Machine learning approaches reveal DNA methylation patterns associated to prediabetes clusters

Introduction Sub-phenotyping of individuals at high risk for type 2 diabetes (T2D) revealed six distinct clusters of which three showed the highest risk of developing T2D and/or complications. These prediabetes clusters could provide an opportunity to improve prediction of complications in the future. However, it requires intensive clinical measurements such as an oral glucose tolerance test. Here, we developed a machine learning (ML) workflow to identify blood-based epigenetic markers to distinguish between prediabetes sub-phenotypes.

Methods: DNA methylation was profiled in blood cells using the Illumina Infinium Human Epic 850K arrays (Version 1 and 2) for different cohorts belonging to clusters C2 (low risk), C3, C5, and C6 (each high risk). A multi-fold elastic net (e-net) was applied to the DNA methylation data, and prediction accuracies were assessed using either semi-supervised clustering (Forest Guided Clustering) or unsupervised clustering (Partition Around Medoids).

Results: In a discovery cohort (n=149), we identified 2,136 CpG sites as predictors for clusters C2, C3, C5, and C6. These were then used to cluster individuals in two independent replication datasets (n=46 & n=146), achieving 97-98% accuracy in clustering. The DNA methylation profiles of these 2,136 sites in blood samples showed sufficient discriminatory capacity to classify individuals at elevated risk for developing severe complications. These methylation sites were found in genes related to insulin signalling, lipid metabolism and inflammation.

Conclusions: Our findings demonstrate the efficacy of ML-based strategies to identify epigenetic biomarkers from high–dimensional DNA methylation data for the classification of distinct prediabetic clusters.

Publication History

Article published online:
28 May 2025