Contrastive-learning of language embedding and biological features for cross modality encoding and effector...

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Identifying bacterial virulence proteins is crucial for understanding pathogenicity and developing therapeutics.
• Protein Language Models (PLMs) show promise in predicting these proteins but face accuracy challenges.
• Existing methods struggle with the sensitivity and specificity required for comprehensive effector prediction.

Purpose of the Study:

• To introduce Contrastive-learning of Language Embedding and Biological Features (CLEF), a novel model for enhanced effector prediction.
• To integrate pre-trained PLM representations with supplementary biological features using contrastive learning.
• To improve the accuracy and sensitivity of predicting bacterial secreted effector proteins.

Main Methods:

• Developed CLEF, a model employing contrastive learning to merge PLM embeddings with diverse biological features.
• Utilized cross-modality biological features to enrich contextualized protein embeddings.
• Evaluated CLEF's performance against state-of-the-art models on enteric pathogens.

Main Results:

• CLEF significantly outperforms existing models in predicting type III, IV, and VI secreted effectors (T3SEs/T4SEs/T6SEs).
• Achieved recognition of all experimentally verified effectors in Enterohemorrhagic Escherichia coli and a high proportion in Salmonella Typhimurium.
• Experimental validation confirmed 12 predicted T3SEs and 11 predicted T6SEs in Edwardsiella piscicida.
• Integrated omics data within the CLEF framework to analyze effector-effector interactions and identify colonization genes.

Conclusions:

• CLEF effectively bridges the gap between in silico PLM capabilities and experimental biological data.
• The model offers a powerful blueprint for tackling complex biological prediction tasks.
• CLEF enhances understanding of microbial pathogenicity and facilitates therapeutic strategy development.