An explainable hybrid CNN-LSTM framework for accurate sequence-based classification of RNA N6-methyladenosine (m6A) modification

N6-methyladenosine (m6A) is one of the most prevalent and functionally significant RNA modifications in eukaryotic transcriptomes, playing critical roles in post-transcriptional gene regulation. Accurate identification of m6A sites remains a considerable challenge due to complex sequence dependencies and limited labeled data. To address these issues, this paper proposed a hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) framework integrated with Shapley additive explanations (SHAP) for feature selection and interpretability. The proposed scheme aims to improve both predictive performance and biological insight by identifying and emphasizing the most informative features in RNA sequences. Firstly, multiple biologically relevant feature extraction methods are employed to encode RNA sequences into numerical representations. Secondly, SHAP is used to quantify each feature's contribution, enabling the selection of the most impactful subset while enhancing model interpretability. Thirdly, the selected data is reshaped and standardized to ensure Conv1D compatibility and capture local sequence motifs effectively. Fourthly, CNN layers extract high-level spatial features, which are then fed into LSTM layers to model long-range dependencies within the sequences. Finally, the output is processed through fully connected layers with a sigmoid activation function to perform binary prediction. The experimental results indicate that the CNN-LSTM model, combined with SHAP-based feature selection, outperforms traditional classifiers and standalone deep learning models. The proposed CNN-LSTM framework, evaluated using 10-fold stratified cross-validation, achieved an accuracy of 87.39 %, sensitivity of 83.25 %, specificity of 91.52 %, and an MCC of 0.7534. The results of the proposed model demonstrate its strong ability to accurately classify RNA m6A modification sites, highlighting its potential for large-scale transcriptome-wide epitranscriptomics analysis. The proposed CNN-LSTM model was rigorously compared with several traditional machine learning classifiers and state-of-the-art deep learning approaches, including DT, SVM, KNN, AdaBoost Classifier, Gaussian NB, and DNN.