Machine learning uncovers thousands of unknown bacterial immune defense systems

Bacterial immune systems protect against viral invaders called phages by precisely targeting specific phage genetic sequences. This precision provided inspiration and building blocks for biotechnologies such as the CRISPR gene editing system.

To discover more of these bacterial immune systems, Peter DeWeirdt and colleagues developed a machine learning model they call DefensePredictor. The model uses information from a protein’s genetic sequence and the sequences of its neighbors in the genome to predict whether the protein is involved in immune defense. When DeWeirdt et al. tested DefensePredictor in 69 diverse strains of E. coli, the model predicted hundreds of previously unknown immune systems and validated their defense function in 42 cases. Their analysis of 1,000 bacterial genomes identified almost 3,000 protein clusters that were different from known bacterial immune systems. The researchers have made DefensePredictor available to all researchers as an open-source software tool.

In a separate study also seeking to discover additional bacterial immune systems, Ernest Mordret et al. developed machine learning models to predict antiphage defense systems at scale and applied their models to over 120 million proteins from bacterial genomes. They identified hundreds of thousands of candidate antiphage families. Together, these two studies reveal that bacterial immunity is far more extensive than previously appreciated, and highlight how such discoveries can inspire powerful biotechnologies.