Predictive modelling using genome-wide DNA methylation data
Year: 2021
Type: other
Abstract: In search of predicative methylation signatures for complex diseases, recent studies have identified valuable DNA methylation-based biomarkers. Most previous studies have built prediction models based on single CpGs, without considering the methylation status of neighboring CpG sites. However, compared to single CpGs, differentially methylated regions give higher confidence and likelihood of biological importance. We tested the hypothesis that the accuracy of Elastic Net prediction models can be improved by using predictors based on co-methylated regions (clusters). We performed both internal cross-validation and external validation using two large DNA methylation datasets in Alzheimer’s disease. Our results showed the prediction performance for cluster-based Elastic Net models performed similarly as models using only the CpGs as predictors. Among different summarization methods for clusters, models that used the most significant CpG to represent each cluster performed the worst. Other summary methods performed similarly. For models that used only the CpGs, the models that used the most significant CpGs performed uniformed better than those using most variable CpGs.
Institution: University of Miami
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
OpenAlex ID: W3198651272