HTGM: A Hybrid Transformer–GNN Framework with Metadata Fusion for Multimodal Prediction and Spatiotemporal Mapping of Antibiotic Resistance Genes in Metagenomic Ecosystems

Abstract

The rapid emergence of antibiotic-resistant microorganisms poses a critical global health challenge, demanding advanced computational tools for accurate detection and monitoring of resistance mechanisms. In this study, we present a novel Hybrid Transformer–Graph Neural Network with Metadata Fusion (HTGM) framework that integrates sequence-level, ecological, and environmental information for robust prediction of antibiotic resistance genes (ARGs) from metagenomic samples. The model employs a Transformer encoder to extract contextual nucleotide-level representations, a GNN module to capture microbial co-occurrence and horizontal gene transfer (HGT) relationships, and a metadata-aware attention mechanism to incorporate environmental and clinical factors such as pH, temperature, and antibiotic usage. Experimental evaluation on soil, marine, and gut microbiome datasets demonstrates that HTGM achieves 96.8% accuracy and 96.5% F1-score, outperforming state-of-the-art baselines such as DeepARG and AMRFinderPlus. Furthermore, HTGM generates interpretable attention maps and spatiotemporal risk visualizations that highlight emerging antibiotic resistance hotspots. The results indicate that the proposed multimodal fusion strategy enhances model generalization, interpretability, and real-world applicability, providing a powerful tool for global antimicrobial resistance (AMR) surveillance and precision public health interventions.

Antibiotic Resistance Genes (ARGs); Metagenomics; Deep Learning; Transformer; Graph Neural Network (GNN); Metadata Fusion; Horizontal Gene Transfer (HGT); Antimicrobial Resistance (AMR) Surveillance; Multimodal Learning; Spatiotemporal Risk Mapping.