We utilized occupational health monitoring data from 5482 workers in a large industrial enterprise in Anhui Province, China, to construct a comprehensive analysis framework for studying the relationship between complex occupational exposure and CEA status. All variables recorded in the occupational health monitoring system are included as candidate features, thereby minimizing prior selection bias to the greatest extent possible. These variables encompass hematological parameters, composite inflammatory indices, and a wide range of occupational exposure indicators, enabling multi-level characterization of host physiology and environmental burden (19).

Hematological measurements capture the global hematopoietic and immune status, including platelet, red blood cell, and white blood cell indices, as well as the relative composition of major white blood cell subgroups. To further summarize systemic immune activation, derived inflammatory markers (NLR, MLR, and SII) were integrated to reflect chronic low-grade inflammation and immune imbalance.

Occupational exposure data covers a highly heterogeneous spectrum of hazards typical of industrial environments. Exposure to particulate matter includes coal dust, crystalline silica, and various inorganic and organic dust, reflecting long-term inhalation risks associated with mining, building materials, and manufacturing processes. Exposure to gaseous and inorganic chemicals includes sulfur and nitrogen oxides, carbon monoxide and carbon dioxide, hydrogen sulfide, ammonia, methane, ozone, chlorine, and related compounds, representing combustion related emissions and confined space exposure. Organic solvents and industrial chemicals encompass aromatic hydrocarbons, halogenated compounds, alcohols, ethers, esters, and other reactive intermediates commonly found in chemical production and material processing. In addition, the dataset also records occupational exposure to corrosive substances, highly toxic substances, and complex industrial mixtures such as coke oven emissions and coal tar. In addition to chemical and particulate hazards, occupational health monitoring systems also systematically record physical stress factors, including noise, extreme heat, arm vibration, ultraviolet radiation, and power frequency electric fields, which constitute sustained non chemical stress factors that can disrupt physiological homeostasis.

A total of 14 algorithms were implemented, covering different method families, including tree based ensemble models (random forest, gradient enhancement, adaptive enhancement, LightGBM, CatBoost), linear and regression based methods (logistic regression, LASSO, partial least squares), kernel and distance based methods (support vector machine with kernel function, k-nearest neighbor), probability and discriminant models (naive Bayes, discriminant analysis), and neural networks (multilayer perceptron). The dataset is randomly divided into a training set (70%) and a testing set (30%). Optimize hyperparameters through 5-fold cross validation using Bayesian optimization (Hyperopt). Evaluate model performance through sensitivity, specificity, and area under the operating characteristic curve (AUC) of the subjects (20, 21). To evaluate robustness, 1000 guided resampling iterations were conducted to obtain the distribution of sensitivity and AUC, and to compare model performance.

Prior to model training, we addressed several key data processing considerations to ensure robustness and fairness. First, given the moderate imbalance between CEA-positive and CEA-negative cases, we applied class weighting schemes for algorithms that support weighted loss functions (e.g., CatBoost and logistic regression), and used stratified sampling throughout training and evaluation steps to maintain representative class distributions.

Second, we evaluated missingness across all features. Variables with less than 5% missing values were imputed using the mean for continuous variables and the mode for categorical variables, while features with greater than 20% missingness were excluded from modeling.

Lastly, during the 1000 iterations of guided resampling, we implemented stratified resampling to preserve the original CEA class proportions across both training and test subsets. This approach minimized bias introduced by artificial class distributions and supported reliable performance comparison across models.

To enhance interpretability, we applied SHAP to quantify the marginal contribution of each feature and characterize potential non-linear exposure–response relationships. Model interpretability was visualized through SHAP summary plots (including beeswarm and bar plots to display global feature importance and direction of effects) and dependence plots (to illustrate non-linear patterns and potential feature interactions).

Human monocytic THP-1 cells were cultured in RPMI-1640 medium (Gibco, Thermo Fisher Scientific) supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin–streptomycin (100 U/mL penicillin and 100 μg/mL streptomycin; Gibco). Human colorectal cancer (CRC) cell lines HT29 and Caco-2 were maintained in high-glucose DMEM (Gibco) containing 10% FBS and 1% penicillin–streptomycin. All cells were incubated at 37 °C in a humidified atmosphere with 5% CO₂. Specifically, the complete culture media composition was described for each cell line, including base medium (RPMI-1640 or high-glucose DMEM), 10% fetal bovine serum (FBS), and 1% penicillin–streptomycin. THP-1 cells were seeded at 5 × 10⁵ cells per well in 24-well plates, while HT29 and Caco-2 cells were seeded at 1.5 × 10⁵ cells per well. After a 12-hour equilibration period, cells were exposed to SiO₂ for indicated durations: 6 hours for RNA extraction, and 24 hours for cytokine detection. For NF-κB inhibition, cells were pretreated with BAY 11-7082 (5 μM) for 1 hour prior to exposure, and the inhibitor was maintained throughout the stimulation period. Conditioned media (CM) were collected by sequential centrifugation (300 × g, then 2,000 × g), filtered through 0.22 μm membranes, and stored at −80 °C. For stimulation assays, CRC cells were treated with a 1:1 mixture of CM and fresh complete medium to ensure consistent serum and nutrient conditions. In this study, we also used 20–50 nm amorphous SiO₂ nanoparticles as model particles to induce inflammatory responses; future studies will include respirable crystalline silica to better simulate occupational exposure conditions.

Amorphous silica (SiO₂) nanopowder (Sigma-Aldrich, 637238; nominal particle size 20–50 nm) was suspended in sterile phosphate-buffered saline (PBS) to prepare a 10 mg/mL stock solution. Immediately prior to each exposure, the suspension was sonicated in an ice-water bath for 10 min to minimize particle agglomeration and ensure homogeneous dispersion.

NF-κB signaling was inhibited using BAY 11-7082 (Sigma-Aldrich, B5556). BAY 11–7082 was dissolved in DMSO to generate a 10 mM stock solution and applied at a final concentration of 5 μM. Cells were pretreated with BAY 11–7082 for 1 h before SiO₂ exposure, and the inhibitor was maintained throughout the exposure period.

For IL-6 neutralization experiments, a monoclonal anti-human IL-6 antibody (R&D Systems, MAB206) was added at a final concentration of 2 μg/mL at the onset of stimulation.

THP-1 cells were seeded in 24-well plates at a density of 5 × 10⁵ cells per well in 500 μL complete RPMI-1640 medium and allowed to equilibrate for 12 h. Cells were then exposed to vehicle or increasing concentrations of SiO₂ (25, 50, or 100 μg/mL). To evaluate the contribution of NF-κB signaling, cells exposed to the highest SiO₂ concentration were additionally co-treated with BAY 11-7082, and BAY 11–7082 alone was included as an inhibitor control. The selected SiO₂ concentrations were determined in pilot experiments to establish a robust inflammatory response while preserving cell viability.

At 6 h after exposure, THP-1 cells were harvested for RNA isolation. Total RNA was extracted using TRIzol reagent (Invitrogen, Thermo Fisher Scientific) and quantified by NanoDrop spectrophotometry. Complementary DNA was synthesized using the PrimeScript RT reagent kit (Takara, RR037A). Quantitative PCR was performed using PowerUp SYBR Green Master Mix (Applied Biosystems, Thermo Fisher Scientific) on a QuantStudio 5 Real-Time PCR system. Cycling conditions consisted of an initial denaturation at 95 °C for 2 min, followed by 40 cycles of 95 °C for 15 s and 60 °C for 60 s, with subsequent melt-curve analysis. Expression levels of IL6, TNFA, and IL1B were normalized to GAPDH and calculated using the 2^-ΔΔCt method.

Culture supernatants were collected 24 h after exposure, centrifuged at 2,000 × g for 10 min to remove cellular debris, and stored at −80 °C. IL-6 concentrations were quantified using a human IL-6 ELISA kit (R&D Systems, DY206) according to the manufacturer’s instructions. At this stage, only IL-6 was quantified at the protein level, as it was prioritized for its established role in inflammatory paracrine signaling and tumor-related CEA regulation. Protein-level analysis of IL-1β and TNF-α was not performed, but their mRNA induction was confirmed by RT-qPCR. Future experiments will expand protein validation to include additional cytokines.

Proteins were extracted using RIPA buffer (Cat:R0010, Solarbio) on ice for 30 minutes, then centrifuged at 12,000 rpm for 15 minutes at 4 °C. The supernatant’s protein concentration was measured with a BCA assay (Cat:PC0020, Solarbio). Equal protein amounts (20 μg per lane) were mixed with 5× SDS buffer (Cat:LT101S, Epizyme Biotech), denatured at 100 °C for 10 minutes, and separated by SDS-PAGE using a 5% stacking gel and an 12% separating gel (Cat:P1200, Solarbio), depending on target protein size, at 80 V for 30 minutes and 120 V for 60–90 minutes.

Proteins were transferred to methanol-activated PVDF membranes via wet transfer at 100 V for 60 minutes on ice. Membranes were blocked with 5% non-fat milk in TBST for 1.5 hours at room temperature, then incubated with primary antibodies (Anti-NF-κB p65 Antibody, Cat:A00284-1, Boster) overnight at 4 °C. After three 5-minute TBST washes, membranes were treated with HRP-conjugated secondary antibodies for 1 hour at room temperature. Following another three TBST washes, protein bands were visualized using ECL and quantified with ImageJ, using GAPDH (Anti-GAPDH Antibody, Cat:A00227-1, Boster) as a reference.

HT29 and Caco-2 cells were seeded in 24-well plates at 1.5 × 10⁵ cells per well and allowed to adhere overnight. Cells were subsequently treated with vehicle or SiO₂ at low, medium, or high concentrations as defined above. Cells were harvested at 12 h for analysis of CEA (CEACAM5) mRNA expression, and culture supernatants were collected at 48 h for measurement of secreted CEA protein. CEA mRNA levels were quantified by RT–qPCR using GAPDH as the reference gene, and secreted CEA concentrations were determined using a human CEA ELISA kit (Abcam, ab99992). Samples were diluted when necessary to fall within the linear range of the standard curve and were analyzed in duplicate.

To generate conditioned medium, THP-1 cells were treated with vehicle or increasing concentrations of SiO₂ as described above. After 24 h, culture supernatants were collected and sequentially clarified by centrifugation at 300 × g for 5 min followed by 2,000 × g for 10 min, and then filtered through a 0.22 μm sterile filter to remove residual cells, debris, and particles. Conditioned media were aliquoted and stored at −80 °C. HT29 and Caco-2 cells were stimulated with conditioned media using a fixed ratio of 50% conditioned medium and 50% fresh complete medium (final 10% FBS) to ensure equivalent nutrient and serum conditions across treatment groups. CEA mRNA expression was assessed at 12 h, and secreted CEA protein levels were measured at 48 h.

To establish causality, conditioned media were generated from THP-1 cells exposed to high-dose SiO₂ in the presence or absence of BAY 11-7082, as well as from cells treated with BAY 11–7082 alone. Conditioned media were collected at 24 h and processed identically to those described above. IL-6 concentrations in conditioned media were quantified by ELISA to confirm effective suppression of inflammatory output following NF-κB inhibition.

CRC cells were then stimulated with conditioned media in the presence or absence of anti-IL-6 antibody or isotype control IgG, which were added at the onset of stimulation and maintained throughout incubation. CEA mRNA expression and protein secretion were analyzed at 12 h and 48 h, respectively.

Core genes associated with SiO₂ exposure were retrieved from the Comparative Toxicogenomics Database (CTD), while colorectal cancer (CRC)–related core genes were obtained from the GeneCards database. The overlapping genes were defined as potential molecular mediators linking environmental exposure to tumor pathogenesis.

Transcriptomic profiles and matched clinical annotations for ovarian cancer were obtained from The TCGA and GEO. Only samples with complete follow-up information were included in the analysis. Genes with a prioritization score greater than four from the preceding integrative analysis were selected for clinical correlation assessment. Survival endpoints included overall survival (CRC) in both TCGA and GEO. For each candidate gene, patients were stratified into high- and low-expression groups based on the median expression value within each cohort. Univariate survival analysis was performed using the Kaplan–Meier method with the log-rank test, while multivariable Cox proportional hazards regression models were applied to estimate hazard ratios (HRs) and 95% confidence intervals (CIs), adjusting for relevant clinical covariates. Genes demonstrating significant survival associations were subjected to functional enrichment analysis.

The preprocessing and analysis of the single-cell RNA sequencing dataset GSE161277 were supported by the KeyanZhiJia Single-Cell Analysis Platform.

Data are presented as mean ± standard deviation (SD). Statistical analyses were performed using GraphPad Prism 9.0. One-way analysis of variance (ANOVA) followed by Dunnett’s or Tukey’s post hoc tests was applied as appropriate. Normality and homogeneity of variance were assessed using the Shapiro–Wilk and Levene tests, respectively, and data were log-transformed when necessary. A two-sided P < 0.05 was considered statistically significant.

To systematically evaluate the robustness and applicability of different machine learning paradigms in predicting carcinoembryonic antigen (CEA) positivity, we constructed and compared 14 representative algorithms across major method families. These include tree based ensemble models (random forest, gradient enhancement, adaptive enhancement, LightGBM, and CatBoost), regression based methods (logistic regression, LASSO, and partial least squares), kernel and distance based methods (support vector machines and k-nearest neighbors), probability models (naive Bayes and discriminant analysis), and neural networks (multilayer perceptron). The dataset is randomly divided into training (70%) and testing (30%) subsets, and hyperparameters are optimized using five fold cross validation. The model performance was evaluated in multiple dimensions, including discriminative ability, predictive consistency, and potential clinical utility.

Clear method stratification has emerged in data segmentation and evaluation metrics. Set based models consistently outperform regression, kernel, and distance based methods in both training and testing sets. During the training phase, LightGBM and CatBoost achieved near saturation discriminative performance with AUC values close to 1.00, followed by adaptive enhancement, gradient enhancement, and random forest. In contrast, logistic regression, k-nearest neighbor, and naive Bayes showed significantly weaker discriminative power, highlighting the limited expressive ability of simpler models in capturing complex exposure biomarker relationships (Figures 1A, B). It is worth noting that the performance differences in the test set have become more pronounced. CatBoost maintains excellent sensitivity, specificity, F1 score, and overall accuracy, with minimal degradation relative to training performance, indicating its strong generalization ability and stable classification behavior (Figures 1C, D). These findings indicate that ensemble learning frameworks not only perform well in model fitting, but also have structural robustness in out of sample prediction.

The analysis of receiver operating characteristics further confirms these trends (Figures 2A, B). In the training set, CatBoost and LightGBM have consistently dominated the ROC space, while in the test set, CatBoost exhibits the smallest curve collapse, highlighting its resilience to sample perturbations (Figures 2A, B). However, relying solely on discriminative performance cannot fully capture the reliability of the model. To address this issue, DCA was used to evaluate potential clinical relevance (Figures 2C, D). Within a wide range of threshold probabilities, CatBoost achieved the highest net benefit in both training and testing queues, outperforming alternative algorithms and default strategies (“treat all” and “do not treat”) (Figures 2C, D). This indicates that CatBoost provides excellent decision level value for identifying high-risk individuals with CEA positivity, supporting its translational relevance to occupational risk stratification (Figures 3A, B). The stability of the model was further verified through residual structure analysis. The reverse cumulative residual distribution indicates that CatBoost consistently exhibits a steep left shift curve in both the training and testing sets, reflecting a significant reduction in the proportion of large prediction errors. Correspondingly, the residual boxplot RMSE of CatBoost, while logistic regression, naive Bayes, and partial least squares models show significant long tail error distributions (Figures 3C, D). Overall, these analyses indicate that the model’s advantage depends not only on average performance metrics, but also on controlled and reproducible erroneous behavior.

Among the 14 machine learning algorithms evaluated, CatBoost demonstrated the best overall performance on the test set in terms of AUC, sensitivity, and specificity. In addition to its strong predictive ability, CatBoost is particularly well-suited for handling complex nonlinear relationships and categorical variables without extensive preprocessing. Its high stability and compatibility with SHAP values for feature attribution made it a practical and interpretable choice for this real-world dataset. The feature attribution using the SHAP framework indicates that CEA positivity is mainly driven by hematological and inflammation related variables, with MLR being the most influential predictor (Figures 4A, B). This discovery emphasizes that systemic immune imbalance is the core determinant of CEA abnormal expression. Among all occupational and environmental exposures, the SHAP contribution of silica dust is the highest, significantly higher than that of nitrogen dioxide, coal dust, and carbon monoxide. SHAP distribution and dependency analysis further indicate that the association between silica exposure and CEA positivity is clearly nonlinear, characterized by threshold like effects and significant inter individual heterogeneity at higher exposure levels (Figure 4C). These patterns indicate that silicon dust may act as a risk amplifier rather than producing uniform linear effects.

To elucidate the potential biological pathways behind this structural association, mediation analysis was conducted to evaluate the role of inflammatory markers in linking silica exposure to CEA elevation (Figure 5). Multiple biomarkers including MLR, neutrophil percentage, lymphocyte percentage, NLR, and SII showed significant indirect effects, indicating that the relationship between silica and CEA is partially mediated through the inflammatory pathway. It is worth noting that MLR accounts for the largest proportion in mediating effects, supporting a model in which monocyte dominated inflammatory activation represents a key mechanistic link between occupational silica exposure and elevated CEA levels.

As shown in Figures 6A, B, SiO₂ exposure strongly activates THP-1 monocytes in a dose-dependent manner. The mRNA levels of IL-6, IL-1 β, and TNF - α gradually increased from the low-dose and medium dose groups to the high-dose group, and multiple pairwise comparisons reached statistical significance (asterisk annotation). Importantly, the addition of inhibitors under high-dose conditions significantly weakened the induction of transcription of all three cytokines, restoring expression to the low-dose range. These findings indicate that SiO ₂ - induced monocyte activation is inhibitor sensitive and regulated, rather than reflecting non-specific transcriptional fluctuations. In contrast, direct exposure of CRC cells to SiO ₂ only resulted in minor changes in CEA expression (Figure 6C). In Caco-2 and HT29 cells, CEA mRNA levels remained close to baseline at low, medium, and high SiO2 levels, and the secreted CEA protein only showed slight upward drift, without a clear dose-response relationship or consistent statistical significance. This pattern supports the explanation that CEA is not strongly regulated by SiO ₂ under these conditions, through direct tumor cell autonomous mechanisms.

We examined how SiO₂ exposure affects NF-κB p65 activation by conducting Western blot analysis on cells treated with different SiO₂ concentrations, with and without an NF-κB inhibitor. Supplementary Figure 1 shows that NF-κB protein levels were baseline in the control group. SiO₂ treatment led to a concentration-dependent increase in NF-κB expression, with the highest levels in the SiO₂-High group, followed by the SiO₂-Mild and SiO₂-Low groups, all significantly higher than the control (P < 0.0001). The NF-κB p65 inhibitor significantly reduced NF-κB p65 upregulation in the SiO₂-High+Inhibitor group (P < 0.01), while the inhibitor alone did not affect basal NF-κB p65 levels. The results suggest that exposure to SiO₂ triggers the NF-κB pathway based on concentration, and this activation can be significantly reduced by using an NF-κB inhibitor.

It is noteworthy that CM derived from SiO ₂ - stimulated THP-1 cells induced significant and graded CEA induction in CRC cells (Figure 6D). In Caco-2 and HT29, CEA mRNA gradually increases with the CM produced by low, medium, and high doses of SiO2. The parallel increase of CEA protein secreted by treated monocytes reaches its maximum level under high CM conditions. The consistent stepwise pattern observed at the transcript and protein levels, as well as dense significant annotations, indicate that the soluble factors released by activated monocytes are sufficient to drive strong CEA responses in CRC cells, and the intensity of this response is proportional to the strength of upstream monocyte activation.

Finally, the intervention experiment supported the causal interruption of the paracrine axis (Figure 6E). High CM produced the strongest increase in CEA mRNA and protein in two CRC cell lines, while the isotype IgG control group remained comparable to the high CM group, indicating that antibody addition alone cannot explain the observed effects. In contrast, CM (high+NFkBi CM) produced under NF - κ B inhibition significantly reduced the increase in CEA driven by high CM, and the neutralization of IL-6 during CRC stimulation also weakened the expression of CEA at the mRNA and protein levels. It is worth noting that in the context of CM control, anti-IL-6 therapy is still close to baseline, supporting pathway specificity and opposing the main direct inhibitory effect of antibody therapy under non inflammatory conditions.

Overall, this set of numbers supports a coherent mechanistic model, in which SiO₂ dose-dependent activates THP-1 inflammatory signaling, direct SiO₂ exposure has little effect on CRC CEA, and monocyte derived conditioned medium can effectively induce CEA in CRC cells in a dose-dependent manner. Partial reversal of NF - κ B inhibition (upstream, monocyte side) and IL-6 neutralization (downstream) suggests that this inflammatory paracrine pathway is partially but not exclusively necessary, consistent with the multifactorial monocyte secretion group driving CRC CEA upregulation.

SiO₂-associated genes were retrieved from the Comparative Toxicogenomics Database (CTD), while CRC–related genes were obtained from the GeneCards database. An intersection analysis identified a total of 42 overlapping genes that were simultaneously associated with SiO₂ exposure and CRC (Figure 7A), suggesting that these genes may serve as potential molecular bridges linking environmental exposure to tumorigenesis. To further characterize the functional relationships among the overlapping genes, a protein–protein interaction (PPI) network was constructed (Figure 7B). Network analysis revealed a highly interconnected interaction pattern among these genes, indicating their involvement in shared biological processes. Based on network topological features, a set of key hub genes was identified, including AKT1, TNF, TGFB1, IL1B, MMP9, MMP2, FN1, ICAM1, PECAM1, and CDH5. These genes exhibited high connectivity within the network and are likely to play central roles in regulating inflammatory responses, extracellular matrix remodeling, and endothelial function.

A risk score model was constructed based on the identified core genes and evaluated in the TCGA-CRC and GSE39582 cohorts to assess its prognostic significance. Multivariable Cox regression analysis demonstrated that the gene signature was significantly associated with overall survival (OS) (Figure 7C) and consistently exhibited robust prognostic performance across both independent cohorts. The concordance index (C-index) further confirmed the predictive accuracy of the model. Kaplan–Meier survival analysis in the TCGA-CRC cohort showed that patients in the high-risk group had significantly poorer overall survival compared with those in the low-risk group (log-rank P < 0.0001) (Figure 7D). Time-dependent receiver operating characteristic (ROC) curves indicated strong discriminatory power for 1-, 3-, and 5-year survival prediction. In addition, decision curve analysis (DCA) demonstrated that the risk model provided a stable net clinical benefit across a wide range of threshold probabilities, highlighting its potential clinical utility.

To provide cell-level evidence supporting the role of SiO₂ in shaping tumor microenvironment susceptibility, we analyzed single-cell RNA sequencing data from the GSE161277 dataset. After stringent quality control, the cells retained for downstream analysis exhibited comparable distributions of detected genes (nFeature_RNA), total UMI counts (nCount_RNA), and mitochondrial gene proportions (percent.mt) across samples (Figure 8A), indicating high data quality and minimal technical bias. Following normalization and dimensionality reduction, highly variable genes were identified to capture dominant biological heterogeneity (Figure 8B). Unsupervised clustering combined with Uniform Manifold Approximation and Projection (UMAP) visualization revealed well-separated cellular populations (Figure 8C). Based on canonical marker genes, these clusters were annotated as major immune and stromal cell types, including T cells, B cells, monocytes, macrophages, neutrophils, fibroblasts, endothelial cells, and epithelial cells, reflecting the complex cellular architecture of the colorectal tumor microenvironment. To directly interrogate inflammation-related transcriptional alterations associated with occupational exposure–related immune imbalance, we examined the expression patterns of key genes involved in immune activation, cytokine signaling, and cell–cell interactions. Comparative analysis between control and CRC samples demonstrated marked upregulation of ICAM1, TGFB1, and IL1B in CRC-derived cells (Figure 8D). Feature plots revealed that these genes were enriched within specific immune and stromal compartments rather than being uniformly expressed, indicating localized microenvironmental activation rather than global transcriptional shifts. Violin plot analysis further confirmed significantly higher expression levels of these genes in CRC samples compared with controls.

Collectively, these single-cell results provide direct cellular evidence that colorectal cancer is characterized by coordinated immune and stromal transcriptional reprogramming, consistent with a state of sustained inflammatory activation. The enrichment of adhesion molecules and cytokine-related genes supports a model in which systemic immune dysregulation—potentially triggered by chronic occupational exposure—contributes to tumor microenvironment remodeling, thereby creating a permissive context for tumor-associated biomarker elevation and disease progression.