Single-molecule direct RNA sequencing reveals the shaping of epitranscriptome across multiple species.

N6-methyladenosine (m6A) is an essential RNA modification that regulates gene expression and influences diverse cellular processes. Yet, fully characterizing its transcriptome-wide landscape and biogenesis mechanisms remains challenging. Traditional next-generation sequencing (NGS) methods rely on short-reads aggregation, overlooking the inherent heterogeneity of RNA transcripts. Third-generation sequencing (TGS) platforms offer direct RNA sequencing (DRS) at the resolution of individual RNA molecules, enabling simultaneous detection of RNA modifications and RNA processing events. In this study, we introduce SingleMod, a deep learning model tailored for precise detection of m6A modification on individual RNA molecules from DRS data. SingleMod innovatively employs a multiple instance regression (MIR) framework, leveraging extensive methylation-rate labels provided by the quantitative NGS-based method, and achieves ROC AUC and PR AUC of ~0.95 for single-molecule m6A prediction. Applying SingleMod to human cell lines, we systematically dissect the transcriptome-wide m6A landscape at single-molecule and single-base resolution, characterizing m6A heterogeneity in RNA molecules from the same transcript. Through comparative analyzes across eight diverse species, we quantitatively elucidate three distinct m6A distribution patterns correlated with phylogenetic relationships and suggest divergent regulatory mechanisms. This study provides a framework for understanding the shaping of epitranscriptome in a single-molecule perspective.