When do protein-truncating variants and copy-number loss agree? A pre-specified, power-aware genome-wide audit of cross-perturbation effect equivalence in 885,362 gene-trait pairs

In the quest to understand how genes influence human health, scientists have long relied on a powerful idea: if you can see what happens when a gene is broken, you can guess what happens when you try to fix it. This logic underpins much of modern drug development. Researchers look for natural variations in the human population that act like accidental experiments. Sometimes, a single letter in the genetic code is changed, effectively breaking a gene's instructions. Other times, a larger chunk of DNA is missing entirely, deleting the gene and its neighbors. For years, the assumption has been that these two types of damage—breaking a gene with a tiny error versus deleting it with a large gap—should produce the same result. If a broken gene causes high cholesterol, then a deleted gene should do the same. This belief allows scientists to treat different types of genetic data as interchangeable maps, using them to predict which drug targets will be safe and effective.

However, a new, massive audit of human genetic data suggests this assumption is far more fragile than previously thought. A team of researchers, led by Beiji Lu, set out to test whether these two ways of damaging a gene truly tell the same story. They did not just look for patterns; they built a rigorous, pre-planned system to check if the effects were actually equivalent, or if the data was simply too noisy to tell the difference. By examining nearly 900,000 pairs of genes and traits, they discovered that for the vast majority of cases, the data is silent. More importantly, they found that the few cases where the two types of damage do agree are often just empty signals, not evidence of a real biological effect. The study concludes that we cannot yet use these genetic maps interchangeably to guide drug development, because the tools we have are not precise enough to distinguish between a true biological match and a statistical coincidence.

The researchers approached this question with the caution of an auditor rather than a hunter. Instead of searching for a few exciting discoveries, they designed a three-stage process with strict rules written down before they looked at the data. Their goal was to create a complete inventory of where protein-truncating variants—tiny errors that break a gene's code—and copy-number losses—large deletions that remove a gene—agree on their effects. They used data from the UK Biobank, a massive collection of health and genetic information from hundreds of thousands of volunteers, along with an external dataset from AstraZeneca to cross-check their findings. The team focused on 57 different physical traits, ranging from blood pressure to bone density, and looked at how nearly 16,000 genes influenced them.

The first stage of their audit revealed a sobering reality: the data is incredibly sparse. When they looked for genes where both types of genetic damage showed a clear, significant effect, they found very few. Out of hundreds of thousands of possibilities, only a tiny fraction provided enough evidence to say anything with confidence. The researchers had hoped to find at least 100 genes where they could see a clear, three-way agreement between different types of genetic evidence, but they found only 14. This lack of data forced them to narrow their focus to a simpler question: do the two types of damage agree on the direction of the effect, or do they disagree?

As they moved to the full genome-wide analysis, the results became even clearer. Of the 885,362 gene-trait pairs they examined, the overwhelming majority—more than 880,000—were classified as "information-insufficient." This means that for most genes, the data was simply too weak to say whether the two types of damage were the same or different. It was not that the effects were different; it was that the signal was too faint to measure. After applying strict statistical corrections to ensure they were not seeing false patterns, they identified 4,252 pairs where the effects were statistically equivalent and 216 pairs where they were clearly different.

Here lies the critical finding that challenges current drug development strategies. The researchers needed to find at least 200 gene-trait pairs where both types of damage were active and pointed in the same direction to prove that the two genetic maps were interchangeable for drug targeting. They found only nine such cases. The vast majority of the "equivalent" pairs were actually cases where neither type of damage had any detectable effect on the trait. In other words, the maps agreed because both showed nothing happening, not because both showed the same thing happening. This distinction is vital: a drug developer needs to know that breaking a gene causes a specific change, not just that breaking it doesn't cause a change.

The study also investigated why the few cases of disagreement existed. They found that genes showing different effects from the two types of damage were almost always located in regions where large chunks of DNA were deleted, often taking out multiple genes at once. This suggests that when a large deletion occurs, the effect is not just about the single gene being studied, but about a cluster of genes being removed together. This "passenger" effect, where the deletion of one gene drags along the effects of its neighbors, creates a confusing signal that a single-gene analysis cannot untangle. The researchers found that genes with these conflicting signals were covered by deletions that removed a median of 10 protein-coding genes, compared to only 2.5 for the genes where the signals agreed.

The team also tried to build a computer model to predict when these two types of genetic damage would disagree, hoping to create a tool that could warn researchers when a drug target might be misleading. They fed the model data about the precision of the measurements and the structure of the genetic deletions. The model worked almost perfectly at separating the different cases, but not because it learned biology. It worked because it learned the limitations of the data itself. The model could tell the difference between a strong signal and a weak one, but it could not tell the difference between a true biological effect and a statistical fluke. Because the "labels" for the model were derived from the same noisy data used to train it, the model was essentially memorizing the noise rather than discovering a rule.

Ultimately, this study serves as a necessary pause in the rush to use genetic data for drug discovery. The researchers did not find a new way to predict drug success; instead, they found that the current method of assuming all genetic damage is the same is not supported by the evidence. The inventory they created is a map of where the data is silent, where it is clear, and where it is misleading. They conclude that to move forward, scientists need new types of data that can isolate single genes from their neighbors and separate true biological effects from the noise of large deletions. Until such data is available, the assumption that a broken gene and a deleted gene are interchangeable remains unproven, and perhaps unprovable with the tools currently at hand. The value of this work lies not in a breakthrough, but in the clarity of its negative result: it tells the scientific community exactly where the road ends and what kind of new vehicle is needed to continue the journey.