Computational host range prediction-The good, the bad, and the ugly

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Antimicrobial resistance necessitates alternative treatments, driving interest in bacteriophages.
• Bacteriophage applications in medicine, agriculture, and biotechnology require accurate host range data.
• Experimental determination of bacteriophage host ranges is time-consuming and labor-intensive.

Purpose of the Study:

• To benchmark the performance of computational tools for predicting bacteriophage host ranges.
• To evaluate the accuracy and precision of machine learning and deep learning approaches for host prediction.
• To identify limitations of current in silico methods for practical bacteriophage applications.

Main Methods:

• Utilized sixteen broad-spectrum bacteriophages with experimentally validated host ranges.
• Evaluated eleven recently developed computational host range prediction tools.
• Assessed prediction accuracy, precision, and sensitivity at species, genus, and strain levels.

Main Results:

• Machine learning and deep learning models demonstrated high accuracy and precision at the species/genus level.
• Strain-level predictions showed moderate sensitivity (<80%) but low precision (<40%).
• Current computational tools are better suited for metagenomics than specific strain targeting.

Conclusions:

• In silico host range prediction is a valuable but developing field.
• Further improvements are needed for computational tools to effectively guide experimental bacteriophage selection.
• Enhanced strain-level prediction accuracy is crucial for practical applications of bacteriophages.