Fusion-m6A: A lightweight hybrid deep learning framework for RNA m6A site prediction

N6-methyladenosine (m6A) is the most common mRNA modification and plays key role in RNA metabolism, gene regulation, and disease. Accurate identification of m6A sites is critical for understanding their functional and biological significance. Although experimental techniques such as Nanopore direct RNA sequencing (DRS) have advanced m6A profiling, they remain costly and laborious. Computational approaches provide scalable alternatives, but many depend on handcrafted features or computationally expensive transformer-based models. We present Fusion-m6A, a hybrid deep learning framework that integrates Word2Vec-based sequence embeddings, convolutional layers for local motif detection, bidirectional gated recurrent unit with attention for capturing long-range dependencies, and auxiliary k-mer features. The fused representations are passed through fully connected layers to predict m6A sites with high accuracy. Benchmarking across multiple human tissues and cell lines shows that Fusion-m6A consistently outperforms state-of-the-art predictors in accuracy and Matthews correlation coefficient. Crucially, the model achieves faster inference and requires substantially less memory, offering a practical and robust solution for large-scale and tissue-specific m6A site prediction. The implementation of Fusion-m6A is publicly available for reproducibility at: https://github.com/waleed551/Fusion_m6A.