Deep Learning-Based DNA Methylation Detection in Cervical Cancer Using the One-Hot Character Representation Technique

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Cervical cancer is a leading global malignancy affecting women.
• Early detection of epigenetic alterations, specifically Deoxyribose Nucleic Acid (DNA) methylation, is crucial for improved clinical outcomes.
• Identifying methylation patterns can enhance diagnostic and prognostic capabilities.

Purpose of the Study:

• To develop and validate a novel deep learning framework for predicting DNA methylation in cervical cancer.
• To assess the efficacy of different encoding strategies and sequence window sizes for methylation prediction.
• To evaluate the framework's performance on promoter regions of key cervical cancer-associated genes.

Main Methods:

• A UNet deep learning architecture was employed, integrated with a one-hot character encoding technique for DNA sequences.
• Monomer and dimer encoding strategies were compared using varying Cytosine-Guanine (CG) site window sizes (100 bp, 200 bp, 300 bp) and sample sizes (5000, 10,000, 20,000).
• Model performance was validated on promoter regions of five cervical cancer-associated genes: miR-100, miR-138, miR-484, hTERT, and ERVH48-1.

Main Results:

• The optimal configuration involved dimer encoding with a 300 bp window and 5000 CG sites, achieving 91.60% accuracy, 96.71% sensitivity, 87.32% specificity, and a 96.53 AUROC score.
• The proposed framework significantly outperformed benchmark models like Convolutional Neural Networks and MobileNet.
• Validation on promoter regions showed 86.27% accuracy in identifying methylated CG sites with an 83.99 AUROC score.

Conclusions:

• The UNet-based deep learning framework demonstrates significant potential as a reliable and scalable tool for early epigenetic modification detection in cervical cancer.
• This approach contributes to advancing biomarker discovery and diagnostic strategies for cervical cancer.
• The study highlights the effectiveness of the proposed encoding and architecture for predicting DNA methylation patterns.