Introduction: Bacteriophage depolymerases are enzymes encoded by bacteriophages that degrade polysaccharides on bacterial surfaces, playing a key role in controlling biofilm-associated infections. Despite the advent of machine-learning-based methods for predicting depolymerases, there is still considerable room for improvement in their accuracy and robustness. Methods: In this study, we propose DPSCA (Depolymerase Prediction with Simple Contrastive Augmentation), a novel deep learning approach that utilizes the pre-trained protein model ESM2 to generate dense vector representations of proteins. We integrate cross-layer contrastive learning to refine the model’s ability to predict phage depolymerases. Results: Experimental evaluations show that DPSCA outperforms existing prediction methods in terms of accuracy. The model demonstrates robust performance when applied to phage genomes, highlighting its versatility. Discussion: DPSCA’s use of contrastive learning significantly enhances prediction reliability, addressing the limitations of current approaches. The case study further underscores its potential for practical application in identifying phage depolymerases within diverse genomic environments. Conclusion: DPSCA is a powerful tool for phage depolymerase prediction, offering improved accuracy and robustness, with significant practical implications for biofilm-related infection control.
Ou et al. (Wed,) studied this question.