CMA-Nano: A DNA Methylation Detection Method for Nanopore Sequencing Data Based on a Cross-Modal Attention Mechanism

DNA methylation, particularly 5-methylcytosine (5mC), plays a pivotal role in gene regulation, development, and disease progression. Nanopore sequencing enables direct methylation detection without chemical treatment, but inherent signal noise and complex multimodal data pose analytical challenges. We present CMA-Nano, a deep learning framework that employs a cross-modal attention mechanism to integrate DNA sequences, nanopore signal statistics, and raw current signals, achieving effective fusion of multimodal features and capturing complex interdependencies. To further enhance feature representation and generalization, CMA-Nano leverages a masked base prediction pretraining strategy. Comprehensive evaluation on Arabidopsis thaliana, Oryza sativa, and human data sets demonstrates that CMA-Nano outperforms existing methods across key metrics. Ablation experiments, transfer learning assessments, and interpretability analyses demonstrate the contributions of cross-modal attention and multimodal fusion to the model's predictive performance and biological relevance. CMA-Nano offers a robust and generalizable solution for accurate DNA methylation detection from nanopore sequencing data.