Achieving precise m6A RNA modification prediction from nanopore signals through integration with pre-trained...

While predicting base modification states directly from nanopore electrical signals offers a cost-effective approach to epitranscriptomic profiling, current computational models remain fundamentally limited. They frequently suffer from suboptimal precision and unacceptably high false-positive rates. More importantly, existing methods fail to effectively exploit inherent RNA sequence features and completely overlook the critical influence of RNA structure. To overcome these limitations, we introduce NanoFM, a deep learning predictor incorporating an RNA structural foundation model. By fusing large-scale model embeddings with raw nanopore signal features, NanoFM achieves accurate, transcriptome-wide methylation prediction. Even when trained on limited empirical data, the framework maintains good generalization across multiple evaluation settings. It successfully identifies methylation sites across previously unseen sites and species, while maintaining high predictive accuracy across diverse RNA biotypes. Systematic evaluation across multiple tasks highlights its broad applicability and performance improvements over state-of-the-art tools. Ultimately, our approach provides insights into the RNA modifications at single-base resolution.