Deconvolution of molecular mechanisms in di-n-butyl phthalate/mono-n-butyl phthalate induced diabetic kidney disease by integrated machine learning and molecular docking Authors: Wenjie Chen, Suyue Hou, Jijia Hu Published: 2025 ABSTRACT AbstractBackgroundThis study investigates the molecular mechanisms by which di-n-butyl phthalate (DBP) and mono-n-butyl phthalate (MnBP)-induced diabetic kidney disease (DKD).MethodsDifferential expression analysis and Weighted Gene Co-expression Network Analysis were used to identify DKD-associated targets. Machine learning, molecular docking, molecular dynamics simulations, and public databases were integrated to explore the interaction between DBP/MnBP and target proteins.ResultsSix core genes were identified: DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2. Among these, DUSP1 and FOSB showed excellent performance in single-gene ROC curves, box plots, public databases, and molecular docking. Molecular docking and molecular dynamics simulations demonstrated a stable binding affinity between DBP/MnBP and the target proteins.ConclusionThis research suggests that DBP/MnBP may promote DKD by targeting these six core genes. The binding capacity and stability of DBP/MnBP with these genes were confirmed by machine learning, molecular docking, and molecular dynamics simulations. These findings provide a direction for future in-depth research on the DBP/MnBP-induced DKD mechanism.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12882-025-04701-3.Keywords: DBP, MnBP, Diabetic kidney disease, Network toxicology, Machine learning, Molecular docking, Bioinformatics Introduction As the primary cause of end-stage kidney disease (ESKD) worldwide, DKD affects about 40% of people with diabetes, presenting a major public health challenge [1]. Beyond established factors like age and smoking [2], a growing body of evidence points to the role of environmental toxins in its development. For example, recent analyses of the NHANES database found a significant link between the risk of DKD and exposure to endocrine-disrupting chemicals (EDCs) [3]. The accumulation of EDC metabolites is strongly correlated with a heightened risk of DKD. Furthermore, a network toxicology study further provides a theoretical basis for how EDCs, involving DBP, contribute to diabetic microvascular complications like DKD [4]. Commonly used plasticizers known as phthalates, particularly DBP, are implicated in various health issues, including kidney disease and diabetes [5]. DBP’s nephrotoxic effects are well-documented [6]. A cross-sectional study showed a positive correlation between urinary DBP metabolites and the albumin-to-creatinine ratio (ACR), and a negative correlation with estimated glomerular filtration rate (eGFR) [7]. DBP exposure also strongly links to Type 2 diabetes and insulin resistance, with supporting evidence from animal and in vitro studies [8]. Co-exposure with other toxins, like benzo[a]pyrene further complicates this by inducing pyroptosis in renal cells and promoting kidney injury [9]. The DBP metabolite, MnBP, has a significant positive association with ACR [10], pointing to a potential connection with diabetic nephropathy. The selection of DBP and MnBP concentrations is justified based on their relevance to DKD pathophysiology, particularly tubular injury and mitochondrial dysfunction. Recent studies highlight the heterogeneity of DKD, including non-albuminuric phenotypes characterized by eGFR decline without albuminuria, often linked to tubular rather than glomerular damage [11, 12]. DBP and MnBP, as environmental toxicants, are known to induce renal tubular injury through oxidative stress and inflammatory pathways, aligning with this DKD subtype [13]. Moreover, mitochondrial dysfunction is a key mechanism in DKD progression [14]. DBP and MnBP exposure disrupts mitochondrial quality control, including dynamics and mitophagy, exacerbating renal damage. Concentrations reflecting human environmental exposure ensure physiological relevance. Additionally, these compounds activate profibrotic and inflammatory pathways, consistent with DKD pathology [15, 16].However, the specific molecular mechanisms behind DBP and MnBP’s effect on this condition are still not fully understood. Network toxicology is a powerful new strategy for understanding toxicant-disease relationships, as it integrates computational biology methods such as transcriptomics, molecular docking, and molecular dynamics to create a systemic map of interactions among external toxins, their genetic targets, and diseases [17]. In recent years, in-silico techniques have become increasingly pivotal in enhancing this framework, offering efficient predictive and validation tools that deepen insights into complex biological systems. For example, enhanced sampling simulations have been applied to identify potent phosphodiesterase-1 inhibitors, demonstrating how molecular dynamics can uncover binding mechanisms relevant to neurological disorders and neuroprotection [17]. Similarly, studies on BRD4-BD1 inhibitors leverage molecular docking and steered molecular dynamics to analyze binding free energies and identify critical amino acid residues, highlighting the utility of these methods in drug discovery for various malignancies [18]. In the realm of infectious diseases, in-silico approaches have been employed to address drug resistance, such as in tuberculosis research, where molecular docking and dynamics simulations validate the efficacy of bioactive compounds like thalirugidine against mutated targets [19]. Additionally, computational interventions in enzyme engineering, including the identification of mutational hotspots in serratiopeptidase to improve pH tolerance, illustrate how molecular dynamics can optimize protein stability and function under extreme conditions [20]. For broader biomolecular interactions, descriptor-based computational models like MDbDMRP enable high-accuracy predictions of drug-miRNA relationships, advancing applications in disease therapy and pharmaceutical development [21]. Furthermore, investigations into phosphodiesterase 10 A inhibitors combine molecular docking, classical molecular dynamics, and steered simulations to identify promising candidates for neurological disorders, reinforcing the value of in-silico techniques in targeted therapeutic design [22]. Collectively, these studies exemplify the integral role of recent in-silico methodologies in network toxicology, facilitating a comprehensive understanding of toxicant-disease interactions and supporting the development of evidence-based interventions.This makes it an ideal tool for investigating how DBP and MnBP affect the kidneys, particularly in DKD [23]. In this field, molecular docking and dynamics simulations are increasingly used to predict how macromolecules and small ligands bind and interact [24, 25]. Machine learning helps streamline the research process by identifying and prioritizing key genes, providing a more precise focus for downstream docking and simulation studies [26, 27]. For instance, researchers like Song et al. have successfully used network analysis, predictive toxicity tools, and molecular docking to pinpoint the targets and mechanisms of endocrine-disrupting chemicals in diabetic microvascular diseases [4]. Materials and methods Acquisition of DKD-related targets To analyze glomerular gene expression in DKD, six transcriptomic datasets (GSE96804, GSE30528, GSE142025, GSE104948, GSE99339, and GSE47183) were retrieved from the NCBI GEO database (https://www.ncbi.nlm.nih.gov/gds) (accessed 23 August 2025). We established a training set by merging GSE96804 (41 DKD, 20 control), GSE30528 (9 DKD, 13 control), and GSE142025 (27 DKD, 9 control). To address batch effects beyond principal components analysis (PCA) visualization, we applied the ComBat algorithm using an empirical Bayes framework to statistically adjust for inter-dataset variations while preserving biological signals, with the correction efficacy further validated by quantifying the reduction in batch-associated variance through statistical metrics such as R-squared values, in addition to PCA plots confirming minimized batch clustering [28]. Furthermore, to mitigate potential confounding errors from normalization biases, all gene expression data were preprocessed with robust multi-array average normalization, and biological replicates were managed by including only unique samples and removing any duplicates to ensure independence and prevent overrepresentation, aligning with best practices for integrating heterogeneous genomic data to enhance model robustness and generalizability. GSE104948 (12 DKD, 26 control) was used for validation, and two additional datasets, GSE99339 (14 DKD, 14 control) and GSE47183 (14 DKD, 17 control), served as independent test sets. The analytical workflow is summarized in (Fig. 1). Flow-chart of this paper Retrieval of chemical components, toxicity prediction, targets for DBP and MnBP The chemical components of DBP and MnBP were obtained from PubChem database(https://pubchem.ncbi.nlm.nih.gov/) [29](accessed 8 September 2025). The canonical SMILES notation for DBP is CCCCOC(= O)C1 = CC = CC = C1C(= O)OCCCC, and for MnBP is CCCCOC(= O) CCCCOC(= O)C1 = CC = CC = C1C(= O)O. The ADMETlab3.0 database (https://admetlab3.scbdd.com/server/screening) [30](accessed 8 September 2025) and ProTox 3.0 database (https://tox.charite.de/protox3/index.php?site=compound_input) [31](accessed 8 September 2025) were applied to predict the toxicity (especially nephrotoxicity) of DBP and MnBP. Targets of DBP and MnBP were acquired from Comparative Toxicogenomics Database (CTD) (https://ctdbase.org/) [32](accessed 8 September 2025), SwissTargetPrediction (http://www.swisstargetprediction.ch) [33](accessed 8 September 2025), PharmMapper (https://www.lilab-ecust.cn/pharmmapper/index.html) [34](accessed 8 September 2025). Targets for DBP and MnBP, retrieved from the three, were integrated and visualized via the ggvenn (v0.1.10) [35] and gridExtra (v2.3) [36] R packages. Differentially expressed genes (DEGs) analysis DEGs analysis was performed on the aforementioned GEO dataset with the limma package (v3.64.3) [37] in R (v4.5.1) [38]. FDR-adjusted p-value < 0.05 and |log₂-fold change| (log₂FC|) > 0.585 (corresponding to 1.5-FC) were considered statistically significant, visualized using the ggplot2 (v4.0.0) [39] and pheatmap (v1.0.13) [40] packages.” Weighted gene co-expression network analysis (WGCNA) Gene expression datasets were preprocessed through log₂ transformation and normalized using the limma package. Low-variance genes were left out (standard deviation > 0.5), and outlier samples were erased, according to hierarchical clustering. Preprocessed expression data were utilized to structure a scale-free network via WGCNA package (v‘1.73’) [41, 42] with a soft-thresholding power of 8, which was selected as it achieved the highest scale-free topology fit index (R² = 0.757). Gene modules were recognized using hierarchical clustering of the topological overlap matrix (TOM) and dynamic tree cutting (minModuleSize = 30, mergeCutHeight = 0.25). Module-trait correlations were evaluated by calculating Pearson correlations between module eigengenes and clinical phenotypes, with significant associations defined as those with a correlation coefficient |R| > 0.5 and a p-value < 0.05. Hub genes were defined as those with high intramodular connectivity (kME > 0.8) within significant modules. Identification of DBP- and MnBP-induced DKD targets The final set of DKD targets associated with DBP and MnBP was identified by taking the intersection of targets from three sources: DEGs, the most significant WGCNA modules, and retrieved DBP/MnBP targets, visualized through the ggvenn (v0.1.10) and gridExtra (v2.3) R packages.The intersection approach is biologically justified as it integrates multiple evidence layers: DEGs reflect expression changes, WGCNA modules indicate co-regulated networks, and targets denote direct interactions, thereby identifying core genes involved in specific toxic pathways. Functional enrichment analysis of DBP- and MnBP-induced disease targets Kyoto Encyclopedia of Genes and Genomes (KEGG) [43] pathway enrichment analysis employed standard significance cutoffs (p-value < 0.05, q-value < 0.05). For all Gene Ontology (GO) [44] categories—Biological Process (BP), Cellular Component (CC), and Molecular Function (MF)—a relaxed threshold (p-value cutoff = 0.10, q-value cutoff = 0.10) was performed to identify potential functional trends due to a lack of significant enrichment under stricter criteria. Mapping protein–protein interaction (PPI) network String database (https://cn.string-db.org/) [45] (accessed 10 September 2025) was adopted to illustrate PPI network (Homo sapiens; medium confidence: interaction score > 0.4) of core genes. Machine learning screening core genes We employed 12 machine learning algorithms—Least Absolute Shrinkage and Selection Operator (Lasso), Support Vector Machine (SVM), Random Forest (RF), Generalized Linear Model Boosting (GlmBoost), Partial Least Squares Regression for Generalized Linear Models (PlsRglm), Stepwise Generalized Linear Model (Stepglm), Ridge Regression (Ridge), Elastic Net (Enet), Gradient Boosting Machine (GBM), Linear Discriminant Analysis (LDA), eXtreme Gradient Boosting(XGBoost), and Naive Bayes Classifier(NaiveBayes)—to construct 137 predictive models. Model performance was evaluated using the area under the Receiver Operating Characteristic (ROC) curve (AUC) on training and validation sets. To prioritize generalizable models and mitigate overfitting, we calculated a penalized score for each model as follows: Penalized Score = Avg Validation AUC − λ × (Train AUC − Avg Validation AUC) where λ = 2 was used to penalize models with large performance gaps between training and validation sets. We selected the top 10 models with the highest penalized scores and validated them on two independent test sets. The robustness of these models was assessed by their AUC values across all sets. Finally, we extracted the feature genes from these top models for further analysis. In our study, we employed 12 distinct machine learning algorithms—Lasso, SVM, RF, GlmBoost, PlsRglm, Stepglm, Ridge, Enet, GBM, LDA, XGBoost, and NaiveBayes—to develop a total of 137 predictive models. To ensure robustness and reproducibility, a comprehensive machine learning pipeline was implemented. Prior to model training, all datasets underwent rigorous preprocessing, including handling of missing values via median imputation for continuous variables and mode imputation for categorical variables, as well as feature standardization using Z-score normalization to mitigate scale-related biases [46]. To address potential class imbalance—a common challenge in clinical prediction tasks—we applied the Synthetic Minority Over-sampling Technique combined with random under-sampling, a strategy shown to improve model generalizability in similar clinical contexts [47]. Hyperparameter optimization for each algorithm was conducted using a grid search approach with 5-fold cross-validation, maximizing the AUC as the evaluation metric [48]. Model performance was assessed based on AUC values calculated on both training and validation sets. To prioritize generalizable models and reduce overfitting, a penalized score was computed for each model as follows: Penalized Score = Avg Validation AUC − λ × (Train AUC − Avg Validation AUC), where λ = 2 was selected to penalize models exhibiting large performance discrepancies between training and validation phases. The top 10 models with the highest penalized scores were subsequently validated on two independent test sets. Finally, to enhance interpretability and clinical relevance, feature genes derived from these top-performing models were extracted and further analyzed [49]. Online database analysis of core genes To validate the core genes, we adopted several online databases. Nephroseq v5 (https://nephroseq.org/resource/login.html) [50, 51](accessed 15 September 2025) was utilized to extract transcriptomic data of core genes from both glomerular and tubule cohorts in DKD-related cohorts, and to analyze their correlation with eGFR. Scatter plots illustrating these correlations were generated using the “ggplot2” package in R. The Kidney Interactive Transcriptomics (https://humphreyslab.com/SingleCell/search.php) [52, 53] (accessed 15 September 2025) was employed to examine the single-cell expression of core genes in DKD and control kidneys. Violin plots were adopted to visualize their expression levels and percentage of expressing cells. The Dsigdb database (https://dsigdb.tanlab.org/DSigDBv1.0/) [54] (accessed 15 September 2025) was utilized to identify approved drugs (D1) targeting each core gene. We applied the Enrichr database (https://maayanlab.cloud/Enrichr/) [55, 56] (accessed 15 September 2025) to perform drug enrichment analysis. Nomogram and ROC curve evaluation across all datasets In the training set, a multivariate logistic regression model was established using the R rms package (v‘8.0.0’) [40] to develop a nomogram for DKD prediction based on core gene expression data. Model performance was assessed through pROC package (v‘1.19.0.1’) [57] evaluating discriminative ability (AUC calculation) for both individual genes and the combined model, along with calibration curves using 1000 bootstrap resamples to verify prediction accuracy. The nomogram was subsequently validated on independent validation and test sets using the same evaluation metrics. SHapley additive explanations (SHAP) analysis To structure a predictive model, we utilized five machine learning algorithms: RF, SVM, XGBoost, GBM, and KNN. The model with the highest AUC was chosen for further analysis. While the nomogram offers an intuitive visualization based on the Cox regression model, its interpretability is constrained by the model’s assumptions. To provide a more robust, model-agnostic interpretation of our high-performing machine learning model, we employed SHAP analysis, which quantifies feature contributions for both global and local interpretability This method quantified the contribution of each feature, delivering both a global ranking of feature importance and local explanations for single predictions via summary, force, and dependence plots. This interpretable framework ensured transparency, identified critical biological features, and demonstrated robust generalizability upon validation with independent datasets. Immune cell infiltration analysis The goal of this analysis is to evaluate the levels of immune cell infiltration across the training, validation, and test datasets (as previously defined) using the CIBERSORT algorithm [58] (https://cibersort.stanford.edu/) (accessed 27 September 2025).The LM22 signature matrix will be used as the reference panel for deconvolution. A heatmap will be generated using the pheatmap package (v1.0.13) to display correlation analysis of immune cells. Box plots were used to illustrate the comparison of immune cells. Lollipop plots were constructed to show the correlation between the targets and the immune cell infiltration levels. Molecular docking Molecular docking was used to predict how DBP/MnBP and FDA-approved drugs bind to target proteins. The protein crystal structures were extracted from the RCSB Protein Data Bank (PDB, https://www.rcsb.org/) (accessed 12 September 2025), while three-dimensional (3D) structures of DBP/MnBP were acquired from the PubChem database. After preparing the proteins by removing unwanted ligands and water molecules using PyMOL, we performed the docking and visualization using the CB-Dock2 web server (https://cadd.labshare.cn/cb-dock2/php/index.php) [59](accessed 13 September 2025). We used CB-Dock2’s blind docking mode with default automated settings: AutoDock Vina 1.1.2 for structural docking and FitDock for template-based docking when homologous templates were available in the BioLiP database. Binding pockets and search boxes were automatically defined by the CurPocket al.gorithm. Ligand structures (DBP/MnBP) were retrieved from PubChem in SDF format, and protein structures were obtained via PDB IDs: 6APX (DUSP1), 6KVF (CXCR2), 6UCL (FOSB), 5VT2 (GDF15), 8WUY (NR4A1), and 5F19 (PTGS2). The top five poses per target remained consistent in binding energy and location across tests. All parameters are reproducible using the same inputs and the public CB-Dock2 server. The 3D structures of the FDA-approved drugs, identified through enrichment analysis using the Enrichr and DSigDB databases, were retrieved from the PubChem database. As described above, these structures were utilized for molecular docking and visualization with the target proteins. Molecular dynamics simulation (MD) The MD were performed on complexes formed between the proteins DUSP1 ,FOSB and PTSG2, and the ligands DBP and MnBP. All molecular dynamics simulations were performed with the following parameters: the system temperature was maintained at 310 K (approximately 37 °C) using the v-rescale thermostat with a coupling constant of 0.1 ps, while the pressure was controlled at 1 bar using the Parrinello-Rahman barostat with a coupling constant of 2.0 ps. The integration time step for the production run was set to 2 fs, and all bonds were constrained using the LINCS algorithm throughout both the equilibration and production phases to ensure numerical stability. The complex structures were initially obtained from CB-Dock2 molecular docking. The initial structures were obtained from CB-Dock2 docking, with a specific portion of the DUSP1 protein (residues 1-367) used to conserve computational resources based on UniProt database(https://www.uniprot.org/) [60] (accessed 20 September 2025). The protein structures were prepared by adding hydrogen atoms with Chem3D 14.0, and the ligands were converted to PDBQT files. The complexes were placed in a cubic box with periodic boundary conditions, solvated with the TIP3P water model, and parameterized with the CHARMM36 force field. NaCl ions (150 mM) were added to simulate physiological conditions. The system was then equilibrated in two steps (NVT and NPT) for 100 ps each, followed by a 100 ns production run using GROMACS 2025.2 [61]. After the simulation, several key parameters were analyzed to assess complex stability: Root Mean Square Deviation (RMSD) and Radius of Gyration (Rg) to measure overall structural stability and compactness; Root Mean Square Fluctuation (RMSF) to identify flexible regions; Solvent Accessible Surface Area (SASA) and hydrogen bond counts to evaluate protein-ligand interactions. To visualize the contribution of dominant residues to the binding free energy, we generated a Free Energy Contribution map from the Molecular Mechanics Poisson-Boltzmann Surface Area analysis. To visualize the conformational stability of the complexes, we constructed both two-dimensional (2D) and three-dimensional (3D) Free Energy Landscape (FEL) plots. Cell line and reagents Human renal proximal tubular epithelial cells (HK-2 cells), obtained from the American Type Culture Collection (Rockville, MD, USA), were maintained in a 1:1 mixture of Dulbecco’s Modified Eagle Medium and Ham’s F-12 nutrient medium containing 5% fetal bovine serum. HK2 cell was routinely checked for mycoplasma contamination while being cultivated at 37 °C and 5.0% CO2. DBP and MnBP were purchased from Aladdin Biochemical Technology Co., Ltd. (Shanghai, China). We divided the experiment into three groups: the DBP treated HK2 cell line group, the MnBP treated cell line group, and the HK2 cell line group without any added drugs. RNA extraction and the quantitative real-time PCR analysis (qRT-PCR) Total RNA was isolated from the harvested HK2 cell using TRIzol reagent (Invitrogen, USA). Related-primers (Table III) were designed and synthesized by Shenzhen Huada Gene Science and Technology Service Co., (Shanghai, China). mRNA levels was normalized to GAPDH RNA. Relative mRNA expression levels was quantified using the 2−ΔΔCt method.The primers are listed in Table 1. Table 1Sequences of primersPrimerSequences (5′→ 3′)DUSP1 FCAGCTTCAGCAGCATCACCTDUSP1 RGTGATGCTGCTGAAGCTGACPTGS2 FTTCAAATGAGATTGTGGGAAAPTGS2 RAGATCATCTCTGCCTGAGTATFOSB FATGGGCTCTCCTGTCAACACFOSB RGTGTTGACAGGAGAGCCCATGDF15 FGAGGAGCTGGTGGTGTTGAAGDF15 RTTCAACACCACCAGCTCCTCNR4A1 FCAGCCACCACCTACACCTACNR4A1 RGTAGGTGTAGGTGGTGCTGCXCR2 FCTTCCTGCTCCTGCTGTACTCXCR2 RAGTACAGCAGGAGCAGGAAG Sequences of primers Western blot analysis (WB) HK-2 cells were washed three times with ice-cold Phosphate-Buffered Saline and homogenized in RadioImmunoPrecipitation Assay Buffer lysis buffer (Solarbio, Beijing, China) containing inhibitors (protein phosphatase inhibitor: protein enzyme inhibitor = 100:1:1) at 4 °C for 30 min, followed by centrifugation to collect supernatants for protein quantification using a bicinchoninic acid assay kit (Elabscience, #E-BC-K318-M) with absorbance measured at 562 nm on a microplate reader; protein concentrations were equalized with lysis buffer, mixed with loading buffer, and denatured at 100 °C for 10 min before loading equal amounts onto 12.5% Sodium Dodecyl Sulfate – Polyacrylamide Gel Electrophoresis gels (EpiZyme Scientific, Shanghai, China) and transferring to Polyvinylidene difluoride membranes (Millipore, USA). The membranes were blocked with 5% non-fat milk for 2 h at room temperature, incubated overnight at 4 °C with primary antibodies from HUABIO: DUSP1 (ET1701-82, 1:1000), PTGS2 (ET1610-23, 1:1000), FOSB (ET1611-75, 1:1000), GDF15 (HA722551, 1:1000), NR4A1 (ET1703-97, 1:1000), CXCR2 (ER1906-87, 1:1000), and β-actin (Affinity Biosciences, AF7018, 1:5000), then rinsed three times with Tris-Buffered Saline with Tween®-20 and probed with Horseradish Peroxidase-conjugated goat anti-rabbit or anti-mouse IgG (1:10000) for 2 h at room temperature. Protein bands were visualized using R-tracker and Super Enhanced Chemiluminescence Detection Reagents (Yeasen, Shanghai, China), quantified with ImageJ software (NIH, USA), and normalized to β-actin as an internal control for relative density analysis. Results Toxicity analysis of DBP and MnBP We assessed the toxicity profiles of DBP and MnBP through an in silico analysis using two predictive platforms: the ADMETlab 3.0 (Table 2) and ProTox 3.0 (Table 3) databases. The ADMETlab 3.0 platform predicted high probabilities of toxicity for both compounds, particularly concerning carcinogenicity and eye irritation. Interestingly, the same database indicated a low probability for nephrotoxicity for both DBP (0.008) and MnBP (0.123). This finding, however, was contradicted by the ProTox 3.0 database, which predicted that both compounds would be active nephrotoxicants. Table 2Results of ADMETlab3.0 toxicityToxicityProbabilityDi-n-butyl phthalateMono-n-butyl phthalateSkin Sensitization0.3560.374Carcinogenicity0.7140.422Eye Corrosion0.1170.427Eye Irritation0.9980.997Respiratory0.0130.095Human Hepatotoxicity0.0060.071Drug-induced Nephrotoxicity0.0080.123Drug-induced Neurotoxicity0.0630.124 Results of ADMETlab3.0 toxicity Table 3Results of pro Tox3.0 predictionToxicityChemicalDi-n-butyl phthalateMono-n-butyl phthalateHepatotoxicityInactiveInactiveNeurotoxicityInactiveInactiveNephrotoxicityActiveActiveRespiratory toxicityInactiveInactiveCardiotoxicityInactiveInactive Results of pro Tox3.0 prediction Identification of DBP- and MnBP-induced targets The molecular structures of DBP and MnBP were sourced from the PubChem database, respectively (Fig. 2A and B). To identify potential protein targets for both compounds, a comprehensive search was performed across three distinct databases: the CTD, SwissTargetPrediction, and PharmMapper. After removing duplicates, a total of 473 potential targets were identified (Fig. 2C). Identification of target proteins for DBP and MnBP. (A) Chemical structure of DBP. (B) Chemical structure of MnBP. (C) Target prediction empolying CTD, PharmMapper, and SwissTargetPrediction Identification of DKD-related targets To enrich the sample size, the GSE142025, GSE30528, and GSE96804 datasets were combined and corrected for batch effects, forming a training set. PCA revealed clear separation between the three datasets before batch correction (Fig. 3A), but they clustered closely together after normalization (Fig. 3B), indicating successful batch effect removal. 440 DEGs were identified in the training set, visualized by a volcano plot (Fig. 3C) and a heatmap (Fig. 3D). We implemented WGCNA to construct a hierarchical clustering dendrogram of the TOM with dynamic tree cut modules (Fig. 3E). Seven modules were generated and visualized in a module-trait relationship heatmap (Fig. 3F). The MEgreen module was selected for further analysis based on its highest absolute Pearson correlation coefficient with DKD (0.69) and the lowest P-value. A Venn diagram showed that the intersection between this module and the DEGs yielded 32 genes (Fig. 3G). Identification of DKD-related target genes. (A) PCA scatter plot of datasets before batch correction (GSE142025, GSE30528, and GSE96804). (B) PCA scatter plot of the integrated datasets after batch correction. (C) Volcano plot showing DEGs. Red dots represent 177 upregulated genes and blue dots represent 263 downregulated genes. Gray dots indicate genes with no significant differential expression. (D) Heatmap of DEGs. The blue color represents downregulated genes, and the red color represents upregulated genes. (E) Gene dendrogram and dynamic tree cut modules. (F) Module-trait correlation heatmap with correlation coefficients and P-values. (G) Venn diagram showing the overlap between DEGs (yellow) and WGCNA modules (purple) Identification and functional enrichment analysis of DBP- and MnBP- induced targets in DKD To uncover the key genes in DBP- and MnBP- induced DKD, we intersected the gene sets for target proteins of DBP and MnBP with those of DKD-related target genes (Fig. 4A). Subsequently, PPI network was assembled and visualized from the STRING database (Fig. 4B). GO enrichment analysis revealed significant enrichment in several key terms. These included response to amphetamine, response to amine, and negative regulation of smooth muscle contraction (BP); nuclear membrane and nuclear outer membrane (CC); and MAP kinase tyrosine/serine/threonine phosphatase activity and adenylate cyclase regulator activity (MF) (Fig. 4C). The KEGG pathway analysis, in addition, highlighted the top three pathways as “IL-17 signaling pathway”, “Serotonergic synapse”, and “Oxytocin signaling pathway” (Fig. 4D). Identification and functional enrichment analysis of DBP- and MnBP- induced targets in DKD. (A) Venn diagram displays the intersection between genes related to DBP- and MnBP-exposure (purple) and those associated with DKD (yellow), yielding seven overlapping genes (1.4%). (B) PPI network of the overlapping genes, with interactions constructed and visualized leveraging the STRING database. (C) GO enrichment analysis of the overlapping genes. The x-axis shows the number of genes, and the color gradient represents the adjusted p-value. (D) KEGG enrichment analysis of the overlapping genes. The x-axis indicates the gene ratio, dot size correlates with the gene count, and the color gradient signifies the adjusted p-value Identification of core genes in DBP- and MnBP-induced DKD By constructing 137 predictive models, we were able to further screen seven candidate targets and identify core genes in DBP- and MnBP-associated DKD. To favor models that could be widely applied and to prevent overfitting, we assigned a penalized score to each model. This score was designed to penalize those with significant performance differences between their training and validation sets. We first selected the top ten models based on their penalized scores (Fig. 5A), and then used two external test sets to evaluate the generalization ability of these models. Across all four datasets, the RF + Stepglm [forward] model showed superior performance, with an AUC of at least 0.949, consistently outperforming the first model (Fig. 5B). The core genes identified were DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2. Their single-gene ROC curves demonstrated excellent performance (AUC > 0.8) in the training set (Fig. 5C), while they performed differently in the validation set and two test sets (Figs.S1A-C). The expression changes in DKD glomeruli relative to a normal control group were visualized using a volcano plot (Fig. 5D). By examining the box plots for each dataset, we can see the expression patterns of core genes in the glomeruli of both DKD and control kidneys (Figs.S2A-D). Interestingly, both DUSP1 and FOSB are downregulated in DKD across every dataset. A heatmap further evaluated the generalization ability of each of the six core genes across different datasets. Notably, both DUSP1 and FOSB maintained an AUC of over 0.85 (Fig. 5E). Identification of core genes in DBP- and MnBP-induced DKD. (A) A heatmap compares the AUC values for various machine learning models. Blue represents the training set performance, while green shows the validation set. The top ten models, based on their penalty scores, are highlighted. (B) This heatmap shows the AUC for the top ten models validated across the training set, a validation set (GSE104948), and two test sets (test1 and test2). (C) Single-Gene ROC Curves of six core genes (DUSP1, PTGS2, FOSB, GDF15, NR4A1, CXCR2) within the training set, generated by the RF + Stepglm [forward] model. The x-axis represents 1-specificity and the y-axis represents sensitivity. (D) Volcano plot: volcano plot shows DEGs with key genes labeled. Red = upregulated genes, blue = downregulated genes. (E) This heatmap displays the AUC values for the six key genes, showing their consistent performance across the training set, validation set, and two test sets Single-cell analysis and renal function of core genes in DKD Using the Kidney Interactive Transcriptomics tool and the single-nucleus RNA sequencing dataset GSE131882 (three early human diabetic kidney samples and three controls), we focused our analysis on specific glomerular cells: podocytes (PODO), mesangial cells (MES), and parietal epithelial cells (PEC), with a secondary focus on endothelial cells (ENDO). CXCR2: This gene has low expression in the glomerulus. In DKD, it shows a slight downregulation in PODO and a slight upregulation in ENDO, while expression is almost nonexistent in MES and PEC (Fig. S3A). DUSP1: In DKD, DUSP1 is slightly upregulated in PODO, significantly downregulated in MES and PEC, and shows no significant change in ENDO (Fig. S3B). FOSB: Under normal conditions, FOSB expression is low in the glomerulus. In DKD, it is significantly upregulated in PODO and PEC, slightly downregulated in MES, and slightly upregulated in ENDO (Fig. S3C). GDF15: Expression of GDF15 is low in all glomerulus-related cells (Fig. S3D). NR4A1: In DKD, NR4A1 is significantly upregulated in PODO, MES, PEC, and ENDO (Fig. S3E). PTGS2: This gene has low expression in the glomerulus in both DKD and control conditions, except for an upregulation in MES in DKD (Fig. S3F). These genes have low expression in the Proximal Convoluted Tubule in both DKD and control kidneys. In summary, DUSP1, FOSB, and NR4A1 are more highly expressed in the glomerulus (in both DKD and control conditions) and show significant changes in DKD, indicating their potential as biomarkers. We then investigated the relationship between these genes’ expression and eGFR, a key indicator of kidney function. We used glomerular data from the “Woroniecka Diabetes Glom” dataset and tubular data from the “Schmid Diabetes TubInt” or “Woroniecka Diabetes TubInt” datasets. CXCR2: This gene showed no correlation with eGFR in the glomerulus (p = 0.559, r = 0.132). However, a negative correlation was observed in the renal tubules (p = 0.007, r = − 0.556) (Fig. S3G). DUSP1: A positive correlation was found between DUSP1 and eGFR in both the glomerulus (p = 0.007, r = 0.561) and the tubules (p = 0.013, r = 0.521) (Fig. S3H). FOSB: FOSB also showed a positive correlation with eGFR in both the glomerulus (p = 0.021, r = 0.489) and the tubules (p = 5.85 × 10 − 4, r = 0.521) (Fig. S3I). GDF15: No correlation with eGFR was found for GDF15 in either the glomerulus (p = 0.282, r = − 0.241) or the tubules (p = 0.569, r = 0.115) (Fig. S3J). NR4A1: While there was no correlation with eGFR in the glomerulus (p = 0.170, r = 0.303), a positive correlation was observed in the tubules (p = 0.014, r = 0.515) (Fig. S3K). PTGS2: This gene did not show a correlation with eGFR in either the glomerulus (p = 0.899, r = − 0.029) or the tubules (p = 0.592, r = − 0.121) (Fig. S3L). Based on our analysis, DUSP1 and FOSB stood out for their remarkable association with eGFR. Their expression levels showed a strong correlation with renal function in both the glomerulus and tubules, highlighting their potential as key indicators for evaluating kidney health. Construction of a nomogram model founded on core genes To improve clinical diagnosis, we built a nomogram model using our core genes. The model’s predictions were highly accurate, as shown by its calibration curve which perfectly aligned with ideal results (Fig. S4A). When creating the model with our training data (Fig. S4B), we found that DUSP1 had the greatest range, suggesting it has the strongest influence on predicting disease risk. To confirm its reliability, we tested the model on various datasets and found its AUC consistently exceeded 0.95 (Figs. S4C-D). This confirms the model’s precision and strong ability to generalize. SHAP-based model interpretation Among the five machine learning models, the GBM model stood out with the highest AUC value, earning its selection as the best model for analysis (Fig. S5A). SHAP analysis applied to GBM models effectively identifies key predictors, such as visceral adipose tissue in metabolic dysfunction-associated fatty liver disease [62] or monocyte-to-lymphocyte ratio in chronic obstructive pulmonary disease exacerbation [63], but the biological interpretation often lacks direct citation of supporting studies on underlying mechanisms; strengthening this link requires integrating references from existing literature, such as known roles of VAT in insulin resistance or MLR in systemic inflammation, to validate SHAP findings with empirical biological evidence and enhance clinical relevance. A SHAP-based interpretation showed that just three top features contributed nearly 80% of the cumulative SHAP value (Fig. S5B). DUSP1 was identified as the main driver of the model, having the greatest mean |SHAP value|, while NR4A1 contributed the least (Fig. 6A). Further analysis revealed that DUSP1, CXCR2, FOSB, and GDF15 had a negative effect on predictions, whereas NR4A1 and PTGS2 had a bidirectional effect (Fig. 6B). Dependence plots also showed a positive correlation between CXCR2 and DUSP1, and FOSB and DUSP1, suggesting potential synergy (Fig. 6C). A force-directed analysis further confirmed that DUSP1 (8.79, Δ = −0.34) and CXCR2 (3.97, Δ = −0.16) served as primary negative regulators, driving prediction values (f(x) = 0.0222) below the benchmark expectation (E[f(x)] = 0.642) (Fig. 6D). SHAP-based model interpretation in the training set. (A) SHAP Summary Plot illustrating the mean |SHAP value| for core genes. (B) The beeswarm plot displaying the distribution of gene expression, with a color gradient indicating the expression levels. (C) SHAP dependence plots showing the relationship between core genes and their SHAP value. Color gradient = expression of CXCR2, DUSP1 and NR4A1. (D) SHAP Force Plot demonstrating how core genes contribute to the prediction of a single sample When the model was applied to independent datasets (the validation set and two test sets), DUSP1 and CXCR2 were consistently identified as major contributors. (Figs. S5C, F, I). DUSP1, CXCR2, and FOSB consistently showed a negative effect on predictions across all datasets, with the same positive correlations holding true (Figs. S5D-E, G-H, J-K). However, some genes like GDF15, PTGS2, and NR4A1 had very low feature values in these datasets, which may be due to the small sample size and warrants further study (Figs. S5C-K). Elevated macrophage M2 cell infiltration and decreased neutrophil infiltration correlated with core genes The analysis, applying the CIBERSORT method, characterized the immune cell profiles in the DKD samples compared to the control group, focusing on their relationship with core genes. A correlation heatmap for 22 types of immune cells revealed several significant relationships (Fig. 7A). Strong positive correlations (R > 0.5) were observed between Naive CD4 + T cells and Activated Dendritic cells, as well as between Activated CD4 + Memory T cells and Activated Dendritic cells. Significant negative correlations (∣R∣>0.5) were noted among several pairs, including Activated Mast cells and Resting Mast cells, Naïve B cells and Memory B cells, and CD8 + T cells and Resting CD4 + Memory T cells. A boxplot analysis illustrating the differences in immune cell infiltration between the experimental (DKD) and control groups (Fig. 7B) highlighted key shifts. The DKD group exhibited a marked increase in M2 Macrophages and Resting Mast cells. Conversely, the DKD patients showed a significant decrease in Monocytes, Activated Mast cells, and Neutrophils. By analyzing the correlation of immune cells, we determined that the core genes may be involved in the progression of DKD by controlling immune cells such as T cells regulatory, monocytes, mast cells, neutrophils, and macrophages (Figs. 7C-H). These findings were further verified in the validation set and the test sets (Figs S6). Among these, CXCR2, DUSP1, NR4A1, and PTGS2 were all associated with immune cell infiltration across all datasets. The findings highlight a specific immune cell imbalance in DKD: there is an increase in M2 macrophages alongside a decrease in neutrophils. Elevated macrophage M2 cell infiltration and decreased neutrophil infiltration correlated with core genes. (A) Correlation analysis of important immune cells in the training set. (B) Box plot displaying comparison of 22 key immune-infiltrating cells in the training set. (C-H) Lollipop plots indicating the correlation between six core genes and immune infiltrating cells in the training set. Statistical significance is denoted as follows: ns: p-value ≥ 0.05.*: p-value < 0.05.**: The p-value < 0.01.***: The p-value < 0.001.****: The p-value < 0.0001 Screening for potential FDA-approved drugs Our comprehensive analysis (single-gene ROC curves, single-cell data, renal function correlation, Nomogram modeling, and SHAP interpretations) identified DUSP1 and FOSB as the two most critical genes from the original six. Among them, DUSP1 was prioritized for drug screening due to its stronger association with renal function impairment (evidenced by SHAP values and single-cell expression in renal injury-related clusters) and higher predictive performance in ROC analysis — making it a more actionable therapeutic target. To identify potential FDA-approved drugs, we followed the workflow in Fig. 8: first, we imported both the two key genes (DUSP1/FOSB) and the initial six genes into Enrichr’s DSigDB library to retrieve associated drugs. We then intersected these drug lists with FDA-approved agents (from Table S1) to generate a preliminary candidate set. Given DUSP1’s critical role in our functional and predictive analyses, we refined the candidates by filtering for drugs that target DUSP1 (a step aligned with our focus on this high-priority gene). Finally, we excluded drugs with reported nephrotoxicity or antifungal properties (to avoid renal safety risks), yielding two final candidates: allopurinol and retinoic acid (all-trans-retinoic acid, per DSigDB annotation) (Fig. S7). Flowchart for screening potential FDA-approved drugs Molecular docking of DBP, MnBP and potential drugs with core genes To understand how DBP and MnBP might interact with core genes, we used CB-Dock2 to perform molecular docking and calculate their binding energy. Our results (Fig. 9A) showed that both DBP and MnBP generally had a good affinity for the genes, with most binding energies below the favorable threshold of -5.0 kcal/mol. Notably, DUSP1 and FOSB showed strong binding with both chemicals. The visualized docking results (Figs. 9B-M) showed the specific binding sites and pockets, suggesting that DBP and MnBP could influence DKD by targeting these genes. Molecular Docking of DBP/MnBP with Targets. (A) This heatmap displays the minimum binding energy following molecular docking. The strength of the binding is represented by the color gradient and the numbers found inside the bubbles. (B) CXCR2-DBP visualized via CB-Dock2 molecular docking. (C) CXCR2-MnBP (D) DUSP1-DBP (E) DUSP1-MnBP (F) FOSB-DBP (G) FOSB-MnBP (H) GDF15-DBP (I) GDF15-MnBP (J) NR4A1-DBP (K) NR4A1-MnBP (L) PTGS-DBP (M) PTGS-MnBP We then tested our potential drug candidates, allopurinol and retinoic acid. Retinoic acid demonstrated a powerful ability to bind to all core genes, while allopurinol showed a more moderate binding affinity for a subset of the genes (DUSP1, PTGS2, NR4A1 and CXCR2) (Fig. S8A). The visualization of these interactions (Figs. S8B-K) suggests that these drugs could serve as potential treatments for DKD by targeting these genes, offering new strategies for patients exposed to DBP/MnBP. MD of the DBP/MnBP-target complexes Based on preliminary analyses—including ROC curve evaluation, online database screening, and molecular docking—we selected FOSB, DUSP1 (1–367), and PTGS2 as representative target proteins, and MD were subsequently performed with DBP and MnBP to comprehensively assess the binding stability and conformational dynamics of the resulting protein-ligand complexes: RMSD (which quantifies the overall structural displacement of the ligand, protein, and complex over time) showed that the FOSB/MnBP system exhibited substantial global conformational rearrangement, with its RMSD gradually increasing in the initial 10 ns before stabilizing between 0.5 and 1.0 nm (Fig. 10D), while the remaining complexes (DUSP1/DBP, DUSP1/MnBP, FOSB/DBP, PTGS2/DBP, PTGS2/MnBP) maintained superior structural integrity from the early stages of the simulation, with RMSD values consistently remaining below 0.45 nm (Figs. 10A–C, E–F); though elevated RMSD may indicate increased conformational flexibility that could facilitate adaptive binding, the preservation of key intermolecular interactions (such as hydrophobic contacts and hydrogen bonds) throughout the simulation supports the reliability of the initial binding pose for inferring functional binding behavior. Rg (a metric reflecting the structural compactness and stability of the complex) showed that all six systems displayed favorable stability throughout the 100 ns simulation: for FOSB/MnBP, Rg fluctuations were slightly more pronounced in the first 25 ns but stabilized thereafter, while the other complexes exhibited Rg variations constrained within 0.2 nm (Figs. 10A–F, middle panels), a consistency that confirms the complexes retained their folded, compact architectures without significant unfolding or aggregation. SASA (which reflects the exposure of the complex to solvent) remained stable with minimal fluctuations across the entire simulation duration for all complexes (Figs. 10A–F, third panels); this indicates that the protein-ligand interfaces were sufficiently shielded from the solvent environment, and the systems successfully maintained structural integrity without excessive exposure of hydrophobic regions that could drive non-specific interactions or instability. RMSF (which characterizes the flexibility of individual protein residues during the simulation) showed that the vast majority of residues in all complexes exhibited RMSF values below 0.3 nm (Figs. 10A–F, fourth panels), a pattern indicative of low residue mobility and overall structural rigidity; notably, key interacting residues located at the protein-ligand binding interface showed a marked reduction in RMSF following ligand binding (Figs. S9A, D, G, J, M, P), a direct observation that demonstrates the stabilizing effect of DBP and MnBP on the protein’s binding pocket—constraining the flexibility of residues critical for intermolecular recognition and reinforcing the specificity of the interaction. Hydrogen bond (H-bond) analysis (which reflects one of the primary intermolecular forces maintaining complex stability) showed that the total number of H-bonds fluctuated between 0 and 7 across all systems (Figs. 10A–F, bottom panels); the DUSP1/DBP complex displayed the lowest and most erratic H-bond numbers, with counts frequently dropping to 0 or 1 (suggesting relatively weak and dynamic binding stability), while the FOSB/MnBP system maintained 3 to 5 persistent H-bonds throughout most of the simulation (consistent with its higher overall stability), and PTGS2/DBP and PTGS2/MnBP exhibited moderate H-bond fluctuations that remained sufficient to support the integrity of the binding interface. Free Energy Landscape (FEL) analysis, plotted against the first two principal components (PC1 and PC2) to identify the system’s stable conformational states, via 2D FEL (Figs. S9B, E, H, K, N, Q) confirmed that all six complexes reached well-defined thermodynamic steady states, represented by deep blue low-energy basins that indicate the dominant conformations adopted by the systems; the 3D FEL (Figs. S9C, F, I, L, O, R) further resolved the topology of these energy basins and the presence of metastable states: the DUSP1/MnBP complex exhibited multiple distinct dominant steady-state conformations (Fig. S9F), a feature that indicates higher conformational flexibility and diversity which may enable it to adapt to different binding microenvironments, while the remaining complexes (including FOSB/DBP, PTGS2/DBP, and PTGS2/MnBP) were characterized by a single, well-confined dominant low-energy basin (Figs. S9C, I, O, R), a pattern that implies greater binding specificity and conformational stability, as the systems are less likely to sample non-functional conformations. Molecular dynamics simulation of the DBP/MnBP-target complexes. (A) The RMSD curve, Rg curve, SASA curve, RMSF curve, and hydrogen bond number curve of DUSP1-DBP; (B) The RMSD curve, Rg curve, SASA curve, RMSF curve, and hydrogen bond number curve of DUSP1-MnBP; (C) The RMSD curve, Rg curve, SASA curve, RMSF curve, and hydrogen bond number curve of FOSB-DBP; (D) The RMSD curve, Rg curve, SASA curve, RMSF curve, and hydrogen bond number curve of FOSB-MnBP; (E) The RMSD curve, Rg curve, SASA curve, and RMSF curve of PTSG2-DBP; (E) The RMSD curve, Rg curve, SASA curve, and RMSF curve of PTSG2-MnBP QRT-PCR AND WB To experimentally validate the transcriptional and translational downregulation of key genes in response to phthalate exposure, we performed qRT-PCR and WB analyses on HK2 cells treated with DBP and MnBP. The qRT-PCR results demonstrated that the mRNA expression levels of DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2 were all significantly lower in both the DBP_HK2 and MnBP_HK2 groups compared to the control HK2 group. Consistent with the mRNA data, WB analysis confirmed that the protein levels of these six genes were also markedly reduced following DBP and MnBP treatment (Figs. 11A-B). . The expression levels of six genes. (A) qRT-PCR expression results of six genes. (B) WB expression results of six genes. *: p-value < 0.05.**: The p-value < 0.01.***: The p-value < 0.001 Discussion DKD, a primary driver of ESKD, has a range of effective therapies, from traditional RAS blockers to modern SGLT2 inhibitors, NS-MRAs, and GLP-1 receptor agonists [1, 64]. While these drugs are pivotal in slowing disease progression, they cannot fully halt it, a limitation stemming from the incomplete understanding of DKD’s pathogenesis. This gap in knowledge has prompted a shift in focus toward the role of environmental toxins like DBP/MnBP [3, 4], which could provide new avenues for prevention and targeted treatment. Our study utilized an integrated bioinformatics framework, combining data from training, validation, and test sets with machine learning and molecular docking, to systematically identify potential targets linking DBP/MnBP exposure to DKD. The toxicity profiles of DBP and MnBP were assessed using two predictive platforms, ADMETlab 3.0 and ProTox 3.0, which yielded divergent predictions regarding nephrotoxicity.The observed variation primarily stems from fundamental methodological differences between these platforms, as ADMETlab 3.0 utilizes comprehensive machine learning algorithms trained on diverse chemical descriptors and biological endpoints to generate multi-parameter ADMET profiles, while ProTox 3.0 relies predominantly on QSAR-based similarity mappings to known toxic compounds and specific protein targets, which may yield divergent risk assessments for novel structures. Furthermore, compound-specific factors significantly contribute to this inconsistency, as our compound’s unique chemical features may reside near the applicability domain boundaries of one model but not the other, leading to varying prediction reliabilities; for instance, ProTox 3.0’s dependency on structural alerts and toxicophores could render it more sensitive to specific functional groups, whereas ADMETlab 3.0’s integrative approach may capture broader physicochemical properties and metabolic pathways. Importantly, such discrepancies are not uncommon in computational toxicology, as evidenced by studies evaluating diverse compounds where prediction tools like ProTox 3.0 and ADMETlab 3.0 have shown varying outcomes due to their distinct training sets and algorithmic frameworks. Initial database predictions of DBP/MnBP’s renal toxicity, followed by network toxicology analysis, helped us pinpoint seven candidate genes. Of these, six core genes—DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2—were found to be particularly significant. Prior studies have established biomarkers such as tumor necrosis factor receptors (TNFR1 and TNFR2) and kidney injury molecule-1 for predicting DKD progression [65], as well as insulin-like growth factor-1 and interleukin-6 for diagnostic purposes in DKD [66]. These biomarkers are linked to inflammation, endothelial dysfunction, and oxidative stress pathways commonly implicated in DKD [67]. In contrast, our study identifies six core genes—DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2—which have not been prominently reported as DKD biomarkers in the existing literature. For example, while inflammatory pathways involving NLRP3 inflammasome activation and cytokines like IL-1β and IL-18 are known in DKD [68], genes such as DUSP1 and FOSB, which regulate stress responses and inflammation, represent novel candidates. This underscores the novelty of our findings, as these genes may offer new insights into DKD mechanisms through pathways distinct from traditionally reported biomarkers, thereby enhancing the biological significance of our results in advancing DKD diagnosis and therapeutic strategies. In addition, Our study identified and validated a core set of genes associated with DKD progression in the context of nephrotoxicant exposure; however, it is important to acknowledge its limitations regarding potential confounding factors. The clinical dataset employed for the initial bioinformatics analysis lacked detailed patient-level information on key variables such as age, diabetes duration, medication use, BMI, and smoking history. Consequently, we were unable to statistically adjust for these potential confounders in our association analysis, and their influence on the observed gene-expression patterns cannot be entirely ruled out. Nevertheless, the primary mechanistic conclusions of our study are robustly supported by the subsequent experimental validation conducted in a controlled in vitro HK-2 cell model. This model system, by its design, is inherently free from the complex interplay of human clinical variables. The consistent and significant downregulation of DUSP1, PTGS2, FOSB, GDF15, NR4A1, and CXCR2 in response to DBP and MnBP exposure in these cells strongly indicates that their altered expression is a direct consequence of the toxicant insult and is intimately linked to the underlying pathophysiology of renal injury, rather than being a mere reflection of demographic or clinical heterogeneity. Future studies incorporating well-annotated clinical cohorts with comprehensive patient metadata will be valuable to further corroborate these findings in a human context and elucidate the interplay between environmental exposures and individual patient factors. Differential expression analysis revealed that these genes were consistently downregulated in DKD samples, with the decrease in DUSP1 and FOSB being especially prominent. The high diagnostic power of these six genes was confirmed by machine learning models. Notably, DUSP1 and FOSB demonstrated strong generalizability in single-gene ROC analysis, leading to the creation of a nomogram to aid clinical diagnosis. Our research also utilized multiple databases—including Nephroseq 5.0, the Kidney Interactive Transcriptome Platform, Dsigdb, and Enrichr—for in-depth characterization of these core genes, leading to the identification of two potential drugs: allopurinol and retinoic acid. Furthermore, SHAP interpretability analysis consistently highlighted DUSP1 as the most influential predictive factor across the datasets. Immune infiltration analysis indicated that core genes might be involved in the occurrence and progression of diabetic nephropathy by regulating the infiltration level of immune cells. The biological plausibility of our findings was underscored by molecular docking results, which revealed a strong binding affinity between DBP/MnBP and the protein products of these genes. To further validate the stability and functional relevance of these interactions, we performed 100 ns molecular dynamics (MD) simulations on the key protein-ligand complexes (DUSP1/DBP, DUSP1/MnBP, FOSB/DBP, FOSB/MnBP, PTGS2/DBP, PTGS2/MnBP). RMSD analysis showed that most complexes maintained structural integrity with values consistently below 0.45 nm, except FOSB/MnBP which stabilized at 0.5–1.0 nm, indicating adaptive conformational flexibility without compromising binding reliability. Rg and SASA measurements confirmed the compactness and solvent shielding of all complexes, while RMSF analysis revealed reduced flexibility at key binding interface residues—directly demonstrating the stabilizing effect of DBP/MnBP on the protein active sites. Hydrogen bond analysis further supported these interactions: FOSB/MnBP maintained 3–5 persistent hydrogen bonds, reflecting robust binding stability, whereas DUSP1/DBP showed more erratic H-bond fluctuations. FEL analysis validated thermodynamic steady states for all complexes, with most systems exhibiting a single dominant low-energy conformation (indicating high binding specificity) except DUSP1/MnBP, which displayed multiple stable states suggestive of conformational adaptability. Collectively, these MD results confirm the structural reliability of the predicted binding modes and provide dynamic insights into the intermolecular forces governing DBP/MnBP interactions with core target proteins, reinforcing their role as key molecular mediators of DKD pathogenesis. In summary, this study suggests the identified core genes are likely key molecular targets in the onset and progression of DKD induced by DBP/MnBP. The core genes identified in this study play multifaceted roles in the development and progression of DBP/MnBP-induced DKD. Based on the molecular docking and dynamics simulations presented in our results, we further discuss the specific molecular interactions between DBP/MnBP and the identified target proteins DUSP1 and FOSB. Our simulations demonstrated stable binding affinities, particularly through hydrophobic interactions and hydrogen bonding within the catalytic pockets of these proteins.These findings align with existing literature on phthalate-protein interactions. For instance, studies have shown that phthalate metabolites like MnBP can modulate gene expression in placental cells, including genes involved in hormonal regulation and inflammation pathways [69]. Specifically, MnBP exposure has been associated with altered expression of PPARG and CYP19A1, suggesting that phthalates may interact with nuclear receptors or transcription factors to disrupt cellular signaling [70]. Additionally, phthalate-induced oxidative stress has been linked to inflammatory responses, potentially through pathways involving DUSP1, which regulates MAP kinase activity, and FOSB, a component of the AP-1 transcription complex [71].The observed interactions in our study may provide a mechanistic basis for these earlier epidemiological and cellular findings. For example, the stable binding of MnBP to DUSP1 could impair its phosphatase activity, leading to dysregulated stress responses, while interactions with FOSB might alter transcriptional programs related to inflammation and cell proliferation.However, it is important to note that the precise residues involved in these interactions require further validation through mutagenesis studies or structural analyses. Future research should also explore whether these interactions are consistent across different cell types and exposure scenarios, as prior studies have indicated sex-specific differences in phthalate effects. Overexpression of DUSP1 (Dual specificity protein phosphatase 1) has been shown to attenuate renal fibrosis in DKD by inactivating the p38 MAPK and ERK1/2 pathways [72]. It also helps restore mitochondrial function by promoting Parkin-mediated mitophagy [73]. Interestingly, other studies suggest that inhibiting DUSP1 can activate autophagy to protect against podocyte injury [74], indicating its function in DKD may be highly dependent on the cellular context and disease stage. According to current scholarship, Traditional Chinese Medicine effectively prevents the progressive deterioration of diabetic kidney disease, mediating through the regulation of PTGS2 (Prostaglandin G/H synthase 2) and the inhibition of ferroptosis [75].Patients with DKD exhibit dysregulated levels of ferroptosis-related molecules, including PTGS2 [76]. Studies in mesangial cells show that suppressing HMGB1 reduces pro-ferroptotic markers like PTGS2. Previous research has demonstrated that inflammatory pathways play crucial roles in renal pathology1, and environmental exposures may contribute to disease progression through various molecular mechanisms [77]. Additionally, studies on pharmacological interventions have shown that modulation of specific pathways can significantly impact disease outcomes in elderly populations with cardiovascular complications [78]. While our current findings highlight the strong binding interactions, we agree that deeper analysis of PTGS2’s role in DKD pathogenesis would strengthen our discussion. We will expand our results and discussion sections to include more detailed analysis of how PTGS2-mediated pathways might contribute to renal inflammation and fibrosis in DKD, particularly focusing on the potential synergistic effects of environmental exposures and metabolic factors. This additional analysis will provide greater context for our binding affinity results and their potential clinical implications. Research suggests that Traditional Chinese Medicine and drugs like the GLP-1 receptor agonist dulaglutide can prevent the deterioration of DKD by modulating PTGS2 and inhibiting ferroptosis [79]. In Human Umbilical Vein Endothelial Cells, high glucose and hypoxia promote inflammation and oxidative stress through the HIF-1α/JMJD1A pathway. FOSB may act as a protective transcription factor that negatively regulates this process [80], potentially offering a novel therapeutic target for DKD. Elevated urinary GDF15 levels are a sensitive prognostic marker for renal decline in DKD patients [81, 82]. Overexpression of GDF15 in mouse models has been shown to significantly suppress renal inflammation, fibrosis, and damage by targeting the AGE/RAGE/TLR4/NF-κB pathway [83]. GDF15, which is often upregulated in DKD, protects the kidney by inhibiting the NEDD4L-IKK/NF-κB pathway [84], while Periostin promotes renal inflammation by suppressing GDF15 expression [85]. NR4A1 drives the progression of diabetic nephropathy through dual mechanisms: inducing Mff-dependent mitochondrial fission and inhibiting Parkin-mediated mitophagy [86]. The deficiency of CXCR2 leads to the inhibition of NF-κB signaling activation, which in turn regulates inflammation and improves DKD outcomes by restoring the endothelial glycocalyx [87]. The IL-8-CXCR1/2 axis may contribute to podocyte damage in DKD [88], while G31P, a CXCL8 antagonist, improves diabetic nephropathy in male diabetic mice and attenuates mesangial injury by inhibiting the activation of JAK2/STAT3 and ERK1/2 [89]. Some small-scale studies have suggested that low-dose allopurinol may provide a protective effect on the kidneys. These studies found that the drug could reduce proteinuria [90], decrease urinary albumin excretion rate, and increase GFR [91], possibly by lowering serum uric acid. Allopurinol also appeared to protect human glomerular endothelial cells from damage caused by high glucose [92]. Larger clinical studies have questioned the therapeutic benefit of allopurinol for kidney disease. A double-blind trial, for instance, found that reducing serum urate with the drug provided no clinically meaningful benefit for kidney outcomes in patients with Type 1 diabetes and early-to-moderate diabetic kidney disease [93]. Similarly, a randomized controlled trial revealed that allopurinol did not slow the decline in eGFR among patients with chronic kidney disease at high risk of progression [94]. These findings were reinforced by a systematic review and meta-analysis, which concluded that allopurinol did not significantly improve renal function or albuminuria in patients with DKD [95]. In contrast, all-trans retinoic acid (RA) shows promise. Acting as an endogenous agonist for RA receptors (RARs) α, β, and γ [96], it works to prevent fibrogenesis in diabetic nephropathy by restoring RAR-α expression and inhibiting the JNK/RAR-α and TGF-β1/Smad3 signaling pathways [97]. Furthermore, RA helps reduce inflammation mediated by the TLR4/NF-κB pathway during diabetic nephropathy [98]. They may potentially prevent DKD induced by DBP and MnBP exposure or serve as potential therapeutic agents for DKD. Despite its valuable findings, our research faces several key limitations. Even with a sizable dataset from six GEO datasets and a rigorous division into training, validation, and test sets, our results warrant further confirmation in larger patient cohorts and through extensive experimental validation. A significant weakness is the lack of information on crucial confounding factors like age, genetics, and lifestyle within the public GEO datasets, variables that could potentially influence our findings. Furthermore, while molecular docking offered support for potential targets and drugs, the specific toxic effects of DBP/MnBP on these genes, as well as the therapeutic actions of the identified drugs, still require direct verification via both in vivo and in vitro experiments. We acknowledge the limitations in our screening strategy, which initially focused on target affinity without pre-filtering for clinical efficacy, as highlighted by the negative evidence from large RCTs like the ALL-HEART study [99] and inconclusive cardiovascular outcomes [100]. However, allopurinol has shown potential in other contexts, such as efficacy in canine visceral leishmaniasis [101], suggesting broader therapeutic applications. We will revise the discussion to strengthen the rationale for allopurinol’s inclusion by balancing this mixed evidence or consider focusing more on retinoic acid as the primary candidate.A promising avenue for future research would be to specifically investigate populations with high DBP/MnBP exposure to improve the study’s reliability and generalizability. Conclusion This study suggests that DBP and MnBP may contribute to DKD by targeting core genes (DUSP1, PTGS2, FOSB, GDF15, NR4A1 and CXCR2). Molecular docking simulations confirmed robust binding affinity between these chemicals and the identified target proteins. These findings provide important insights into the molecular mechanisms of phthalate-induced kidney injury and support the potential reuse of drugs such as allopurinol and retinoic acid for DKD treatment. Further experimental validation and clinical studies are needed to verify these mechanisms and explore targeted intervention strategies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author contributions Wenjie Chen wrote the main manuscript text. Wenjie Chen and Suyue Hou wereinvolved with designing the study and data analysis. Jijia Hu edited and reviewed the manuscript.All authors approved the final manuscript as submitted and agreeto be accountable for all aspects of the work. Funding The present study was supported by the National Natural Science Foundation of China 82300767 to Jijia Hu. Data availability The data that support the findings of this study are available from the authors. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data that support the findings of this study are available from the authors.