Hybrid deep learning-based rapid broad-Spectrum antimicrobial susceptibility prediction from whole-genome assemblies

Antimicrobial resistance (AMR) is a global public health threat. Mortality and poor treatment outcomes are the key consequences of AMR. Conventional antimicrobial susceptibility testing (AST) is slow, limited in coverage, and dependent on laboratory infrastructure, creating delays in clinical decision-making. In this study, we developed a hybrid deep learning model for broad-spectrum antimicrobial resistance prediction by analyzing 699 bacterial genome assemblies and paired antimicrobial susceptibility outcomes across 22 antibiotics. Genome assemblies were encoded using 6-mer frequency and antimicrobial susceptibility phenotypes were engineered into genome-antibiotic pairs for binary prediction. The proposed model integrates convolutional neural networks (CNNs) for local sequence feature extraction, bidirectional long short-term memory (BiLSTM) networks to capture long-range genomic dependencies, and an attention mechanism to improve interpretability. Model evaluation achieved an accuracy of 0.772 and AUROC of 0.77 at a resistance decision threshold of 0.55, with balanced accuracy of 0.697 and AUPRC of 0.489. The results demonstrate variable predictive performance across antibiotics and organism groups. This study demonstrates that a hybrid CNN-BiLSTM-Attention model can rapidly predict antimicrobial resistance from genome-derived k-mer features while incorporating organism and antibiotic metadata for broad-spectrum AST prediction. This framework offers a scalable way to predict susceptibility from genome data and can help advance the development of AMR decision-support tools for clinical use.