mIDEA: An Interpretable Structure-Sequence Model for Methylation-Dependent Protein-DNA Binding Sensitivity

DNA methylation dynamically reshapes protein-DNA interaction landscapes, yet the mechanistic principles governing methylation-dependent recognition remain incompletely understood. Currently, there is no predictive framework that integrates methylation-aware binding specificity with 3D structure-based residue interactions for quantitatively assessing the impact of DNA methylation on protein-DNA interactions. To bridge this gap, we introduce mIDEA (Methylation-informed Interpretable protein-DNA Energy Associative model), a structure-aware, residue-level biophysical framework that predicts and interprets protein-methylated-DNA interactions by optimizing an amino acid-nucleotide energy matrix explicitly accounting for the modulatory effect of 5-methylcytosine. We validate mIDEA across diverse human transcription factors with distinct CpG methylation preferences. Using only structural templates and prior biochemical knowledge, mIDEA accurately recapitulates both "methyl-plus" and "methyl-minus" effects of DNA methylation on protein-DNA binding. Incorporating quantitative methylation-sensitive binding data further enables mIDEA to capture position-specific methylation effects. When integrated with whole-genome methylation profiles, the model enhances in vivo binding-site prediction by reducing false positives while preserving sensitivity compared to models trained on unmethylated sequences. Overall, mIDEA provides a principled framework for elucidating the molecular mechanisms underlying methylation-modulated protein-DNA recognition. The mIDEA source code is available at: https://github.com/LinResearchGroup-NCSU/IDEA_DNA_Methylation.