Emerging research links the gut, skin, and oral microbiomes to allergies, with serine proteases (SPs) identified as potential allergens. This study leverages deep learning and pre-trained protein language models (pLMs) to uncover allergenic SPs in metagenomic data. First, we develop a model to identify the catalytic serine residue in serine hydrolases, demonstrating how pLMs capture structural information. Next, we create a deep learning framework to detect candidate SP allergens across gene catalogs, using the conserved catalytic triad to identify homologs in gut and oral sites despite low sequence identity. Our model predicts a putative SP allergen resembling V8 protease, a known trigger for protease-activate... More
Emerging research links the gut, skin, and oral microbiomes to allergies, with serine proteases (SPs) identified as potential allergens. This study leverages deep learning and pre-trained protein language models (pLMs) to uncover allergenic SPs in metagenomic data. First, we develop a model to identify the catalytic serine residue in serine hydrolases, demonstrating how pLMs capture structural information. Next, we create a deep learning framework to detect candidate SP allergens across gene catalogs, using the conserved catalytic triad to identify homologs in gut and oral sites despite low sequence identity. Our model predicts a putative SP allergen resembling V8 protease, a known trigger for protease-activated receptor 1. It also identifies a cysteine protease similar to Der f 1 from dust mites. Immunization with these proteases induced allergic responses, validating their allergenic potential experimentally. This approach uncovers candidate allergens beyond traditional methods, offering new targets for allergy research. A record of this paper's transparent peer review process is included in the supplemental information.
