Phage Protein Classification Using Advanced Sequence Encoding: Performance Comparison of Deep Learning and Classical Machine Learning Approaches

Abstract

Bacteriophages, or phages, play a crucial role in microbial diversity, environmental microbiology, and phage therapy. Accurate classification of phage virion proteins (PVPs) is essential for understanding phage-host interactions and developing novel antimicrobial therapies. Traditional machine learning techniques such as support vector machine (SVM) and random forest (RF) have been widely applied for PVP classification but face challenges in handling high-dimensional sequence data. Recent advances in deep learning, particularly in the areas of transformers and autoencoders, have demonstrated superior performance in biological sequence analysis by capturing intricate sequence patterns. This study compares the performance of classical machine learning classifiers with deep learning models for PVP classification. We employ a transformer-based model (ESM2-650M) and an autoencoder for sequence encoding and feature extraction, contrasting their effectiveness against RF and SVM classifiers. Our dataset, derived from protein sequences in FASTA format, was processed using ESM2 tokenization for deep learning-based models. The results reveal that the transformer model significantly outperforms all classifiers, demonstrating strong predictive capability. The autoencoder model, while improving upon traditional methods, still requires further optimization. This research underscores the effectiveness of deep learning approaches in PVP classification and highlights the potential of transformers in computational virology, paving the way for enhanced phage protein analysis in synthetic biology and therapeutic applications.