HGTM: A Structural and Contextual Deep Learning Model for Viral DNA Classification

The exponential growth of genomic data has highlighted the need for advanced computational models to classify DNA sequences with high accuracy and biological interpretability. Traditional approaches such as CNN-LSTM architectures often struggle to capture both long-range dependencies and structural relationships within nucleotide sequences. To address these limitations, we propose a Hybrid Graph-Transformer Model (HGTM) that integrates Graph Neural Networks (GNNs) with Transformer encoders for DNA sequence classification. DNA sequences are tokenized into k-mers and embedded using pre-trained representations, while graph-based modeling encodes structural dependencies between nucleotides. A multi-head self-attention mechanism fuses features from both the GNN and Transformer components, enabling the model to leverage local structural motifs as well as longrange contextual relationships. Experiments on viral DNA datasets, including SARS and MERS, demonstrate that the proposed HGTM achieves superior accuracy and interpretability compared to conventional CNN, SVM, and Bi-LSTM approaches. This hybrid framework provides a scalable and biologically informed solution for genomic sequence classification, with potential applications in pathogen surveillance, drug discovery, and precision medicine.