Accurate detection and quantification of single-base m6A RNA modification using nanopore signals with multi-view deep learning

N6-methyladenosine (m6A) is a crucial epitranscriptomic mark. While Nanopore Direct RNA Sequencing (DRS) enables transcriptome-wide detection, most existing methods neglect or lack the capacity to effectively process the intrinsically variable-length raw signals generated by DRS reads. Here, we present MultiNano, a multi-view deep learning framework that converts variable-length raw signals into image-like feature representations, effectively resolving the length inconsistency problem. By integrating raw signal and basecalling features, MultiNano enables accurate and comprehensive transcriptome-wide detection of m6A modifications. Our model achieved state-of-the-art (SOTA) performance in various tasks, including site-level prediction, read-level prediction, cross-species transfer learning, and modification rate estimation. Furthermore, the false positive control strategy implemented in MultiNano significantly enhances the model's robustness and predictive accuracy, offering a powerful alternative to traditional thresholding-based filtering algorithms. Collectively, our approach provides novel insights for the absolute quantification and single-base resolution of RNA modifications.