Quantifying the genetic separability of disease subtypes

Many diseases are not single, uniform conditions but rather collections of different subtypes, much like how "headache" can describe a tension headache, a migraine, or a cluster headache. Doctors often classify these subtypes based on symptoms, where the disease appears in the body, or how a patient responds to treatment. These distinctions matter because they can guide prognosis and treatment choices. However, a patient's symptoms do not always reveal their specific subtype clearly, and sometimes an incorrect classification can delay the right care. Scientists have long wondered if genetics could help sort these subtypes apart. Since every person carries a unique set of genetic instructions that influence their risk for disease, it seems logical that the genetic profiles of people with different subtypes might look different. The central question is whether these genetic differences are strong enough to reliably tell one subtype from another, or if the subtypes are so genetically similar that their DNA offers little help in distinguishing them.

A team of researchers has developed a new statistical tool to answer this question, moving beyond simple comparisons to measure the true potential of genetics to separate disease subtypes. They call their method GenSep. Instead of just testing how well current genetic tests work, GenSep calculates a theoretical maximum: the best possible discrimination that could ever be achieved if scientists knew the exact genetic risk for every single person without any error. This theoretical limit acts as a ceiling. By comparing this ceiling to the performance of existing genetic prediction tools, the researchers can determine if the problem is that the subtypes are fundamentally too similar to be separated by genetics, or if the current tools are simply not good enough yet.

The researchers applied this framework to eighteen pairs of disease subtypes using data from the UK Biobank, a massive medical database containing genetic and health information from hundreds of thousands of volunteers. They looked at pairs ranging from type 1 and type 2 diabetes to different forms of inflammatory bowel disease and thyroid disorders. For each pair, they first calculated the genetic ceiling. The results showed that for many subtypes, the ceiling was quite high. For instance, the genetic difference between type 1 and type 2 diabetes was so distinct that, in theory, a perfect genetic test could separate them with near-perfect accuracy. Similarly, the genetic profiles for hypothyroidism and hyperthyroidism, as well as Crohn's disease and ulcerative colitis, showed very strong separation. In these cases, the genetics clearly support the clinical distinction between the subtypes.

However, when the researchers tested the actual genetic prediction tools currently available, the results were starkly different. These tools, known as polygenic risk scores, aggregate the effects of thousands of genetic variants to estimate a person's risk. In practice, these scores performed poorly at distinguishing the subtypes. Across the eighteen pairs, the best current tools achieved a median accuracy that was barely better than random guessing. While the theoretical ceiling for separating type 1 and type 2 diabetes was nearly perfect, the actual tools managed to capture only a tiny fraction of that potential. In fact, for most of the eighteen pairs, the current tools recovered less than five percent of the genetic separation that theoretically exists. This gap reveals that the subtypes are often genetically distinct enough to be told apart, but our current methods are not yet precise enough to do it.

The study also investigated whether using larger datasets to build these genetic tools would solve the problem. The researchers tested tools trained on data from other large studies, including the FinnGen project in Finland and massive international consortiums with hundreds of thousands of participants. Surprisingly, even though these external datasets were two to four times larger than the data used to train the internal tools, they did not significantly improve the ability to distinguish the subtypes. The larger sample sizes did not translate into better discrimination. This suggests that simply adding more data is not the complete solution; the tools themselves may need fundamental improvements, or the genetic signals might be harder to capture than anticipated due to differences in how genetic patterns vary across populations.

To understand how much larger the studies would need to be to reach the theoretical limits, the researchers ran simulations to project future performance. For the three pairs with the highest genetic separation—type 1 versus type 2 diabetes, hypothyroidism versus hyperthyroidism, and Crohn's disease versus ulcerative colitis—they calculated the number of patients required to build a tool that could match the genetic ceiling. The numbers were staggering. To reach a high level of accuracy for type 1 versus type 2 diabetes, they would need roughly 29,000 training cases for each subtype. For hypothyroidism versus hyperthyroidism, the requirement jumped to about 161,000 cases per subtype. For Crohn's disease versus ulcerative colitis, the simulation suggested a need for nearly two million cases per subtype. These figures indicate that while the genetic potential exists, bridging the gap between current performance and that potential will require enormous amounts of data and likely better prediction methods.

The findings held up when the researchers tested them in other large cohorts, such as the All of Us program in the United States and the Million Veteran Program, which includes participants of diverse ancestry. The theoretical ceilings for most subtypes were consistent across these different groups and ancestries, suggesting that the genetic separation is a real biological feature rather than an artifact of a specific dataset. However, there were exceptions where the definition of the disease mattered. For example, the genetic separation between type 1 and type 2 diabetes appeared much lower in one dataset because the way the disease was recorded there mixed the two subtypes together, blurring the genetic signal. This highlighted that how a disease is defined and recorded is just as critical as the genetics themselves.

Ultimately, this work provides a clear map for the future of precision medicine regarding disease subtypes. It separates the problem into two categories: those where the genetic differences are too small to be useful, and those where the differences are large but our tools are currently too weak to detect them. For the latter group, which includes several major disease pairs, the path forward is clear. The genetic architecture supports distinguishing these subtypes, but achieving that in practice will require much larger studies and more sophisticated prediction models. The study does not claim that genetic testing will immediately replace clinical diagnosis, but it establishes a realistic benchmark. It shows that for some conditions, the genetic blueprint is ready to be read, we just need better readers.