed 733 studies used to create polygenic risk scores from 2008 to 2017 and found 67 percent included participants of only European ancestry, while just 3.8 percent of the studies included people of African, Latino/Hispanic, or Indigenous ancestries. When the researchers applied 25 PRS models derived in mostly European populations to individuals of African ancestry, the scores had lower predictive ability by 68 percent on average. This point was illustrated for 11 conditions in an elegant paper recently published in the New England Journal of Medicine. Lifetime risk reductions for people of European ancestry were often about two times higher than for those of African ancestry. Genevieve Wojcik, genetic epidemiologist and assistant professor at Johns Hopkins University Bloomberg School of Public Health, takes it a step further, saying that PRS models “capture societal elements that aren’t necessarily biological … such as pollution, access to healthy foods, or education”—all of which may interact with our biology and influence our risks.
To this point, the genetic testing behemoth Ambry suspended its cancer PRS tests in May due to “limited data across ethnic populations.” Will companies like Orchid or Genomic Prediction follow suit? Noor Siddiqui, Orchid’s CEO, told me via email that she couldn’t answer questions as “the embryo screening product is not available right now,” suggesting a pause in its testing not described on its website. When I emailed Jennifer Eccles, head of genetic counseling at Genomic Prediction, she said, “We provide results only where we have validated relative risk reduction for that particular population, for that particular disease.” She noted the company does provide a PRS for certain conditions in people of African and Asian ancestry. “It is a continuous effort of ours to obtain more genetically diverse datasets for this purpose.” Translation: Genomic Prediction has parents of non-European ancestries in mind, but they will have to wait for the same level of information as people of European backgrounds.
These tests aren’t necessarily bad or inaccurate. Rather, the science is incomplete—a puzzle with many missing pieces. Beyond a company’s pictures of parents smiling at chubby babies, what’s important to understand is that current PRS tests are of limited use, and most babies are born healthy despite risks for many conditions. Before using a PRS test for embryo screening, Jamal urges prospective parents to ask themselves: “What are my goals, and what problem am I trying to solve by learning this information? What am I worried about that this will help me prevent? Is this [test] the right tool for that job?” Correlation has never been causation; a crystal ball this is not.