MeDiCNet: Integrating Multi-scale Dynamic Convolution and Enhanced Position-Aware Transformer for DNA Methylation...

Area of Science:

• Epigenetics and Genomics
• Computational Biology
• Bioinformatics

Background:

• DNA methylation regulates gene expression and is implicated in various biological processes and diseases.
• Current computational methods struggle to model both fine-grained sequence motifs and long-range dependencies across different methylation types.
• Predicting multiple DNA methylation sites (N6-methyladenine, 5-hydroxymethylcytosine, N4-methylcytosine) requires advanced computational approaches.

Purpose of the Study:

• To introduce MeDiCNet, a unified deep-learning framework for predicting multiple types of DNA methylation sites.
• To develop a method that effectively models both local sequence patterns and global dependencies across diverse biological contexts.
• To provide a robust, efficient, and interpretable tool for large-scale, cross-type epigenomic analysis.

Main Methods:

• MeDiCNet employs multi-scale dynamic convolution for local pattern extraction and a Transformer encoder with enhanced positional attention for global context.
• The framework integrates nucleotide identity, dynamic convolution, and Transformer-based positional encoding (rotary and clipped relative embeddings).
• A gated fusion module adaptively combines these feature streams for accurate classification of methylation sites.

Main Results:

• MeDiCNet achieved improved accuracy (ACC) by up to 8.1% and Matthews correlation coefficient (MCC) by up to 0.10 across seventeen benchmark datasets.
• Demonstrated high performance, including 94.82% accuracy on the F. vesca 6mA dataset and AUC > 0.98 on the Mus musculus 5hmC dataset.
• Unsupervised analysis confirmed MeDiCNet's ability to recover biologically authentic motifs with high fidelity, using significantly fewer parameters than comparable large language models.

Conclusions:

• MeDiCNet effectively captures complex local and global sequence features for DNA methylation prediction.
• The framework offers a robust, efficient, and interpretable solution for cross-type epigenomic analysis.
• MeDiCNet advances the prediction of N6-methyladenine, 5-hydroxymethylcytosine, and N4-methylcytosine sites across diverse species.