Machine learning models can improve genetic prediction of type 1 diabetes by as much as 10%, results from a University of California, San Diego study show.
Researchers used the machine learning model T1GRS to improve the gold standard polygenic genetic risk score used to predict who is likely to develop a condition called GRS2.
Type 1 diabetes is an autoimmune disease that affects approximately 2 million people in the United States. Although this is a multifactorial disease, genetics plays a large role, with approximately 50% of a person’s susceptibility being due to genetics.
“The natural history of type 1 diabetes suggests that the disease occurs in genetically susceptible individuals exposed to environmental triggers, resulting in the development of islet-specific autoantibodies and autoreactive T cells, leading to a progressive loss of insulin secretory function, but the underlying pathogenesis is not fully understood,” wrote first author Kyle Galton, Ph.D., associate professor of pediatrics at the University of California, San Diego School of Medicine, and colleagues. natural genetics.
The GRS2 polygenic risk score has been widely tested and can be used to predict which newborns are at increased risk of developing type 1 diabetes. Although early prediction does not necessarily stop the disease, it may prevent emergencies such as diabetic ketoacidosis at the time of diagnosis, give families time to prepare, and allow for the use of treatments that delay the onset of symptoms.
In this study, Gaulton and colleagues conducted a genome-wide association study in 20,355 type 1 diabetics and 797,363 non-diabetic Europeans, and further analyzes around the MHC region in 10,107 diabetics and 19,639 non-diabetics.
“There is a ‘block’ of collaborative genetic information in the MHC that is extremely rich in patients with type 1 diabetes,” said co-first author Dr. Emily Griffin, a postdoctoral fellow in the Galton lab. “Having them doesn’t mean you’ll get diabetes, but it does mean you’re very unlikely to get diabetes if you don’t have them.”
In total, 160 risk signals were identified, and the team trained the T1GRS model to predict who is likely to develop type 1 diabetes based on their genetics. This model was able to improve the predictions of the GRS2 model by up to 10% in both populations of European American and African American ancestry.
Overall, the new score correctly indicated that about 89 out of 100 people had type 1 diabetes, while correctly reassuring about 84 out of 100 that they did not have the disease.
“Our results highlight the value of combining results from genetic association studies with machine learning techniques to improve prediction of complex diseases,” the authors conclude.
