Trans-ARG: Predicting Antibiotic Resistance Genes with a Transformer-Based Model and Pretrained Protein Language Model

This is the authors' abstract. We don't add key points for this paper.

Antibiotic resistance presents an emerging challenge to global health by reducing the efficacy of antibiotics administered to treat bacterial infections. Traditional approaches for detecting antibiotic resistance genes (ARGs) involve alignment-based sequence similarity techniques, which are time-consuming and resource-intensive. This signifies the necessity for developing advanced computational methods to detect ARGs early. To address this challenge, we introduce Trans-ARG, a novel multi-head attention transformer-based model designed to predict potential ARGs. Our approach leverages the pretrained ESMFold model to extract embeddings from protein sequences, capturing intricate structural and functional information. Protein sequences labelled as ARG or non-ARG were used as the dataset. We formulate our problem as a binary classification task, where the extracted embeddings serve as inputs to our transformer model with output results predicting ARGs. The transformer network is excellent at handling sequential data and capturing long-range dependencies, and the use of multi-head attention within the network improves he model’s capacity to comprehend relationships within the data from various perspectives. Additionally, we implemented a five-fold cross-validation strategy to ensure robust performance during training. Our experimental results demonstrate that Trans-ARG significantly outperforms standard existing baseline methods, presenting an accuracy of 90.96% and an AUC score of 97.08% on the test dataset. The high-efficiency performance of Trans-ARG is attributed to integrating embedding features obtained from the pretrained ESMFold model and effectively utilizing the transformer’s architecture. This integration allows Trans-ARG to generalize well across diverse protein sequences, making it a valuable tool for ARG prediction. Future research may explore applying this approach to predict antibiotic resistance categories, further enhancing our understanding of antibiotic resistance.