UniMethylNet: A Universal DNA Methylation Site Prediction Network Integrating a Neural Network and an Attention Mechanism

DNA methylation plays a crucial role in biological processes. However, existing prediction methods often suffer from limited generalization ability due to the scale and diversity constraints of the training samples, as well as insufficient recognition of significant interspecies differences in methylation patterns. To address these challenges, this study proposes a novel methylation site prediction model, namely, UniMethylNet, for robust identification of different methylation types (4mC, 5hmC, and 6 mA) across 12 species. UniMethylNet incorporates a Position Linear Layer to precisely capture local patterns, while utilizing a Bidirectional Long Short-Term Memory network to model long-term dependencies. UniMethylNet also employs a Channel-Spatial Dual Attention module for adaptive feature weighting and multiscale focusing, enabling one to effectively extract methylation-related features. Experimental results on 20 public data sets demonstrate that UniMethylNet achieves a mean accuracy of 87.78% and a mean area under the receiver operating characteristic curve of 93.01%, significantly surpassing existing models and exhibiting superior cross-species and cross-type generalization. Overall, UniMethylNet provides a powerful tool for DNA methylation site prediction, offering a quantitative approach for in-depth exploration of the conservation and specificity of epigenetic regulation by capturing the underlying conserved sequence motifs across diverse biological contexts.