Semantic search using protein large language models detects class II microcins in bacterial genomes

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Class II microcins are antimicrobial peptides with potential as novel antibiotics.
• Discovery of new class II microcins is limited by their short length and sequence divergence.
• Existing sequence-based methods like BLAST struggle to identify divergent microcins.

Purpose of the Study:

• To investigate the utility of protein large language models (LLMs) for detecting class II microcins in bacterial genomes.
• To compare the efficacy of LLM-based embeddings against sequence-based methods for microcin discovery.
• To identify novel putative class II microcins using LLM embeddings.

Main Methods:

• Generating numerical embeddings for proteins using LLMs.
• Applying these embeddings to search bacterial genome assemblies for microcins.
• Comparing the performance of embedding-based detection with BLAST searches.
• Analyzing genomic data from *Escherichia coli*, *Klebsiella* spp., and *Enterobacter* spp.

Main Results:

• LLM embeddings significantly outperform BLAST in reliably detecting known class II microcins.
• Highly sequence-divergent microcins exhibit small distances in the embedding space.
• Novel putative class II microcins were identified in *E. coli*, *Klebsiella* spp., and *Enterobacter* spp. genomes.
• These novel microcins were previously missed by sequence-based search methods.

Conclusions:

• Protein LLM embeddings provide a robust and sensitive method for discovering class II microcins.
• This approach overcomes limitations of sequence-based methods, enabling the identification of novel antimicrobial peptides.
• LLM-based discovery holds promise for expanding the repertoire of potential novel antibiotics to combat resistance.