Integrating machine learning and experiments to elucidate the potential molecular mechanisms of methylparaben-induced Alzheimer’s disease: evidence from a Tau hyperphosphorylation cell model Authors: Hui E Zhang, Meng Li Xiao, Jin Jin Ji, Yu Rong Cheng, Fang Lu Published: 2025 ABSTRACT AbstractBackgroundAlzheimer’s disease (AD) was a progressive neurodegenerative disorder characterised by an insidious onset and gradual cognitive decline. It remained a significant global health challenge. Methylparaben (MEP), a preservative commonly used in cosmetics and food processing, had been associated with the development and progression of AD.MethodsFirst, we acquired the initial three-dimensional (3D) structure of MEP from PubChem (CID: 7456), followed by structural optimization via energy minimization using Chem3D software to complete its 3D structural characterisation. This was followed by systematic target prediction across the SwissTargetPrediction, SEA, GeneCards and OMIM databases. We then constructed protein-protein interaction (PPI) networks using STRING and visualised them in Cytoscape to identify core targets. Molecular docking simulations using CB-Dock2 elucidated the binding affinities between MEP and the key proteins. Experimental validation combined Gene Expression Omnibus (GEO) database analysis with quantitative reverse transcription polymerase chain reaction (qRT-PCR) to quantify transcriptional changes in SK-N-SH neural cells.ResultsA total of 153 potential targets associated with MEP and AD were identified. Ten core targets were determined through screening using the STRING platform and Cytoscape software, including HIF1A, IGF1R, PDGFRB, PTK2, VCAM1, CXCL12, ERBB2, ESR1, JAK2 and BCL2L1. Furthermore, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses revealed that the core MEP targets in AD primarily concentrate on the following key signalling pathways: Neuroactive ligand-receptor interactions, EGFR tyrosine kinase inhibitor resistance, HIF-1 signalling pathway and gamma-aminobutyric acid (GABA) synapse. Molecular docking simulations using CB-Dock2 confirmed a high binding affinity between MEP and these core targets. To investigate the mechanism of action of MEP, we validated the findings using clinical datasets and the human neuroblastoma cell line SK-N-SH. Upregulation of ten transcriptional expressions was observed, suggesting that MEP might influence cognitive function in patients with AD.ConclusionThis study elucidated the potential molecular mechanisms of MEP in the progression of Alzheimer’s disease-related tau pathology, offering new insights for the prevention and intervention of degenerative diseases that might be triggered by excessive exposure to MEP environments.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12883-025-04582-x.Keywords: Methylparaben, Alzheimer’s disease-related tau pathology, Machine learning, Experimental validation Introduction Alzheimer’s disease (AD) was recognised as the most prevalent form of dementia. It was a progressive neurodegenerative disorder that poses a significant challenge to global public health in contemporary society [1, 2]. This devastating condition was characterised by a gradual deterioration of cognitive function, persistent memory impairment and a progressive loss of independence when performing daily activities. It profoundly compromised patients’ quality of life and imposes a substantial socio-economic burden on affected families [3, 4]. According to the latest estimates from the World Health Organization (WHO), the global prevalence of AD currently affected approximately 50 million individuals. Epidemiological projections indicated that this figure was likely to double by 2050 [5]. These staggering statistics highlighted the significant social and economic burdens that were expected to emerge in the coming decades. The etiopathogenesis of AD was characterized by its multifactorial nature, involving an intricate interplay between genetic predisposition, environmental determinants, and lifestyle factors [6, 7]. Contemporary research had established that the hallmark neuropathological features of AD primarily include the cerebral accumulation of β-amyloid plaques and the formation of neurofibrillary tangles composed of hyperphosphorylated tau proteins [8]. These pathological alterations precipitated neuronal apoptosis and synaptic dysfunction, ultimately culminating in progressive cognitive deterioration [9]. Furthermore, emerging evidence highlighted the critical role of neuroinflammation, oxidative stress mechanisms, and blood-brain barrier dysfunction in accelerating AD progression [10]. Numerous studies had also indicated that a healthy diet is a key factor in reducing the risk of Alzheimer’s disease [11]. Methylparaben (MEP) was a compound of particular interest in this context, given its status as a ubiquitous environmental endocrine disrupter. It was widely used as a preservative in personal care products, pharmaceuticals and food formulations. It also occurred naturally in various fruits and vegetables [12, 13]. A substantial body of evidence suggested that this chemical compound might act as an endocrine-disrupting agent, with documented associations with diverse health complications including metabolic disorders, oncological conditions, and digestive pathologies [14]. Notably, epidemiological studies had identified correlations between prenatal MEP exposure and cognitive impairment in offspring [15], while cross-sectional research in the United States had implicated MEP as a potential risk factor for cognitive decline in the geriatric population [16]. However, the precise molecular mechanisms through which MEP might directly induce or exacerbate AD pathogenesis remain inadequately validated. Our investigation sought to address this knowledge gap by exploring the potential association between MEP exposure and AD development, with particular emphasis on elucidating the underlying biological mechanisms. The following was the specific analysis process (Fig. 1). Flowchart for studying the effects of Methylparaben on Alzheimer’s disease (Schematic of the study workflow: MEP 3D structure acquisition/optimization → multi-database target screening and intersection analysis (153 common targets) → PPI network construction and core target identification (10 genes) → GO/KEGG enrichment analysis → machine learning-based diagnostic validation (10 algorithms) and SHAP interpretation → molecular docking → in vitro qRT-PCR verification with SK-N-SH cells (three groups) to elucidate the molecular mechanisms of MEP-induced AD-related tau pathology) Materials and methods Network toxicological analysis To systematically evaluate the toxicological properties of MEP, we employed two recognized tools with established reliability in toxicological prediction. First, Swiss-ADME (a free web tool for in silico ADME property prediction, https://swissadme.ch/), which was used to assess the impact of MEP on the blood-brain barrier (BBB). Second, Protox 3.0 (a platform for in silico toxicity prediction, https://tox.charite.de/protox3/index.php?site=home), which was applied to classify the toxicity category of MEP. These tools were selected for their ability to integrate multidimensional chemical and biological data, thereby ensuring the scientific validity of MEP toxicity predictions. For details on the predictive methodology of these tools, please refer to Reference [17, 18]. Collection of MEP targets The initial three-dimensional (3D) structure of MEP was downloaded from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/, CID:7456). Structural optimization was performed using Chem3D software (version 14.0.0.17). The MM2 force field was applied for energy minimization until the root mean square (RMS) gradient reached 0.01 kcal/(mol·Å), to obtain a stable and low-energy 3D conformation. The optimized structure was used for subsequent analyses (e.g., molecular docking). Subsequently, targets were obtained using SMILES in the SwissTargetPrediction and SEA databases. By removing all duplicate targets, we ensure that only potential MEP targets unique to the ‘Homo sapiens’ species were retained [19, 20]. Selection of Alzheimer’s disease-related target network The keyword ‘Alzheimer’s disease’ was used to identify targets related to AD in the OMIM and GeneCards databases [21]. An intersection analysis was then performed using a Venn diagram generated with R software to obtain the shared targets between AD and MEP [22]. Construction of protein interaction network and screening of core targets We used STRING (https://cn.string-db.org/, version 12.0) to analyze the common targets between MEP and AD, with the species restricted to Homo sapiens. A high-confidence interaction threshold (minimum score > 0.9) was set based on three core rationales: (1) alignment with the “Highest confidence” level of the STRING scoring system, which integrates multi-dimensional evidence (e.g., gene co-expression, literature mining) to minimize false-positive interactions; (2) the need to filter non-specific associations for precise identification of AD-related core targets, avoiding network redundancy; and (3) consistency with standard protocols in neurodegenerative disease-related network toxicology studies [23, 24]. After constructing the PPI network (145 nodes, 596 edges), the data were imported into Cytoscape (version 3.7.0) for topological analysis [25]. Core regulatory genes were identified through a standardized workflow. First, calculating the degree centrality of all nodes using the Network Analyzer module (undirected network, weight threshold consistent with STRING score > 0.9). Second, initial screening of candidate nodes with degree values in the top 20%. Third, secondary sorting via the CytoHubba module’s Maximum Cluster Centrality (MCC) algorithm (default parameters: clustering coefficient threshold = 0.1, shortest path calculation = Dijkstra’s algorithm), and finally, integrating the two ranking results to determine the core targets [26]. GO and KEGG pathway analysis To elucidate the functional profiles of target genes, we performed functional enrichment analyses using the clusterProfiler package in R, which enables simultaneous GO annotation and KEGG pathway analysis through integrated biological database mining [27]. For statistical significance, we set the threshold of P < 0.05 (adjusted P-value, Padj < 0.05) to filter enriched GO terms and KEGG pathways, ensuring only biologically meaningful functional items were retained for subsequent analysis. Machine Learning-Based core gene screening To screen diagnostic markers for MEP-related AD, we built a multi-algorithm machine learning framework with rigorous data processing and parameter optimization. Gene expression data came from GEO’s GSE5281 (74 normal, 87 AD samples) and GSE138260 (19 normal, 17 AD samples), undergoing preprocessing: k-NN imputation (k = 5) for < 5% missing values (excluding > 5%), RMA normalization to eliminate batch effects, z-score scaling (mean = 0, SD = 1) for 10 core targets, and stratified partitioning into 70% training (n = 138) and 30% validation (n = 59) sets.Leveraging gene expression profile data from the validation cohort, ten classic machine learning algorithms (including Partial Least Squares (PLS), Random Forest (RF), Decision Tree Stacking (DTS), Support Vector Machine (SVM), Logistic Regression, K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Neural Network, and glmBoost) were employed to develop predictive models. Overfitting was avoided via 5-fold cross-validation (training set stability), regularization (L2 for Logistic Regression/SVM, pruning for XGBoost/GBM), and early stopping (10 epochs for Neural Network, 5 rounds for XGBoost). Primary evaluation used AUC (> 0.7 for good performance), with accuracy, precision, recall, and F1-score as secondary metrics to assess core genes’ diagnostic value [28]. Model interpretation We used the Shapley additive explanations (SHAP) algorithm to quantify the contribution of each feature to the predictive outcomes. This method assigns a SHAP value to each feature, making it possible to assess its influence on the model’s predictions in an interpretable way. Molecular Docking The optimized 3D structure of MEP was uploaded into the CB-Dock2 platform, charged with H atoms, and then subjected to free energy minimization [29]. In addition, the protein crystal structures obtained from the PDB database, were processed by adding hydrogen atoms and removing all water molecules. The binding affinity energy (kcal/mol) was computed using CB-Dock2, which projected the optimal docking model with the least amount of energy. Cell culture and qRT-PCR analysis Human neuroblastoma cell line SK-N-SH was obtained from Puri Biotech (Wuhan, China, Item No. CL-0214). Cells were cultured in DMEM-F12 (Thermo Fisher Scientific, Massachusetts, USA, Item No. 11320033) supplemented with 10% heat-inactivated fetal bovine serum (TransSerum FQ Fetal Bovine Serum, Beijing, China, Item No. C0232) and 1% penicillin-streptomycin (New Cell & Molecular Biotech, Jiangsu, China, Item No. C125C5) [30]. To clarify the regulatory effect of methylparaben (Selleck, Shanghai, China, Item No. S3985, Purity: 99.95%) on the transcriptional expression of core genes related to AD-related tau pathology model, we detected the mRNA expression levels of the ten core genes using quantitative reverse transcription polymerase chain reaction (qRT-PCR). Three groups of parallel samples were set up in the experiment: the normal control group (NC group, SK-N-SH neural cells without any pathological intervention), the AD model group (AD group, an AD-related tau pathology cell model constructed by intervention with 20 nmol/L okadaic acid (MedChemExpress, America, Item No. HY-N6785, Purity: ≥ 99.19%) for 24 h), and the MEP intervention group (SK-N-SH cells that were first treated with 20 nmol/L okadaic acid to establish an AD model and then treated with 3 µg/L MEP for another 24 h) [31, 32]. Subsequently, total RNA was extracted after 24 h of treatment with MEP. Reverse transcription was performed using SuperScript™III First-Strand Synthesis SuperMix (Thermo Fisher, America, Item No. 11752-050) [33]. Quantitative real-time polymerase chain reaction (qRT-PCR) was employed to analyze the mRNA expression of ten genes (Table 1). qRT-PCR was conducted using Power SYBR® Green Master Mix (Applied Biosystems, America, Item No. 4367659) [34]. Amplification conditions were as follows: 95 ◦C for 1 min, followed by 40 cycles of 95 ◦C for 15 s, and 63 ◦C for 25 s. Dissociation curve analysis was performed to confirm primer specificity, and β-actin served as an internal control. The expression level of the target RNA was calculated based on the threshold cycle (Ct) value using the comparative Ct method, where the relative expression level R = 2 −ΔΔCt. Here, ΔCt = Cttarget gene−Ctreference gene (with β-actin serving as the reference gene), and ΔΔCt = ΔCtexperimental group − ΔCt control group [35, 36]. Table 1qRT-PCR primers and conditionsGeneGenbank AccessionPrimer Sequences(5’to3’)Size(bp)Annealing (℃)β-actinNM_001101.3GATGACCCAGATCATGTTTGAGAC11260GGAGTCCATCACGATGCCAGTHIF1AAF208487GCAGTTCCGCAAGCCCTGAAAG11160CAGTGGTGGCAGTGGTAGTGGTIGF1RNM_000875.5CCTGCACAACTCCATCTTCGTG12560CGGTGATGTTGTAGGTGTCTGCPDGFRBNM_002609.4TGCAGACATCGAGTCCTCCAAC10960GCTTAGCACTGGAGACTCGTTGPTK2NM_005607.4GACCTGAGCGAGTTCATCAAG19860CTTGGTGCTGATGTCCTTGGTVCAM1NM_001078.4GCTGTCCTGATGCTGTTTCTG18760CAGGTGATGTTGCTGATGTTGACXCL12NM_002096.6TTGCTGCTTTAGCTGCCTTC21260CAGATCCGCAGCTCTGAAACERBB2NM_004448.4AGCCAACGTGTTCAGTGAAA19560TGGTCATAGGGCACGTAGGTESR1NM_000125.4GCTTACTGACCAACCTGGCAGA12960GGATCTCTAGCCAGGCACATTCJAK2NM_004972.4CAGAGCCTACAGCAAGATGG20360GGTCTTGGTGATGTCCTTGABCL2L1NM_001191.3GAGCTGGTGGTTGACTTTCT17860CAGGTATGCACCCAGAGTGA qRT-PCR primers and conditions Statistical analysis All results of this study were analyzed using the GraphPad Prism 8 software (version 8.01) for t-tests on two groups and one-way ANOVA on three groups. P < 0.05 or lower is considered statistically significant. All data were presented as the mean ± standard deviation (SD). Results Toxicity analysis of MEP Swiss-ADME analysis indicated that MEP has a significant impact on the blood-brain barrier (BBB) (Supplementary Table S1 and Fig. 2A), while Protox 3.0 classified MEP as a Category IV toxic substance (Fig. 2B). These findings were consistent with previous literature reports on MEP-mediated human toxicity, and lay the foundation for our further in-depth investigations into the toxic effects of MEP on the human body in a systematic and profound manner. The results of toxicity analysis (A) Swiss-ADME analysis showing MEP’s ability to affect blood-brain barrier (BBB) integrity; (B) Protox 3.0 prediction: MEP LD50 = 2000 mg/kg, toxicity Category IV, average similarity = 100%, consistent with known toxicological data Identification of targets of MEP-induced AD A total of 209 MEP targets from the SwissTargetPrediction and SEA databases were subjected to screening. Furthermore, a total of 15,976 Alzheimer’s disease targets were identified from the GeneCards and OMIM databases. Following the elimination of duplicate targets, a total of 153 overlapping targets were obtained (Fig. 3). Venn diagrams of intersection genes (Venn diagram (R software) showing 209 MEP targets (SwissTargetPrediction/SEA) and 15,976 AD targets (GeneCards/OMIM), with 153 overlapping targets as core research objects) Interaction network of potential targets and acquisition of core genes We constructed a protein-protein interaction (PPI) network using the STRING database, yielding 145 nodes and 596 edges. Simultaneously, we performed an in-depth analysis of the topological properties of network nodes using Cytoscape software (Fig. 4A). Based on degree values, we ultimately identified HIF1A, ESR1, ERBB2, CXCL12, VCAM1, PDGFRB, JAK2, PTK2, IGF1R, and BCL2L1 as core targets influencing MEP in AD (Fig. 4B). PPI network of targets associated with AD (A) STRING-derived PPI network (Homo sapiens, confidence > 0.9): 145 nodes, 596 edges (node size = interaction strength, edge thickness = confidence); (B) Top 10 core targets (Cytoscape: Degree + MCC algorithm): HIF1A, IGF1R, PDGFRB, PTK2, VCAM1, CXCL12, ERBB2, ESR1, JAK2, BCL2L1 (darker color = higher core degree) Functional enrichment analysis GO and KEGG enrichment analyses were performed to functionally interpret the shared targets (P < 0.05, Padj < 0.05). A total of 1508 enriched GO biological functions were identified, including 1263 biological processes (e.g., positive regulation of kinase activity, regulation of membrane potential), 78 cellular components (e.g., synaptic membrane, postsynaptic membrane), and 167 molecular functions (e.g., neurotransmitter receptor activity, hydro-lyase activity). Additionally, 28 significantly enriched KEGG pathways were detected, with key pathways including neuroactive ligand-receptor interaction (Padj = 9.0 × 10⁻5) and GABAergic synapse (Padj = 0.002). These findings suggest that candidate genes play a critical role in AD (Fig. 5). Enrichment analysis of potential targets (A) GO enrichment (P < 0.05, Padj < 0.05): 1263 biological processes, 78 cellular components, 167 molecular functions (e.g., kinase activity regulation, synaptic membrane localization); (B) KEGG enrichment: 28 pathways, including neuroactive ligand-receptor interaction (Padj = 9.0 × 10⁻⁵) and GABAergic synapse (Padj = 0.002) Clinical dataset validation and predictive model analysis To further validate the clinical relevance and predictive value of the core genes, this study first conducted empirical analysis using a clinical dataset (merged GSE5281 and GSE138260). Results showed (Fig. 6) that there were significant differences in the expression levels of the ten core genes (HIF1A, ESR1, ERBB2, CXCL12, VCAM1, PDGFRB, JAK2, PTK2, IGF1R, and BCL2L1) between AD samples and normal tissue samples (P < 0.05, P < 0.01), which initially confirmed the close association of these core genes with the pathological progression of AD. The expression of core genes in AD (Box plot of 10 core genes’ mRNA levels (GEO datasets GSE5281/GSE138260). Red boxes = AD samples, blue boxes = normal samples. Statistical significance: *P < 0.05, **P < 0.01, ***P < 0.001) Subsequently, a multi-algorithm machine learning prediction framework was constructed based on gene expression profile data. Ten classic machine learning algorithms, including Partial Least Squares (PLS), Random Forest (RF), Decision Tree Stacking (DTS), Support Vector Machine (SVM), Logistic Regression, K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Neural Network, and glmBoost, were employed to validate the diagnostic potential of the aforementioned core genes. It is important to clarify that the data used in this predictive model analysis are derived from predictive research, which differs from the clinical sample data in Fig. 6. Receiver Operating Characteristic (ROC) curve analysis showed that their Area Under the Curve (AUC) values were all greater than 0.70 (Fig. 7), confirming the reliable diagnostic potential of these core genes for AD. The results indicate that ten core genes exhibit optimal predictive capability under the SVW model (AU = 0.889). Comparison of receiver operating characteristic (ROC) curves for different machine (ROC curves of 10 algorithms (PLS, RF, SVM, etc.). AUC > 0.7 for all models; SVM performs best (AUC = 0.889, 95% CI: 0.801–0.977). Abscissa = 1-specificity, ordinate = sensitivity) To clarify the contribution of each core gene to the predictive model, the Shapley Additive Explanations (SHAP) algorithm was used for model interpretability analysis. Results showed (Fig. 8A) that IGF1R (SHAP value = 0.200) and HIF1A (SHAP value = 0.073) were the two key factors with the greatest impact on the prediction results. A violin plot displayed the expression distribution characteristics of the core genes under different conditions (Fig. 8B), where the width represents data density and colors distinguish expression levels. A scatter plot revealed the distribution of SHAP values of key genes and their direction of influence on prediction results, with a color gradient intuitively reflecting the regulatory effect of IGF1R and HIF1A expression levels (Fig. 8C). Force-directed analysis further indicated (Fig. 8D) that IGF1R (8.89, Δ = 0.194) acted as a major positive regulatory factor, while HIF1A (9.77, Δ = −0.198) served as a major negative regulatory factor. Together, they drove the predicted value (f(x) = 0.313) below the baseline expectation (E[f(x)] = 0.521). The consistent results from the above clinical validation and predictive model analysis suggest that IGF1R and HIF1A have potential diagnostic value for AD and are expected to become novel biomarkers or therapeutic targets, providing references for subsequent AD-related research and the development of treatment strategies. Identification of core genes in MEP-mediated tau pathogenesis of AD (A) SHAP feature importance: IGF1R (SHAP = 0.200) and HIF1A (SHAP = 0.073) are key; (B) Violin plot: gene expression distribution (width = data density, color = expression level); (C) SHAP scatter plot: gene expression vs. SHAP value (color gradient = regulatory direction); (D) Force-directed analysis: IGF1R (Δ = 0.194, positive regulator) and HIF1A (Δ = −0.198, negative regulator) Molecular Docking analysis Molecular docking analyses were conducted to investigate the interactions between MEP and ten core targets: HIF1A (PDB ID: 5L9V), IGF1R (PDB ID: 3LW0), PDGFRB (PDB ID: 1H9O), PTK2 (PDB ID: 6YOJ), VCAM1 (PDB ID: 1IJ9), CXCL12 (PDB ID: 4LMQ), ERBB2 (PDB ID: 8U8X), ESR1 (PDB ID: 8BZC), JAK2 (PDB ID: 8BXH) and BCL2L1 (PDB ID: 7JGW). The docking results obtained from CB-Dock2 revealed binding energies ranging from − 5.0 to −6.0 kcal/mol between MEP and target proteins (Supplementary Table S2). Generally, a binding energy of less than 0 indicates binding activity, whereas values below − 5.0 kcal/mol indicate strong binding activity, with lower values denoting stronger interactions. The results demonstrated strong affinities between MEP and the ten core targets, indicating spontaneous binding and suggesting their critical role in the molecular mechanisms of MEP-mediated tau pathogenesis of AD. The docking results were visualised in 2D representations using CB-Dock2 (Fig. 9). The optimal docking results of HIF1A, IGF1R, PDGFRB, PTK2, VCAM1, CXCL12, ERBB2, ESR1, JAK2, and BCL2L1 (2D docking schematics (CB-Dock2) of MEP with targets (PDB IDs: 5L9V, 3LW0, etc.). Dashed lines = intermolecular interactions (hydrogen bonds, hydrophobic forces). Binding energies: −5.0 to −6.0 kcal/mol) qRT-PCR The qRT-PCR results (Fig. 10) showed that compared with the AD-related tau pathology model group, the mRNA expression levels of the ten genes in the MEP intervention group were significantly increased (*P < 0.05, **P < 0.001). In contrast, the mRNA expression levels of these genes in the normal control group (NC group) were lower than those in both the MEP-mediated tau pathogenesis of AD group and the MEP intervention group. These results were highly consistent with the validation results of the clinical dataset, further confirming the key role of genes in the MEP-mediated tau pathogenesis of AD. qRT-PCR was used to detect the regulatory effects of MEP on the ten core genes (Bar graph of 10 core genes’ relative mRNA levels (2⁻ΔΔCt method, β-actin as internal control). Groups: NC (Control), AD (OA-induced), MEP (AD + 3 µg/L MEP). *P < 0.05, **P < 0.001 (error bars = SD, n = 3)) The core significance of this finding lay in the following aspects: it verified at the cellular and molecular level that MEP could affect the AD-related tau pathological process by upregulating the transcriptional expression of these genes, thereby providing direct experimental evidence for the hypothesis that MEP exposure is associated with the risk of AD. Meanwhile, these ten genes might serve as core molecular targets linking MEP to AD-related tau pathology, laying a key experimental foundation for further exploring the molecular mechanism of MEP-induced AD-related tau pathology and developing targeted intervention strategies. Discussion In this study, we integrated network toxicology, machine learning, molecular docking, and in vitro experiments to systematically explore the potential molecular mechanisms by which MEP might contribute to AD progression. Through multi-database target screening and topological analysis, we identified 153 overlapping targets between MEP and AD, and further confirmed 10 core targets (HIF1A, IGF1R, PDGFRB, PTK2, VCAM1, CXCL12, ERBB2, ESR1, JAK2, and BCL2L1) that might mediate the pathological effects of MEP. Functional enrichment analysis revealed that these core targets are primarily involved in key pathways such as neuroactive ligand-receptor interactions and GABAergic synapses, which are closely associated with neuronal function and cognitive processes. Clinical dataset validation and machine learning modeling further highlighted the diagnostic potential of ten core targets, while qRT-PCR experiments also confirmed that MEP can upregulate the transcriptional expression of these ten genes in AD cell models with tau protein hyperphosphorylation, providing the evidence for the link between MEP exposure and AD-related tau pathological processes. HIF1A, a key transcription factor in hypoxic response, regulated neuronal survival, neuroplasticity, and energy metabolism—processes critical for maintaining normal cognitive function [37–40].Abnormal activation of HIF1A had been implicated in AD-related energy metabolism disorders and neuronal damage. IGF1R, as the primary receptor for IGF-1, modulated neuroplasticity, learning, and memory [41–43], and its dysregulation might disrupt neuronal homeostasis in AD. Other core targets also played pivotal roles in AD pathogenesis: PDGFRB and PTK2 regulate neuroplasticity and synaptic function [44–48]. VCAM1 mediated neuroinflammation, a key driver of AD progression [49–53]. CXCL12 participated in neuroprotection and neuroregeneration [54–57]. ERBB2 and ESR1 regulated neuronal survival and amyloid-β (Aβ) deposition [58–61].JAK2 was involved in multiple pathological cascades including Aβ aggregation and tau hyperphosphorylation [62], and BCL2L1 inhibited neuronal apoptosis [63]. Collectively, these targets reflected the multi-system dysfunction of the “neuro-glio-vascular unit” in AD, suggesting that MEP might accelerate AD pathology through synergistic interactions with these molecular pathways. Despite the valuable insights gained, this study had several limitations that need to be addressed. First, the in vitro AD model used in this study has inherent limitations: OA-induced SK-N-SH cells primarily simulated tau hyperphosphorylation but did not fully recapitulate the dual pathological features of AD, namely the deposition of Aβ plaques and neurofibrillary tangles. As a result, the model might not capture the full complexity of AD pathogenesis, and the observed effects of MEP on ten core genes might be specific to the tau-related pathway rather than Aβ-mediated processes. Second, SK-N-SH cells are immortalized tumor-derived cells, which differ from primary neurons in terms of physiological characteristics and responsiveness to external stimuli, potentially limiting the translational relevance of the findings to native neuronal environments. To address these limitations and enhance the robustness of the conclusions, future research should consider the following directions. First, the findings should have been validated using more mature AD cell models, such as APP/PS1 double-transgenic cell lines or Aβ oligomer-treated primary cortical neurons, which could simulate both Aβ deposition and tau hyperphosphorylation, thereby comprehensively evaluating MEP’s effects on multiple AD pathological cascades. Second, in vivo studies should have been conducted using AD model mice (e.g., APP/PS1 transgenic mice) to investigate the effects of MEP exposure on cognitive function and pathological changes in the brain, providing preclinical evidence for the association between MEP and AD. Third, the dose-effect relationship between MEP exposure and AD risk should have been explored, as well as potential intervention strategies to mitigate MEP-induced neurotoxicity. Additionally, future studies (at that time) should have incorporated clinical cohort data to analyze the correlation between MEP exposure levels (e.g., in serum or cerebrospinal fluid) and AD incidence or cognitive decline, further confirming the clinical relevance of the findings. In conclusion, this study provided preliminary evidence that MEP might contribute to AD progression by regulating core targets, and identified key signaling pathways involved in this process. The limitations of the in vitro model highlight the need for more comprehensive experimental systems in future research, which would help clarify the exact molecular mechanisms of MEP-induced AD and inform the development of targeted prevention and intervention strategies for AD associated with environmental chemical exposure. Conclusion This study indicated that MEP might impact the pathogenesis of AD-related tau pathology by targeting specific genes and associated signaling pathways. Molecular docking demonstrated that MEP exhibited significant binding specificity to its target proteins. Quantitative real-time polymerase chain reaction (qRT-PCR) analyses further confirmed that ten key genes could serve as critical molecular links bridging MEP exposure and the development of AD-related tau pathology. Collectively, these findings laid a solid experimental foundation for subsequent investigations into the underlying mechanisms driving MEP-induced AD-related tau pathology. Future research should aim to validate the correlation between MEP exposure levels and AD risk in clinical or population-based studies, while also exploring potential targeted interventions to mitigate the adverse effects of MEP on brain tissue homeostasis. Supplementary Information Supplementary Material 1 Supplementary Material 2 Acknowledgements We sincerely acknowledge the PubChem (https://pubchem.ncbi.nlm.nih.gov/), SwissTargetPrediction (https://swisstargetprediction.ch/), SEA (https://sea.bkslab.org/), GeneCards (https://www.genecards.org/), OMIM (https://www.omim.org/), STRING (https://cn.string-db.org/, v12.0) and GEO (https://www.ncbi.nlm.nih.gov/geo/) team for developing and maintaining these freely accessible databases. This study benefits greatly from their high-quality open data. We also thank the developers of CB-Dock2 (http://183.56.231.194:8001/cb-dock2/php/blinddock.php), SwissADME (https://swissadme.ch/), Protox 3.0 (https://tox.charite.de/protox3/index.php? site=home) and Cytoscape (v3.7.0, https://cytoscape.org/) for making these analytical tools publicly available, facilitating our data processing and statistical analyses. Authors’ contributions Hui E Zhang: Writing-original draft. Meng Li Xiao: Methodology. Jin Jin Ji and Yu Rong Cheng: Visualization, Conceptualization. Fang Lu: Visualization, Methodology, Supervision. Funding This study was supported by the excellence clinical research program in research-oriented wards parallel project (Grant number: BRWEP2024Z014170102). Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate There are no human subjects in this article and informed consent is not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.