A Comparative Analysis of Deep Learning Architectures for Protein Host Classification

Abstract

Influenza A virus (IAV) is an ongoing global health threat because of its high mutation rate, antigenic variation, and interspecies transmissibility. Virus host range prediction is thus essential for the early surveillance of zoonotic spillover and outbreak preparation. In this work we present and compare three deep learning architectures: Deep 1D CNN, Enhanced BiLSTM and a hybrid model of CNN-BiLSTM for direct host-species classification from raw HA protein sequences. The model comparisons showed that the Deep 1D CNN was the best in terms of the overall performance and resulted in validation accuracy and test accuracy of 0.8984 and 0.9016 respectively, which outperformed not only the CNN-BiLSTM Hybrid (0.9005) but also the Enhanced BiLSTM (0.8973).These findings indicate that the raw sequence-level patterns associated with host specificity can be well captured by the convolutional feature extraction, though the performances of the hybrid architectures are comparable. Our proposed deep learning model obviates the requirement of manual feature engineering and offers a scalable and accurate tool for viral host prediction, thus helping to enhance genomic surveillance and public health preparedness.