In January 1977, five landmark health surveys, led by the famous Framingham Heart Study, reported a “striking” revelation about HDL cholesterol—a.k.a. the “good cholesterol.” The higher the HDL in a person’s bloodstream, researchers had found, the lower the risk of heart attack. This held true for every age group and both sexes. In fact, HDL was the only reliable predictor of heart disease risk in people over 50, which is the age group people who have heart attacks are likely to fall into.
In observation after observation ever since, the relationship between HDL cholesterol and heart health has been so outrageously robust it’s hard to imagine that HDL doesn’t play a fundamental role in preventing the disease process. This has led drug companies to spend billions of dollars developing and testing HDL-raising drugs, with the expectation that heart attacks will be prevented, lives saved, and investments recouped many times over.
And yet those drugs have universally and spectacularly failed. “Dashing hopes,” as a 2016 New York Times headline proclaimed, “a cholesterol drug had no effect on heart health.” The question is: why? One likely possibility is that despite all signs to the contrary, HDL cholesterol plays no mechanistic role in heart disease: that, put simply, the two have no causal relationship. Maybe high HDL doesn’t protect us from heart attacks. Maybe it’s just a sign of good heart health, a marker.
Anyone who follows the constant flip-flops in health news—what’s good for us this week, invariably, seems bad the next—may have gathered that epidemiology, the branch of medicine that searches for the causes of disease, is poorly equipped to resolve these kinds of fundamental questions. The problem is encompassed in four words that have justly earned cliché status: correlation is not causation. The fact that two phenomena or trends are correlated in time does not mean one causes the other. Arguably, the most important question in all of medicine and public health is how to tell which correlations are causal and which are not.
Now health researchers are wielding a new tool they hope will let them determine the true causes of chronic disease. And it comes through a surprise route: genetics. Researchers say that by employing innate genetic differences between people—an inborn susceptibility to alcohol, say, or to higher cholesterol levels in the arteries—they can now mimic, at much less effort and expense, the kinds of large trials that would be necessary to determine if an HDL-lowering medicine is really beneficial. The new technique, called Mendelian randomization, is already being used by drug companies to make billion-dollar decisions about which drugs to pursue.
Here’s how it works, using HDL as the example. At the moment of conception, some of us inherit specific variants of genes that boost our HDL levels. If HDL really protects against heart disease, then people with more of these HDL-raising variants should have lower rates of heart disease and live longer than those who get other variations. If so, it suggests that elevating HDL through drugs or diet is an excellent idea. Before randomly assigning people to different HDL-raising drugs or diets in huge, costly studies, the equally random lottery that determines which gene variants we inherit can be used to gauge whether such trials would be worth the risk and investment.
“We’ve all been recruited into an experiment, without knowing, it at conception,” says George Davey Smith, an epidemiologist at the University of Bristol in the UK, who has championed the new method as an invaluable tool for untangling causality and correlation throughout medical research.
The method is already settling long-standing and critical questions about heart disease—among other things, putting to rest 40 years of uncertainty about HDL. In 2012, a large international collaboration reported that despite the impressive correlation between HDL and heart disease, those of us born with genes that naturally raise our HDL cholesterol have no fewer heart attacks than people without those genes. In short, though HDL is inversely correlated with heart disease, it plays no causal role. And that’s why the drugs failed. They went after a target that turned out to be just a bystander.
Correlation is not causation
The prime movers in what might be a pharmaceutical-industry and public-health revolution are Davey Smith and his Bristol colleague Shah Ebrahim, physicians who bonded in the 1990s over their taste in rock music (Velvet Underground and Captain Beefheart) and, later, their shared disillusionment with epidemiology’s failure to zero in on the causes of common health problems. In 2000, Davey Smith and Ebrahim wrote a lengthy editorial for the International Journal of Epidemiology, asking whether the time had come for the entire field to “call it a day.”
They noted that epidemiology has an admirable track record in identifying the causes of infectious diseases from AIDS to Zika, in part because hypotheses can be tested in laboratories and in the field, and because governments have been easily convinced to do such tests. But chronic conditions like heart disease and cancer present an entirely different challenge. Epidemiology can provide correlations between diet, lifestyle, and disease, but that’s all it does. It generates hypotheses about possible causes, and little more than that. What concerned Davey Smith and Ebrahim was that epidemiologists had taken to jumping the scientific gun: they were advising people how to live and eat on the basis of mere hypotheses, without doing the rigorous (and very expensive) trials that might determine whether they were right.
In medicine, the experiments that rigorously test hypotheses are known as randomized controlled trials. Subjects are randomly assigned (often by the thousands or tens of thousands) to an intervention (a drug or diet) or a control (a placebo) and then followed for long enough (years, if necessary) to determine the effect unambiguously. In these trials, randomization is absolutely critical: it minimizes the chance that other characteristics of the subjects—behavioral, cultural, educational, socioeconomic—will influence the outcome. But such trials can be exorbitantly, if not prohibitively, expensive. They can be unethical. And like any scientific experiment, even when done with meticulous care, and after costing hundreds of millions of dollars, they can still get the wrong answer.
By 2000, Davey Smith had an inkling that the lottery of genetics could provide a solution. He had been studying the relationship between heart disease and an amino acid called homocysteine. High homocysteine levels, like low HDL, are often associated with heart disease, but is homocysteine a culprit or just an innocent bystander? Researchers were beginning to do clinical trials using folic acid (a B vitamin) to lower levels of the amino acid. Davey Smith realized that if people inherited higher or lower homocysteine levels naturally, this could help settle the causality question. If their heart disease risks were also different, that would be very telling.
What was exciting is that almost everything else about these two groups would be determined by chance, just as in a randomized trial. The randomization wouldn’t only hold true for the behaviors, exposures, and life stories that followed each person’s birth. It would also apply, in theory, to all other genetic influences. Each person with homocysteine-raising or homocysteine-lowering gene variants would bring to the experiment a genome that was otherwise thoroughly shuffled, since the combination of genes we inherit from our mother and father is a matter of chance.
The devil, Davey Smith knew, would be in the details. He realized there were many ways Mendelian randomization could fail. Just one is a phenomenon called pleiotropy—when particular genes affect more than one trait. Imagine, for instance, that homocysteine- or HDL-related variants also had subtle effects on intelligence. That could shape people’s socioeconomic status and overall health. It would “confound” the effort to isolate homocysteine’s or HDL’s role in heart disease by introducing a new causal factor.
In 2000, Davey Smith attended a conference with Ebrahim in northern India. Over the course of a four-hour taxi ride, Davey Smith briefed Ebrahim on his thinking. “He had to explain it three times before I actually understood what he was trying to say,” Ebrahim says. Three years later, the two Bristol researchers introduced Mendelian randomization in a 22-page article in the International Journal of Epidemiology, explaining not only how to use it but all the possible ways they could imagine to use it incorrectly.
Before their idea could be put to use, the field of genetics, too, needed to clean up its act. Researchers working to locate the genetic basis of human traits had been battling their own epidemic of spurious associations. Finally, though, studies became large enough (routinely involving hundreds of thousands of people) to unambiguously detect genetic variants that contribute to virtually every imaginable aspect of human existence, from physical characteristics and personality to risk of disease. Now, says Davey Smith, scientists have “literally tens of thousands of genetic variants” with which to work.
Today, Mendelian randomization may have settled some of the thorniest issues of cause and effect facing heart doctors—including not just HDL (not causal) and homocysteine (not causal) but also the LDL particles that carry "bad" cholesterol (causal) and even C reactive protein, a measure of the inflammatory process that got considerable media attention as a potential cause of heart disease and now, also, seems to be no more than an association.
Now the pharmaceutical industry is working with researchers from Bristol and elsewhere to predict the results of planned clinical trials before they’re complete, and so de-risk the exorbitantly expensive drug development process. This means not just predicting which drugs are likely to work or not, but how big an effect to expect and how large a trial is really necessary to find out.
“Companies are putting up hundreds of millions to test these drugs,” says Brian Ference, an interventional cardiologist and public health researcher who runs the Center for Naturally Randomized Trials at Cambridge University in the UK. If these companies choose the wrong target, or don’t use enough subjects—if the trial isn’t “powered” correctly—“they’ll not only lose their investment but lose years of work,” he says. “And without question that has been happening repeatedly in clinical trials in heart disease space for 40 years.”
Just too easy
As researchers realize the potential of Mendelian randomization to shed light on what’s truly causal and what’s just association, excitement is spreading to other disciplines with their own cause-and-association questions—even into fields such as social science, psychology, and economics. By identifying genetic variants that predispose individuals to be fatter or leaner, that influence how they metabolize alcohol and thus whether or not they are likely to become drinkers, researchers can begin to isolate the effect of these “exposures” on everything from depression, schizophrenia, and other psychopathologies to earning power and even academic achievement.
An explosion of studies—more than 250 so far this year—have tried to answer such questions about the human condition, sometimes using hundreds of gene variants tied to the trait being studied: whether being taller, for instance, causes men to earn more (apparently yes, with an extra two inches being worth $1,000 to $2,000 a year), whether neuroticism leads to smoking or vice versa (still unclear), whether wearing glasses turns people into bookworms (apparently the opposite—it’s the studying that causes myopia), and whether cannabis use can contribute to schizophrenia (leaning toward yes).
Davey Smith, Ebrahim, and their fellow pioneers of Mendelian randomization now seem as worried about how the technique will be misused as they are excited about its promise. Even when it’s done right, says Davey Smith, it’s easy to overinterpret results. For instance, Mendelian randomization can tell researchers whether a gene variant working since the moment of conception increases or decreases disease risk—the lifelong impact of that gene. But that’s a different question, says Davey Smith, from “whether taking a drug or changing your diet [at] age 60” is beneficial. Finding that out would still require a clinical trial.
What may worry Davey Smith and others most is that as genetic databases have multiplied, tying genes to virtually any imaginable biological or even behavioral variable, studies of cause and association have become almost effortless.
The University of Bristol hosts a platform called MR-Base that lets anyone carry out virtual experiments without collecting any new data.
“You can do these studies now, sitting at your desk, in 10 minutes. It’s just too easy,” says Ference. “Because of the flood of studies coming out, it may very well fall into disrepute.”