Prediction of DNA N4-Methylcytosine Sites Based on a Position-Aligned Multi-branch Fusion Network

While DNA N4-methylcytosine (4mC) plays regulatory roles in various organisms, its endogenous presence in mammalian genomes remains debated. Accurate identification of putative 4mC sites in eukaryotic genomes is impeded by scarce validated data and complex sequence dependencies. Although the mouse is a vital mammalian model for epigenetic studies, computational predictors capable of reliably screening murine 4mC candidate sites remain lacking. To address this, we propose TriAlignNet-4mC, a position-aligned triple-branch neural network designed for the precise and interpretable prediction of benchmark mouse 4mC candidate sites. Our framework unifies local biochemical properties, long-range contextual dependencies, and duplex-connectivity relations by integrating physicochemical descriptors, DNABERT embeddings, and duplex-connectivity-aware relational graph representations. Via position-aware alignment and late fusion, TriAlignNet-4mC generates comprehensive feature representations that markedly enhance prediction stability. Extensive benchmarking on a Mus musculus dataset demonstrates the model's highly competitive performance. In independent testing, TriAlignNet-4mC achieved a sensitivity of 0.7937, a specificity of 0.8375, an accuracy of 0.8156, and a Matthews correlation coefficient of 0.6319, highlighting its robust generalization. Furthermore, ablation studies confirm the complementary contributions of the 3 branches: the contextual Transformer branch enhances sensitivity, while the graph-based structural branch improves specificity. Overall, rather than relying on exhaustive manual feature engineering, TriAlignNet-4mC is highly competitive with existing advanced predictors, exhibiting a distinct and robust trade-off between sensitivity and specificity.