Combining network toxicology with machine learning and experiment validation to analyze the molecular mechanism and core target screening of aristolochic acid nephropathy

Aristolochic acid nephropathy (AAN), a rapidly progressive kidney disease caused by aristolochic acid (AA), frequently progresses to end-stage renal disease and lacks effective therapy. The molecular targets that couple AA exposure to renal injury remain further defined. This study aimed to identify candidate targets of AAN and to clarify their diagnostic and mechanistic relevance through an integrated network toxicology, machine learning, and experimental approach. AA targets were predicted by ligand-protein similarity and pharmacophore modeling using ChEMBL, SEA, and SwissTargetPrediction (SMILES from PubChem CID 2236). AAN-associated targets were retrieved from OMIM (phenotype-mapped entries), TTD (clinical/research targets), and GeneCards (relevance score ≥ 1, gene-information-fitness ≥ 40) using the search terms “aristolochic acid nephropathy”, “AAN” and “Chinese herbal nephropathy”, and intersected with AA targets. Two surrogate GEO datasets (GSE1009: human glomeruli, 3 diabetic nephropathy vs. 3 controls; GSE37171: peripheral blood, 75 uremia vs. 40 controls) were combined post-ComBat batch correction and used only for hypothesis-generating feature selection, despite not being anatomically or etiologically matched to AAN. Candidate targets were ranked by LASSO logistic regression and SVM-RFE. Genes selected by both algorithms with corroborating bootstrap stability frequencies, were retained. Diagnostic performance was assessed by ROC analysis and leave-one-dataset-out (LODO) cross-validation. Immune infiltration was profiled by ssGSEA and CIBERSORT. AA-target binding was analyzed by molecular docking and dynamics simulations. Targets were validated in an AAN mouse model and in AA-treated HK-2 cells at a single 40-µM, 48-h condition. 290 AA-AAN overlapping targets were enriched in fatty-acid β-oxidation, PPAR signaling, and arachidonic acid metabolism. Fatty acid binding protein 3 (FABP3), cAMP-specific phosphodiesterase 4B (PDE4B), and Solute carrier family 1 member 3 (SLC1A3) were jointly selected by LASSO and SVM-RFE. The three-gene panel achieved an AUC of 0.981 (merged cohort) and 0.936 under LODO validation. ssGSEA and CIBERSORT concordantly showed CD8⁺ T-cell depletion and monocyte/macrophage enrichment. Molecular docking indicated AA binding to all three targets with energies between −7.0 and −8.6 kcal/mol, while molecular dynamics simulation confirmed stable, energetically favorable complexes for PDE4B and FABP3. Upregulation of all three targets was confirmed in AAN mice and AA-treated HK-2 cells. FABP3, PDE4B, and SLC1A3 emerge as preliminary candidate biomarkers and putative mediators of AAN. As surrogate datasets and limited experimental models were used, these results require validation in authentic AAN cohorts and in targeted perturbation studies.