Deep learning-based identification of N6-methyladenine sites via hybrid feature fusion and SHAP-driven feature selection

N6-methyladenine (6 mA) is a critical epigenetic modification involved in gene regulation, genome stability, and cellular adaptation. Accurate computational identification of 6 mA sites is essential for elucidating epigenetic mechanisms and advancing disease-related research. However, existing methods are often constrained by limited feature representations and poor interpretability, hindering predictive performance and generalisation across diverse genomic contexts. To address these challenges, we propose a novel deep neural network (DNN)-based framework that integrates optimal hybrid features for robust and interpretable 6 mA site identification. The proposed framework employs a comprehensive multi-feature extraction strategy to capture complex sequence-level patterns in DNA, which are subsequently fused into a unified hybrid representation. To enhance model efficiency and interpretability, SHapley Additive exPlanations (SHAP) are applied for feature importance analysis, enabling the selection of the most discriminative features for downstream classification. The optimised feature set is then fed into a DNN classifier for accurate 6 mA site prediction. Evaluated using 5-fold cross-validation, the proposed model achieved training accuracies of 97.95% and 96.15% on the F. vesca and R. chinensis datasets, respectively. On independent test datasets, the model achieved accuracies of 97.10% (F. vesca) and 95.32% (R. chinensis), demonstrating strong generalisation. These results establish the proposed framework as an accurate, reliable, and interpretable computational tool for genome-wide identification of 6 mA sites, with broad applicability to epigenetic research and beyond.