Abstract
Background: Cervical cancer is among the most prevalent malignancies in women worldwide, and early detection of epigenetic alterations such as Deoxyribose Nucleic Acid (DNA) methylation is of utmost significance for improving clinical results. This study introduces a novel deep learning-based framework for predicting DNA methylation in cervical cancer, utilizing a UNet architecture integrated with an innovative one-hot character encoding technique. Methods: Two encoding strategies, monomer and dimer, were systematically evaluated for their ability to capture discriminative features from DNA sequences. Experiments were conducted on Cytosine-Guanine (CG) sites using varying sequence window sizes of 100 bp, 200 bp, and 300 bp, and sample sizes of 5000, 10,000, and 20,000. Model validation was performed on promoter regions of five cervical cancer-associated genes: miR-100, miR-138, miR-484, hTERT, and ERVH48-1. Results: The dimer encoding strategy, combined with a 300-base pair window and 5000 CG sites, emerged as the optimal configuration. The proposed framework demonstrated better predictive performance, with an accuracy of 91.60%, sensitivity of 96.71%, specificity of 87.32%, and an Area Under the Receiver Operating Characteristic (AUROC) score of 96.53, significantly outperforming benchmark deep learning models, including Convolutional Neural Networks and MobileNet. Validation on promoter regions further confirmed the robustness of the model, as it accurately identified 86.27% of methylated CG sites and maintained a strong AUROC of 83.99, demonstrating its precision-recall balance and practical relevance during validation in promoter-region genes. Conclusions: These findings establish the potential of the proposed UNet-based approach as a reliable and scalable tool for early detection of epigenetic modifications. Thus, the work contributes significantly to improving biomarker discovery and diagnostics in cervical cancer research.