Endogenous labeling empowers accurate detection of m6A from single long reads of direct RNA sequencing

Although plenty of machine learning models have been developed to detect m6A RNA modification sites using the electric current signals of ONT direct RNA sequencing (DRS) reads, the landscape of m6A on different RNA isoforms is still a mystery due to their limited capacity to distinguish the m6A on individual long reads and RNA isoforms. The primary challenge in training the model with single-read accuracy is the difficulty of obtaining the training data from individual DRS reads that comprehensively represent the m6A on endogenous RNAs. Here, we endogenously label the methylated m6A sites on single ONT DRS reads by APOBEC1-YTH induced C-to-U mutations, strategically positioned 10-100 nt away from the known m6A sites on the same reads. Adopting a semi-supervised leaning strategy, we obtain 700,438 reliable 5-mer single-read level m6A signals, providing a comprehensive representation of m6A on endogenous RNAs. Leveraging this dataset, we develop m6Aiso, a deep residual neural network model that not only accurately identifies and quantifies known m6A sites but also reveals unknown, subtly methylated m6A sites responsive to METTL3 depletion. Analyzing m6Aiso-determined m6A on single reads and isoforms uncovers distance-dependent linkages of m6A sites along single molecules, as well as differential methylation of identical m6A sites on different isoforms. Moreover, we find wide-spread functionally important dynamic changes of m6A sites on specific isoforms during epithelial-mesenchymal transition (EMT). The pivotal utilization of the endogenous labeling strategy empowers m6Aiso to achieve remarkable precision in pinpointing m6A on individual molecules, underscores its effectiveness in elucidating the intricate dynamics and complexities of m6A across RNA isoforms. ### Competing Interest Statement The authors have declared no competing interest.