GCN-LSTM: A Graph-Based Deep Learning Method for Metagenomic Virus Classification

Viruses, as the most common organisms in the natural world, have significant impacts on both the environment and human society. Metagenomic sequencing of the virome can provide comprehensive information about viral sequences within microbial communities. Therefore, it is crucial to classify viral and host sequences from metagenomic data. Currently, the analysis of viral sequences heavily relies on one-hot encoded text data, which poses challenges due to its large volume and difficulty in extracting deep features. In this study, we propose a graph-based deep learning approach called GCN-LSTM for the classification of viral sequences in metagenomic data. GCN-LSTM utilizes a graph encoding technique to represent viral and host sequences, and employs graph convolution to extract features. Furthermore, it employs an LSTM neural network with self-attention mechanism to achieve the classification of viral and host sequences. Experimental results demonstrate that the performance of the proposed model improves with increasing sequence length, and outperforms the commonly used DeepVirFinder method.