Abstract
Differential methylation is a key epigenetic process contributing to cancer development. Most DNA methylation prediction methods rely on DNA sequences from the background reference genome, neglecting individual genetic variation, which limits their ability to capture methylation differences. To address this, we propose CMC-WDTK, a deep learning framework that combines a weight-sharing dual-branch Transformer with a Kolmogorov‒Arnold network (KAN) to integrate sequences flanking CpG sites and adjacent single nucleotide variation (SNV) information to predict methylation changes between DNA sequences. CMC-WDTK captures global and local features of both reference and variant sequences and models high-dimensional relationships, offering accurate predictions of methylation changes. CMC-WDTK accurately predicted DNA methylation changes in eight real datasets (AUC greater than 0.8 for all datasets), with strong generalizability across datasets. Method comparison and ablation analyses further confirm that CMC-WDTK outperforms existing approaches and that its full architectural design is essential for achieving robust and accurate methylation-change prediction across datasets. Additionally, it identified a repeated cytosine and guanine sequence motif that promotes increased methylation. CMC-WDTK is the first computational tool used to predict methylation changes between sequences, offering significant advancements in understanding and comparing DNA methylation across diverse datasets and biological conditions.