Abstract
BACKGROUND: N6-methyladenosine ([Formula: see text]) plays a crucial role in enriching RNA functional and genetic information. While deep learning has advanced [Formula: see text] site identification, current state-of-the-art methods, particularly those based on large-scale Pre-trained Language Models (PLMs), often suffer from high computational complexity and over-parameterization, limiting their scalability for genome-wide analysis. RESULTS: In this work, we propose MKE-ResNet, an ultra-lightweight end-to-end framework designed to balance predictive performance with extreme computational efficiency. MKE-ResNet integrates a deep residual network with an innovative Multi-Kernel Efficient Channel Attention (MKE-ECA) module to adaptively capture multi-scale sequence patterns. Extensive experiments on 22 datasets demonstrate that MKE-ResNet achieves competitive performance against complex SOTA models (including MST-m6A). Notably, our model exhibits exceptional stability against sequence noise perturbation and demonstrates superior generalization capability on independent test sets (leading in 8 out of 11 cases in terms of AUROC). Furthermore, interpretability analysis reveals that MKE-ResNet moves beyond the static central motif to capture dynamic sequence patterns resembling RBP binding motifs (e.g., SRSF1 and PUM2). The source code and datasets are available at https://github.com/Gxttk/MKE-Resnet. CONCLUSIONS: MKE-ResNet provides a biologically interpretable and computationally efficient solution for [Formula: see text] site identification, offering a pragmatic tool for large-scale epitranscriptome screening.