A Region-Aware Structured Framework Improves Prediction of Gene Expression from DNA Methylation

DNA methylation is a key epigenetic modification that plays an important role in gene expression regulation and disease development. Inferring gene expression from DNA methylation provides a computational strategy for cross-omics integration and facilitates the exploration of regulatory relationships between epigenetic modifications and transcription. However, the regulatory effects of methylation on gene expression often exhibit complex characteristics, and methylation in different functional regions of a gene may follow distinct regulatory patterns. Existing methods typically lack the capacity to model such region-aware nonlinear relationships. In this study, we propose RSMethy-Net, a neural network framework based on region-aware encoding for predicting gene expression from DNA methylation data. The model incorporates grouped region encoding modules for different gene functional regions to capture their latent regulatory patterns and characterize methylation-expression associations under a nonlinear predictive framework. We systematically evaluated RSMethy-Net across 6 cancer cohorts. Experimental results demonstrate that RSMethy-Net outperforms multiple baseline methods in predictive performance. Furthermore, by integrating region design with model interpretability analyses, the framework can quantify the contributions of different gene regions to predictions, providing insight into methylation-expression associations under a nonlinear predictive setting.