The rapid rise of antibiotic-resistant bacteria has created an urgent need for alternative treatments,such as bacteriophage therapy,which relies on phages to selectively eliminate specific bacterial pathogens.However,...The rapid rise of antibiotic-resistant bacteria has created an urgent need for alternative treatments,such as bacteriophage therapy,which relies on phages to selectively eliminate specific bacterial pathogens.However,the success of phage therapy hinges on accurately matching phages to their bacterial hosts,and predicting these phage–host interactions(PHIs)remains a significant challenge.The existing computational approaches for PHI prediction have major limitations.For example,the PredPHI model achieves less than 70%accuracy,and most methods predict only broad host categories(e.g.,genus levels)rather than pinpointing the exact bacterial strain.To address these gaps,we present a novel data-driven framework that predicts PHIs at the individual host strain level using only protein amino acid sequence data as input.Our approach extracts informative features from protein sequences—including amino acid composition(AAC),key chemical element abundance(AC),and molecular weight(MW)—and employs a one-dimensional convolutional neural network(1D CNN)tailored for sequential protein data.To further increase the prediction accuracy and mitigate class imbalance,we train multiple CNN models on different subsets of the data and integrate their outputs via a weighted ensemble strategy.This ensemble 1D CNN model achieves an AUC of 0.968,an AUPR of 0.976,an accuracy of 88%,and a sensitivity of 0.911,with a Matthews correlation coefficient(MCC)approximately 40%greater than that of PredPHI.Overall,this data-driven bioinformatics approach provides a powerful tool for rapid phage–host matching and custom phage design,potentially accelerating the development of targeted phage therapy.展开更多
基金supported by the Beijing Genomics Institute under grant numbers 62203134,6220020329 and 62176164Scientific Foundation for Youth Scholars of Shenzhen University 868-000001033385.The author gratefully acknowledges the support provided by the“Design of phage tail protein based on adversarial machine learning”project and the“Construction of artificial phage model based on deep generative adversarial network learning”project.And the datasets is provided by BGI-Shenzhen.
文摘The rapid rise of antibiotic-resistant bacteria has created an urgent need for alternative treatments,such as bacteriophage therapy,which relies on phages to selectively eliminate specific bacterial pathogens.However,the success of phage therapy hinges on accurately matching phages to their bacterial hosts,and predicting these phage–host interactions(PHIs)remains a significant challenge.The existing computational approaches for PHI prediction have major limitations.For example,the PredPHI model achieves less than 70%accuracy,and most methods predict only broad host categories(e.g.,genus levels)rather than pinpointing the exact bacterial strain.To address these gaps,we present a novel data-driven framework that predicts PHIs at the individual host strain level using only protein amino acid sequence data as input.Our approach extracts informative features from protein sequences—including amino acid composition(AAC),key chemical element abundance(AC),and molecular weight(MW)—and employs a one-dimensional convolutional neural network(1D CNN)tailored for sequential protein data.To further increase the prediction accuracy and mitigate class imbalance,we train multiple CNN models on different subsets of the data and integrate their outputs via a weighted ensemble strategy.This ensemble 1D CNN model achieves an AUC of 0.968,an AUPR of 0.976,an accuracy of 88%,and a sensitivity of 0.911,with a Matthews correlation coefficient(MCC)approximately 40%greater than that of PredPHI.Overall,this data-driven bioinformatics approach provides a powerful tool for rapid phage–host matching and custom phage design,potentially accelerating the development of targeted phage therapy.
