Tracing the impact of metal mixtures on primary membranous nephropathy: An integrative exposome-machine learning approach

Disentangling the nonlinear and interactive effects of chemical mixtures remains a central challenge in environmental health, hindering etiologic understanding of immune-mediated diseases such as primary membranous nephropathy (PMN). We developed a multi-stage pipeline integrating exposure clustering, penalized feature selection, interpretable machine learning, and model-based causal inference framework to dissect metal impacts on immune-renal dysfunction. Among 795 adults in Guangxi, China, we quantified 21 plasma metals and three endpoints: PMN diagnosis, estimated glomerular filtration rate (eGFR), and uric acid. Clustering revealed Zn-dominant profiles enriched in PMN patients versus Fe-centered patterns in controls. Tree-based models with SHAP interpretation identified Ba and Tl as being positively associated with PMN, while Co and Ni showed inverse associations. Mo was further linked to elevated uric acid and reduced eGFR, indicating coordinated relationships across metabolic and renal biomarkers rather than definitive evidence of toxicity. Interaction analyses revealed complex, nonlinear relationships, including Ni Cu antagonism and Co Ni dose-modulated associations, highlighting the collective dynamics of metal mixtures in influencing PMN risk. As the first study to couple plasma metal exposomics with interpretable and causal-inspired machine learning in PMN, this work identifies novel exposure patterns and generates etiologic hypotheses, offering a transferable analytic framework for environmental health research. • Plasma metal exposure pattern in PMN patients were markedly altered, with Zn emerging as the dominant signature across reconstructed clusters. • Elastic net screening and tree-based models accurately predicted PMN risk, with SHAP interpretability highlighting key exposure contributors. • Nonlinear association patterns and model-based strong differentiating signals for Ba, Mo, and Tl suggested divergent modes of influence within the immune-renal system. • Co and Ni showed consistent negative associations with PMN and renal biomarkers, suggesting unexpected modulatory roles under complex mixtures.