Exo-Tox: Identifying Exotoxins from secreted bacterial proteins

Area of Science:

• Biochemistry
• Bioinformatics
• Microbiology

Background:

• Bacterial exotoxins are secreted proteins linked to diseases, necessitating accurate identification for drug discovery and safety.
• Current toxin predictors are generalized, limiting their effectiveness in identifying specific bacterial exotoxins.
• Protein Language Models (PLMs) offer potential for improved toxin prediction by analyzing protein sequence context.

Purpose of the Study:

• To develop a specialized predictor for bacterial exotoxins.
• To improve the accuracy of bacterial exotoxin identification compared to existing methods.
• To leverage PLMs for enhanced toxin prediction.

Main Methods:

• Developed Exo-Tox, a predictor trained on curated bacterial exotoxin and non-toxic protein datasets.
• Utilized PLMs to generate protein embeddings for feature representation.
• Compared Exo-Tox performance against BLAST and generalized toxin predictors.

Main Results:

• Exo-Tox achieved a Matthews correlation coefficient > 0.9, outperforming existing methods.
• The predictor demonstrated robust performance across varying protein lengths and signal peptide presence.
• Limited transferability to bacteriophage and non-secreted proteins was observed.

Conclusions:

• Exo-Tox reliably identifies bacterial exotoxins, addressing a gap in current prediction tools.
• Domain-specific training data and specialized models are crucial for accurate classification.
• The Exo-Tox model, data, and guidelines are publicly available.