NEW YORK – A research team including collaborators from the Chinese University of Hong Kong and Sanford Burnham Prebys Medical Discovery Institute has crafted an algorithm to predict whether patients with type 2 diabetes will develop kidney disease based on their DNA methylation profiles.
The researchers used data from 1,271 patients with type 2 diabetes in the Hong Kong Diabetes Register to build the algorithm, which was described in a paper published this month in Nature Communications.
According to Ronald Ma, a professor in the department of medicine and therapeutics at the Chinese University of Hong Kong and one of the authors on the paper, the researchers — who have been working on identifying diabetes and kidney disease biomarkers for years — wanted to find ways for clinicians to better identify which patients were at risk of developing kidney disease and progressing to kidney failure.
The team identified patients with type 2 diabetes within the Hong Kong Diabetes Register who later developed kidney disease and compared them with people with diabetes who didn't develop kidney disease to form the basis of its study design, Ma said.
Using Illumina's Infinium HumanMethylation450K BeadChip array, the researchers profiled DNA samples to detect methylation changes and looked for biomarkers associated with increased future risk of kidney disease and failure, Ma said. The team tried to find methylation markers related to overall kidney function as well as markers related to how quickly kidney disease deteriorates, and found that "kidney function turned out to be a very strong determinant of the methylation profile, and that methylation status of certain markers may affect kidney function," he said.
Kevin Yip, a professor in the cancer genome and epigenetics program at San Diego-based Sanford Burnham Prebys and another author on the paper, added that the team measured more than 450,000 sites in the genome of these patients because one "can't expect that one site or two sites can explain everything" in such a complex disease. They investigated single sites and combinations of sites and how they interacted to determine which were related to kidney function and had the best predictive value.
By looking at multiple sites, the team was able to bypass certain limitations. For instance, some sites on the genome may not be associated with kidney function strongly by themselves but in combination with other sites may have a stronger correlation, Yip noted.
The team used machine learning methods to identify the predictive markers and determine which of those markers would be most useful. They were able to find "a number of markers" predictive of kidney disease in patients with diabetes, Ma said.
He noted that after finding the methylation markers, the team still had to control for issues that could impact the model, such as different samples having a different composition of blood cell types, which can affect methylation signals. In the Nature Communications paper, the team explains that it has utilized 64 sites for baseline kidney function and 37 for the rate of decline in kidney function in its final model after controlling for potential issues.
The team used its model to predict the risk of developing end-stage kidney disease within five years and achieved an area under the receiver operating curve of 0.94. When it excluded patients that had a low baseline estimated glomerular filtration rate, meaning they had a high risk of developing end-stage kidney disease within five years, the area under the receiver operating curve was 0.88. In both cases, the model's performance was comparable to the performance of current clinical risk equations, the researchers wrote.
The researchers also validated the algorithm on a separate cohort of 326 Native American patients with type 2 diabetes. Yip noted that the markers for baseline kidney function performed well in this cohort, but that the performance of the markers for kidney function decline could be improved.
Though performance of the markers alone is already "fairly similar" to traditional clinical information like blood pressure, adding the markers to clinical factors improves prediction further, Ma said, adding that it "hopefully will provide a new way to help predict in a more accurate manner who's most at risk" and target those patients for more intensive treatment.
The researchers are also further analyzing their data to understand why some of these regions affect kidney function and potentially apply the information to developing future treatments for kidney disease, he said.
The team also has plans to further develop its model as a potential test and has filed patents relating to this use, he noted. A similar test, KidneyIntelX, is offered in the US by in vitro diagnostics firm Renalytix to determine which patients with type 2 diabetes and chronic kidney disease are at low, intermediate, or high risk for rapid progressive decline in kidney function.
Ma's hospital has already implemented a risk model for kidney disease based on clinical factors, but he believes his team's algorithm would be an improvement on that method. The researchers have also provided an online calculator allowing other clinicians to input methylation data and clinical factors and receive a readout on kidney function or the rate of decline, he said. In Ma's view, the algorithm can potentially be implemented into clinical practice after further validation studies, but the cost and difficulty of gathering methylation data, as well as the need for standardized pipelines for data analysis, may be a deterrent. Right now, methylation arrays are primarily used as research tools and not for use on individual samples or as potential diagnostic tools, and those are obstacles to clinical implementation, he said.
Yip added, however, that now that the researchers know what sites to use, gathering the methylation data could be cheaper and simpler because there's no need to look at 450,000 sites for each patient.