MSCF-DM: center-site-aware attention and gated multi-scale fusion network for cross-species DNA methylation site prediction

DNA methylation plays important roles in the regulation of gene expression, cell differentiation, chromatin stability, and disease development and progression. Existing deep learning methods are typically developed separately for a single species or dataset, making it difficult to simultaneously capture the shared patterns across different methylation types and the differences among species, while also limiting their unified application across multiple datasets. To address these limitations, we propose MSCF-DM, a multi-scale species-conditioned feature fusion network for predicting DNA methylation sites across multiple species. The model integrates sequence encoding, enhanced nucleotide composition, positional encoding, and species information. Residual multi-scale dilated depthwise-separable convolutions are employed to extract local sequence patterns under different receptive fields, while a bidirectional gated recurrent unit is used to capture contextual dependencies on both sides of candidate sites. In addition, MSCF-DM introduces a species-conditioned feature modulation mechanism that adaptively adjusts the feature responses of the convolutional and recurrent branches according to species information. Feature-wise gating and center-aware attention pooling are further used to integrate local features, global contextual information, and representations of key candidate sites. Experiments were conducted on 17 multi-species benchmark datasets covering three types of DNA methylation modifications, namely 4mC, 5hmC, and 6 mA. The results demonstrate that MSCF-DM achieves strong overall performance in terms of accuracy, sensitivity, specificity, Matthews correlation coefficient, AUC, and F1 score, and outperforms the compared state-of-the-art methods on most datasets, achieving the best overall macro-averaged AUC performance across the benchmark datasets. These results indicate that MSCF-DM provides favorable prediction stability and adaptability across the studied species, and achieves consistent performance on the multi-species benchmark datasets. In addition, MSCF-DM maintains a lightweight computational profile with low parameter and computational requirements. The proposed method offers a new framework for the unified modeling of different methylation types and may also provide computational support for related studies of epigenetic functions.