Unraveling the adverse outcome pathways of metabolic dysfunction-associated steatotic liver disease triggered by environmental mixtures via biological knowledge-driven machine learning.

Environmental pollutant mixtures are potential risk factors for metabolic dysfunction-associated steatotic liver disease (MASLD), yet their joint effects in complex co-exposure scenarios and toxicological mechanisms remain incompletely elucidated. This cross-sectional study quantified 54 urinary environmental pollutants via mass spectrometry among 1036 participants from the Yinchuan elderly cohort. Logistic regression and mixed-exposure models evaluated epidemiological associations and joint effects. Biological knowledge-driven machine learning algorithms extracted mechanistic features. Bayesian weighted quantile sum regression showed that neonicotinoid and metal(loid) mixtures were significantly associated with elevated MASLD odds (OR = 1.73 (1.37, 2.20) and OR = 1.51 (1.22, 1.84), respectively, per quartile increase), with acetamiprid and zinc as the leading contributors. Interaction analysis identified significant synergistic effects between thiamethoxam and lead, and between nitenpyram and selenium. Machine learning models incorporating biological knowledge-graph features retained ITGAM (integrin alpha-M, CD11b) and LIF (leukemia inhibitory factor) as key target proteins and GO:0006558 and GO:0032125 as core pathways related to inflammatory signaling, DNA damage, and apoptosis. An adverse outcome pathway framework was constructed for six pollutants that were significantly positively associated with MASLD in logistic regression and retained by machine learning models (thallium, 3-hydroxycarbofuran, acetamiprid, zinc, dinotefuran, and thiamethoxam), linking estrogen receptor agonism and NR1I3 (constitutive androstane receptor, CAR) suppression to downstream DNA damage, oxidative stress, and apoptosis as key mechanistic routes. This study provides an integrative approach connecting population-level mixture exposures with mechanistic hypotheses, offering a scientific basis for the health risk assessment of environmental chemicals. SYNOPSIS: Epidemiological models combined with biological knowledge-driven machine learning identified key pollutant-protein-pathway associations, and an adverse outcome pathway framework was constructed to link these findings to mechanistic hypotheses.