AI system predicts amyotrophic lateral sclerosis (ALS).
Diagnosehelfer achieved an 87 percent hit rate based on genetic factors alone
Early diagnosis: Researchers have developed an AI system that can use a DNA sample to predict the risk of amyotrophic lateral sclerosis (ALS). In a test with 3,000 ALS patients and 7,000 healthy controls, the artificial intelligence achieved a hit rate of almost 87 percent. The system also identified 922 gene variants that play an important role in ALS. This could help to find out more about the genetic causes of the neurodegenerative disease.
Amyotrophic lateral sclerosis (ALS) affects around six to eight people per 100,000 in Germany, making it the most common degenerative disease of the motor nervous system in adults. Because the nerves responsible for muscle control are increasingly losing their function, those affected suffer from progressive muscle atrophy and paralysis. In the late stage, swallowing and breathing are also affected. There is no cure, but starting therapy early can at least slow the progression.
Complex genetics
The problem, however, is that ALS is often only recognized relatively late – also because the causes have only been partially clarified. Studies suggest that at least half of the disease is due to hereditary factors. However, the risk genes identified so far through genome-wide comparative studies only explain around ten percent of this heredity. So far, this has made it almost impossible to predict whether a person with a corresponding predisposition will develop ALS.
“In many diseases that are hereditary, there are overlapping, so-called additive effects of genetic factors – for example in schizophrenia,” explains Alexander Schönhuth from Bielefeld University. “The more of these factors the genome contains, the more likely it is that people will get sick. We can therefore recognize the genetic disposition from the genes. With ALS, on the other hand, it is much more complicated.”
AI system unravels genetic interactions
But Schönhuth and his colleagues from the Dutch Research Center for Mathematics and Computer Science in Amsterdam may now have found a way of gene-based prognosis for ALS as well. For their method, they used an AI system that not only evaluates the presence of risk variants, but also takes into account the hierarchies and interactions between the affected genes. The artificial intelligence is based on so-called capsule networks, neural networks that were originally developed for image analysis.
“The great advantage of this method is that overlapping processes can also be recorded,” explains Schönhuth. “Our AI method clearly shows which genes and their processes are particularly important for the development of ALS.” In their study, they used this method to analyze the genetic data of 3,000 ALS patients and 7,000 people who did not have ALS .
ALS predicted with 87 percent accuracy
The result: The AI system called “DiseaseCapsule” was able to correctly identify almost 87 percent of ALS patients based on the genome data alone. Compared to the clinical standard, the AI can thus predict much better which person will develop ALS based on their genetic disposition. “With our method, there are 28 percent fewer people who are misdiagnosed,” the researchers explain.
In the analyses, the AI system also identified 922 gene changes that are crucial for the ALS diagnosis and that probably play an important role in the disease. 644 of these genes do not contribute to the ALS probability via classic additive effects, but are linked to the disease via more complex interactions. “These connections need to be investigated further,” says Schönhuth. “If we learn more about the genes, we also learn more about the processes.”
The researchers hope that the genes identified by their AI system will help advance research into the causes of ALS. They also see AI tools like DiseaseCapsule as an opportunity to speed up and improve diagnosis of the disease. This could directly benefit those affected. However, their AI system is still in the development stage and cannot be used in clinical practice. (Nature Machine Intelligence, 2023; doi: 10.1038/s42256-022-00604-2 )
