Host–pathogen interactions and genetic selection for resilience to post-weaning diarrhea in piglets

Abstract Rapid etiological diagnosis of diarrheal disease is challenging in multi-pathogen settings. Host blood transcriptomes provide a pathogen-agnostic signal, but models must be accurate and interpretable. We analyzed peripheral blood transcriptomes from pigs experimentally challenged with Lawsonia intracellularis, enterotoxigenic Escherichia coli , transmissible gastroenteritis virus, porcine deltacoronavirus, and Clostridium perfringens type C, plus healthy controls. Six machine learning algorithms were compared. The best model, a deep neural network (DNN), underwent feature selection to build 50-, 20-, and 5-gene classifiers. Performance was evaluated by receiver operating characteristic analysis and standard metrics. Interpretability was achieved using SHapley Additive exPlanations (SHAP) and a 20-gene Kolmogorov–Arnold network (KAN) that provided explicit gene-to-node functions. The DNN outperformed other algorithms (AUC: 0.690, accuracy: 73.2%, recall: 91.3%, F1: 0.792). A 20-gene DNN preserved this performance (AUC: 0.690, 95% CI: 0.507–0.854) and was chosen as the optimal model, whereas a 5-gene model slightly increased AUC but reduced recall. Pathway enrichment of the 20 genes implicated immune and metabolic pathways, including PI3K–Akt, cytokine–cytokine receptor interaction, and AMPK signaling. SHAP consistently identified ACTG1, SLC5A1, and ATP1A1 as dominant contributors across pathogens. The 20-gene KAN achieved comparable performance (AUC: 0.71) and yielded simple linear mappings showing negative effects of ACTG1/SLC5A1 and positive effects of ATP1A1 on the internal risk signal. A compact, biologically coherent 20-gene host signature enables accurate and interpretable prediction of diverse diarrheal pathogens, with ACTG1, SLC5A1, and ATP1A1 emerging as core cross-pathogen markers and promising diagnostic candidates.