RMNet: An RNA m6A Cross-species Methylation Detection Method for Nanopore Sequencing

N6-methyladenosine (m6A) is the most prevalent RNA modification in eukaryotic cells, influencing RNA lifecycle processes. Existing m6A detection methods, such as wet-lab techniques and statistical approaches, are time-consuming, labor-intensive, or require control samples, while machine learning models often lack cross-species applicability. This study aims to develop RMNet, a robust cross-species m6A detection method using nanopore sequencing.

RMNet employs Conformer and RNN architectures, integrating signal and alignment features from nanopore sequencing data. Contrastive learning enhances differentiation between m6A and non-m6A sites. The model was trained and tested on datasets from synthesized RNA, Arabidopsis, and human samples, using a single set of model weights.

RMNet achieved state-of-the-art performance with accuracies of 99.7% for synthesized RNA, 78.8% for Arabidopsis, and 88.9% for human datasets. It outperformed existing methods (m6Anet, DENA, and RedNano) across six metrics, including AUC and AUPR, demonstrating robust cross-species generalization.

RMNet's ability to detect m6A sites across diverse species with a single model addresses limitations of species-specific models. Its high sensitivity and feature representation enable applications in cancer research, neurodevelopmental studies, and plant biology. Limitations include higher error rates in human datasets for thymine-rich k-mers, likely due to complex secondary structures.

RMNet provides an efficient, powerful tool for cross-species m6A detection, advancing epitranscriptomics research with potential applications in precision medicine and agricultural science.