Polygenic prediction of treatment response for precision medicine is vulnerable to collider bias

Imagine a world where doctors could look at your DNA and know exactly which medicine would work best for you. This is the promise of precision medicine, a field that aims to move beyond the "one-size-fits-all" approach to healthcare. Central to this vision are polygenic risk scores, which are essentially a single number calculated from thousands of tiny genetic variations. This number estimates a person's likelihood of developing a disease, such as heart disease, long before any symptoms appear. Because these scores can predict risk so well, researchers have begun using them to design clinical trials, hoping to find patients who are genetically destined to benefit the most from a specific treatment. If a drug works wonders for people with a high genetic risk but does little for those with a low risk, it would be a massive breakthrough, allowing doctors to target therapies with surgical precision.

However, a new study suggests that the excitement surrounding these genetic predictions might be based on a statistical illusion. The researchers, led by Matthew Tudball and colleagues, investigated a pattern seen in several major trials for cholesterol-lowering drugs. In these trials, people with higher genetic risk scores for heart disease appeared to get significantly more benefit from the medication than those with lower scores. This finding was so compelling that it is currently driving the design of new, expensive trials intended to treat only the high-risk group. The authors of this new study propose that this dramatic difference in benefit is not a biological reality, but rather a trick of the numbers caused by how the trials selected their participants. They argue that the rules for entering the trial created a hidden bias that made the genetic score look like a predictor of success when it was actually just a marker for other factors.

To understand the problem, one must look at how these trials were built. To ensure the studies were efficient and safe, researchers set strict rules for who could join. For example, a trial might only accept people who already had heart disease and still had high cholesterol despite taking medication. In the general population, a person's genetic risk for heart disease is independent of their other risk factors, such as smoking or diabetes. But inside a trial with strict entry rules, this independence breaks down. If a person has a low genetic risk, they usually need a heavy burden of other bad habits or conditions to get sick enough to qualify for the trial. Conversely, a person with a high genetic risk might qualify even if they have very few other risk factors. This creates a strange, artificial relationship within the trial group: the higher the genetic risk, the lower the burden of other risk factors.

This artificial relationship is the heart of the issue. The researchers found that the other risk factors—like smoking or diabetes—actually do change how well a person responds to treatment. Because the trial rules forced a link between genetics and these other factors, the genetic score began to act as a stand-in for them. When the high-risk group showed better results, it wasn't because their genes made the drug work better; it was because they happened to have fewer other risk factors that might have dulled the drug's effect. The study shows that even when a drug is assigned randomly, as it is in a proper clinical trial, this selection bias can distort the results. The randomization balances the treatment groups, but it cannot undo the bias introduced by the very criteria used to let people into the study in the first place.

The team tested this idea using computer simulations. They built a virtual world where they knew for a fact that genetic risk had zero effect on how well the drug worked. They then applied the same strict entry rules used in the real-world trials. The result was startling: the simulations produced the exact same pattern of "genetic benefit" that the real trials reported. In these fake trials, people with high genetic risk appeared to have a hazard ratio of 0.77 for major vascular events compared to 0.95 for those with low genetic risk, purely because of the selection bias. This happened even though the drug worked exactly the same for everyone in the simulation. The study also looked at real-world data from over 175,000 people in the Helix Research Network and nearly 93,000 in the UK Biobank. In these large groups, where people were not selected by strict trial rules, there was no evidence that genetic risk changed how well statins worked. But when the researchers artificially restricted these groups to look like the trial participants, the fake "genetic benefit" suddenly appeared again.

The implications of this finding are significant for the future of personalized medicine. The study does not suggest that polygenic risk scores are useless. They remain powerful tools for predicting who is likely to get sick, which helps in planning trials to ensure enough people have events to study. However, the authors warn that using these scores to predict who will respond best to a drug is far more dangerous. The apparent success of targeting high-risk patients in past trials may have been a statistical mirage. If new trials are designed based on the assumption that high genetic risk guarantees a better response, they may be built on a false premise. The researchers conclude that what looks like a biological interaction between genes and drugs may often be nothing more than a side effect of how patients are chosen for the study. Before doctors can confidently prescribe treatments based on genetic profiles, they must first ensure that the rules of the game are not rigging the outcome.