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

Area of Science:

• Epigenetics and Genomics
• Computational Biology
• Mammalian Genetics

Background:

• DNA N4-methylcytosine (4mC) has regulatory roles, but its presence in mammalian genomes is debated.
• Accurate identification of 4mC sites is challenging due to limited data and sequence complexity.
• A lack of reliable computational tools hinders 4mC site prediction in the key mammalian model, Mus musculus.

Purpose of the Study:

• To develop a precise and interpretable computational predictor for identifying mouse 4mC candidate sites.
• To address the limitations of existing methods in screening 4mC sites within mammalian genomes.
• To provide a robust tool for advancing epigenetic studies in Mus musculus.

Main Methods:

• Proposed TriAlignNet-4mC, a position-aligned triple-branch neural network.
• Integrated local biochemical properties, long-range contextual dependencies (DNABERT), and duplex-connectivity via graph representations.
• Employed position-aware alignment and late fusion for comprehensive feature representation and prediction stability.

Main Results:

• TriAlignNet-4mC demonstrated highly competitive performance on Mus musculus datasets.
• Achieved high prediction metrics in independent testing: 0.7937 sensitivity, 0.8375 specificity, 0.8156 accuracy, and 0.6319 MCC.
• Ablation studies confirmed the distinct contributions of the Transformer (sensitivity) and graph-based (specificity) branches.

Conclusions:

• TriAlignNet-4mC offers a robust and interpretable method for predicting mouse 4mC sites, outperforming manual feature engineering.
• The model provides a favorable trade-off between sensitivity and specificity, enhancing its utility in epigenetic research.
• This tool advances the computational screening of 4mC candidate sites in mammalian genomes, particularly Mus musculus.