Machine learning and free energy clustering reveal PAH protein binding linked to AD risk

Summary

This study develops a computational framework integrating bioinformatics, machine learning, and ΔG clustering to prioritize polycyclic aromatic hydrocarbons (PAHs) for Alzheimer’s disease (AD)-associated neurotoxicity. PAH targets were predicted from ChEMBL/STITCH databases; AD-related differentially expressed genes (DEGs) were identified via WGCNA and differential expression analysis of GEO datasets. Protein-protein interaction (PPI) networks, GO/KEGG enrichment, and XGBoost feature selection identified PARP1, PTPN1, and ITGA4 as candidate core PAH targets enriched in neuroinflammation, microglial activation, lipid metabolism, and atherosclerosis pathways. Molecular docking produced ΔG heatmaps for clustering 16 PAHs into eight toxicity-similarity categories. Category-average ΔG values correlated linearly with literature LD₅₀/BMDL data (ρ = 1, p = 0.0417), yielding an empirical relationship BMDL = 1.723 × ΔG + 22.602. Zebrafish motility assays provided preliminary support (Spearman ρ = −1.0, p = 0.167; n = 3). This pipeline provides initial insights into PAH mechanisms and potential therapeutic targets, pending experimental validation.

Graphical abstract

Highlights

• •Computational framework prioritizes PAHs for AD neurotoxicity
• •Predicted PAH targets: PARP1, PTPN1, ITGA4 via PPI/XGBoost
• •ΔG clustering categorizes 16 PAHs and strongly correlates with LD₅₀/BMDL data
• •Zebrafish assays support PAH neurotoxicity predictions

Medicine; Neuroscience; Bioinformatics

Introduction

Alzheimer’s disease (AD) is prevalent worldwide and is the leading cause of dementia in older adults (age ≥65 years).¹ China alone has 15.07 million dementia patients over 60, including 9.83 million with AD.² Recent data suggest that the global prevalence of dementia will triple by 2050. Notably, this estimate is three times higher when using a biological definition of AD rather than a clinical one.³ The main hypotheses for AD’s pathogenic mechanism are the β-amyloid hypothesis, the tau protein abnormal phosphorylation hypothesis, and the neuroinflammatory hypothesis.¹ Recent research recognizes environmental pollutants as promoters of neurological dysfunction, with ecological risk factors potentially influencing AD progression significantly.^4,5 Identifying modifiable risk factors is crucial to reducing the growing burden of AD in older adults and aging societies.⁶

Polycyclic aromatic hydrocarbons (PAHs) are persistent organic pollutants of global concern.⁷ Numerous experimental animal studies have shown that PAH analogs have toxic effects on the nervous system.^8,9 In humans, PAH exposure is linked to AD pathophysiology, including the buildup of amyloid-related proteins.^10,11 The US Environmental Protection Agency (US EPA) Priority Pollutant List includes only 16 PAHs. In fact, over 100 other PAHs are produced in the environment.¹² Although several studies have found associations between some PAHs and AD,^6,13,14 research on the AD-associated toxicity of a specific type of PAH alone is lacking. The scarcity of analytical methods for isolating and determining these individual PAH and their potential isomers, along with limited commercial standards, makes studying these compounds difficult.¹⁵ This gap hampers research on PAH neurotoxicity, especially in AD toxicity studies. Consequently, we need innovative approaches to thoroughly and efficiently assess the potential health risks of uncharacterized toxic PAHs that are increasingly present in our environment.

We present a promising strategy for assessing AD-associated toxicity in PAHs. This approach integrates machine learning, molecular docking, and ΔG clustering, and is validated by data from the literature and experimental data. When evaluating a molecule’s toxicity through molecular docking, identifying its target is crucial. Normally, a molecule may interact with multiple targets in an organism, forming a complex network of associations.¹⁶ Consequently, a toxicant’s protein target may be a group rather than a single entity. Identifying applicable targets for toxic substances is vital in assessing toxicity using molecular docking. In our study, we analyzed PAH targets of AD using weighted gene co-expression network analysis (WGCNA), community network analysis, and other methods. The most critical factor was the use of machine learning to identify a set of genes that exhibited significant changes in the gene expression profiles of human AD patients. We then used these genes to identify valid targets for PAH.

Molecular docking and molecular dynamics simulation techniques are essential in computational biology and play a vital role in biological research. These techniques provide detailed insights into molecular interactions and visually explain their mechanisms. This is particularly crucial in drug development, where these methods have become essential tools for elucidating biological mechanisms.¹⁷ This approach can also explore the material basis of toxicant toxicity.^18,19 However, molecules’ multi-target effect limits the reliability of molecular docking ΔGs. Assessing pharmacological or toxicological effects based on the docking score of a single target provides insufficient clarification.²⁰ Docking alone is rarely used for toxicity prediction.²¹ AD-associated molecular mechanisms are so complex that a single target often cannot determine the overall course of AD.^22,23 The correlation between the stability of the ligand-receptor binding conformation and the substance’s overall toxicity is determined by the combined action of multiple PAH receptors in AD. To accurately assess the multi-target AD toxicity of individual substances, we replaced individual PAH target ΔG with the average molecular docking ΔG of multiple core targets through clustering analysis. Then, we compared the average molecular docking ΔG score from clustering multi-targets to the magnitude of quantifiable PAH neurotoxicity from the literature. This comparison aimed to explore the fit function between the PAH benchmark dose lower confidence limit (BMDL) and the average molecular docking ΔG. Specifically, we sought to determine whether the molecular binding ΔG of multi-targets correlates with the quantifiable toxicity parameters reported in the literature. The efficacy of this correlation was subsequently validated through empirical studies utilizing zebrafish as a model organism. This would confirm the feasibility of our molecular docking ΔG clustering strategy.

This study integrates bioinformatics, machine learning, and molecular docking to first identify PAH targets (ChEMBL/STITCH), AD-associated DEGs (GEO/WGCNA), and their PPI network intersections (STRING/Cytoscape), yielding core proteins PARP1, PTPN1, and ITGA4 enriched in neuroinflammatory pathways. Subsequently, ΔG clustering of docking scores informs a linear BMDL∼ΔG model calibrated on literature toxicity data and preliminarily validated in zebrafish motility assays, thus enabling rapid, hypothesis-generating prediction of PAH neurotoxic potency in AD contexts.

Results

Data collection and processing

The workflow strategy is shown in Figure 1. The ChEMBL data predicted 30 target proteins for benzo[a]pyrene, phenanthrene, fluoranthene, and pyrene. After removing non-human proteins, 17 remained (see Table S2). The STITCH database yielded 23 target proteins for PAHs (Table S3). Thirty seven proteins were screened as potential targets for the four PAHs. In the GeneCards database, proteins with a target relevance score ≥ the overall median (2.87) were considered related to AD, resulting in 7,153 proteins (Table S4). Our screening criteria identified six groups of AD-related data from the GEO database. Three datasets—GSE44771, GSE44770, and GSE33000—showed close principal components (Figure 2A). We performed batch correction before conducting differential gene expression and WGCNA analyses (Figure 2B). Dimensionality reduction analyses of the AD and control group data are shown in Figure 2C. The DEGs in the AD dataset compared to the control group are displayed in a volcano plot (Figure 2D). All DEGs are listed in Table S5, with 1,380 upregulated and 1,467 downregulated genes.

Figure 1.

Flowchart for constructing a predictive model for AD-associated toxicity of PAHs

AD, Alzheimer’s disease; DEGs, differentially expressed genes; PAHs, polycyclic aromatic hydrocarbons; GO, gene ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Figure 2.

AD-associated GEO microarray dataset and its DEGs screening

(A) Principal component analysis (PCA) of six microarray datasets.

(B) PCA of three sets of GSE microarray datasets after calibration for different batch effects.

(C) Dimensionality reduction analyses of AD and control microarray datasets.

(D) Volcano plot for DEGs. |log2 (fold change)| ≥ 0.07 and adjusted p value (padj) ≤ 0.001.

Identifying AD-associated DEGs using WGCNA

To identify module genes associated with AD, we conducted WGCNA using a scale-free topology with a soft thresholding cutoff of five to power the gene correlations (Figure 3A). Figures 3B and 3C display the gene clustering and module correlations, respectively. Figure 3D illustrates the correlation between modules and AD. We selected gene modules with r > 0.4 as candidate modules associated with AD. The Gray60 module showed the strongest positive correlation with AD (r = 0.46), while the black module exhibited the strongest negative correlation (r = −0.45). Figure 3E depicts the correlation between gene significance (GS) and module membership (MM) in the AD-associated modules. All hub genes (1,164 in total) are listed in Table S6. Figure 3F presents the intersection of DEGs and WGCNA hub genes in the AD samples. We compiled a list of AD-associated genes in Table S7.

Figure 3.

Identifying AD-associated hub genes using weighted correlation network analysis

(A) Utilizing a scale-free topology model to determine the optimal β value, selecting β = 5 as the soft threshold based on average connectivity and scale independence.

(B) Hierarchical clustering dendrogram showing module identifiers. The gene dendrogram was obtained by average linkage hierarchical clustering; colored rows below indicate module assignments determined by the dynamic tree cut.

(C) Visualizing correlations between eigengenes. Dendrogram branches (meta-modules) combine positively correlated gene sets. The heatmap shows adjacencies in the Eigengenes network. Each row and column represents a module gene (indicated by color). Blue indicates low adjacency (negative correlation), while red indicates high adjacency (positive correlation).

(D) Relationship diagram between module genes and AD. Each row corresponds to module eigengenes, and the column corresponds to AD. Cells are filled with the corresponding correlation and p value. Redder colors indicate stronger positive correlations between phenotypic traits and module eigengenes, while greener colors indicate stronger negative correlations.

(E) Correlation between gray60 (or black) module membership and gene significance for AD. A highly significant correlation (0.79 or 0.76) between gene significance and module membership implies that hub genes in the turquoise module tend to be highly correlated with AD.

(F) Intersection of hub genes with DEGs. In the WGCNA analysis, genes with a correlation coefficient, R > 0.4 for module-trait relationships and a module membership (kME_MM) > 0.8 were identified as hub genes in modules strongly associated with AD.

Interaction network of PAHs potential targets and their downstream genes

A Venn analysis was performed on 1,061 AD-associated proteins, 7,153 proteins obtained from the GeneCards database, and 37 target proteins for PAHs. Only one protein, GAD2, was shared among these sets (Figure 4A; Table S8). Twenty-five proteins were found at the intersection of GeneCards source proteins and predicted PAH target proteins. We utilized the STRING database to investigate the relationship between predicted PAH targets and AD-associated DEGs. We entered 1,061 AD-associated DEGs and 26 possible PAH target proteins into STRING to construct the PPI network. Due to its lack of linked genes, the TDO2 gene was removed when transferring the PPI network from STRING to Cytoscape. PAHs’ predicted target proteins and their directly interacting DEGs formed a complex topological network of 145 genes (Figure 4B). Using the Leading eigenvector algorithm, we further divided this network into six sub-networks (Figure 4C). We then performed GO and KEGG enrichment analyses for all 145 genes (Figures 4D and 4E). Our GO results suggest that potential targets of PAHs and their downstream genes are involved in AD-related processes. These include reactive oxygen species metabolism, cytokine production regulation, neuroinflammation, microglial activation, and other related pathways. Not only that, in KEGG signaling pathway enrichment, lipid and atherosclerosis, NF-κB signaling pathway, cocaine addiction, serotonergic synapse, and other pathological pathways associated with AD were also enriched.

Figure 4.

Interaction network of PAHs’ potential targets and their downstream proteins

(A) Venn diagram analysis of AD-associated DEGs, AD-associated GeneCards library genes, and potential target protein genes for PAHs.

(B) Protein-protein interaction networks (PPI) analysis between predicted PAH targets and AD-associated DEGs. Orange indicates predicted PAH targets; blue and yellow indicate AD-associated DEGs.

(C) Using the leading eigenvector algorithm, modularity, and community structure in PPI from predicted PAH targets and AD-associated DEGs.

(D) GO enrichment analysis of six modular genes, including biological process, cellular component, and molecular function analysis.

(E) KEGG enrichment analysis of six modular genes.

Screening of core target proteins for PAHs brain injury

To prioritize PAH targets by AD relevance without data leakage, XGBoost used AD-only expression data from independent datasets (GSE44771/70/33000 batch-corrected; GSE15222/118553) to identify disease feature genes. PAH linkage was established post hoc via STRING PPI networks (Figure 4B). Building on this, differential expression analysis and WGCNA identified 25 potential PAH target proteins. However, it remains to be determined which of these proteins play a central role in PAH-related brain injury. This requires further investigation of their downstream genes. Therefore, we utilized XGBoost to assess the importance of the top five downstream genes of PAH target proteins in the disease, using three distinct datasets. These data included the GSE44771, GSE44770, and GSE33000 gene sets after removal of batch effects, as well as GSE15222 and GSE118553 (Figure 5), which differed significantly from the above gene sets. The predictive accuracy of these models is crucial when screening for characterized genes in gene sets, especially given the use of neural networks. The AUROC score was used to evaluate model phenotypes. In the gene sets GSE44771, GSE44770, and GSE33000, where batch effects were eliminated, XGBoost achieved an AUROC greater than 95 (Figure 5B). GSE15222 and GSE118553 had AUROC scores above 85 (Figures 5D and 5F). We screened a total of 15 genes and assumed the upstream proteins that these 15 characterized genes can relocate to interact with PAH as the target core proteins of PAHs affecting AD (Figure 6A). Moreover, to find the most central PAHs target proteins, the connection free energies greater than −8 in the Sankey diagram were hidden from the connection line segments. Eight proteins in total: APP, PTPN1, IL2, ITGA4, PARP1, NQO1, TNF, and GAD2 were identified as core target proteins for PAH-induced brain injury.

Figure 5.

Feature importance with three different microarray gene expression datasets

(A) The top five genes, ranked by feature importance score, in the three batch-corrected microarray gene expression datasets.

(B) Overall prediction score of the XGBoost model in the three batch-corrected microarray gene expression datasets.

(C and E) The top five genes were identified by feature importance score in GSE15222 or GSE118553, which are associated with AD.

(D and F) Overall prediction score of the XGBoost model in GSE15222 or GSE118553 microarray datasets.

Figure 6.

Molecular docking of PAHs with core target proteins in AD

(A) Sankey diagram showing PAHs, their predicted target proteins, and downstream DEGs.

(B) Heatmap depicting ΔG between PAHs and their predicted target proteins.

Molecular docking of PAHs with core target proteins in brain injury

Molecular docking analyses were performed to investigate the interactions between PAHs and eight core target proteins. All eight docking results, generated using AutoDock, showed low binding energies, indicating a strong affinity between the PAHs and the targets. We displayed the docking results of the 16 PAH molecules and these eight genes in a weighted network (Figure 6A). The free energy of molecular docking of 16 PAHs with core target proteins are demonstrated in a heatmap (Figure 6B). PARP1, PTPN1, and ITGA4 exhibited lower Gibbs free energies with most of the 16 PAHs, suggesting these three proteins have a higher probability of spontaneously binding. This highlights their essential role in the molecular mechanism of PAH-induced brain toxicity and damage. A visualization of the lowest binding energy between the four types of PAHs and PARP1, PTPN1, and ITGA4 proteins was created using Pymol (Figure S1).

Molecular dynamics simulation

RMSD measures the average distance between a set of atoms and serves as a metric for the degree of conformational change in a protein. The radius of gyration describes the compactness of a protein—the smaller the radius, the better the compactness and the more stable the protein structure. In this study, we calculated the free energy topography of PAH-target PARP1 proteins using RMSD and radius of gyration, with results displayed in Figure 7. Both Benzo(a)pyrene and phenanthrene show energy wells at 0.15, but PARP1 exhibits a tighter structure when bound to Benzo(a)pyrene, indicating a more significant structural difference (Figures 7A and 7B). The energy profiles of phenanthrene and pyrene appear similar, except that Pyrene stabilizes the PARP1 structure at an RMSD of 0.13 (Figures 7C and 7D). During molecular dynamics simulations, the binding of all four substances to PARP1 significantly altered its structure.

Figure 7.

Free energy landscape (FEL) of PARP1 binding with four PAHs

(A–D) Shown in order: benzo[a]pyrene, phenanthrene, fluoranthene, and pyrene. RMSD (root-mean-square deviation) measures the stability of a structure during simulation. Gyrate computes the radius of gyration of a molecule and the radii of gyration about the x, y, and z axes. GYRATE calculates the radius of gyration (Rg) of the group of atoms, including all effects due to atoms passing through periodic boundaries. A stable folded protein maintains a steady Rg value.

Calculation of the toxicity parameter BMDL for PAHs

Data on PAH neurotoxicity at different dose treatments are plentiful but limited in individual experiments with more than four dose levels. We found only two sets of animal experiments, and one of these sets of cellular experiments could be applied in this study. The original data used to calculate the BMD are summarized in Table S9. Data for assessing the digging behavior of brandling worms (Eisenia fetida) induced by phenanthrene-²⁴ and fluorene-contaminated soil²⁵ were directly adopted in this work. Data on the neurotoxicity of anthracene and benz[a]anthracene for hippocampal neuronal cell lines from Tosin et al.²⁶ were modified based on the methodology described in the previous section. Figure 8 shows the results of the BMD calculation. Figures 8A and 8B display the BMD of fluorene and phenanthrene on brandling worms in soil. In assessing the BMD of fluorene, we chose the Hill model based on the AIC being at or near the minimum value and the graphical visual fit being close to the actual situation. Figures 8C and 8D illustrate the BMD of benz[a]anthracene and anthracene affecting the MAO of hippocampal neuronal cell lines. BMDL values in Figure 8 for the four PAHs were obtained from the model fit data in Table S9.

Figure 8.

Calculated BMDL values for four PAHs

(A and B) Original data from neurotoxicity experiments using fluorene- and phenanthrene-contaminated soils on brandling worms.

(C and D) Original data from neurotoxicity studies of benzo[a]anthracene and anthracene on hippocampal neuronal cell lines. Data are presented as mean ± SD of incidence of neurotoxicity at each dose; error bars represent SD.

We repeated the cluster analysis of PARP1, PTPN1, ITGA4, and PAHs docking free energy values, excluding the interference of proteins with slightly poor binding affinities. Based on the classification results, we calculated the average docking scores of the clustered proteins and small molecules (Figure 9A). We divided the PAHs into eight categories. The lowest docking scores were for benzo[a]pyrene-containing categories, whereas the highest was for naphthalene.

Figure 9.

Models for predicting AD toxicity in PAHs

(A) Linear regression of molecular ΔG against BMDL. Cluster dendrogram of average ΔG bind between PAHs and PARP1, PTPN1, and ITGA4.

(B) Linear regression of molecular ΔG against LD₅₀.

(C) Representative swimming trajectories of individuals under different treatment groups (control group CK, naphthalene 1,342 μM, anthracene 45.0 μM, and pyrene 13.7 μM). Each circle represents the swimming path of an individual.

(D) Correlation analysis between average ΔG and swimming distance. The blue line represents the linear regression fit, and the gray shading represents the 95% confidence interval. Spearman correlation coefficient (rho) and p value are indicated in the figure.

Estimating BMDLs for different PAH categories was challenging because BMDLs were calculated for only four of the 16 PAHs. There should be a particular pattern in the distribution of toxicity data for similar substances; however, predicting such a pattern is challenging. As a result, we initially examined the LD₅₀ toxicity data, which contains a somewhat greater volume of information (Figure 9B). Table S10 displays the collected PAH toxicity data. Among them, the 150 mg LD₅₀ for naphthalene was not used for two reasons. First, while naphthalene’s intravenous LD₅₀ is twice that of phenanthrene, its intraperitoneal LD₅₀ is only one-fifth of phenanthrene’s. Second, Ka et al.’s study didn’t report any mouse deaths at doses ranging from 200 to 600 mg/kg in mice.²⁷ We then investigated the distribution of the LD₅₀ data within the PAH category using a linear fitting approach (Figure 9B, rho: 1, p value: 0.0417). Then, based on this result, we hypothesize that the distribution of BMDLs across different PAH classes may also follow a linear pattern. Our results show that BMDLs of benz[a]anthracene and anthracene are 7.856 and 10.096, respectively. Their average G values for PARP1, PTPN1, and ITGA4 proteins are −9.4 and −8.1, respectively. We estimate the slope of the BMDL∼ ΔG to be 1.723. Jang et al. assayed BMD using transcriptional profiling in the mouse hippocampus. The most sensitive pathways were “synaptic long term potentiation” and “G-protein coupled receptor signaling,” both yielding benzo(a)pyrene median BMDL values of 5.2 mg/kg bw per day.²⁸ The average G value for benzo[a]pyrene and PARP1, PTPN1, and ITGA4 proteins is −10.1. We calculated the intercept of the BMDL∼ ΔG linear equation to be 22.602. Then we established a linear equation to estimate the BMDL value for PAH classification: y = 1.723x + 22.602.

In brief, clustering analysis of PAH docking free energies to PARP1, PTPN1, and ITGA4 grouped the 16 PAHs into eight toxicity-similarity categories (Figure 9A). Consequently, we used category-average ΔG values as the predictor variable in linear regression against literature BMDL/LD₅₀ data (Figure 9B). This process yielded an empirical relationship y = 1.723x + 22.602 (ρ = 1, p = 0.0417), enabling rapid PAH neurotoxicity estimation.

Effects of naphthalene, anthracene, and pyrene on zebrafish motility

Our preliminary experiments determined LC10 values of 1408 μM for naphthalene, 54.1 μM for anthracene, and 18.6 μM for pyrene. To study zebrafish swimming impairment, we treated the specimens with slightly lower concentrations: 1342 μM naphthalene, 45.0 μM anthracene, and 13.7 μM pyrene. During the 18-min observation period, the total swimming distances of 30 zebrafish were 558.86 mm in the naphthalene group, 311.71 mm in the anthracene group, and 270.30 mm in the pyrene group.

To compare the relationship between the ΔG of three PAHs and the impairment of zebrafish swimming ability, we calculated the effect of each substance on zebrafish swimming distance at a single concentration. Given the small sample size and the nonnormal distribution of the data, we performed statistical tests using Spearman’s correlation analysis. We found that as the free energy values increased, the swimming distance of the zebrafish decreased; however, the statistical p value was not significant (Figures 9C and 9D). It is worth noting that a Spearman ρ of 1 reflects perfect rank agreement between predicted and observed rankings within this small sample. Nonetheless, small sample sizes (such as n < 10) tend to have low statistical power, often resulting in p > 0.05 even if a true effect exists, which prevents rejecting the null hypothesis of no correlation. Consequently, although our zebrafish experiments show a perfect Spearman correlation (ρ = 1), the limited sample size and non-significant p value (p > 0.05) mean that the results only offer preliminary evidence.

Discussion

In this study, we first screened four PAH compounds (benzo[a]pyrene, phenanthrene, fluoranthene, and pyrene) for protein interactions through ChEMBL and STITCH databases. We then collected AD-related genes from GeneCards and analyzed microarray data from the GEO database, focusing on human cases over 65 years old with AD. Through differential expression analysis and WGCNA, we identified key DEGs related to AD in aging. Using STRING and Cytoscape, we constructed an interaction network that revealed the enrichment of PAHs and AD-related differentially expressed genes (DEGs) in several key pathways. Our workflow and key findings are illustrated in Figures 1, 2, 3, and 4.

In GO enrichment analysis, DEGs related to PAH-targeting proteins were enriched in neuroinflammatory response, neuron death, and microglial cell activation. Previous studies have shown that elevated inflammatory markers are present in AD patients. The identification of AD risk genes related to innate immune function suggests that neuroinflammation plays a crucial role in AD development. Microglia serve as key players in this process. Our community network GO analysis also revealed that reciprocal DEGs of PAH target proteins are involved in neuroinflammation and microglial cell activation. Additionally, KEGG analysis identified three closely AD-associated pathways: lipid and atherosclerosis, drug metabolism—cytochrome P450 and NF-κB signaling pathway. In recent decades, the dynamic balance of lipids has become a focus of AD research.²⁹ The P450 enzyme system directly metabolizes activated PAHs³⁰ and is associated with age-related neurodegenerative pathologies.³¹ Neurodegeneration can be mediated through the NF-κB signaling pathway. Our GO and KEGG analysis data showed that the PAH target proteins and their downstream genes screened in this work are also associated with AD. Consequently, PAHs have the potential to exert neurotoxic effects as environmental pollutants.

This study employed the Leading eigenvector algorithm to disentangle a complex network involved of PAH target proteins and AD-associated differentially expressed genes (DEGs). The algorithm’s core is defining a modularity matrix,³² calculating the eigenvector of the modularity matrix corresponding to the most significant positive eigenvalue, and then separating vertices into sub-community networks. In Figure 4C, six proteins were screened in the largest sub-network: PTPN1, PARP1, NOX4, IL2, ITGA4, and TNF. These six proteins interact with PAHs and have the potential to influence the AD process by adjusting the function of downstream genes or proteins. However, their downstream DEGs may also determine which proteins most appropriately assess PAHs’ AD toxicity.

We identified feature genes in DEGs of AD patients over 65 using the XGBoost method, as shown in Figures 5A, 5C, and 5E. The receiver operating characteristic curves (Figures 5B, 5D, and 5F) evaluated the XGBoost model’s ability to distinguish between categories. The model produced 15 feature genes across the three datasets. These genes may interact with PAH target proteins in PPI analysis (Figures 4B and 5A). Using Autodock, we calculated the binding free energies of the 15 PAHs and these target proteins (Figure 5B). We found that PARP1, PTPN1, and ITGA4 were the main target proteins with binding free energies less than 0.8. Notably, these three proteins appeared in the largest sub-network in Figure 4C. Machine learning identified feature genes, and the PPI community network analysis results point to the same group of PAH target proteins, predicting that these PAH targets may have a critical impact on AD toxicity.

The link between PARP1 and AD has been well summarized by Mao et al.³³ Specifically, PARP1 is expressed throughout the brain in patients with AD, particularly in the frontal and temporal lobes.³⁴ Additionally, Aβ enhances PARP1 activity, producing reactive oxygen species (ROSs) and/or reactive nitrogen species (RNSs).³⁵ Consequently, activated PARP1 promotes Aβ deposition and Tau protein tangle formation, exacerbating AD symptoms.³⁶ Given that the molecular docking free energies of all the PAHs we studied with PARP1 are less than −8, we hypothesized that binding of various PAHs to PARP1 would also enhance its activity, leading to AD. Furthermore, in the presence of AD pathology, activated PARP1 promotes DNA binding of NF-κB in microglia.³⁷ In Figure 6A, RELA in DEGs downstream of PARP1 is a subunit of NF-κB. Therefore, we believe that the molecular mechanism of AD toxicity by PAHs may involve the PAHs-PARP1-RELA pathway. Moreover, PARP-1 also has a regulatory effect on FOXO1-mediated transcription.³⁸ FoxO transcription factors are essential in aging and age-related metabolic and neurodegenerative diseases.³⁹ Hence, the PAHs-PARP1-FOXO1 pathway (Figure 6A) is also considered an important pathway involved in the neurotoxicity of PAHs.

Figure 6 shows that, in addition to PARP1, PTPN1, and ITGA4 are also linked to AD-associated toxicity from PAHs. Previous studies have shown that PTPN1 signaling rescues tau pathology in AD⁴⁰ and serves as a target for anti-AD treatment with hederagenin.⁴¹ Additionally, the therapeutic potential of ITGA4 inhibition interferes with AD progression and cognitive impairment.⁴² The synergistic effects of integrin function with the extracellular matrix (ECM) in synaptic plasticity and memory are crucial in the brain, and the role of integrins in brain diseases has been well-documented.⁴³ However, whether PAHs can directly interact with PTPN1 and ITGA4 and thus affect memory remains unclear. Notably, the genes PARP1, PTPN1, and ITGA4 all play essential roles in AD, and their interactions with PAHs have very low binding free energies. Therefore, these binding free energies may provide a basis for preliminarily estimating the AD-related toxicity of PAHs within this framework.

There is another question: is the magnitude of the free energy of binding PAHs to these target proteins consistent with the AD-associated toxicity of PAHs? We sampled the interaction sites of four PAHs with PARP1, PTPN1, and ITGA4 (Figure S1) and found that benzo[a]pyrene, phenanthrene, fluoranthene, and pyrene bind in the same protein-binding pocket. When we constructed the free energy landscapes of the four PAHs with PARP1 (Figure 7), we found that benzo[a]pyrene can make PARP1 change significantly and is relatively stable (dark blue color is located in the lower right corner). Figure 9B suggests an approximately linear trend between LD₅₀ values and PAH categories, in which stronger intermolecular binding stability tends to be associated with lower LD₅₀ concentrations in this dataset. Benzo[a]pyrene could lead to more pronounced AD-related toxicity due to its capacity to induce more significant protein denaturation.

To further verify whether PAHs’ ΔG can be used to predict the AD toxicity of PAHs, we use BMDL^44,45 as the response variable of ΔG and explore their relationship. In Figures 8A and 8B, the BMDLs of fluorene and phenanthrene from brandling worms in soil were 0.474 and 0.583, respectively. Their molecular docking free energies with PARP1 are −8.3 and −8.6. These two data points may not differ significantly, for the two substances are also clustered in Figure 6B. In the cluster analysis, where only PARP1, PTPN1, ITGA4, and PAHs docking free energy values were analyzed (Figure 9A), fluorene and phenanthrene were in one module, further indicating that the AD toxicity of these two compounds might be essentially the same.

Interestingly, in Figures 8C and 8D, the BMDLs of benz[a]anthracene and anthracene are 7.856 and 10.096, respectively, and their average G docking to PARP1, PTPN1, and ITGA4 proteins are −9.4 and −8.1. The trends of these two datasets remain consistent. In Figure 9A, the free energy of the module where benz[a]anthracene is located is −8.63, while for anthracene, it’s −8.03. These values also align with the BMDL values for both chemicals. We speculate that molecular docking scores of PAHs to these AD-related proteins may provide preliminary information on PAH-associated toxicity in AD. The reference docking target can be PARP1 alone or a combination of PARP1, PTPN1, and ITGA4. To further determine if predicting the AD toxicity of PAHs with their binding ΔG is theoretically feasible. We propose a hypothesis: the BMDL values for AD toxicity of PAHs have an increasing relationship based on the classification itself. When this condition is valid, the slope of neurotoxicity can be calculated using two values reported in the literature. The intercept can then be calculated based on the BMDL value of 5.2 mg/kg for Benzo(a)pyrene. This linear equation could be explored as a first approximation to estimate BMDL values for the 16 PAHs in this dataset. However, the reliability of the parameters of this linear equation remains to be verified. Based on existing research results, we suggest using ΔG clustering analysis to assess the neurotoxicity similarity of PAHs within the same cluster group.

To obtain an initial indication of how our model (Figure 9A) might relate to in vivo responses, we selected three substances—naphthalene, anthracene, and pyrene—from Figure 9A, whose docking free energy scores decrease in order. We found that their effects on the swimming distance of zebrafish gradually increase. Phototaxis assays, which involve automatically tracking the movement of larvae under different lighting conditions, are widely used to screen for neurotoxic effects.⁴⁶ These findings suggest that naphthalene, anthracene, and pyrene can elicit zebrafish neurotoxic responses under our assay conditions, and their relative effects show a trend consistent with the ΔG ranking, even though the statistical p value exceeds 0.05 (Figure 9D). It is important to note that, to ensure the statistical significance and predictive capability of the model, the compound data for naphthalene (see Result 3.7 and Table S10) were explicitly excluded from the LD₅₀-ΔG linear correlation analysis in this study. The purpose of this data processing was to maintain a monotonic LD₅₀-ΔG relationship for exploratory modeling. It is worth noting that in the subsequent zebrafish experiments, naphthalene was included as a characteristic compound. The results confirmed that the linear relationship between swimming distance and ΔG remained valid, as naphthalene was successfully included as a characteristic compound in these experiments. Zebrafish validation further demonstrated perfect correlation between predicted and observed effects (Spearman’s ρ = 1). However, the p value was non-significant (p > 0.05) due to limited sample size, thus providing only preliminary evidence of translational validity.

We recognize that the substantial leap from in silico binding affinity (ΔG) to complex in vivo neurotoxic outcomes (e.g., BMDL, LD₅₀, zebrafish behavior) overlooks key determinants of a compound’s overall AD risk. Notably, absorption, distribution, metabolism, and excretion (ADME) properties, blood-brain barrier penetration, metabolic activation/detoxification pathways (critical for PAHs), and compensatory biological mechanisms play pivotal roles in actual toxicity. Our model does not assume ΔG is the primary or sole determinant but posits it as a proxy for an integrative toxicological property—specifically, the potential for PAH-core target engagement that initiates downstream AD-relevant cascades (e.g., PARP1-NF-κB). Preliminary correlations with empirical data (e.g., Spearman’s ρ = 1 in zebrafish, linear LD₅₀-ΔG fit) support this proxy role, though causality remains unproven without ADME integration. Future models should incorporate quantitative structure-activity relationships (QSAR) for BBB permeability and CYP450 metabolism to refine predictions. This ΔG-based clustering thus serves as a high-throughput screening tool, prioritizing PAHs for targeted in vivo validation rather than standalone risk assessment.

Overall, our study integrates bioinformatics, machine learning, and ΔG clustering to propose a rapid screening framework for PAH neurotoxicity linked to AD. Through this pipeline, we identified PARP1, PTPN1, and ITGA4 as core targets enriched in AD-relevant pathways (neuroinflammation, microglial activation, and lipid metabolism). Building on docking-derived ΔG clustering (Figure 9A), category-average binding energies were consistent with an approximate linear relationship with literature LD₅₀/BMDL data (ρ = 1, p = 0.0417; Figure 9B), from which we derived an exploratory, hypothesis-generating relationship BMDL = 1.723×ΔG + 22.602 within this compound set. This relationship was qualitatively consistent with zebrafish motility assays (Spearman’s ρ = −1.0, p = 0.167; n = 3), although the evidence was underpowered. While promising, these preliminary findings offer novel mechanistic insights into PAH-AD interactions and computational tools for environmental neurotoxicant prioritization, pending comprehensive experimental validation of target causality, ADME integration, and model robustness.

Limitations of the study

The data presented in this paper have certain limitations. Firstly, the main findings are primarily derived from bioinformatics analysis, molecular docking, and a limited number of zebrafish behavioral experiments. They lack comprehensive functional validation in cellular and mammalian models, which means the causal relationships are not fully established. Further validation through in vitro binding assays, protein functional interventions, and exposure studies in AD animal models is necessary. Secondly, the BMDL-ΔG linear model was developed using limited dose-response data from only a few PAHs. Applying it to all 16 PAHs might overstate the model’s robustness and applicability. Future research should include additional dose gradients and multi-endpoint toxicology data, and consider nonlinear or multifactorial models. Third, although batch effects were corrected when combining multiple GEO chip datasets, population differences and platform variations could still introduce biases. To improve the generalizability and translational relevance of the findings, future studies should incorporate larger-sample RNA-seq or single-cell datasets alongside epidemiological evidence.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Chao Chen (chenchao@bnuz.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All data reported in this article will be shared by the lead contact upon request.
• •All original code has been deposited at (https://github.com/pdpmb/PAHs) and is publicly available at (https://doi.org/10.5281/zenodo.18849519) as of the date of publication.
• •Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.

Acknowledgments

This work was supported in part by the National Key Research and Development Project (no. 2023YFC3905105), the National Customs Administration Applied Technology Project (no. 2023HK135), and the National Natural Science Foundation of China (82370509 and 82370510).

Author contributions

Conceptualization, C.C. and D.S.; methodology, C.C. and D.S.; formal analysis, Y.H. and D.S.; software, Y.H.; validation, Y.N.; investigation, Y.N. and M.C.; resources, M.C.; data curation, Y.N.; visualization, Y.H.; writing – original draft, D.S.; writing – review and editing, C.C. and W.Z.; funding acquisition, M.C. and W.Z.; project administration, M.C.; supervision, W.Z.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
GEO datasets	https://www.ncbi.nlm.nih.gov/gds	GSE15222, GSE44771, GSE44770, GSE48350, GSE33000, GSE118553
Software and algorithms
ComBat (sva package)	https://cran.r-project.org/web/packages/COMBAT/index.html	v3.48.0
tinyarray	https://cran.r-project.org/web/packages/tinyarray/index.html	v2.3.3
Rtsne	https://cran.r-project.org/web/packages/Rtsne/index.html	v0.17
WGCNA	https://fuzzyatelin.github.io/bioanth-stats/module-F21-Group1/module-F21-Group1.html	v1.70-5
STRING (PPI)	https://string-db.org/	N/A
Cytoscape	https://cytoscape.org/	v3.9.1;
XGBoost	https://github.com/dmlc/xgboost	v1.6.0.1
AutoDockTools	https://autodock.scripps.edu/	v1.5.6
Dockstring (modified)	https://github.com/dockstring/dockstring	v 0.3.4
pheatmap	https://cran.r-project.org/web/packages/pheatmap/index.html	v1.0.12
PyMOL	https://www.pymol.org/	v3.0
MD simulation	https://conda-forge.org/	AMBER99SB + SPC
The code used in this work	https://github.com/pdpmb/PAHs	https://doi.org/10.5281/zenodo.18849519
Other
Zebrafish larvae	Instrumentation and Service Center for Science and Technology	8 hpf embryos, cultured to 5 dpf
Behavioral system	Instrumentation and Service Center for Science and Technology	Viewpoint Zebra Lab ZEB6194

Experimental model and study participant details

Animals

The experiment used a concentration gradient method to prepare working solutions of pyrene (12.5 – 400 μM), anthracene (31.25 – 1000 μM), and naphthalene (50 – 1600 μM). Healthy 8 hours post-fertilization (hpf) zebrafish embryos were selected, and 3 mL of the corresponding concentration working solution was added according to the experimental groups. The control group (n=30) was added with an equal amount of embryo culture medium (sodium chloride (NaCl) 0.8 g, potassium chloride (KCl) 0.04 g, sodium dihydrogen phosphate (Na₂HPO₄) 0.00358 g, potassium dihydrogen phosphate (KH₂PO₄) 0.006 g, calcium chloride (CaCl₂) 0.144 g, magnesium sulfate heptahydrate (MgSO₄·7H₂O) 0.246 g, sodium bicarbonate (NaHCO₃) 0.35 g, 1L). The experimental groups (n=30/group) were cultured in a constant-temperature, humidity-controlled, light-controlled incubator at 28°C, with medium changes at 72 hpf. We continued observations until 120 hpf. Using a ZEISS AXIO ZOOM.V16 microscope, we recorded phenotypic characteristics and mortality rates in each zebrafish group to determine the maximum non-lethal concentration (MNLC) of the three polycyclic aromatic hydrocarbons.

Zebrafish embryos at 8 hpf were treated with the MNLC, detoxified at 108 hpf, and then cultured under dark, constant temperature conditions until 5 days post-fertilization (dpf). Developmentally normal 5 dpf zebrafish larvae were selected and transferred to 96-well plates according to group (n=12), with 200 μL of embryo culture medium added to each well. To ensure the accuracy of behavioral experiments, samples were acclimated in a behavioral observation chamber for 30 minutes, with strict control of environmental parameters to avoid external interference. Subsequently, measurements were performed using the Viewpoint Zebra Lab (model: ZEB6194) juvenile fish behavioral analysis system. A 5-minute pre-adaptation period was conducted before measurement. The experimental parameters were set as follows: background at 25 pixels, swimming speed between 1 and 3 mm/s, data acquisition interval of 1 minute, and light intensity at 30%. Each measurement lasted 18 minutes, including three complete light-dark cycles (each cycle comprising 3 minutes of light and 3 minutes of darkness).

All animal experiments were approved by the Experimental Animal Ethics Committee of Beijing Normal University at Zhuhai (approval No. CBNUZ2023005, approved on [2023-3-14]).

Method details

Predicting target proteins of PAHs for neurological abnormalities

Four types of PAHs with well-understood molecular mechanisms involved in neurological abnormalities were selected as representatives of PAHs: benzo[a]pyrene, phenanthrene, fluoranthene, and pyren.⁴⁷ PAH target proteins were then screened using the ChEMBL (https://www.ebi.ac.uk/chembl/) and STITCH (http://stitch.embl.de/) databases. In ChEMBL, proteins with a confidence score above 90% were considered predicted PAH target proteins. From STITCH, human-origin proteins were selected as potential PAH targets.

GEO data cleaning

GEO data for genes related to AD was screened using the following criteria: sample counts > 100, expression profiling by array, and cortical tissue. After discarding data without original probe information, six datasets were obtained: GSE15222, GSE44771, GSE44770, GSE48350, GSE33000, and GSE118553. Dual-channel microarrays (GSE44771, GSE44770, and GSE33000) were selected for the discrepancy and WGCNA analyses. The remaining non-dual-channel biochip data was used for subsequent machine learning. As late-onset AD typically begins after the age of 65, clinical samples from individuals aged 65 years and over were selected for subsequent bioinformatics analyses.

Differentially expressed genes (DEGs) analysis

Differences between batches were eliminated using ComBat in the R package sva (version 3.48.0). Principal component analysis (PCA) results, both before and after batch correction, were presented using the R package tinyarray (version 2.3.3). After combining the data, patients and controls were analyzed for dimensionality reduction using Rtsne (version 0.17). DEGs in the AD dataset were screened using thresholds of log2 ≥ 0.07 (a 5% trend in expression change) and adjusted P-value ≤ 0.001. DEG expression patterns were visualized as volcano maps using ggplot2 (version 3.4.0) and EnhancedVolcano (version 1.16.0) R packages.

Weighted gene co-expression network analysis (WGCNA)

WGCNA (version 1.70-5) was used to obtain insight into relationships between clustered genes in the AD dataset and AD traits. Only genes with expression in the top 90% of the AD dataset were retained, and GSM1423944 was removed due to anomalies in the gene clustering process. Scale-free co-expressed gene networks were constructed using the one-step network construction function of the “WGCNA” package. A soft-threshold power of β = 5 was used as the weight value. After module eigengenes (MEs) for each cluster, associations between MEs and clinical features (R > 0.4) were calculated. Finally, Gene Significance (GS) and Module Membership (MM) measurements were used to identify genes highly associated with AD (kME_MM > 0.8) and associated module members.

Constructing protein-protein interaction (PPI) networks and screening core targets

DEGs in AD were intersected with AD-associated genes identified through WGCNA to obtain AD-associated DEGs. This gene set, along with the predicted target proteins of PAHs, was then input into the STRING database. For the PPI network construction, the parameters were as follows: species was restricted to “Homo sapiens,” the “Minimum Required Interaction Score” was set to “High Confidence > 0.7,” and the “FDR Strictness Value” was set to “High.” These parameters enabled the analysis of possible downstream proteins that interact with PAHs. STRING results were then imported into Cytoscape software (version 3.9.1) to analyze network structure. PAH-predicted target proteins and their directly interacting DEG gene sets were used for further community network analysis and Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis. The community network was split using the Leading eigenvector algorithm with the clusterMaker2 plugin.⁴⁸ For GO enrichment analyses, pAdjustMethod was set to “BH” and pvalueCutoff = 0.01 and qvalueCutoff = 0.05 were selected as filtering parameters. For KEGG enrichment analysis, qvalueCutoff = 0.05.

Screening key downstream genes of PAH target proteins using machine learning methods

To identify critical downstream genes of PAH-predicted target proteins, we first assessed each gene's importance score using the XGBoost package (version 1.6.0.1), capitalizing on its robust feature importance metrics. Before this, we validated XGBoost's superior performance for gene expression prediction by benchmarking it against three established models—Random Forest, Gradient Boosting Machine (GBM), and Ridge Regression—as shown in Table S1. Specifically, we utilized gene expression features (from GSE15222) to predict AD values. Across these models, XGBoost demonstrated the highest classification AUC and competitive Accuracy, thereby confirming its enhanced generalization and at least comparable—or superior—predictive power relative to established approaches. Additionally, when performing data analysis, hyperparameters were tuned via grid search (e.g., max_depth = 5, nrounds = 100, eta = 0.1, and gamma = 0.25), with implementation details available in the referenced GitHub repository (https://github.com/pdpmb/PAHs).

Genes of interest were screened from PAH-predicted target proteins and their directly interacting DEGs. Seventy percent of the data was used as a training dataset, and 30% as an independent test dataset, with the model trained using k-fold cross-validation (k = 5). Optimization parameters were chosen based on the smaller root mean square error value. Additionally, the top 5 feature genes obtained from two independent datasets, GSE15222 and GSE118553, were included in the evaluation. GSE48350 was excluded from the XGBoost analysis due to its small sample size.

Molecular docking of PAHs to core targets

Core targets of PAHs affecting AD were screened by using key genes isolated by XGBoost as bait molecules to capture the core predicted target proteins from the community network. Molecular docking was then employed to analyze intermolecular interactions between PAHs and these core target proteins, predicting binding modes and affinities. Molecular crystal structures of core target proteins were downloaded from the RCSB Protein Data Bank (https://www.rcsb.org/). Water molecules and original ligands were removed using Chimera, with only one chain being retained and hydrogens and charges being added. Pdbqt files were prepared using AutoDockTools 1.5.6, and molecular docking was performed using a modified version of Dockstring.⁴⁹ The docking pocket size was set to 30 nm for proteins containing the original ligand and 70 nm for those without.

A network of PAHs, PAH targets, and featured essential genes obtained using XGBoost was visualized with Cytoscape software. Docking scores were displayed using pheatmap (1.0.12), and final docking results were visualized using PyMOL (Version 3.0 Schrödinger, LLC). Cluster analysis of binding free energies of PAHs to their core target proteins was performed using hclust in R. Euclidean distances were used to calculate the distances between harmonized quantiles. After generating the ΔG-based cluster dendrogram [binding free energy (ΔG)], the average ΔG for molecules in each cluster was calculated based on their cluster assignments.

Molecular dynamics simulation

All molecular dynamics simulations were performed using 2023.4-conda_forge software, with the AMBER99SB protein force field and SPC water model being employed. Box boundaries were localized at 1 Å and all systems were initially energy-minimized using a rapid descent algorithm to reduce unreasonable contact. MD models were generated after pre-equilibration using Nvt and Npt. Each system was simulated 10⁷ times with dt set to 0.002 (2 fs). RMSD and radius of gyration were used to map morphology, and a final energy heat map was created using Python code.

Benchmark dose (BMD) calculation

To verify whether the free energy of molecular docking binding of PAHs to core targets can be used for toxicity assessment, molecular docking energies (binding free energies, ΔG) were compared with benchmark dose values. The BMD online tool (https://www.epa.gov/bmds) was used to calculate confidence intervals for estimated BMDs. The BMDL (benchmark dose lower confidence limit) was used to assess the toxicity of the compound. The Benchmark Response (BMR) was set as a 5% or 10% change in adverse reaction response rate relative to controls, depending on whether the response data were continuous or dichotomous. Five dose levels for continuous variables were collected to calculate BMDL values. Due to the limited data available in most studies—only four dose levels of continuous variables—the raw data were estimated using ORIGIN from literature bar graphs. This approach assumes that continuous variables representing cellular effects are composite representations of dichotomous variables for cell groups. The BMD online tool was then used to make a preliminary assessment of the BMDL for cellular effects of dichotomous variables, since the prerequisite for BMD analyses in this category is four dose levels.

Next, we established an essential assumption. We assumed that the toxicity magnitude of a given substance is consistent with others within the same categories in the cluster dendrogram and that a linear relationship exists between the BMDLs of PAHs from different categories. Due to the scarcity of accessible BMD data, Median lethal dose (LD₅₀) data for 16 PAHs were collected to assess whether our assumption of a linear relationship between the toxicity data was reasonable. The T3DB database⁵⁰ was reviewed in the literature and utilized for LD₅₀ data collection. After the LD₅₀ data indicated a linear relationship, linear equations for BMDLs of 16 PAHs were extrapolated.

Quantification and statistical analysis

All analyses were performed in R (version 4.0.1) and Python (version 3.6). Microarray data were batch-corrected using ComBat (sva, v3.48.0), explored by PCA (tinyarray, v2.3.3) and t-SNE (Rtsne, v0.17), and DEGs between AD and controls were identified by linear modeling with thresholds of ∣log⁡₂fold‑change∣≥0.07 and Benjamini–Hochberg–adjusted P ≤ 0.001, visualized by ggplot2 (v3.4.0) and EnhancedVolcano (v1.16.0). Co-expression modules were constructed using WGCNA (v1.70-5), with a soft-threshold power β = 5; Pearson correlations between module eigengenes and clinical traits were computed, and modules with |R| > 0.4 were retained. PPI networks and community structures were obtained from STRING (Homo sapiens, confidence > 0.7) and Cytoscape (v3.9.1), and functional enrichment (GO, KEGG) was analyzed in R using over-representation tests with BH FDR correction (GO: P < 0.01 and q < 0.05; KEGG: q < 0.05). Key downstream genes were selected by XGBoost (v1.6.0.1) models (70% train/30% test, 5-fold cross-validation, grid-searched hyperparameters), with performance summarized by AUC, accuracy, and RMSE. Docking of PAHs to core targets used structures from RCSB PDB, prepared in Chimera and AutoDockTools (v1.5.6), run with a modified Dockstring workflow; binding free energies (ΔG) were clustered by hierarchical clustering (hclust) and visualized using pheatmap and PyMOL. Molecular dynamics simulations were carried out with GROMACS (2023.4-conda_forge, AMBER99SB force field, SPC water), and stability metrics (RMSD, radius of gyration) were summarized as mean ± SD. Benchmark dose (BMD) and BMDL values were derived for PAHs using the US EPA BMDS online tool, and linear relationships between ΔG, BMDL, and LD50 data (T3DB) were assessed by simple linear regression. Unless otherwise stated, all tests were two-sided with significance defined as P < 0.05 or FDR-adjusted q < 0.05; for each analysis, the exact n (number of independent human samples, docking pairs, simulations, or PAHs), definition of center (typically mean) and dispersion (SD, SEM, or 95% CI), and full statistical details are provided in the figure legends, main Results, or Supplementary Tables.

Published: March 11, 2026