A computational method to predict methylation of CpG sites in Chr21 using sequence features from WGS with 450 K array features

DNA methylation is a vital epigenetic mechanism influencing cell differentiation, disease progression, and therapeutic development. However, accurately detecting CpG methylation sites remains challenging due to their dynamic nature. This, in addition, to the high cost of current experimental techniques that were expanded to cover the whole genome and the low cost experimental techniques with limited coverage. This study aims to develop an efficient, cell type independent model to predict CpG methylation sites using cost-effective and widely available data sources. Sequence features were extracted from whole-genome sequencing (WGS) data and integrated them with features from the Infinium HumanMethylation450 (450 k) array. This is to predict both binary methylation states and continuous beta values at single base resolution on human chromosome 21 (chr21). Eleven machine learning and ensemble methods were evaluated, with random forest (RF) achieving the best performance of 75 % balanced accuracy for classification. The model was further applied to beta value regression and categorized using a novel five-level “EGYPT Methylation Categorization” system, achieving 63.2 % of categorical prediction accuracy. These results demonstrated that integrating inexpensive genomic data with robust machine learning techniques can effectively approximate methylation patterns, offering a scalable and transferable approach for epigenetic analysis and precision medicine.