Unimeth: A unified transformer framework for accurate DNA methylation detection from nanopore reads

Abstract

Nanopore sequencing enables direct detection of DNA modifications from native DNA. However, accurate methylation calling across species, sequence contexts, modification types and chemistries remains challenging. We present Unimeth, a transformer-based framework that jointly processes raw signals and basecalled sequences in read patches and predicts all target methylation sites within each patch. Unimeth uses a three-phase training strategy that combines signal pre-training, methylation fine-tuning and site-level calibration using methylation frequency information. We evaluated Unimeth for 5mC and 6mA detection using public and in-house datasets spanning 14 species, three nanopore chemistries and wild-type, mutant and enzyme-treated samples. Unimeth improved plant 5mC detection in non-CpG contexts, reduced false-positive calls in low-methylation samples and maintained high 5mCpG performance in mammalian datasets. For 6mA, Unimeth reduced background calls while preserving signals for Fiber-seq nucleosome and gene-level analyses. Unimeth provides a unified framework for nanopore-based methylation detection across methylation types and biological contexts.