Predicting Antibiotic Resistance Genes and Bacterial Phenotypes Based on Protein Language Models

Boqian Wang¹Renjie Meng²Zhong Li³Mingda Hu¹Xin Wang¹Yunxiang Zhao¹Zili Chai¹Yuan Jin¹Junjie Yue¹Wei Chen²Hongguang Ren^1*

• ¹Beijing Institute of Biotechnology, Fengtai, Beijing, China
• ²National University of Defense Technology, Changsha, China
• ³Chinese PLA General Hospital, Hainan, China

Antibiotic resistance is becoming a global public health threat. To better understand antibiotic resistance genes (ARGs) and guide clinical antibiotic use against resistant bacteria, precise prediction of bacterial resistance genes and phenotypes is crucial. Although highthroughput DNA sequencing technology provides a powerful foundation for this purpose, current identification methods still lack precision and require manual intervention and verification. By integrating bacterial protein sequence information with ProtBert-BFD and ESM-1b protein language models, and further utilizing data augmentation and LSTM deep learning techniques, we developed a novel resistance gene prediction model. This model outperforms existing methods in accuracy, precision, recall, and F1-score, significantly reducing both false negatives and positives in resistance gene prediction, which provides a robust technical tool to identify ARGs. Additionally, we successfully applied this model to predict bacterial resistance phenotypes, offering a potential tool for clinical treatment guidance and related research.