MCT-ARG: Identification and classification of antibiotic resistance genes based on a multi-channel Transformer model

The global dissemination of antibiotic resistance genes (ARGs) poses a serious public health threat. Existing ARG prediction methods based on sequence homology or shallow machine learning often fail to capture non-sequence features and lack interpretability, limiting the discovery of novel resistance mechanisms (Gullberg et al., 2014). We propose MCT-ARG, a multi-channel Transformer framework that integrates protein primary sequences, predicted secondary structure, and relative solvent accessibility (RSA) to construct comprehensive multimodal representations for ARG prediction and mechanistic insight. By leveraging multi-head self-attention, MCT-ARG effectively models long-range dependencies across heterogeneous modalities. Additionally, a dual-constraint regularization strategy—combining entropy minimization and local continuity enforcement—improves attention focus on functionally relevant residues. Benchmark evaluations show that MCT-ARG achieves superior binary classification performance (AUC-ROC = 99.23 %, MCC = 92.74 %) and multi-class classification across 15 antibiotic categories (accuracy = 92.42 %, macro-AUC-PR = 99.65 %), maintaining robustness under class imbalance (MCC = 90.97 %). Interpretability analyses reveal that MCT-ARG learns evolutionarily conserved and functionally critical regions aligned with known catalytic motifs and active sites. Overall, MCT-ARG surpasses current state-of-the-art models in predictive accuracy, robustness, and interpretability. The framework and curated ARG database provide a valuable resource for resistance surveillance, functional annotation, and the rational design of novel antimicrobial agents. Code and data are available at https://github.com/nanbei45/MCT-ARG/tree/master.