Multi-Omics and Explainable Artificial Intelligence Reveal Peimine Modulates Macrophage Polarization Homeostasis via CCR3/ITK/S1PR5/PRKCQ and MMP8/UGCG/KLF5 Axes in Pneumonia

Background: Pneumonia is driven by disordered immune inflammation dominated by dysregulated macrophage polarization. Peimine, the major bioactive steroidal alkaloid of Fritillariae Bulbus, exerts anti-inflammatory and immunomodulatory effects, whereas its direct targets and the metabolic routes through which it may influence macrophage polarization in pneumonia remain unclear. Objectives: This paper explores, and prioritizes for experimental testing, how peimine may regulate macrophage polarization in pneumonia, by combining public transcriptomic data with explainable machine learning. Methods: Potential human protein targets of peimine were predicted from three independent databases and pooled as a union. Two GEO datasets (GSE42830, GSE65682; 80 controls and 200 patients) were corrected for batch effects, screened for pneumonia DEGs and deconvoluted with CIBERSORT to estimate M0/M1/M2 macrophage fractions. WGCNA identified macrophage-related modules and hub genes, which were intersected with peimine targets and DEGs. Ten shared genes entered 113 classifiers under a 7:3 stratified split, with normalization and feature selection fitted inside the training data; SHAP was used to rank features of a gradient-boosting surrogate model, and AutoDock Vina docking was used to generate structural hypotheses. Results: A total of 238 peimine targets and 2991 pneumonia DEGs were obtained, and the estimated M0, M1 and M2 fractions were all higher in pneumonia samples. The brown module was positively associated with M0 (r = 0.68) and M2 (r = 0.28) fractions and contained 4 M0-related and 6 M2-related shared genes, whereas no M1-related shared gene was identified. The RF + Lasso hybrid model achieved a test-set AUC of 0.999 in an internal random split. SHAP ranked UGCG, MMP8 and KLF5 highest among the M0 features and CCR3, ITK, S1PR5 and PRKCQ highest among the M2 features; all seven were differentially expressed, with single-gene AUCs of 0.763–0.983. Docking predicted favorable binding of peimine to CCR3, ITK, S1PR5, PRKCQ and MMP8 (estimated binding energies −10.6 to −7.6 kcal/mol), and pathway annotation linked UGCG and S1PR5 to sphingolipid and glucosylceramide metabolism. Conclusions: These computational analyses do not demonstrate efficacy; they generate the testable hypothesis that peimine may act through two complementary modules, a CCR3/ITK/S1PR5/PRKCQ module associated with M2-like programmes and an MMP8/UGCG/KLF5 module associated with M0-derived inflammation and matrix remodeling, with a possible sphingolipid metabolic link. Experimental validation is required.