#22 · ML / AI methods

Protein language models uncover antimicrobial and antiphage genes in microbial genomes

The common thread is applying protein and genomic language model representations to prokaryotic sequence data for functional discovery. GeoARG fuses protein structural geometry with protein language models via distillation to detect remotely homologous antibiotic resistance genes from sequence. DefensePredictor uses protein language model embeddings to classify proteins as defensive, with predictions experimentally validated. A separate study applies protein and genomic language models to predict antiphage function from genomic context and sequence, with 12 new defense systems validated experimentally. EEMC contributes a resource of 78,213 bacterial and archaeal genomes from extreme-environment metagenomes and isolates, with protein language models predicting antimicrobial peptides. What is new is the use of learned sequence representations to identify divergent defense and resistance genes, paired with searchable genome collections.