By Matt Phillion
Conditions like preterm birth, preeclampsia, and gestational diabetes impact 45 million women every year across the world. That’s triple the rate of cancer, yet we lack a reliable way to identify patients who are at risk.
A team of physicians are working on groundbreaking research into pregnancy-related cell-free RNA (cfRNA) to develop a simple blood test that would enable early detection of these conditions. Drs. Michal Elovitz, Thomas McElrath, and Stephen Quake are working alongside Mirvie on this test, which uses a small blood sample to analyze tens of thousands of molecular transcripts from the mother, baby, and placenta. Made possible with the help of state-of-the-art machine learning, the test aims to help detect conditions earlier and therefore promote better outcomes for mothers and babies.
“The motivation is simple: Our ability to treat patients in obstetrics is behind other fields,” says McElrath, maternal-fetal medicine attending at Brigham and Women’s Hospital and faculty at Harvard Medical School. “If you look at the molecular therapeutics available in oncology or cardiology, big subfields of internal medicine, it’s pretty impressive. That same advancement is not shared by obstetrics.”
The field has improved in treating conditions like preterm birth and preeclampsia, but the ability to predict who’s at risk is “ripe for exploration and innovation,” says McElrath.
There are any number of reasons why such innovations haven’t been explored as much as other fields: The latent inequality in women’s health is one, but the innate caution about doing research on pregnancy and pregnant patients is another.
“We’re beginning to realize that that’s contributed to the problem as well,” says McElrath.
Access to samples
Another limitation that had slowed down previous research was a lack of appropriate samples for testing.
“There’s not a lot of blood or biobanked specimens from pregnant patients,” explains McElrath, who became involved in part because he runs a biobank in the Boston area. Between his organization and several others who have become involved, the research now has access to several thousand samples. “Once you start to have those kinds of numbers, you can start to look at statistically significant variations. If you’re talking about relatively uncommon conditions in relatively small numbers, it’s hard to predict a statistical fluke, but once the sample numbers get high enough, you’re able to get a real signal instead of random noise.”
An important factor to this project is that researchers are not just looking at tissue samples from the placenta, but also equally important transcripts from the mother’s side.
“The mother is not neutral in this, although that has been an implicit assumption in past research,” McElrath says. “The signal we’re getting from the mother may be a signal of a pending pathology. We’re looking at both placental and maternal responses.”
This is particularly important because of how singular human pregnancy is. Whereas other areas of medical research can make use of other mammals for testing, obstetrics cannot.
“Human pregnancy is unique among all mammals—we don’t have a lab animal that is uniquely demonstrative of what’s happening in humans,” says McElrath. “Humans have a very particular form of placental development.”
And because of this, researchers are seeing and learning about many things for the first time.
“It’s about developing a test, but also the underlying biology in a system that is hard to approach,” says McElrath. “We’re just beginning to come to terms technologically with how we conduct research, even if it’s just observational research. There’s still a lot that is unique to humans—no other species gets preeclampsia, for example.”
The research has led to some interesting discoveries the team is following up on.
“One piece that validates some of what we’ve suspected is that many of the pathologies we are observing clinically really have their roots much earlier in pregnancy,” says McElrath. “No one in the field would be surprised at that, but it is a renewed, constant reminder that obstetrics deals with diseases as they present, but the disease starts much earlier.”
This opens the door to developing prophylaxis and more productive therapeutics, given that the research helps identify who is heading into these risk conditions, he says.
“This wasn’t really possible earlier,” says McElrath. “For these patients who are high risk, we can start looking at other therapies. The ability to stratify these conditions earlier fosters that initiative.”
Starting earlier means the opportunity for even more innovation in treatment.
“Once you start to be able to identify who is at risk, once you start to understand what pathological systems are at play, you begin to generate insights into what to do next,” McElrath notes.
It’s almost a self-obvious statement, he explains. We understand, for example, that cancer begins development before it is clinically diagnosed. But this research helps clarify the lead time involved, enabling treatment weeks or months ahead of the traditional diagnosis.
“Innovations like this one allow highly personalized prenatal care,” says McElrath. “It’s looking at the intersection of molecular genetics and histological patterns to develop a unique care plan.”
Obstetrics traditionally treats patients with the same broad strokes, McElrath says. But imagine how things could change over the next few years: If a patient is predicted to be at risk for an outcome, they could then be funneled to an appropriate clinic or connected to a clinical trial. A patient who is predicted not to be at risk, meanwhile, could have the freedom to work with the family practitioner or midwife of their choice.
The machine learning component
Machine learning will come into play in a number of ways with testing and analysis of samples, McElrath says. For one, it will assist with sorting through thousands of transcripts in ways human researchers could not.
“There’s a process to this type of work—classical statistical inquiry assumes you have many more times the number of subjects than you have variables,” he says. “With a lot of this research, it’s the opposite. You have thousands of variables but only a couple hundred or a few thousand subjects.”
Thus, you have more variables than subjects to consider, or a “p greater than n problem,” and classical statistics breaks down in that circumstance. “You basically overwhelm the system,” says McElrath.
To prevent this, you need techniques that will start to identify which of the variables are most likely to yield a true signal, and at the same time, iterate them to try to identify whether they consistently display that predictive capability.
“That’s a lot of what you find in machine learning,” says McElrath. “We’re looking for pattern recognition among the proteins and other samples to try to reconstruct if those patterns can be predictive given the limited number of subjects.”
A lot of common techniques with machine learning come into play here. The computer doesn’t care whether it’s looking at proteins, faces, or sounds—it comes down to numbers.
“As these tests become deployed and associated with clinical outcomes and other variables, we can cycle back and improve the predictive capability of individual biomarkers,” says McElrath. Truthfully, the process “isn’t that complicated when you break it down. Machine learning is elegant in some of its techniques. We’ve been beholden to classical statistics for so long, but it’s not so mystifying as people tend to believe when they hear about artificial intelligence.”
Where the research leads
While the team is using 21st-century analytics, bodies like the FDA will require additional steps to demonstrate value in terms of how the test is run and what it predicts.
“These techniques are fascinating, but the bottom line is, does it work if you bring in a new population under analysis?” says McElrath. “It’s going to take another investigation using real-world subjects.”
It’s also important to make the expectations clear, he explains.
“A lot of folks are looking for a test for conditions of pregnancy with an analog to testing for infectious disease, but in those cases, because you’re looking for something unique or different, it’s easy” to test for, McElrath says. “But for the conditions we’re looking for in pregnancy, you’re looking at underlying biomarkers. What these tests are going to look like is more along the lines of what we have in the cancer field. You have a biomarker increased, it needs to be followed, and additional downstream observation is needed. We’re doing a disservice if we say it will be a simple yes/no that you’re going to have a preterm birth.”
The conditions being investigated in pregnancy health do not lend themselves to yes/no answers, but rather a search for indicators to help predict future potential risks.
“Nobody expects big screening tests like mammograms to be yes or no, but we use them and they are informative,” says McElrath. “What do we have now? Pregnancy history, which is a poor predictive test, and it does not help first-time pregnancies. We use familiar solutions, not better solutions.”
McElrath says that now is the time for the field to take advantage of advancements that could lead toward better care.
“We can do better. A lot of fields are all doing better,” he says. “Obstetricians need to avail themselves of these innovations that we’ve been a little cautious in adopting.”
Matt Phillion is a freelance writer covering healthcare, cybersecurity, and more. He can be reached at firstname.lastname@example.org.