BERT-T6: Toward High-Accuracy T6SS Bacterial Toxin Identification Using a Protein Language Model.

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• Type VI secretion system effectors (T6SEs) are critical bacterial virulence factors.
• Accurate identification of T6SEs is essential for understanding bacterial pathogenesis and host interactions.
• Existing computational predictors for T6SEs need performance improvements.

Purpose of the Study:

• To systematically evaluate sequence-based features and protein language model embeddings for T6SE prediction.
• To develop and validate a novel, high-performance computational predictor for T6SEs.

Main Methods:

• Evaluation of various sequence-based features and embeddings, including ProtBert.
• Development of BERT-T6, a predictor fine-tuned from ProtBert using transfer learning.
• Incorporation of imbalance-aware training strategies for classification.

Main Results:

• ProtBert embeddings demonstrated superior performance for T6SE representation.
• BERT-T6 achieved state-of-the-art performance on independent test sets.
• Key performance metrics included mean accuracy of 0.959, sensitivity of 0.909, and F1-score of 0.907.

Conclusions:

• Protein language models combined with transfer learning offer a powerful approach for T6SE prediction.
• BERT-T6 provides a valuable and accurate tool for identifying T6SEs.
• This predictor will aid in further research on bacterial virulence mechanisms.