PLM-ARG: antibiotic resistance gene identification using a pretrained protein language model

Area of Science:

• Microbiology
• Bioinformatics
• Artificial Intelligence

Background:

• Antibiotic resistance is a critical global health and environmental issue.
• Current methods for identifying antibiotic resistance genes (ARGs) struggle with novel or dissimilar sequences.
• This under-recognition poses a challenge for assessing the full scope of antibiotic resistance.

Purpose of the Study:

• To develop an AI-powered framework for simultaneous identification and classification of ARGs.
• To address the limitations of sequence similarity-based methods in ARG detection.
• To provide a robust tool for analyzing antibiotic resistance in diverse biological contexts.

Main Methods:

• Utilized a pretrained large protein language model (PLM).
• Trained the model on a comprehensive dataset of over 28,000 ARGs and 29 resistance categories.
• Employed a 5-fold cross-validation strategy and an independent validation set for performance assessment.

Main Results:

• The PLM-ARG framework achieved high accuracy, with a Matthew's correlation coefficient (MCC) of 0.983 ± 0.001 during cross-validation.
• Achieved an MCC of 0.838 on an independent validation set, significantly outperforming existing ARG prediction tools (51.8%-107.9% improvement).
• Demonstrated utility by annotating resistance in the UniProt database and evaluating ARG impact on environmental microbiota.

Conclusions:

• PLM-ARG offers a powerful and accurate solution for identifying and classifying ARGs.
• The framework enhances the ability to detect and understand antibiotic resistance.
• PLM-ARG is available via GitHub and a webserver for academic use.