Protein language models uncover carbohydrate-active enzyme function in metagenomics.

Area of Science:

• Microbiology
• Bioinformatics
• Enzymology

Background:

• Functional annotation of uncharacterized microbial enzymes from metagenomic data is challenging.
• Traditional methods like sequence homology often fail to identify remote homologs or structurally conserved enzymes.
• CAZyLingua was developed to address this gap using protein language models (pLMs).

Purpose of the Study:

• To develop and evaluate CAZyLingua, an annotation tool for carbohydrate-active enzyme (CAZyme) families and subfamilies.
• To improve the accuracy and scope of functional annotation for microbial enzymes.
• To leverage pLMs for enhanced CAZyme classification.

Main Methods:

• Development of CAZyLingua, a novel annotation tool utilizing protein language models (pLMs).
• Performance evaluation against state-of-the-art methods, including hidden Markov model-based and purely sequence-based approaches.
• Application of CAZyLingua to metagenomic datasets from mother/infant pairs and patients with Crohn's disease and IgG4-related disease.

Main Results:

• CAZyLingua achieved high precision and recall, comparable to HMM-based methods and superior to sequence-based methods.
• Identified over 27,000 putative CAZymes missed by other tools in a mother/infant metagenomic catalog, including horizontally-transferred enzymes.
• Uncovered disease-associated CAZymes in Crohn's disease and IgG4-related disease, noting an expansion of carbohydrate esterases (CEs) in IgG4-related disease.
• Functionally validated a CE17 enzyme overabundant in Crohn's disease, confirming its activity on acetylated manno-oligosaccharides.

Conclusions:

• CAZyLingua effectively enhances existing functional annotation pipelines for CAZymes.
• pLMs enable the discovery of novel CAZyme diversity and enzymatic functions relevant to health and disease.
• The tool contributes to a deeper understanding of biological processes related to host health and nutrition.