Enhanced Caudovirales Bacteriophage Protein Recognition with Convolutional Neural Networks and Self-Attention Mechanisms

Accurate classification of bacteriophage proteins (also known as phage proteins) is crucial for understanding their functions, particularly for key proteins of Caudovirales phages such as the portal protein, the large terminase subunit (TerL), and the small terminase subunit (TerS). To address the challenges in identifying these proteins, we propose a novel method named CNN-SANs, which integrates Convolutional Neural Networks (CNN) with Self-Attention Mechanisms (SANs). Our approach uses one-hot encoded protein sequence data as input, where CNNs extract local features through convolutional and pooling layers to identify key patterns. Subsequently, self-attention mechanisms dynamically adjust the focus on different positions within the sequences to capture long-range dependencies and global context. Experimental results demonstrate that applying CNN-SANs to metagenomic data significantly enhances Precision and F1 Score in protein identification. Compared to traditional methods, CNN-SANs exhibits superior performance in accurately classifying Caudovirales phage proteins, proving its effectiveness as a powerful tool for bacteriophage identification.