Stephen Hsu’s startup Genomic Predictions analyzes genetic data to predict the chance of diseases like diabetes and cancer — and forecast IQ
For years, hopeful parents pursuing in vitro fertilization (IVF) treatment have had the option of screening embryos for severe heritable diseases like cystic fibrosis, hemophilia, and Tay-Sachs disease. These rare and often deadly conditions, known as monogenic disorders, can be easily identified through genetic screening because they arise due to a mutation on a single gene. For doctors, diagnosis is a simple positive or negative.
But the diseases that are most likely to shadow the average person’s life — cancer, heart disease, diabetes — are polygenic, meaning that they result from interactions between thousands of genetic signals. In the past, this has made these diseases — which kill millions of Americans each year — all but impossible to screen for with genetic tests.
But Genomic Prediction, a New Jersey-based company that analyzes genetic data using machine learning, is hoping to change that. Taking advantage of the new troves of genetic sequences that have accumulated over the past decade, the company is offering what is known as polygenic risk scores, a screening process that attempts to establish the statistical probability of a person developing diseases like diabetes or hypertension throughout their life.
It’s easy to imagine that this new technology could be the first step in a dystopian science fiction scenario: designer babies.
More controversially, however, Genomic Prediction is also offering IVF patients the option of screening embryos for projected cognitive ability. While the company says that at this stage it will only inform parents about the risks of potential intellectual disability — defined as 25 points below the average IQ — it’s easy to imagine that this new technology could be the first step in a dystopian science fiction scenario: designer babies.
Genomic Prediction was co-founded in 2017 by Stephen Hsu, a professor of theoretical physics at Michigan State University. A decade ago Hsu noticed that the continually decreasing cost of genetic sequencing and genotyping was so steep and rapid that it might allow him to solve a fundamental problem that had fascinated him since childhood.
“When I first learned about evolution in biology class, I became interested in a particular mathematical, theoretical question,” Hsu told OneZero. “If traits are heritable, and you have access to all of a person’s DNA, how much can you predict about them?”
Starting in 2010, Hsu began carving out some time from his regular research schedule, and along with a team of colleagues, began thinking about how to build tools for predicting complex traits from DNA. They decided on a machine learning approach, training an algorithm on data comprised of genotypes (the genetic makeup of a person), and corresponding phenotypes (their physical characteristics). They believed that once trained, the algorithm would be able to predict the likelihood of a complex trait by identifying a combination of genetic signals in the DNA.
In 2017, Hsu and his collaborators released a report showing that they could successfully use this approach to predict height from a genotype within roughly one inch. That same year, Hsu co-founded Genomic Prediction and quickly began using the technology to predict the risk from polygenic diseases.
Now the company offers polygenic risk scores, which assign a probability to diseases that might arise in a customer given their genetic makeup. Unlike monogenic disorders like cystic fibrosis, “we can’t just look at a person and say yes or no, you’re going to have breast cancer,” Hsu told me. “But what we can do is tell someone that they are in the 99th percentile of people who have risk for breast cancer, which might mean that they have a 50% chance of developing it in their lifetime.”
At this stage, Genomic Prediction is offering polygenic risk scores to help prospective parents using IVF — which usually involves the creation of multiple embryos — to select those embryos that increase the probability of having a healthy child. But Hsu sees polygenic screening becoming a part of routine medical checkups. “At some point in the near future, when you go for your physical, your doctor will ask you for blood to get a genotype for you,” he said. “And then the doctor will get a report from us that will inform them on how best to look after you.”
While Hsu believes that polygenic risk scores will become part of standard clinical health care within a few years, other experts have their doubts. Cecile Janssens, professor of epidemiology at Emory University who has been studying genetic prediction for over a decade, told me that she was skeptical of how much meaning Hsu assigns to the output of these algorithms. “Prediction is extremely difficult, and probabilities are hard to interpret,” she explained. “Would you not choose an embryo because it’s risk of getting breast cancer was increased from 8% to 14% or its risk of diabetes was 18% instead of 11%? Would you even know what that all means?”
Hsu acknowledges that at this stage, polygenic risk scores are most useful in assessing statistical outliers — those cases where an embryo is far more likely to develop a certain disease than other embryos — but expects significant improvement as more genomic data is collected. “For some diseases, we only have between 10,000 to 100,000 cases, but if we start to build larger and larger databases, that could cause a significant material advance in the quality of the predictor,” he explained.
More controversial is the fact that Genomic Prediction’s polygenic risk analysis includes a panel for cognitive ability. According to Catherine Bliss, a professor of sociology at the University of California San Francisco and the author of two books on genomics, predicting for cognitive ability turns on the presupposition that intelligence is a highly heritable trait, which is, to say the least, a historically fraught and controversial claim.
As decades of research has shown, the science of cognitive heredity is always in flux, mostly due to the complex way intelligence interacts with our social environments, including education. There are also problems with standard intelligence metrics like IQ tests, which have been shown to be culturally and racially biased. “I don’t think predicting for cognitive capacity is a good idea until we have a better grasp on this gene-environment relationship,” Bliss told me.
Hsu has been embroiled in controversy concerning genetics and IQ before, due to his work for a Shenzen-based research institute, BGI, which allegedly mapped the genome of unusually intelligent people in order to try and isolate an intelligence gene. But he told me that at this stage, the company’s polygenic test for intelligence will only indicate which embryo is a genetic outlier in terms of cognitive impairment, which correlates to an IQ 25 points below average. “If you’re a parent going through IVF and five embryos are fine, but one of them has scored in the bottom percentile for a cognitive score, I think that you deserve the option of knowing this,” he said.
When it comes to the social implications of this technology, Hsu is more concerned about accessibility and how these technologies will exacerbate class inequality. In addition to being able to give their child a private education and “even buy their way into Stanford,” Hsu says, wealthy people may soon be able to design their children better genes. While Hsu worries that a kind of genetic caste system looms on the near horizon, he hopes that predictive technology might also be used for socially progressive ends, or what he calls a “redistribution of genetic endowments.”
“Maybe let people from disadvantaged groups select from 20 embryos, instead of seven,” he said, which presumably raise the possibility of getting a healthy embryo that scores high on cognitive ability. “If this was done over a long period of time, maybe it would eventually catch that group up.”
Underpinning this vision is Hsu’s fundamental belief that genetics can be used to not only predict and select for traits in individuals, but to engineer positive social change. When I suggested that what he was arguing for amounted to an updated version of eugenics — the attempt to “improve” humanity scientifically through selective breeding — Hsu agreed, but distinguished between the “coercive” genetic engineering of the past and “opt-in” vision he was proposing.
“If I tell a couple that they are carriers of a disease and that there is a one in four chance that their kid is going to die a horrible death, and allow them to select the healthy embryo, is that eugenics? I guess it is,” Hsu said. “But if I give them the option to make to make a choice, and don’t coerce them, I don’t think there’s ethically anything wrong with it.”
“I can predict a situation where the government of Singapore approaches us and say that they want to be able to rank order embryos by intelligence… I could imagine a scenario where we say okay.”
While few would argue that selection in some circumstances is beneficial — and though it has long been the practice for prospective parents to screen for severe genetic disorders like cystic fibrosis and muscular dystrophy — in the U.S., embryonic screening is largely unregulated and takes place in the private IVF sector. Legally there is nothing stopping Hsu from adding cosmetic traits like hair color, height, and skin tone to the polygenic screening. He could also offer parents the option to test for high IQ and heritable personality traits like a predisposition to violence.
At this stage, Genomic Prediction, a company that raised $4.5 million in private investment this January, does not offer such services. When I asked Hsu how he draws that moral line, he replied that the company tries not to get “too far out in front of American society.”
But Hsu believes that attitudes around prediction and selection will change over time, as they already have, and points out that they already vary between cultures. In South and East Asia, for instance, there is more cultural acceptance of the idea of selecting for embryos in order to enhance desirable characteristics like intelligence, rather than just avoid genetic diseases.
“I can predict a situation where the government of Singapore approaches us and say that they want to be able to rank order embryos by intelligence,” Hsu told me. I asked him to predict how he might respond to such a request. “I could imagine a scenario where we say okay.”
Update: This story has been corrected to indicate Stephen Hsu’s correct title. He is a professor of theoretical physics.