EvoPlantMeth: an interpretable deep learning framework reveals DNA methylation evolution and functional elements in plants

DNA methylation is a fundamental epigenetic modification in plants. While deep learning has improved epigenomic predictions, current models lack a broad evolutionary perspective and struggle to capture plant-specific sequence patterns and contextual dependencies. To address this, we assembled a phylogenetically curated DNA methylation atlas across 94 plant species and developed EvoPlantMeth, an interpretable deep learning framework trained on over 1.82 billion tokens. By integrating local DNA sequences with both the methylation status and physical distance of adjacent cytosines, EvoPlantMeth achieves high cross-species generalizability and reliably predicts continuous methylation levels across all sequence contexts. Using gradient-based methods, we interpreted the model’s learned features to uncover key epigenetic determinants. Specifically, we identified distinct local sequence preferences and quantified the influence range of neighboring methylation. Crucially, we established a framework to screen functional regulatory elements by integrating two quantitative parameters: epigenetic plasticity, which uses predictive variance to capture dynamic variation, and regulatory impact, which employs gradient sensitivity to measure influence on neighboring methylation states. Applying this approach to alfalfa ( Medicago sativa ), we successfully identified functional epigenetic loci governing regenerative capacity. Ultimately, EvoPlantMeth provides a computational framework for characterizing DNA methylation across diverse species, offering insights into plant epigenetic evolution.