DefensePredictor: A machine learning model to discover prokaryotic immune systems.

Area of Science:

• Microbiology
• Bioinformatics
• Immunology

Background:

• Antiphage defense systems are crucial for bacterial survival against viral infections.
• These systems have implications for understanding eukaryotic innate immunity and developing biotechnologies like CRISPR-Cas9.
• Current discovery methods, like genomic colocalization, are limited to identifying systems within defense islands.

Purpose of the Study:

• To develop a machine learning model, DefensePredictor, for identifying antiphage defense systems.
• To discover previously unknown antiphage defense systems in prokaryotes.
• To explore the uncharacterized defense repertoire in prokaryotic genomes.

Main Methods:

• Utilized protein language model embeddings for classifying proteins.
• Developed DefensePredictor, a machine learning model for defense protein prediction.
• Applied DefensePredictor to 69 diverse *Escherichia coli* strains and 1000 prokaryotic genomes.

Main Results:

• Predicted hundreds of novel antiphage defense systems in *E. coli*.
• Experimentally validated 42 of the predicted systems.
• Identified nearly 3000 protein clusters without homology to known systems across 1000 prokaryotic genomes.

Conclusions:

• DefensePredictor is effective in discovering novel antiphage defense systems.
• A vast, uncharacterized repertoire of bacterial defense mechanisms exists.
• This work will advance the discovery of antiphage systems, revealing links between prokaryotic and eukaryotic immunity and accelerating biotechnology.