Protein and genomic language models uncover the unexplored diversity of bacterial immunity
Summary
The bacterial pangenome contains a vast diversity of antiphage systems, whose overall extent is still unknown. In this study, we developed complementary machine learning approaches to systematically predict antiphage function from genomic context, protein sequence, or their combination, achieving up to 99% precision and 92% recall. We validated these models experimentally in Escherichia and Streptomyces with the discovery of 12 antiphage systems. Applied to over 32,000 bacterial genomes, t
Content
# Protein and genomic language models uncover the unexplored diversity of bacterial immunity
*Published: 2026 Apr 2*
The bacterial pangenome contains a vast diversity of antiphage systems, whose
overall extent is still unknown. In this study, we developed complementary
machine learning approaches to systematically predict antiphage function from
genomic context, protein sequence, or their combination, achieving up to 99%
precision and 92% recall. We validated these models experimentally in
Escherichia and Streptomyces with the discovery of 12 antiphage systems. Applied
to over 32,000 bacterial genomes, these models expand the predicted antiphage
repertoire, with ~1.5% of bacterial genomes devoted to defense and more than 85%
of predicted protein families remaining uncharacterized. We provide an
interactive catalog of more than 19,000 candidate operon families for
experimental follow-up. Together, these findings show that most molecular
diversity in bacterial immunity remains uncharacterized and provide a foundation
for its systematic exploration.
DOI: 10.1126/science.adv8275