Integrated network toxicology and computational profiling of acetyl tributyl citrate-related mechanisms in Alzheimer's disease

Abstract

Background

Acetyl tributyl citrate (ATBC), an eco-friendly plasticizer, exhibits poorly characterized neurotoxic effects.

Objective

We integrated network toxicology, machine learning, and molecular docking to elucidate molecular mechanisms underlying the link between ATBC exposure and Alzheimer's disease (AD) pathogenesis.

Methods

Potential action targets of ATBC were screened from ChEMBL, TargetNet, and SwissTarget Prediction databases; disease-associated targets were derived from differential expression analysis of GEO datasets. Overlapping candidates underwent protein-protein interaction network construction (STRING) and subsequent Gene Ontology (GO)/Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. Machine learning employing SHAP prioritized pivotal targets, while molecular docking and dynamics simulations validated binding affinities.

Results

We identified 68 shared targets, of which five were designated as critical (CCKBR, RAF1, GABRG2, STS, RAPGEF3). GO enrichment revealed that ATBC compromises neuronal function and synaptic plasticity by perturbing glial cell differentiation, synaptic transmission, benzodiazepine receptor activity, and serine/threonine kinase activity. KEGG analysis implicated neuroactive ligand-receptor interactions, calcium, FoxO, and PI3K-Akt signaling pathways. Molecular simulations confirmed stable compound-target binding.

Conclusions

This integrative computational approach elucidates mechanisms underlying plasticizer-associated neurotoxicity in AD, establishing a framework for investigating neurological impacts of environmental contaminants.

Keywords

Alzheimer's disease ATBC machine learning molecular docking molecular dynamics simulation network toxicology

Introduction

Acetyl tributyl citrate (ATBC), a non-phthalate plasticizer, has attracted significant attention for use in food-contact materials, biomedical devices, and products for children. The extensive adoption is based on its advantageous toxicological profile and enhanced physicochemical properties.¹ However, concerns have been expressed about its significant migration potential. Research indicates that migration levels of polyvinyl chloride (PVC) cling film into food can reach up to 2.3 mg/kg,^2,3 which is roughly 10 times higher than the migration of di(2-ethylhexyl) phthalate (DEHP).⁴ Human exposure to ATBC can occur through multiple pathways, such as skin contact, ingestion of contaminated food sources, inhalation of contaminated air and dust, consumption of medications coated with ATBC, and interaction with medical instruments.^5,6 High concentrations of ATBC have been identified in indoor environments, which raises concerns about the risk of long-term exposure.⁷ Recent evidence suggests that low-dose ATBC exposure may be linked to alterations in lipid metabolism and neuroinflammation,⁸ raising concerns about its potential role in cognitive decline and neurodegenerative disorders. Notably, the study by Zheng et al.⁹ explored ATBC's toxicological mechanisms using network toxicology and molecular docking approaches, demonstrating that ATBC may accelerate cellular senescence through key pathways, thereby establishing a foundation for investigating its effects on neurodegenerative diseases, including Alzheimer's disease (AD).

AD is a degenerative neurological disorder marked by a deterioration in cognitive functions and a memory decline.¹⁰ Amid global population aging, AD poses a significant social and economic burden¹¹; however, its pathogenesis remains incompletely understood, and no curative treatments are available.¹² Importantly, although aging is the primary risk factor for AD,¹³ the disease involves distinct pathological features beyond general cellular senescence, including the accumulation of amyloid-β (Aβ) plaques, neurofibrillary tangles formed by hyperphosphorylated tau protein, synaptic dysfunction, and progressive neuronal loss.^14,15 These AD-specific molecular cascades necessitate targeted investigations separate from general aging studies. The role of environmental pollutant exposure as a potential risk factor for AD has garnered increasing attention. While this previous work⁹ has established the impact of ATBC on general aging processes, whether ATBC specifically affects AD-related molecular networks and pathogenic mechanisms remains unexplored. Therefore, a systematic evaluation of ATBC's neurotoxic potential and AD-specific pathogenic mechanisms is essential to address this critical knowledge gap in understanding environmental contributions to AD risk.

Network toxicology is an emerging paradigm that uses bioinformatics, large-scale computational analytics, and multi-omics profiling to construct “compound-target-disease” network architectures. This approach allows for systematic investigation of the pleiotropic pathways by which toxicants cause hazardous consequences.¹⁶ Computational elucidation of protein-ligand interactions necessitates synergistic integration of molecular docking and molecular dynamics (MD) simulations. The former characterizes binding energetics and interaction geometries between small molecules and their cognate protein targets.¹⁷ While the latter assesses the temporal stability and conformational dynamics of these complexes. The integration of these methods offers an effective approach for the swift evaluation of toxicity and the underlying mechanisms of toxicants. Moreover, leveraging the analytical capabilities of machine learning models enables a thorough systems-level examination of biological complexity by integrating them.

Therefore, this study systematically identifies the potential core targets of ATBC implicated explicitly in AD pathogenesis through an integrated network toxicology and machine learning approach, and validates their binding activities via molecular docking. The findings are anticipated to yield new insights into the relationship between ATBC exposure and AD pathogenesis, offering fresh perspectives for the prevention and treatment of this condition.

Methods

Acquisition of AD-related targets

Four transcriptomic datasets from the NCBI GEO database (GSE5281, GSE132903, GSE48350, and GSE36980) were selected to investigate AD. GSE5281 and GSE132903 served as the discovery cohort for initial feature identification, with subsequent validation performed using GSE48350 and GSE36980. To mitigate potential batch effects across the datasets, a multi-stage normalization pipeline was implemented. This pipeline first employed Surrogate Variable Analysis (SVA) to address latent confounding variables in the discovery cohort, followed by ComBat harmonization to correct for residual batch effects. Post-correction principal component analysis (PCA) confirmed the efficacy of this normalization strategy, demonstrating improved sample clustering within batches. The complete analytical workflow is visually represented in Figure 1.

Figure 1.

Work flow.

Acquisition of chemical components and targets of acetyl tributyl citrate

ATBC underwent thorough characterization through data integration. Physicochemical properties and biological parameters were gathered from PubMed, and the canonical 2D structural descriptor (SMILES: CCCCOC(=O)CC(CC(=O)OCCCC)(C(=O)OCCCC)OC(=O)C) was obtained from PubChem. Target prediction in Homo sapiens was conducted using the ChEMBL Database (ligand-receptor interactions, https://www.ebi.ac.uk/chembl/), SwissTargetPrediction (chemical genomics, http://www.swisstargetprediction.ch), and TargetNet (small-molecule-target interactions, http://targetnet.scbdd.com/). All anticipated targets were limited to the human proteome.

Differential gene expression analysis

Transcriptome data were analyzed using the “limma” package. Differentially Expressed Genes (DEGs) were identified using a threshold of |log2FC|> 1.5 and an FDR-adjusted p value < 0.05. ggplot2 was used to visualize results.

Weighted gene co-expression network analysis

A scale-free co-expression network was constructed utilizing the WGCNA package. Following sample quality control and hierarchical clustering-based outlier exclusion, the optimal soft-thresholding power was determined via dynamic tree-cutting to attain a scale-free topology (R² > 0.85). Gene modules were subsequently delineated through TOM-based hierarchical clustering (minModuleSize = 30, mergeCutHeight = 0.25) and correlated with phenotypic variables via their eigengenes (Pearson's |R| > 0.5, p < 0.05). Hub genes were characterized by elevated intramodular connectivity (kME > 0.8).

Identification of ATBC-AD-related targets

Core targets putatively mediating the interplay between ATBC and AD were delineated through concordance analysis of DEGs, WGCNA hub genes, and ATBC-predicted targets, as depicted in Venn diagrams.

Functional enrichment analysis

Using the clusterProfiler program, we conducted KEGG pathway and GO (biological process, cellular component, molecular function) studies to clarify the mechanistic functions of ATBC in AD pathogenesis (p < 0.05).

Machine learning: validation of ATBC-AD-related genes

Three complementary machine learning methods were employed to identify the most predictive hub genes accurately. A LASSO regression model was fitted using the glmnet package, with a binomial distribution. Optimal λ values were established via 10-fold cross-validation to reduce prediction error and eliminate redundant features. Second, the SVM-RFE algorithm, utilizing the e1071 package, was used to systematically eliminate features with the lowest weights via 10-fold cross-validation, thereby reducing the root mean squared error and identifying the most discriminating genes. Finally, a random forest analysis with 500 decision trees was conducted to assess variable importance in the high-dimensional dataset. Genes with an importance score exceeding 10 were identified as biologically significant. The convergence of genes identified by these three distinct methodologies, each grounded in different mathematical frameworks and offering unique benefits for feature selection, was ultimately recognized as a collection of ATBC- and AD-related genes. The research findings were represented through an Upset plot. The diagnostic efficacy of these key genes was evaluated by generating ROC curves for each gene and calculating AUC values. Five genes with the highest AUC values were selected for subsequent analysis and validated using an independent dataset.

SHAP analysis

This study utilized eleven supervised machine learning algorithms: Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine with Kernel (SVM), Logistic Regression (LR), k-Nearest Neighbors (KNN), Partial Least Squares Regression (PLS), AdaBoost (Adaptive Boosting), Artificial Neural Network (ANN), Naive Bayes Classifier (NB), Linear Discriminant Analysis (LDA), and Least Absolute Shrinkage and Selection Operator (LASSO). The algorithms facilitated the development of predictive models, markedly improving their accuracy and generalization capacity. Shapley Additive Explanations (SHAP) analysis was deployed to evaluate the optimal model's performance while quantifying individual feature contributions to its predictive output, thereby elucidating the underlying inferential architecture.

Molecular docking

Molecular docking was conducted to interrogate the binding modes between ATBC and candidate target proteins. Requisite 3D protein architectures were retrieved from UniProt, and the ligand structure was retrieved from PubChem (SDF format). The preprocessing workflow involved excision of solvent and phosphate moieties and PDB conversion in PyMOL 2.6, followed by hydrogenation of all structures, specification of ligand rotatable bonds, and docking grid assignment using AutoDock 1.5.6. Docking was performed with AutoDock Vina, and the resultant poses were visualized in Discovery Studio 2019 and PyMOL. Binding affinity was quantified via the calculated binding energy (ΔG), with <0 kcal/mol indicating spontaneous association and < −5 kcal/mol denoting a stable complex.

Molecular dynamics simulation

One-hundred-nanosecond molecular dynamics (MD) simulations were performed using Gromacs (v2022.3), parameterizing the protein and ligand with the amber14sb and GAFF2 force fields, respectively. Each complex was solvated within a cubic TIP3P water box and neutralized with counterions, maintaining a minimum distance of 1.0 nm from the periodic boundary.¹⁸ Long-range electrostatics were computed using the Particle Mesh Ewald (PME) method with a 1.0 nm cutoff. The SHAKE algorithm constrained all hydrogen bonds, permitting a one fs integration time step with the Verlet leap-frog integrator. The solvated system was subjected to a two-stage energy minimization protocol (3000 steepest descent steps, then 2000 conjugate gradient steps). Production simulations were conducted using an NPT ensemble at constant temperature (310 K) and pressure (1 atm). Trajectories were subsequently analyzed for RMSD, RMSF, hydrogen bonds, Rg, SASA, and Gibbs Free Energy using native Gromacs utilities.

Results

Identification of AD-related target genes

Batch effects were mitigated by merging the GSE5281 and GSE132903 datasets and then normalizing the gene expression matrix. The effectiveness of this normalization was validated by PCA, which revealed a more pronounced clustering pattern in the normalized dataset relative to the original (Figure 2(a), b). Subsequent differential expression analysis identified 991 genes with significantly altered expression in AD, as illustrated in volcano plots and heatmaps (Figure 2(c), d). We identified the optimal soft-thresholding power (β) for constructing WGCNA. Systematic evaluation of power values (1–20) revealed that β=4 was the minimum value meeting the scale-free topology criterion (R² ≥ 0.85). Hierarchical clustering of the Topological Overlap Matrix (TOM) identified 30 distinct gene modules. The modules were color-coded to enhance visualization (Figure 2(e)). Analysis of the module-trait relationships revealed significant associations between specific modules and AD (p < 0.05) (Figure 2(f)). The process of integrating DEGs via conventional analysis and WGCNA module genes, after duplicate removal, identified 1885 genes linked to AD (Figure 2g).

Figure 2.

Identification of target genes associated with AD. (a,b) PCA demonstrating successful batch effect correction between the GSE5281 and GSE132903 datasets before (a) and after (b) integration. (c) Volcano plot identifying differentially expressed genes (DEGs) based on expression fold-change and statistical significance. (d) Heatmap depicting hierarchical clustering of samples based on DEG expression profiles, illustrating relative gene expression levels across samples. (e) Weighted Gene Co-expression Network Analysis (WGCNA) gene clustering dendrogram and corresponding module assignments. (f) Module-trait correlation heatmap, where each cell displays the correlation coefficient between a module eigengene (ME) and the AD phenotype, with the corresponding p-value in parentheses. (g) Venn diagram identifying core target genes at the intersection of DEGs and the most significant AD-associated WGCNA module.

Identification of ATBC-related targets in AD

Putative biological targets of ATBC were delineated by integrating predictions from three computational platforms: ChEMBL, TargetNet, and SwissTargetPrediction. The molecular structure of ATBC, used as input for these predictions, was obtained from the PubChem database. Following data consolidation and removal of duplicate entries, 668 unique potential targets were identified (Figure 3(a)). An in-depth investigation contrasting these ATBC targets with genes known to be related to AD identified 68 potential major targets involved in ATBC-modulated AD pathogenesis (Figure 3(b)). GO and KEGG pathway enrichment analyses (Figure 3(c)-(e)) provided detailed functional insights into these key targets, illuminating potential mechanisms of action. GO analysis revealed significant enrichment of: gliogenesis, response to peptide, and regulation of postsynaptic membrane potential (biological processes); GABA-A receptor complex, synaptic membrane integral component, and neuron projection membrane (cellular components); and benzodiazepine receptor activity, polypeptide Ser/Thr phosphotransferase capacity, and γ-aminobutyric acid-mediated chloride ion permeability (molecular functions). KEGG analysis identified significant involvement in several crucial pathways, including neuroactive ligand-receptor interaction; various signaling cascades, such as the Rap1 signaling pathway, calcium signaling pathway, and FoxO pathway; vascular smooth muscle contraction; insulin resistance; adherens junction; and EGFR tyrosine kinase inhibitor resistance. The findings indicate that ATBC may contribute to AD pathogenesis through modulating synaptic function and neuronal signaling pathways, disrupting cell adhesion, and causing metabolic disturbances.

Figure 3.

Identification of disease targets associated with ATBC in AD. (a) Target prediction was conducted utilizing CHEMBL, TargetNet, and SwissTargetPrediction. (b) A Venn diagram illustrates the comparison of genes associated with ATBC exposure and AD, revealing 68 overlapping genes, constituting 2.7% of the total. (c) PPI network illustrates interactions among overlapping genes. Nodes represent genes and edges represent predicted interactions. (d) Gene Ontology (GO) enrichment identifies genes that overlap in biological processes (BP), cellular components (CC), and molecular functions (MF). The X-axis represents gene count, while the gradient scale indicates -log10(p-value), indicating the level of statistical significance. (e) KEGG analysis displays the top 10 pathways ranked by p-value and their associated proteins, with lines representing the connections between pathways and genes.(f) KEGG analysis shows enriched pathways for overlapping genes. X-axis = gene ratio, dot size = gene count, color gradient = -log10(p-value).

Identification of key ATBC-AD-related genes through machine learning

We performed an in-depth analysis of Hub Genes using three machine learning methods: LASSO regression (Figure 4(a), b), SVM-RFE (Figure 4(c), d), and Random Forest (Figure 4(e), f). ROC curves and AUC values were generated by integrating each algorithm's results (Figure 4g) to assess the diagnostic potential of these core genes. Based on the AUC value ranking (Figure 4h), we ultimately identified five key ATBC-AD-related genes: CCKBR, RAF1, GABRG2, STS, and RAPGEF3. The expression levels of these five genes were then validated in independent datasets (GSE48350, GSE36980) (Figure 4(i)).

Figure 4.

Machine learning approaches and validation of genes associated with ATBC-AD. Thirty-three genes were discerned utilizing the LASSO regression technique; (a) a plot illustrating the bias-variance trade-off; (b) a coefficient plot based on the L1 norm. Additionally, seventeen genes were detected through the SVM-RFE method. (c) A graphical representation showing the relationship between the number of features and the accuracy achieved via 10-fold cross-validation; (d) A plot depicting the relationship between the number of features and the error rate obtained through 10-fold cross-validation. Furthermore, thirty-one genes were identified through random forest analysis employing 500 decision trees, with importance scores surpassing 3. (e) A plot demonstrating the connection between random forest error and the quantity of trees; (f) a depiction of feature importance. (g) Upset Venn diagrams consolidating findings from the three methodologies; (h) An evaluation of the diagnostic capabilities of ten key genes through ROC curve analysis and AUC values. Expression levels of ATBC-AD-Related Genes in the Validation Cohort. **p < 0.01; ***p < 0.001.

SHAP explanatory analysis

This study employed 11 supervised machine learning algorithms—RF, GBM, SVM, LR, KNN, PLS, AdaBoost, ANN, NB, LDA, and LASSO—to predict ATBC-related genes (Figure 5(a)). Evaluation of the area under the curve (AUC) revealed that the PLS regression model exhibited the best overall predictive performance. To elucidate the mechanism of the PLS regression model, we calculated SHAP values using the R package shapviz and ranked the genes by their contribution. The SHAP analysis results, illustrated via waterfall plots, force plots, bar charts, swarm plots, and scatter plots (Figure 5(b)-(f)), revealed significant insights: the bar chart demonstrated that RAPGEF3 and GABRG2 contributed most substantially to the PLS regression. At the same time, STS exhibited the smallest effect (Figure 5(b)). The swarm plot further confirmed that RAPGEF3 is a major driver of sample classification, and collectively, these five genes have a positive impact on the predictive outcomes (Figure 5(c)). The scatter plot revealed a complex interplay between gene expression and model predictions. The expression of RAPGEF3 and RAF1 was positively correlated with SHAP values, indicating a facilitative role in model predictions. In contrast, CCKBR expression was associated with negative SHAP values, suggesting an inhibitory effect. There exists a negative correlation between GABRG2 and STS. The waterfall and force plots quantified the aggregate contributions of the five genes to the prediction scores, providing a detailed interpretation of sample classification (Figure 5(d), f).

Figure 5.

SHAP explanatory analysis. (a) PLS model was selected from eleven supervised machine learning algorithms trained for the prediction of ATBC-AD-Related Genes . The predictive utility of the five genes in the resulting signature was subsequently interrogated using SHAP. Bar plots quantified the magnitude of each gene's contribution (b), while a beeswarm plot identified RAPGEF3 as having the predominant sample-level impact (c). The SHAP summary plot further specified these contributions, where positive and negative values denote an increasing or lowering effect on the prediction, respectively (f). Finally, waterfall plots illustrated the basis of classification for individual samples (d), and scatter plots visualized the impact of key genes through their corresponding SHAP values (e).

Molecular docking

Molecular docking simulations predicted high-affinity interactions between ATBC and its five machine learning-identified protein targets (CCKBR, RAF1, GABRG2, STS, and RAPGEF3), with favorable binding energetics observed for all complexes (binding energies < −5 kcal/mol). Analysis of the resulting conformations confirmed the formation of stable docking poses within the binding pocket of each protein (Figure 6). These in silico findings provide a structural rationale for the direct engagement of ATBC with these key AD-related proteins.

Figure 6.

Molecular docking. Diagram depicting the molecular docking of ATBC with genes associated with ATBC-AD.

Molecular dynamics simulation verification

MD simulations were conducted to elucidate the dynamic interaction patterns and evaluate the binding stability of ATBC with five potential target proteins. The equilibrium state of the systems was assessed by analyzing the Root Mean Square Deviation (RMSD) (Figure 7(a)). The results indicated that the RMSD curves for the complexes formed by ATBC with RAF1, CCKBR, and GABRG2 reached equilibrium in the mid-to-late stages of the simulation and remained relatively stable thereafter, converging below 2.6 Å, 2.7 Å, and 3.3 Å, respectively. This finding suggested the formation of stable conformations. In contrast, the STS and RAPGEF3 systems exhibited significant overall fluctuations, with RMSD values of approximately 7 Å and 6.5 Å, respectively, revealing their conformational instability.

Figure 7.

Analysis of ATBC with target protein via molecular dynamics. (a-c) Time evolution of structural stability parameters: (a) Root mean square deviation (RMSD), (b) radius of gyration (Rg), and (c) solvent-accessible surface area (SASA) for all systems. (d) Time evolution of the number of hydrogen bonds. (e) Root mean square fluctuation (RMSF) per residue showing regional flexibility. (f) Free energy landscapes projected onto RMSD and radius of gyration (Rg), illustrating conformational sampling and stability. The scale represents free energy in kJ/mol, where lower values correspond to more stable conformational states.

Subsequently, the flexibility of residue-specific regions within each system was analyzed using Root Mean Square Fluctuation (RMSF) (Figure 7(e)). In the RAF1, CCKBR, and GABRG2 complexes, the majority of residues fluctuated within a narrow RMSF range, indicating a more rigid structure. Conversely, the STS and RAPGEF3 complexes displayed substantially higher flexibility in specific regions, which may be associated with their function or instability. Furthermore, the Radius of Gyration (Rg, Figure 7(b)) and the Solvent-Accessible Surface Area (SASA, Figure 7(c)) remained relatively constant across all systems, suggesting that the protein backbones maintained a compact conformation without significant unfolding throughout the simulations. Analysis of hydrogen bonds (Figure 7(d)) provided further evidence for differences in binding strength: the more stable systems generally formed a greater number of more persistent hydrogen-bond networks.

To gain energetic insights into the dynamic behavior of the complexes, three-dimensional Free Energy Landscape (FEL) plots were constructed based on RMSD and Rg (Figure 7(f)). In these plots, regions shaded in blue represent the system's stable-state conformations. The energy landscapes for RAF1, CCKBR, and GABRG2 featured deeper, more localized blue potential wells, indicating that their binding is thermodynamically more favorable and that they more readily access and maintain low-energy stable states.

In summary, the MD simulations consistently revealed differential dynamic behaviors of ATBC upon binding to various target proteins, as assessed by conformational stability, residue flexibility, key interactions, and energy distributions. These findings provide robust theoretical support for elucidating the selectivity and affinity differences of ATBC binding.

Discussion

This study integrates network toxicology, machine learning, and molecular docking strategies to evaluate the potential molecular mechanisms by which ATBC may contribute to AD. We successfully identified 68 common targets of ATBC and AD, and further screened five core targets (CCKBR, RAF1, GABRG2, STS, and RAPGEF3) using machine learning algorithms. SHAP analysis of our machine learning model identified the principal molecular targets of ATBC, which functional enrichment analysis showed are significantly concentrated in neuroactive ligand-receptor interactions and key signaling cascades, including the Rap1, calcium, and FoxO pathways. Providing a structural basis for these findings, subsequent molecular docking and dynamics simulations confirmed stable, high-affinity binding between ATBC and its core protein targets. Collectively, these results establish a mechanistic framework linking ATBC exposure to perturbations in critical neurosignaling pathways.

Compared with previous studies investigating the toxicological effects of ATBC, our findings complement and extend the study by Zheng et al.,⁹ which used network toxicology and molecular docking approaches to examine ATBC's effects on the general aging process. Their analysis identified 32 aging-related targets, including EGFR, STAT3, and BCL-2, with pathways enriched in cellular senescence and proliferation, highlighting ATBC's involvement in cellular and systemic aging processes and its associations with oxidative stress and apoptosis. Although both studies employed computational toxicology frameworks, the present study specifically focuses on the pathogenesis of AD and explores molecular mechanisms that may be more relevant to AD. Notably, both studies identified the FoxO and PI3K–Akt signaling pathways as key regulatory mechanisms, suggesting a potential role of ATBC in neuronal survival and proliferation. However, whereas the study by Zheng et al. mainly emphasized biological processes related to cellular aging, our results indicate that ATBC is involved in neuroactive ligand–receptor interactions and the Rap1 signaling pathway, suggesting that ATBC may influence AD pathogenesis by modulating neuronal signaling. In addition, machine learning methods, including LASSO, random forests, and support vector machines, were integrated to improve the stability of target selection and feature identification. Collectively, these findings suggest that the neurotoxic effects of ATBC may not be limited to the acceleration of general aging processes but could also be associated with AD-related pathological cascades. This finding highlights the importance of disease-specific toxicological evaluation and suggests that ATBC may represent a potential, modifiable environmental factor warranting further investigation in the context of AD.

The primary targets discovered in this study are important for AD-related neurological systems. The Cholecystokinin B Receptor (CCKBR) serves as the B-type receptor for cholecystokinin (CCK). It is primarily found in critical brain regions linked to memory and learning, including the hippocampus and entorhinal cortex.¹⁹ The CCK/CCKBR signaling system is crucial for regulating neuronal excitability, synaptic transmission, anxiety, and cognitive processes such as learning and memory.²⁰ Research involving animal models of AD such as 3xTg AD mice, consistently demonstrates a notable reduction in CCK mRNA levels and CCKBR protein expression in critical brain regions, including the hippocampus, as the disease progresses.^21,22 The downregulation is linked to cognitive deficits and impaired neuroplasticity in animal models, suggesting that reduced CCKBR levels may be an important component of AD pathology and a potential biomarker for disease progression.²³ However, extrapolating these findings from animal models to human patients requires additional confirmation. CCKBR may potentially contribute to the control of learning and memory by augmenting synaptic connections between hippocampus and cortical neurons, thereby facilitating the establishment of long-term potentiation (LTP).²⁴ Upon binding of CCK to CCKBR, multiple complex intracellular signaling networks may be activated downstream, including the PKA/CREB-BDNF signaling pathway. In this process, protein kinase A (PKA) can, once activated, phosphorylate and activate the transcription factor CREB. When pCREB is activated, it enters the nucleus and promotes transcription of BDNF and its receptor.^25,26 The BDNF-TrkB signaling is considered one of the key mechanisms for maintaining synaptic plasticity, promoting LTP formation, and consolidating memory.²⁷ Therefore, downregulation of CCKBR function may attenuate the activity of these pathways, leading to impaired LTP and memory deficits, consistent with typical synaptic dysfunction observed in AD patients.²⁸ Notably, animal studies have shown that CCKBR-specific agonists (e.g., HT-267) can exert rescue effects in aged AD mice, effectively improving learning and memory and enhancing hippocampal LTP levels.^22,29 These indicate that reactivating or enhancing CCKBR signaling may help compensate for functional deficiencies in the endogenous CCK/CCKBR system, thereby providing therapeutic benefits; however, further validation through preclinical studies and clinical trials is still necessary.

Gamma-Aminobutyric Acid Type A Receptor Subunit Gamma-2 (GABRG2) encodes the γ2 subunit of the GABA-A receptor,³⁰ which plays a vital role in regulating the equilibrium between excitation and inhibition within the brain. Studies indicate that an early characteristic of AD is an imbalance of excitation and inhibition (E/I) in neural circuits, manifested by network hyperexcitability,³¹ with dysfunction of GABRG2 and the GABAergic system potentially contributing to this imbalance. Existing research has found that the Aβ protein impairs GABAergic neurotransmission through various mechanisms,³² while GO enrichment analysis reveals significant enrichment of “GABA-A receptor complexes” and “GABA-gated chloride channel activity,” further supporting this hypothesis. On the one hand, Aβ may influence the function and survival of GABAergic interneurons, leading to the degeneration of inhibitory neurons³³; on the other hand, it may interfere with the functionality and stability of GABA-A receptors at the postsynaptic membrane, thereby disrupting the integrity of inhibitory neural circuits.³⁴ In addition, disruption of this balance leads to prolonged hyperexcitability, which may induce excitotoxic effects, including increased calcium influx and oxidative stress. These effects damage synaptic integrity, accelerate neuronal death, and impair neural circuit functions critical for precise information encoding and processing.^34,35 Ultimately, these alterations may lead to a significant decline in learning and memory capabilities.

The Raf-1 proto-oncogene, serine/threonine kinase (RAF1), is a crucial kinase within the MAPK signaling pathway. It functions as an intermediary between the upstream RAS protein and the downstream MEK kinase,³⁶ playing a significant role in the regulation of cell proliferation, differentiation, and apoptosis.³⁷ Although current research on RAF1 primarily focuses on its role in cancer, direct evidence of its specific function in the pathophysiology of AD remains lacking. However, based on its central role in the MAPK/ERK pathway and the established association of this pathway with AD,³⁸ the following hypotheses can be proposed: (1) RAF1 may be linked to the hyperphosphorylation of tau protein. Hyperphosphorylation of the tau protein is a significant pathological characteristic of AD, leading to the formation of neurofibrillary tangles. Research demonstrates that the MAPK pathways, particularly ERK, are abnormally activated in vulnerable neurons of AD patients. This suggests the involvement of the MAPK signaling network in the pathophysiology of AD, with RAF1 potentially contributing to tau hyperphosphorylation.³⁹ (2) The role of RAF1 in neuroinflammation and cellular stress responses. Aβ deposits in AD patients’ brains can activate microglia, leading to increased proinflammatory cytokines and oxidative stress. This activation, in turn, activates different inflammatory signaling pathways.⁴⁰ As part of the MAPK pathway, RAF1 may contribute to the release of inflammatory factors and to oxidative stress induced by Aβ⁴¹; however, this hypothesis requires further experimental validation.

Steroid sulfatase (STS) is a crucial enzyme that hydrolyzes sulfate esters of steroid hormones, catalyzing the conversion of DHEA sulfate, estrone sulfate, and others into their free bioactive forms.⁴² As a crucial enzyme regulating the biological activity of neurosteroids, its functional state directly impacts the balance of steroid hormones in the brain. STS may indirectly contribute to the pathological processes of AD by regulating levels of neurosteroids with neuroprotective functions.⁴³ Existing studies indicate that in the brain tissue of AD patients, the gene expression level of STS is significantly downregulated,⁴⁰ and this decrease in expression may lead to reduced local concentrations of neurosteroids with important biological functions, such as DHEA. Consequently, this may compromise neuronal survival, synaptic plasticity, and cognitive function. Additionally, neurosteroids can inhibit excessive microglial cell activation and reduce the release of proinflammatory cytokines.⁴⁴ Therefore, the decrease in neurosteroid levels caused by STS downregulation may weaken the brain's endogenous anti-inflammatory defense mechanisms, rendering it more vulnerable to sustained and excessive inflammatory responses when exposed to pathological stimuli such as Aβ.

Rap Guanine Nucleotide Exchange Factor 3 (RAPGEF3) encodes the EPAC1 protein (Exchange Protein Directly Activated by cAMP 1). EPAC1 is a crucial guanine nucleotide exchange factor (GEF) that regulates cell adhesion, ion channel activity, exocytosis, and synaptic plasticity.^45,46 Research indicates that the EPAC1 protein, encoded by the RAPGEF3 gene, may significantly influence hippocampus-dependent memory retrieval.⁴⁷ BACE1 is a key proteolytic enzyme for amyloid-β protein precursor (AβPP),⁴⁸ and its activity may influence the generation of Aβ. EPAC1 may regulate BACE1 expression through the Rap1 and miR-124 pathways,⁴⁹ consequently influencing Aβ production and the pathological processes of AD. Neuroinflammation is an important pathological feature of AD. Existing studies show that EPAC1 may regulate inflammatory factors, such as TNF-α and IL-1β, through the PDE4/PDE4D pathway,⁵⁰ thereby accelerating neuronal damage; however, the specific molecular mechanisms require further validation. Additionally, EPAC1 can modulate the expression and function of synaptic proteins (e.g., PSD95, GluA3) by regulating signaling pathways involved in synaptic plasticity (e.g., Rap1, RhoGTPase),^51,52 leading to synaptic dysfunction and memory impairment.

GO and KEGG pathway enrichment analyses delineate several core mechanisms through which ATBC interferes with AD progression. GO enrichment analysis reveals that ATBC is closely associated with various biological processes and molecular functions, particularly gliogenesis, synaptic transmission, benzodiazepine receptor activity, and regulation of threonine kinase activity; dysregulation of these processes or components may lead to neuronal dysfunction and impaired synaptic plasticity. Furthermore, KEGG enrichment analysis indicates that ATBC plays a significant role in neuroactive ligand-receptor interaction, calcium signaling pathways, insulin resistance, FoxO signaling, and the PI3K–Akt signaling pathway, mechanisms associated with cell survival, proliferation, and anti-apoptosis. Notably, enrichment of the “Rap1 signaling pathway” and the “calcium signaling pathway” has significant pathological implications. The former is involved in AβPP processing,⁵³ while the latter is directly related to neuronal calcium homeostasis and survival.⁵⁴ ATBC may disrupt these intricate signaling pathways by targeting RAPGEF3, leading to synaptic failure. Of particular note is the “insulin resistance” pathway, which links the potential hazards of ATBC to the metabolic mechanisms theory of AD. We hypothesize that ATBC may exacerbate cerebral insulin resistance by interfering with downstream insulin receptor signaling (e.g., Akt activation), thereby accelerating AD pathology. Additionally, the “focal adhesion” pathway suggests that ATBC may affect neurite growth and stability.⁵⁵ In summary, ATBC may affect neuronal function through multiple pathways, including interfering with neurotransmitter transmission, activating proinflammatory processes, and influencing key pathways involved in normal neurophysiological function, thereby exacerbating AD-related neuronal damage and pathological progression. This finding provides important computational evidence for understanding the detrimental effects of ATBC on AD.

The results from the functional, pathway, and enrichment analyses of the primary targets indicate that, although this study does not provide direct evidence of an association between ATBC and AD, there remains the potential for pathogenic risk. Furthermore, although there is currently a lack of direct epidemiological evidence linking ATBC to AD, ATBC is utilized across multiple sectors, including food packaging, pharmaceuticals, cosmetics, and toy manufacturing,⁵⁶ making long-term human exposure unavoidable. Therefore, we speculate that it might play a role in the development of AD, and further clinical or fundamental research is expected to yield conclusive validation in the future.

This study highlights the advantages of computational biology methods in network toxicology for predicting health risks from environmental pollutants, systematically identifying “target-pathway” networks as focal points for future experiments. However, limitations arise from reliance on predictive data from public databases (such as ChEMBL and TTD), which may have incomplete coverage or species bias, thereby compromising the comprehensiveness of the results. Future research should integrate multi-omics data (e.g., transcriptomics, proteomics) for cross-validation or conduct large-scale longitudinal epidemiological studies and relevant animal models to validate these predictions for preventive and therapeutic strategies. Additionally, limitations in data characteristics, such as variations in transcriptomic data from different brain regions, hindered the identification of an appropriate test set to evaluate the machine learning model's performance. While cross-validation shows promising results on the training set, the absence of an independent test set prevents assessment of the model's generalization to new data. Although we performed external validation with a distinct dataset, results should be interpreted cautiously due to potential differences from real-world data. Future research should focus on using larger, more diverse datasets and multiple validation methods to thoroughly evaluate the model's performance and reliability.

Conclusion

In conclusion, our research combines molecular docking, network toxicology, machine learning, and molecular dynamics simulation to elucidate the molecular pathways and mechanisms underlying ATBC-induced AD toxicity. After screening, we identified 68 potential targets strongly linked to ATBC-induced AD. We further prioritized these targets to the five most critical ones (CCKBR, RAF1, GABRG2, STS, and RAPGEF3) using machine learning algorithms. Further analysis indicates that ATBC may modulate the neurotransmitter system, the Rap1/calcium signaling pathway, and insulin signaling through these targets. Ultimately, these interactions may contribute to the onset and progression of AD by impairing synaptic function, inducing tau pathology, and disrupting cerebral energy metabolism. This study not only fills a critical knowledge gap regarding the pathogenic effects of ATBC on AD but also lays the groundwork for future investigations.

Footnotes

Acknowledgements

The authors have no acknowledgments to report.

ORCID iDs

Tao Peng

Peili Xu

Xiaofang Guo

Jian Lin

Mengfan Zhang

Xinghua Liu

Jianglin Ye

Xingdong Lin

Ethical considerations

Not applicable

Consent to participate

Not applicable

Consent for publication

Not applicable

Author contribution(s)

Tao Peng: Data curation; Methodology; Writing – original draft.

Peili Xu: Conceptualization; Supervision; Writing – review & editing.

Xiaofang Guo: Conceptualization; Supervision; Writing – review & editing.

Jian Lin: Supervision; Writing – review & editing.

Mengfan Zhang: Conceptualization; Writing – original draft.

Xinghua Liu: Data curation; Funding acquisition; Project administration; Writing – review & editing.

Jianglin Ye: Data curation; Supervision; Writing – review & editing.

Xingdong Lin: Funding acquisition; Supervision; Writing – review & editing.

Funding

This study was supported by the National Key R&D Program of China (grant numbers. 2022YFC3501402), the Special Research Project of the Guangdong Provincial Department of Education: Research on Improving Diabetic Cognitive Impairment with Yizhiren-Derived Nanoparticles (grant numbers. 2023KTSCX025), and the Research Project of the Guangdong Provincial Administration of Traditional Chinese Medicine: Research on the Common Mechanism of Tonifying Kidney and Benefiting Essence Method in the Treatment of Alzheimer's disease and Osteoporosis (grant numbers. 202504211654045920).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The data supporting the findings of this study are available from the corresponding author upon reasonable request.

References

Kim

Choi

, et al. Pharmacokinetic properties of acetyl tributyl citrate, a pharmaceutical excipient. Pharmaceutics 2018; 10: 177.

Zygoura

Paleologos

Kontominas

. Migration levels of PVC plasticisers: effect of ionising radiation treatment. Food Chem 2011; 128: 106–113.

Goulas

Riganakos

Ehlermann

, et al. Effect of high-dose electron beam irradiation on the migration of DOA and ATBC plasticizers from food-grade PVC and PVDC/PVC films, respectively, into olive oil. J Food Prot 1998; 61: 720–724.

Bernard

Masse

Boeuf

, et al. Medical devices used in NICU: the main source of plasticisers’ exposure of newborns. Sci Total Environ 2023; 858: 159994.

Zhang

Jie

, et al. Characterizing the obesogenic and fatty liver-inducing effects of Acetyl tributyl citrate (ATBC) plasticizer using both in vivo and in vitro models. J Hazard Mater 2023; 445: 130548.

García Ibarra

Sendón

Bustos

, et al. Estimates of dietary exposure of Spanish population to packaging contaminants from cereal based foods contained in plastic materials. Food Chem Toxicol 2019; 128: 180–192.

Subedi

Sullivan

Dhungana

. Phthalate and non-phthalate plasticizers in indoor dust from childcare facilities, salons, and homes across the USA. Environ Pollut 2017; 230: 701–708.

Zhang

Zhu

, et al. Acetyl tributyl citrate exposure at seemingly safe concentrations induces adverse effects in different genders of type 2 diabetes mice, especially brain tissue. Toxics 2023; 11: 877.

Zheng

Peng

Shen

, et al. Efficient analysis of toxicity and mechanisms of Acetyl tributyl citrate on aging with network toxicology and molecular docking strategy. Toxicology 2025; 510: 154009.

10.

Chen

Chu

JMT

Chang

RCC

, et al. The complement system in the central nervous system: from neurodevelopment to neurodegeneration. Biomolecules 2022; 12: 337.

11.

Liu

Zheng

, et al. Amyloid, tau, pathogen infection and antimicrobial protection in Alzheimer's disease -conformist, nonconformist, and realistic prospects for AD pathogenesis. Transl Neurodegener 2018; 7: 34.

12.

Grobler

van Tongeren

Gettemans

, et al. Alzheimer's disease: a systems view provides a unifying explanation of its development. J Alzheimers Dis 2023; 91: 43–70.

13.

Kukreja

Kujoth

Prolla

, et al. Increased mtDNA mutations with aging promotes amyloid accumulation and brain atrophy in the APP/Ld transgenic mouse model of Alzheimer's disease. Mol Neurodegener 2014; 9: 16.

14.

Dong

Zhou

Tang

, et al. CSF sTREM2 in neurological diseases: a two-sample Mendelian randomization study. J Neuroinflammation 2022; 19: 79.

15.

Temprano-Carazo

Contreras

Saura

, et al. Spatial memory training counteracts hippocampal GIRK channel decrease in the transgenic APP(Sw,Ind) J9 Alzheimer's disease mouse model. Int J Mol Sci 2022; 23: 13444.

16.

Huang

. A novel strategy for the study on molecular mechanism of prostate injury induced by 4,4'-sulfonyldiphenol based on network toxicology analysis. J Appl Toxicol 2024; 44: 28–40.

17.

Pinzi

Rastelli

. Molecular docking: shifting paradigms in drug discovery. Int J Mol Sci 2019; 20: 4331.

18.

Mark

Nilsson

. Structure and dynamics of the TIP3P, SPC, and SPC/E water models at 298K. J Phys Chem A 2001; 105: 9954–9960.

19.

Asim

Wang

Chen

. Shedding light on cholecystokinin's role in hippocampal neuroplasticity and memory formation. Neurosci Biobehav Rev 2024; 159: 105615.

20.

Huang

Bello

. Neuropeptide cholecystokinin: a key neuromodulator for hippocampal functions. Neural Regen Res 2025; 20: 1991–1992.

21.

Asim

Wang

Waris

, et al. Cholecystokinin neurotransmission in the central nervous system: insights into its role in health and disease. Biofactors 2024; 50: 1060–1075.

22.

Zhang

Sui

Jendrichovsky

, et al. Cholecystokinin B receptor agonists alleviates anterograde amnesia in cholecystokinin-deficient and aged Alzheimer's disease mice. Alzheimers Res Ther 2024; 16: 109.

23.

Plagman

Hoscheidt

McLimans

, et al. Cholecystokinin and Alzheimer's disease: a biomarker of metabolic function, neural integrity, and cognitive performance. Neurobiol Aging 2019; 76: 201–207.

24.

Chen

Wong

, et al. Cholecystokinin release triggered by NMDA receptors produces LTP and sound-sound associative memory. Proc Natl Acad Sci U S A 2019; 116: 6397–6406.

25.

Guo

Cheng

Wang

, et al. FFPM, a PDE4 inhibitor, reverses learning and memory deficits in APP/PS1 transgenic mice via cAMP/PKA/CREB signaling and anti-inflammatory effects. Neuropharmacology 2017; 116: 260–269.

26.

Xue

Wang

Gong

, et al. PKA-CREB-BDNF signaling regulated long lasting antidepressant activities of Yueju but not ketamine. Sci Rep 2016; 6: 26331.

27.

Lin

Mahgoub

, et al. A synaptic locus for TrkB signaling underlying ketamine rapid antidepressant action. Cell Rep 2021; 36: 109513.

28.

Bartolotti

Bennett

Lazarov

. Reduced pCREB in Alzheimer's disease prefrontal cortex is reflected in peripheral blood mononuclear cells. Mol Psychiatry 2016; 21: 1158–1166.

29.

Jacobsen

Redwine

, et al. Early-onset behavioral and synaptic deficits in a mouse model of Alzheimer's disease. Proc Natl Acad Sci U S A 2006; 103: 5161–5166.

30.

Fatemi

Folsom

. GABA Receptor subunit distribution and FMRP-mGluR5 signaling abnormalities in the cerebellum of subjects with schizophrenia, mood disorders, and autism. Schizophr Res 2015; 167: 42–56.

31.

Ali

Islam

Constanti

. The fate of interneurons, GABA(A) receptor sub-types and perineuronal nets in Alzheimer's disease. Brain Pathol 2023; 33: e13129.

32.

Cisternas

Taylor

Perkins

, et al. Vascular amyloid accumulation alters the gabaergic synapse and induces hyperactivity in a model of cerebral amyloid angiopathy. Aging Cell 2020; 19: e13233.

33.

Ulrich

. Amyloid-β impairs synaptic inhibition via GABA(A) receptor endocytosis. J Neurosci 2015; 35: 9205–9210.

34.

Ren

Yao

Yan

, et al. Amyloid β causes excitation/inhibition imbalance through dopamine receptor 1-dependent disruption of fast-spiking GABAergic input in anterior cingulate cortex. Sci Rep 2018; 8: 302.

35.

Harney

Jones

. Pre- and postsynaptic properties of somatic and dendritic inhibition in dentate gyrus. Neuropharmacology 2002; 43: 584–594.

36.

Deng

Feng

, et al. Machine learning models identify ferroptosis-related genes as potential diagnostic biomarkers for Alzheimer's disease. Front Aging Neurosci 2022; 14: 994130.

37.

Bayraktar

Lam

Altay

, et al. Revealing the molecular mechanisms of Alzheimer's disease based on network analysis. Int J Mol Sci 2021; 22: 11556.

38.

Iqbal

Liu

Gong

, et al. Mechanisms of tau-induced neurodegeneration. Acta Neuropathol 2009; 118: 53–69.

39.

Planel

Tatebayashi

Miyasaka

, et al. Insulin dysfunction induces in vivo tau hyperphosphorylation through distinct mechanisms. J Neurosci 2007; 27: 13635–13648.

40.

Giovagnoni

Ali

Eijssen

LMT

, et al. Altered sphingolipid function in Alzheimer's disease; a gene regulatory network approach. Neurobiol Aging 2021; 102: 178–187.

41.

Zhang

Gou

, et al. Metabolic reprogramming of microglia enhances proinflammatory cytokine release through EphA2/p38 MAPK pathway in Alzheimer's disease. J Alzheimers Dis 2022; 88: 771–785.

42.

Reed

Purohit

Woo

, et al. Steroid sulfatase: molecular biology, regulation, and inhibition. Endocr Rev 2005; 26: 171–202.

43.

Steckelbroeck

Nassen

Ugele

, et al. Steroid sulfatase (STS) expression in the human temporal lobe: enzyme activity, mRNA expression and immunohistochemistry study. J Neurochem 2004; 89: 403–417.

44.

Noorbakhsh

Ellestad

Maingat

, et al. Impaired neurosteroid synthesis in multiple sclerosis. Brain 2011; 134: 2703–2721.

45.

Kawasaki

Springett

Mochizuki

, et al. A family of cAMP-binding proteins that directly activate Rap1. Science 1998; 282: 2275–2279.

46.

Enserink

Christensen

de Rooij

, et al. A novel Epac-specific cAMP analogue demonstrates independent regulation of Rap1 and ERK. Nat Cell Biol 2002; 4: 901–906.

47.

Ostroveanu

van der Zee

Eisel

, et al. Exchange protein activated by cyclic AMP 2 (Epac2) plays a specific and time-limited role in memory retrieval. Hippocampus 2010; 20: 1018–1026.

48.

Pan

Liu

Ren

, et al. Epac: a promising therapeutic target for vascular diseases: a review. Front Pharmacol 2022; 13: 929152.

49.

Arrazola Sastre

Luque Montoro

Gálvez-Martín

, et al. Small GTPases of the Ras and Rho families switch on/off signaling pathways in neurodegenerative diseases. Int J Mol Sci 2020; 21: 6312.

50.

Yang

Liu

, et al. PDE4 Regulates tissue plasminogen activator expression of human brain microvascular endothelial cells. Thromb Res 2012; 129: 750–753.

51.

Olabarria

Pasini

Corona

, et al. Dysfunction of the ubiquitin ligase E3A Ube3A/E6-AP contributes to synaptic pathology in Alzheimer's disease. Commun Biol 2019; 2: 111.

52.

Zhang

Musheshe

van der Veen

, et al. The expression of Epac2 and GluA3 in an Alzheimer's disease experimental model and postmortem patient samples. Biomedicines 2023; 11: 2096.

53.

Maillet

Robert

Cacquevel

, et al. Crosstalk between Rap1 and Rac regulates secretion of sAPPalpha. Nat Cell Biol 2003; 5: 633–639.

54.

Sukumaran

Nascimento Da Conceicao

Sun

, et al. Calcium signaling regulates autophagy and apoptosis. Cells 2021; 10: 2125.

55.

Chrzanowska-Wodnicka

Burridge

. Rho-stimulated contractility drives the formation of stress fibers and focal adhesions. J Cell Biol 1996; 133: 1403–1415.

56.

Alves

Giovanoulis

Nilsson

, et al. Case study on screening emerging pollutants in urine and nails. Environ Sci Technol 2017; 51: 4046–4053.