DeepSANet: A deep learning approach for hierarchical geographical source attribution of Salmonella

Microbial foodborne pathogens pose serious threats to public health and the food industry, with Salmonella being particularly widespread and hazardous. Tracing its geographical origins is crucial for outbreak control and prevention. However, existing genomic approaches lack required precision. To address this, we propose a deep source attribution network (DeepSANet) for hierarchical geographical source attribution of Salmonella, as the first to introduce deep learning to this task. In DeepSANet, a Swin Transformer is used to generate deep representations for genomic data. Then, we introduce the parallel hierarchical predictor (PHP) module to achieve simultaneous predictions of multi-level geographical origins. Finally, an adaptive hierarchical transfer (AHT) loss is designed to leverage the label hierarchy, enhancing prediction accuracy and consistency across multiple granularities. We conducted experiments on a public Salmonella enterica serovar Enteritidis genome dataset. DeepSANet outperforms existing methods, achieving source attribution accuracies of 91.88 %, 87.05 %, and 80.83 % at the region, subregion, and country levels, respectively. To comprehensively evaluate the generalizability of the proposed method, we constructed a large-scale dataset for Salmonella hierarchical geographical source attribution based on the EnteroBase, comprising globally distributed isolates across diverse serovars. Results show that DeepSANet achieves over 90 % source prediction accuracy across all geographical levels, demonstrating generalization capability to diverse serovars. Notably, above performance was achieved using only 3,002 core genome multilocus sequence typing (cgMLST) loci as features, highlighting the model's efficiency and practicality. These findings suggest that deep learning holds promise for supporting foodborne pathogen surveillance.