FourierAN-m¹A: A Novel Fourier-Domain Adaptive Feature Extraction Framework for m¹A RNA Modification Prediction

RNA modifications play a pivotal role in regulating gene expression. Among them, N¹-methyladenosine (m¹A), as a crucial post-transcriptional modification, requires accurate site identification to elucidate its regulatory mechanisms. However, existing prediction models still suffer from limitations in feature extraction, long-range dependency modeling, and interpretability. Furthermore, there remains a lack of systematic investigation into modification-loss variants. To address these gaps, we developed a high-quality data set encompassing both m¹A modifications and their loss-of-modification variants and proposed a novel deep learning framework named FourierAN-m¹A. The model integrates convolutional neural networks, adaptive Fourier neural operators (AFNO), and a Transformer encoder. By introducing a dynamic filtering mechanism based on complex-valued neural networks and employing a learnable adaptive mask, it jointly learns frequency-domain and spatial-domain features of RNA sequences. Experimental results reveal that the proposed model outperforms existing methods in prediction accuracy and exhibits biological interpretability, providing a high-performance approach for precise identification of m¹A sites.