Comprehensive Assessment of BERT-Based Methods for Predicting Antimicrobial Peptides.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Antimicrobial peptides (AMPs) are vital for combating drug-resistant bacteria, necessitating advanced prediction methods.
• Natural language processing (NLP) techniques, particularly BERT-based models, are increasingly applied for AMP recognition.
• Existing NLP methods for AMP prediction vary significantly in their underlying models, data, and feature encoding strategies.

Purpose of the Study:

• To comprehensively survey and evaluate current BERT-based methods for antimicrobial peptide (AMP) prediction.
• To establish a benchmark for comparing the predictive performance of various computational AMP prediction tools.
• To introduce a novel, high-performing AMP prediction method, iAMP-bert, based on advanced protein language models.

Main Methods:

• A comprehensive survey of existing BERT-based antimicrobial peptide (AMP) prediction tools was conducted.
• An independent benchmark dataset was created to evaluate the predictive capabilities of surveyed tools.
• A new method, iAMP-bert, was developed using the ESM-2 pretrained model for AMP prediction.

Main Results:

• LM_pred (BFD) demonstrated superior performance among the surveyed BERT-based tools.
• Cross-validation experiments using identical datasets revealed performance variations and highlighted the need for retraining.
• The proposed iAMP-bert method significantly outperformed existing approaches in predicting antimicrobial peptides (AMPs).

Conclusions:

• While BERT-based methods show promise for antimicrobial peptide (AMP) prediction, accuracy improvements are still needed.
• The novel iAMP-bert method, leveraging the ESM-2 model, represents a significant advancement in AMP recognition.
• iAMP-bert offers a publicly accessible and highly effective tool for antimicrobial peptide (AMP) prediction.