Noninvasive early detection and grading of pneumoconiosis via plasma proteomics and machine learning: PRSS3 as a potential biomarker

Abstract

Background

Coal-dust, a persistent airborne pollutant, induces dose-related pulmonary fibrosis; however, plasma biomarkers for pre-clinical toxicity remain lacking.

Methods

We enrolled 158 participants, including 28 healthy controls (HCs), 30 dust-exposed workers (DEWs), and 100 patients with coal workers’ pneumoconiosis (CWP) at different stages (n _CWP−I=40, n _CWP−II=30, n _CWP−III=30). Plasma proteomic profiling was performed via data-independent acquisition (DIA) mass spectrometry. Differentially expressed proteins were identified and functionally annotated. Key proteins were selected and multiple machine learning algorithms were employed to construct and validate predictive models.

Results

We identified 1,239 plasma proteins, including 645 high-confidence candidates. Functional enrichment revealed significant associations between disease progression and pathways such as PPAR signaling, cholesterol metabolism, Epstein-Barr virus infection, and the pentose phosphate pathway. These alterations converge on dysregulated lipid metabolism, chronic inflammatory signaling and virus-induced immune evasion, suggesting a metabolic-immune axis that orchestrates early fibrotic progression. We successfully constructed the first plasma proteomics-based machine learning models for pneumoconiosis grading and early screening. Notably, a single biomarker, PRSS3, demonstrated exceptional performance in distinguishing DEW patients from early-stage pneumoconiosis patients (CWP-I), achieving an area under the curve (AUC) of 1.00 and an accuracy of 1.00 in the training set and an AUC of 1.00 with an accuracy between 0.93 and 1.00 in the validation set.

Conclusion

This study establishes innovative machine learning-based models for the grading and early screening of pneumoconiosis via plasma proteomics. The identification of PRSS3 as a potential biomarker highlights the clinical utility of our approach. These findings provide a foundation for noninvasive diagnostic strategies and future translational research in occupational lung diseases.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12014-026-09591-7.

Introduction

Pneumoconiosis is an umbrella term for a group of occupational lung diseases characterized by diffuse pulmonary fibrosis resulting from long-term inhalation and retention of pathogenic mineral dusts in occupational settings. The most common form is coal-workers’ pneumoconiosis (CWP), caused by exposure to mixed dust containing both silica and coal, and is typified by chronic inflammation and nodular lesions in the lung parenchyma. Other frequently encountered entities include silicosis and asbestosis [1, 2]. Early-stage disease is usually asymptomatic or accompanied only by mild, nonspecific symptoms. With disease progression, respiratory and systemic manifestations emerge or worsen, dominated by cough, sputum production, chest pain, and dyspnea, and may be accompanied by wheezing, hemoptysis, and constitutional symptoms. Radiological findings typically include small rounded opacities, large opacities, and/or pleural abnormalities. The condition is progressive and irreversible, often culminating in respiratory failure and cor pulmonale, and markedly impairs patients’ quality of life [3, 4].

In 2021, there were 62,866 new cases and 18,323 deaths globally due to pneumoconiosis, placing a significant economic burden on society and making it one of the most severe occupational diseases worldwide [5]. The etiology of CWP is complex, its pathogenesis remains unclear, and it has a long latency period of 10–20 years, making early detection challenging and limiting treatment options [6]. Currently, diagnosis relies primarily on a well-documented history of occupational dust exposure and chest X-ray findings, lacking sensitive and specific biomarkers to dynamically monitor disease progression. Therefore, early detection and diagnosis are crucial for disease prevention, and removal from exposed environments plays a critical role in halting disease progression.

Omics technologies, including genomics, metabolomics, and proteomics, are increasingly being utilized in pneumoconiosis research, uncovering critical genes, proteins, and associated signaling pathways. These advancements underscore the pivotal role of omics approaches in elucidating the disease’s underlying mechanisms. Genomic studies have identified differentially expressed molecules such as Transforming Growth Factor Beta-Activated Kinase 1 (TAK1) and Interleukin-6 (IL-6) in pneumoconiosis [7, 8]. Furthermore, Tandem Mass Tag (TMT)-based quantitative proteomics has revealed numerous differentially expressed proteins, including CD14, PSAP, GM2A, COL1A1, ITGA8, and CLDN5 [9]. Subsequent Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses demonstrated that these proteins are predominantly involved in pathways related to phagosome formation, lysosomal function, complement and coagulation cascades, glutathione metabolism, cell adhesion, and ECM-receptor interactions [10].

The integration of machine learning with omics data significantly enhances the efficiency of screening for biomarkers with diagnostic potential. For instance, a high-throughput metabolomics study employing machine learning algorithms identified propyl benzoate as a promising biomarker [9].

Investigating omics profiles across different disease stages is crucial for early detection and intervention. Studies indicate that early-stage pneumoconiosis involves alterations in pathways such as glycolysis/gluconeogenesis, autoimmunity, carbon metabolism, and phagocytosis, with microfibril-associated glycoprotein 4 being implicated. In contrast, advanced stages are more strongly associated with processes like autoimmunity and cell adhesion, linked to factors such as calcium hippocampal binding protein [11]. Additional research has shown that disease progression is characterized by proteomic changes, including significant enrichment of proteins within the TGF-β signaling and autoimmunity-related pathways. Concurrently, metabolomic analyses have revealed disturbances in amino acid metabolism and unsaturated fatty acid biosynthesis, indicated by metabolite level changes (e.g., TFR). Multi-omics integration further underscores that ABC transporters represent a commonly dysregulated pathway during pneumoconiosis progression [12].

Therefore, this study aims to construct a plasma proteomic profile of pneumoconiosis to distinguish different disease states and to establish disease grading and early screening models by integrating machine learning approaches. These findings are expected to provide new strategies for the precise diagnosis and early intervention of pneumoconiosis.

Methods

Ethical approval

Baseline questionnaires and clinical data were collected after obtaining voluntary informed consent from all participants. The study protocol was approved by the Institutional Ethics Committee of Sinopharm Tongmei General Hospital (Approval No.: 2017-25−1; 2019-02) and conducted in accordance with the principles of the Declaration of Helsinki.

Pneumoconiosis cohort

A total of 158 participants were enrolled from the Sinopharm Tongmei General Hospital between February 2019 and July 2021, consisting of 28 healthy controls (HCs), 30 dust-exposed workers (DEWs), and 100 CWP patients. CWP diagnosis followed China’s National Diagnostic Criteria (GBZ70-2015) and the 1980 International Labor Organization Classification [13–15]. Chest CT, X-ray, and physical examinations were conducted for all CWP patients. CWP patients were categorized into stages I, II, and III based on opacity profusion on X-rays and were independently evaluated by three certified physicians. Participants were matched for demographics and duration of coal dust exposure.

Inclusion criteria

1. CWP group: Male coal miners; age ≥ 40 years; history of coal dust exposure ≥ 10 years with a confirmed diagnosis of coal workers’ pneumoconiosis, and here are the diagnostic criteria for CWP Staging:

CWP-I: A total small opacity profusion of category 1, involving at least two lung zones.

CWP-II: (1) A total small opacity profusion of category 2 affecting more than four lung zones; or (2) a total small opacity profusion of category 3 involving at least four lung zones.

CWP-III: (1) Presence of one or more large opacities with the longest dimension ≥ 20 mm and the shortest dimension > 10 mm; or (2) a total small opacity profusion of category 3 affecting more than four lung zones; or (3) a total small opacity profusion of category 3 involving more than four lung zones, accompanied by the presence of large opacities.

• (2)DEW group: Male coal miners; age ≥ 40 years; history of coal dust exposure ≥ 10 years without a diagnosis of pneumoconiosis.
• (3)HC group: Healthy male individuals with no history of coal dust exposure; age ≥ 40 years.

All participants were required to fully meet the respective group criteria to be included [14]. Final diagnoses were determined by a consensus panel of senior occupational disease specialists, integrating occupational history, imaging findings, clinical data, and, where applicable, response to therapy.

Exclusion criteria

Participants were excluded if they met any of the following criteria:

1. Coexisting severe pulmonary diseases, such as other interstitial lung diseases, lung cancer, or tuberculosis;

2. Significant comorbidities including congestive heart failure, rheumatic systemic diseases, hepatic fibrosis, renal fibrosis, or other fibrotic disorders;

3. Patients with a history of other systemic malignancies, or previous radiotherapy or chemotherapy.

We excluded the aforementioned conditions through a comprehensive diagnostic evaluation, including assessment of occupational exposure history, imaging studies, characteristic clinical manifestations, and response to diagnostic treatment.

Sample preparation of proteomic analysis

Sample collection

Peripheral venous blood (3–5 mL) was collected from each participant after fasting by an experienced nurse via ethylenediaminetetraacetic acid (EDTA) anticoagulant tubes. The samples were subsequently centrifuged at 2,000 × g for 10 min at 4 °C. The supernatant was aliquoted and stored at − 80 °C until further analysis.

Protein extraction

The plasma samples were thawed on ice. A 5 µL aliquot of plasma was mixed with 45 µL of lysis buffer (8 M urea, 1% protease inhibitor cocktail, 50 µM PR-619) by vortexing, followed by brief centrifugation at 300 × g. The protein concentration was determined using a BCA assay kit (Beyotime Biotechnology).

Trypsin digestion

Protein digestion was performed with the MagicOmics-MMB8X Micro/Universal Proteomics Enzymatic Digestion Kit (Qinglian Bio, Beijing, China) following the manufacturer’s instructions. Briefly, 20 µL of protein solution (1–100 µg) was added to an 8-tube strip containing MMB beads and incubated at 37 °C with 1 000 rpm shaking for 30 min. After adding 45 µL Binding Buffer, the mixture was shaken at 1 000 rpm for 15 min at room temperature. The beads were then separated on a magnetic rack, washed three times with Wash Buffer, and resuspended in 20 µL digestion buffer containing trypsin at an enzyme-to-protein ratio of 1:20 (w/w). Digestion proceeded at 37 °C with 1 000 rpm shaking for 4 h, followed by the addition of 5 µL Quench Buffer to terminate the reaction. The supernatant was collected, lyophilized, and stored at − 80 °C until LC-MS/MS analysis.

LC-MS/MS proteomic workflow

Peptide prefractionation (High-pH RP-HPLC)

Peptide pellets were re-dissolved in 1 mL of buffer A (2% acetonitrile, pH 9.0, adjusted with ammonia), vortexed, and centrifuged twice at 12 000 × g for 5 min. The clarified supernatant was loaded onto an Agilent 300Extend C18 column (10 mm × 250 mm, 5 μm) maintained at 35 °C. Elution was monitored at 214 nm with an automated fraction collector (1 tube/min). A 12-step high-pH gradient was applied using buffer B (98% acetonitrile, pH 9.0, adjusted with ammonia) at the following acetonitrile concentrations: 6%, 9%, 12%, 15%, 18%, 21%, 25%, 30%, 35%, 40%, 45% and 50%. Twelve fractions were collected, pooled accordingly, and lyophilised to dryness. Prior to LC-MS/MS analysis, iRT standard peptides were added to each fraction.

Spectral library generation (DDA)

High-pH fractions were reconstituted in 10 µL of solvent A (0.1% formic acid in water), vortex-mixed, centrifuged (14 000 × g, 20 min, 4 °C) and 1 µg of the supernatant was loaded onto a 25 cm × 150 μm i.d. fused-silica capillary column packed in-house with ReproSil-Pur C18-AQ, 1.9 μm particles (Qinglian Biotech, Beijing). The column was operated at 600 nL min⁻¹ on an EASY-nLC 1200 system (Thermo Fisher Scientific) with the following gradient (0–60 min): 0 min 8% B; 5 min 12% B; 35 min 30% B; 44 min 40% B; 45 min 95% B; 60 min 95% B. Solvent A: 0.1% formic acid in water; solvent B: 0.1% formic acid in 80% acetonitrile.

Eluting peptides were analysed on a Q Exactive HF-X mass spectrometer (Thermo Fisher Scientific) using a Nanospray Flex ion source (2.1 kV, 275 °C). Full-scan spectra were acquired at 120 000 resolution (m/z 200) over 350–1 500 m/z with an AGC target of 3 × 10⁶ and a maximum injection time of 80 ms. The top-40 most-abundant precursors were isolated (1.6 m/z window) and fragmented by HCD (NCE 27%). MS/MS spectra were recorded at 15 000 resolution (m/z 200), AGC 5 × 10⁴, maximum injection time 45 ms. Raw files were used to generate the spectral library.

Quantitative proteomics (DIA)

Peptide powders were re-dissolved in 0.1% formic acid (FA) in water containing iRT standard peptides. One microgram of the resulting mixture was injected for LC-MS/MS analysis. Separation was performed on a 25 cm × 150 μm i.d. column packed with ReproSil-Pur C18-AQ, 1.9 μm particles (Qinglian Biotech, Beijing) using an EASY-nLC 1200 system (Thermo Fisher Scientific) at 600 nL min⁻¹. The gradient was: 0 min, 8% B; 5 min, 12% B; 35 min, 30% B; 44 min, 40% B; 45 min, 95% B; held at 95% B until 60 min. Solvent A: 0.1% FA in water; solvent B: 0.1% FA in 80% acetonitrile.

MS data were acquired on a Q Exactive HF-X mass spectrometer (Thermo Fisher Scientific) equipped with a Nanospray Flex ion source operated at 2.1 kV. The instrument was run in data-independent acquisition (DIA) mode with a full-scan range of m/z 350–1 500 at 60 000 resolution (m/z 200), AGC target 3 × 10⁶, and maximum injection time 50 ms. Peptides were fragmented using stepped collision energies of 25.5%, 27%, and 30%.

Data analysis

All raw files were analyzed via Spectronaut software (Biognosys, version 17.0) [16, 17]. A DDA spectral library was constructed using data from the fractionated pooled samples (12 fractions in total). All searches were performed against the UniProt Homo sapiens Swiss-Prot protein sequence database (downloaded on 2022-03−17; 20,361 target sequences). The fixed modification was set as carbamidomethyl (C), and variable modifications included oxidation (M) and acetylation (protein N-terminal). Other parameters were set to the software defaults. The digestion enzyme was specified as trypsin, max missed cleavages were set as 2, with peptide lengths ranging from 7 to 52 amino acids. For library generation, each peptide was required to contain at least three fragment ions, and a maximum of the six best fragments were retained. Quality control was enforced at both the protein and peptide levels via a target-decoy strategy with a false discovery rate (FDR) threshold ≤ 1%.

Quantification was based on the MS2 peak area. Normalization was performed via either Global Median or LOESS (locally estimated scatterplot smoothing). Missing values were handled as follows: peptides with signal- noise ratio < 5 or a Q-value > 0.01 were considered missing; proteins with > 30% missing values across all samples were excluded; and the remaining missing values were imputed via the k-nearest neighbors (KNN, k = 5) complementation algorithm. Protein abundance was calculated as the median intensity of all quantified peptides corresponding to the same protein.

The retention time was calibrated via iRT standard peptides, ensuring a deviation of < 2 min. A “Normal Report” template was used to output results at both the protein group and precursor levels. Differential expression was assessed by limma (v3.50.3) with Benjamini–Hochberg false discovery rate (FDR) correction. Proteins with |log₂ fold-change (FC)| ≥ 0.263 (linear FC ≥ 1.2 or ≤ 1/1.2) and FDR < 0.05 were considered differentially expressed. For trend analysis across ordered groups, one-way analysis of variance (ANOVA) (FDR < 0.05, |log₂FC| ≥ 0.263) was applied, followed by fuzzy c-means clustering using the Mfuzz package (v2.48.0). Pathway enrichment was performed with Metascape (integrating GO, KEGG and MSigDB databases) against the identified protein background; only pathways with FDR < 0.05 and containing ≥ two differentially expressed proteins were retained. Networks were visualised in Cytoscape (v3.8.1).

WGCNA

In this study, a weighted co-expression network was constructed using the Weighted Co-expression Network Analysis (WGCNA) package (v1.72-1) in R (v4.3.0). Briefly, prior to network inference, only proteins within the top 75% of the median absolute deviation (MAD) were retained. The expression matrix was log₂-transformed and normalized via quantile normalization. Pairwise protein correlations were computed via the biweight midcorrelation (bicor). A soft threshold power β = 6 was selected with the “pickSoftThreshold” function under the condition of a scale-free topology fit index R² > 0.85, and the correlation matrix was transformed into a signed weighted adjacency matrix. A topological overlap matrix (TOM) was subsequently calculated, and its dissimilarity (1 – TOM) was used for clustering.

Average-linkage hierarchical clustering was performed on basis of the TOM dissimilarity matrix. Modules were initially identified via the dynamic tree cut algorithm (minModuleSize = 15; other parameters set to default). Health status was treated as an ordinal trait (HC = − 1, DEW = 0, CWP-all = 1), and the Spearman correlation between each module eigengene and health status was computed. Three modules with p < 0.01 were retained for further analysis. Within each retained module, the top 50% of proteins ranked by intramodular connectivity (kME) were selected for subsequent functional annotation. KEGG enrichment analysis was performed for each module using all detected proteins as the background. Only terms containing at least two proteins per pathway were retained.

Machine learning

All steps were completed within the training set, whereas the independent test set was solely used for final validation. The serum proteomics data from 158 patients were stratified and randomly split into a training set and an independent test set at an 8:2 ratio on the basis of disease category.

One-way ANOVA (with the disease category as the dependent variable) was performed on the quantified proteins in the training set, and differentially expressed proteins with p < 0.05 were retained. Then, least absolute shrinkage and selection operator (LASSO) multinomial regression with 10-fold cross-validation was applied to the above proteins to identify those with nonzero coefficients. Under a random forest framework (using “caret::rfe” with 5-fold cross-validation), recursive feature elimination was performed on the nonzero coefficient proteins to ultimately identify the optimal protein set. Based on the optimal proteins, six algorithms were compared via 5-fold cross-validation on the training set, with overall accuracy as the metric. The six algorithms were XGBoost, LightGBM, support vector machine (SVM), random forest, linear discriminant analysis (LDA), and decision tree.

The trained model based on the selected proteins was directly applied to the independent samples in the test set to evaluate its discriminative performance. The overall accuracy was calculated, and a confusion matrix along with ROC curves and the area under the curve (AUC) were generated. The R package “pROC” was used to plot the ROC curves and estimate the AUC values. Notably, an AUC value between 0.7 and 0.9 indicates good model performance, while a value > 0.9 suggests excellent performance.

Statistical analysis

Statistical analysis was performed using Graphpad Prism 9.0.0 (GraphPad Software, Inc). Categorical data were compared by χ² or Fisher’s exact test (expected frequency < 5 in > 20% of cells). Continuous data were tested for normality; normally distributed variables were analysed with one-way ANOVA and Tukey’s post-test, otherwise with Kruskal–Wallis and Dunn’s test. P < 0.05 was considered statistically significant.

Results

Cohort characteristics and plasma proteomic profiling

This study included 158 male participants, comprising 28 in the HC group, 30 in the DEW group, 40 in the CWP-I group, 30 in the CWP-II group, and 30 in the CWP-III group. Baseline characteristics, including age, sex, height, weight, body mass index (BMI), health status, smoking history, dust exposure history, and pulmonary function parameters, were recorded. Significant differences (P < 0.05) were detected in health status, dust exposure, height, and pulmonary function indicators among the groups, while no significant differences (P > 0.05) were detected in the other parameters (Table 1).

Table 1.

Clinical baseline characteristics of the study participants

	HC	DEW	CWP-I	CWP-II	CWP-III	p	Test
n	28	30	40	30	30
Age (yr) (mean (SD))	63.11 (10.38)	63.53 (9.46)	64.70 (10.00)	69.73 (11.01)	66.97 (11.35)	0.087
Sex = male (%)	28 (100.0)	30 (100.0)	40 (100.0)	30 (100.0)	30 (100.0)
Height (cm) (median [IQR])	171.50 [170.00, 175.00]	170.50 [170.00, 173.00]	170.50 [168.00, 175.00]	169.00 [162.50, 170.75]	170.00 [167.00, 172.75]	0.006	nonnorm
Weight (kg) (mean (SD))	68.55 (9.16)	68.07 (10.76)	69.96 (11.92)	65.90 (9.08)	63.96 (10.48)	0.158
BMI (mean (SD))	23.19 (2.65)	23.42 (3.34)	23.94 (3.70)	23.69 (3.00)	22.54 (3.24)	0.469
Health status (%)						< 0.001
0	28 (100.0)	30 (100.0)	0 (0.0)	0 (0.0)	0 (0.0)
1	0 (0.0)	0 (0.0)	40 (100.0)	0 (0.0)	0 (0.0)
2	0 (0.0)	0 (0.0)	0 (0.0)	30 (100.0)	0 (0.0)
3	0 (0.0)	0 (0.0)	0 (0.0)	0 (0.0)	30 (100.0)
Smoking status (%)						0.881	exact
0	4 (14.3)	4 (13.3)	4 (10.0)	3 (10.0)	3 (10.0)
1	13 (46.4)	19 (63.3)	25 (62.5)	20 (66.7)	20 (66.7)
2	11 (39.3)	7 (23.3)	11 (27.5)	7 (23.3)	7 (23.3)
Smoking duration (yr) (median [IQR])	25.50 [16.75, 40.00]	40.00 [20.00, 44.25]	31.00 [20.00, 40.50]	30.00 [28.50, 39.00]	30.00 [20.75, 40.00]	0.775	nonnorm
Cigarettes per day (CPD) (median [IQR])	20.00 [20.00, 20.00]	15.00 [10.00, 20.00]	20.00 [6.75, 20.00]	20.00 [12.50, 20.00]	20.00 [10.00, 20.00]	0.369	nonnorm
Smoking Index (SI) (median [IQR])	31.50 [16.75, 40.00]	23.25 [10.31, 41.50]	26.50 [9.19, 38.50]	30.00 [20.00, 37.50]	24.25 [13.12, 40.00]	0.961	nonnorm
Dust-exposed (yr) (median [IQR])	0.00 [0.00, 0.00]	25.00 [20.00, 30.00]	30.00 [20.00, 30.25]	29.50 [22.25, 30.00]	29.50 [24.25, 30.00]	< 0.001	nonnorm
FVC (%Pred) (median [IQR])	3.86 [3.52, 4.13]	3.87 [3.69, 4.06]	3.76 [3.42, 4.12]	3.54 [3.32, 3.67]	3.69 [3.47, 4.06]	0.009	nonnorm
FVC (L) (mean (SD))	3.81 (0.72)	3.70 (0.57)	3.09 (0.77)	2.82 (0.89)	2.78 (0.88)	< 0.001
FVC (L/%Pred) (mean (SD))	0.98 (0.14)	0.96 (0.13)	0.83 (0.18)	0.81 (0.22)	0.76 (0.24)	< 0.001
FEV1 (%Pred) (mean (SD))	3.06 (0.45)	3.06 (0.28)	2.92 (0.45)	2.66 (0.43)	2.87 (0.45)	0.001
FEV1 (L) (median [IQR])	2.70 [2.50, 3.08]	2.83 [2.33, 3.04]	1.96 [1.62, 2.65]	1.62 [1.25, 2.54]	1.38 [1.09, 1.99]	< 0.001	nonnorm
FEV1 (L/%Pred) (median [IQR])	0.92 [0.85, 1.02]	0.90 [0.77, 0.99]	0.74 [0.51, 0.90]	0.72 [0.49, 0.88]	0.48 [0.38, 0.76]	< 0.001	nonnorm
FEV1/FVC (median [IQR])	0.74 [0.72, 0.76]	0.74 [0.66, 0.79]	0.69 [0.56, 0.73]	0.66 [0.51, 0.72]	0.55 [0.45, 0.65]	< 0.001	nonnorm

DEW, dust-exposed workers; HC, healthy controls; CWP, Coal workers’ pneumoconiosis. Health status: 0 = HC or DEW; 1 = CWP-1; 2 = CWP-II; 3 = CWP-II. Smoking status: 0 = never-smoker (no passive exposure); 1 = ex-smoker (mild/passive exposure); 2 = current smoker (heavy or sustained passive exposure). FEV₁: forced expiratory volume in the first second, FVC: forced vital capacity, FEV₁% predicted: percentage of predicted FEV₁, FVC% predicted: percentage of predicted FVC, FEV₁/FVC ratio: expressed as a percentage. Data are presented as median [inter-quartile range] or mean (standard deviation). “Non-norm”: indicates non-normally distributed data; data were analyzed via non-parametric tests. All other quantitative variables were compared via one-way ANOVA. “exact”: indicates the use of Fisher’s exact test; all other qualitative variables were analyzed via the chi-square test. P < 0.05 was considered statistically significant

The study workflow is illustrated in Fig. 1. First, a cohort of 158 subjects was established, including HCs (n = 28), DEW patients (n = 30), and patients with CWP at different stages (n _CWP−I = 40, n _CWP−II = 30, n _CWP−III = 30), and baseline characteristics were collected. Proteomic analysis, which included plasma sample collection, tryptic digestion, and mass spectrometry detection, was subsequently performed. Data preprocessing was conducted via software such as Spectronaut [16]. The quantitative proteomic data were subjected to quality control (Supplementary Fig. 1), differential expression analysis, enrichment analysis, and WGCNA to identify key proteins. Finally, feature proteins were selected via LASSO regression, and RFE was applied to construct predictive models. Multiple machine learning methods, including XGBoost, LightGBM, SVM, random forest, LDA, and decision tree, have been employed to develop and validate the models.

Fig. 1.

Study flowchart. A total of 158 plasma samples were collected, including 28 healthy control (HC) samples, 30 dust-exposed worker (DEW) samples, 40 stage I coal worker (CWP-I) samples, 30 stage II (CWP-II) samples, and 30 stage III (CWP-III) samples. All the samples were subjected to proteomic analysis, data processing, and predictive model construction and validation. DEW: dust-exposed workers, HC: healthy controls, CWP: coal workers’ pneumoconiosis

A total of 1,239 proteins were identified across the 158 plasma samples, among which 646 were high-confidence proteins (Supplementary Figs. 1 A–B). To ensure data reliability, stringent quality control was implemented. The number of proteins identified per sample was approximately 500 (Supplementary Fig. 1B), and high Pearson correlation coefficients were observed between samples (Supplementary Fig. 1 C). Principal component analysis (PCA) revealed minimal separation between the HCs and DEW patients, while both groups were clearly distinct from the CWP patients. Furthermore, the differences among the three CWP stages were relatively small (Supplementary Fig. 1D).

Proteomic alterations across disease progression

To clarify the differences and common biological processes among the different groups, we performed differential proteomic analysis and KEGG enrichment analysis.

First, comparative analysis between the DEW and HC groups was conducted to reveal the proteomic characteristics of dust-exposed workers, which revealed the early characteristics of the disease. A total of 55 DEPs were identified, 27 of which were upregulated and 28 were downregulated. The downregulated proteins were enriched primarily with cell adhesion molecules. The upregulated proteins were significantly enriched in the Rap1 signaling pathway, cytokine-cytokine receptor interaction, focal adhesion, protein digestion and absorption (e.g., SLC3A2 and PRSS3), and the PI3K-Akt signaling pathway (Fig. 2 A). Compared with those in the HC group, the DEPs in the DEW group were associated mainly with environmental information processing-related pathways, suggesting that early dust exposure may lead to impaired cell adhesion and activation of pro-inflammatory and survival signaling pathways.

Fig. 2.

Analysis of differentially expressed proteins and pathway enrichment across comparison groups. (A) Volcano plot of differentially expressed proteins and KEGG pathway enrichment analysis between DEW and HC groups. A total of 55 differentially expressed proteins were identified (27 up- and 28 down-regulated proteins). (B) Volcano plot and KEGG enrichment of differentially expressed proteins between CWP-all and HC groups. A total of 175 differentially expressed proteins were identified (87 up- and 88 down-regulated proteins). (C) Volcano plot and KEGG enrichment of differentially expressed proteins between CWP-all and DEW groups. A total of 155 differentially expressed proteins were identified (67 up- and 88 down-regulated proteins). (D) Volcano plot and KEGG enrichment of differentially expressed proteins between CWP-I and DEW groups. A total of 148 differentially expressed proteins were identified (64 up- and 84 down-regulated proteins). DEW: dust-exposed workers, HC: healthy controls, CWP: coal workers’ pneumoconiosis. KEGG: Kyoto Encyclopedia of Genes and Genomes

Next, we compared all CWP patients with HCs to uncover the overall proteomic alterations associated with pneumoconiosis, which revealed the overall state of CWP. We identified 175 DEPs, including 87 upregulated and 88 downregulated DEPs. The downregulated proteins were enriched mainly in the pentose phosphate pathway and glycolysis/gluconeogenesis. The upregulated proteins were significantly enriched in the NF-κB signaling pathway, peroxisome proliferator-activated receptor (PPAR) signaling pathway, cholesterol metabolism, and Epstein-Barr Virus (EBV) infection (Fig. 2B). These findings indicate systemic inflammatory activation, disordered lipid metabolism, and EBV-related molecular abnormalities in CWP patients.

To further investigate proteomic dynamics during disease progression, we compared the proteomic profiles between all CWP groups and the DEW group, as well as between the CWP-I and DEW groups. In the CWP vs. DEW comparison, 155 DEPs were identified (67 up- and 88 down-regulated DEPs). The downregulated proteins were enriched in the pentose phosphate pathway, growth hormone synthesis and action, peroxisomes, FcγR-mediated phagocytosis, dilated cardiomyopathy, and hypertrophic cardiomyopathy, suggesting that myocardial impairment may occur as early as stage I pneumoconiosis [18]. The upregulated proteins were enriched in cell adhesion molecules, the PPAR signaling pathway, cholesterol metabolism, estrogen signaling pathway, and EBV infection (Fig. 2 C). These results suggest that during pneumoconiosis progression, immune phagocytosis, the growth hormone axis, and antioxidant capacity are weakened, and glucose metabolism is suppressed. In contrast, cell adhesion, dyslipidemia, estrogen signaling, and EBV-related mechanisms are significantly enhanced.

Similarly, 148 DEPs were identified (64 up- and 84 down-regulated DEPs) in the CWP-I vs. DEW comparison. The downregulated proteins were significantly enriched in the Ras signaling pathway, dilated cardiomyopathy, hypertrophic cardiomyopathy, transcriptional misregulation in cancer, osteoclast differentiation, lysosome, cytokine-cytokine receptor interaction, axon guidance, and the Rap1 signaling pathway. The upregulated proteins were enriched mainly in the PPAR signaling pathway, cholesterol metabolism, and EBV infection (Fig. 2D). These results indicate that upon progression from dust exposure to CWP-I, a broad range of signal transduction, cardiovascular disease, and cancer-related pathways are downregulated, whereas upregulation of lipid metabolism abnormalities and EBV-related mechanisms remains a prominent feature. Detailed information on the identified and differential proteins is provided in Supplementary Tables 1–2.

WGCNA identifies disease-associated co-expression modules

WGCNA results for all samples

To investigate co-expression patterns among proteins, identify functionally related modules, and associate them with clinical traits, we performed WGCNA involving HC, DEW, and all CWP samples. After preprocessing the proteomic data, correlation coefficients between all the gene pairs were calculated to construct a co-expression similarity matrix, which was then transformed into an adjacency matrix via power transformation. The TOM was computed, and hierarchical clustering was applied on the basis of the TOM. Genes were assigned to different modules via a dynamic tree-cutting algorithm. WGCNA identified seven co-expression modules, containing 232, 74, 61, 42, 34, 21, and 19 proteins, respectively (Fig. 3 A).

Fig. 3.

Analysis of differentially expressed proteins and pathway enrichment among the HC, DEW, and CWP-all groups via WGCNA. (A) Seven WGCNA modules identified from differentially expressed proteins across the HC, DEW, and CWP-all groups. Key parameters: MAD 75%, bicor correlation, minModuleSize = 15. (B) Correlations between module eigengenes and clinical traits. High expression of proteins in the green and brown modules was positively correlated with pneumoconiosis progression, whereas high expression in the turquoise module was negatively correlated with pneumoconiosis progression. (C) Heatmap of protein expression and pathway enrichment results for the green, brown, and turquoise modules—most relevant to disease progression—after stringent filtering (top 50% connectivity used for protein selection). DEW: dust-exposed workers, HC: healthy controls, CWP: Coal Workers’ Pneumoconiosis, WGCNA: weighted gene co-expression network analysis

Multiple algorithms were used to evaluate correlations between module protein expression and clinical features, enabling the analysis of intramodular relationships and identification of hub genes (Fig. 3B). Spearman correlation analysis revealed that high expression of proteins in the green and brown modules was positively correlated with pneumoconiosis progression, while high expression in the turquoise module was negatively correlated with pneumoconiosis progression. Expression heatmaps were generated, and KEGG enrichment analysis identified key pathways: proteins negatively associated with disease progression were enriched in the phospholipase D signaling pathway, growth hormone synthesis, secretion and action, hepatocellular carcinoma, the Ras signaling pathway, and inflammatory mediator regulation of TRP channels. Proteins positively associated with disease progression were enriched mainly in the PPAR signaling pathway, cholesterol metabolism, vitamin digestion and absorption, fat digestion and absorption, and EBV infection (Fig. 3 C). These results suggest that as the disease progresses, lipid metabolic dysregulation, vitamin and fat digestion and absorption, and EBV-related mechanisms are significantly increased, whereas multiple signal transduction pathways are broadly suppressed. Detailed data for the comparisons among the HC, DEW, and all CWP groups are provided in Supplementary Table 3.

WGCNA reveals the most disease-related modules

We further performed WGCNA and constructed a staging discrimination model among the CWP-I, CWP-II, and CWP-III groups. WGCNA identified eight co-expression modules, each color-coded. Spearman correlation analysis revealed that high expression of proteins in the blue module was positively correlated with pneumoconiosis progression, whereas high expression in the brown and black modules was negatively correlated with disease progression (Supplementary Fig. 3).

Expression heatmaps and KEGG enrichment analysis indicated that proteins positively associated with disease stage were enriched mainly in the osteoclast differentiation pathway (e.g., CSF1R), suggesting that destructive bone inflammatory processes may occur in advanced disease stages. Proteins that were negatively correlated with progression were enriched primarily in pathways related to metabolic regulation and the immune response, including the PPAR signaling pathway, cholesterol metabolism, vitamin digestion and absorption, and fat digestion and absorption. These findings suggest that the apolipoprotein family may play a protective role by regulating lipid metabolism and energy balance, and that its decreased expression could lead to impaired lipid clearance and energy supply disruption, thereby promoting disease progression.

Additionally, negatively correlated proteins were enriched in the following pathways: cysteine and methionine metabolism, the glucagon signaling pathway, pyruvate metabolism, propanoate metabolism, central carbon metabolism in cancer, glycolysis/gluconeogenesis, and the HIF-1 signaling pathway, indicating that glycolysis may be associated with a proinflammatory phenotype during pneumoconiosis progression.

Immune-related pathways, including EBV infection, neutrophil extracellular trap formation, legionellosis, protein processing in the endoplasmic reticulum, antigen processing and presentation, and influenza A, were also significantly enriched. These results suggest that chronic viral infection and overactivation of innate immunity may accelerate disease progression.

In summary, positively regulated proteins are involved primarily in osteoclast differentiation, while negatively regulated proteins are associated mainly with lipid metabolic homeostasis, the immune stress response, and glycolytic reprogramming.

Machine learning models for disease diagnosis, staging, and early screening

A multiprotein classifier for diagnosing pneumoconiosis

To achieve noninvasive stratification of pneumoconiosis and identify characteristic biomarkers, we further constructed a multiclass prediction model via machine learning methods (Supplementary Fig. 2). A training set was established, including HC (n = 22), DEW(n = 24), and CWP-all (n = 80) samples, and 45 feature proteins were selected from the differentially expressed proteins via LASSO regression (Supplementary Fig. 2 A). RFE was then applied to ultimately identify 8 proteins for model construction (Supplementary Fig. 2B). Six machine learning algorithms (XGBoost, LightGBM, SVM, random forest, LDA, and decision tree) were employed in parallel to develop the classification model, with hyperparameters optimized via cross-validation. Model performance was evaluated on an independent test set, including HC (n = 6), DEW (n = 6), and CWP-all (n = 20) samples, through multi-model comparison, with overall classification performance assessed using metrics including accuracy, sensitivity, and specificity (Supplementary Fig. 2 C). On the basis of the AUC‒ROC curve, XGBoost was identified as the optimal model, and its confusion matrix indicated satisfactory predictive performance (Supplementary Fig. 2D). The eight selected proteins were RAB4A, PRSS3, CSF1R, LRRC56, EIF4A2, BNC2, CLEC3B, and TRIM17, whose expression profiles across all the samples are shown in Supplementary Fig. 2E. The PCA results revealed clear separation among the healthy control, dust-exposed, and pneumoconiosis groups, indicating that the eight-protein-based classification model has strong discriminative power (Supplementary Fig. 2 F).

A prognostic model for stratifying disease severity

Similarly, machine learning methods including LASSO regression, RFE, and LightGBM, were employed to construct a staging model discriminating among CWP-I, CWP-II, and CWP-III (Fig. 4 A). Six feature proteins were identified that effectively distinguished different pneumoconiosis stages: LMTK2 (lemur tyrosine kinase 2), PRSS3 (serine protease 3), BRF1 (BRF1 RNA polymerase III transcription initiation factor), HYI (hydroxypyruvate isomerase), GLIPR2 (GLI pathogenesis-related 2), and FKBP4 (FK506 binding protein 4) (Fig. 4B).

Fig. 4.

Machine learning-based development and validation of a staging model discriminating among CWP stages (I, II, and III). (A) Forty-one feature proteins selected by the least absolute shrinkage and selection operator (LASSO) regression from differentially expressed proteins among the CWP stages in the training set. (B) Recursive feature elimination (RFE) identified an optimal 6-protein feature set for modeling (the red marker indicates 6 features). (C) Performance accuracy of multiple machine learning models using these 6 proteins in both the training (5-fold cross-validation) and test sets. (D) Confusion matrix of the LightGBM model on the test set. (E) Heatmap of the expression levels of the 6 proteins across all the samples. (F) Principal component analysis (PCA) results demonstrating the discriminative ability of the 6 proteins among the CWP stages (I, II, and III). DEW: dust-exposed workers, HC: healthy controls, CWP: coal workers’ pneumoconiosis

As shown in Fig. 4C, the accuracies of the different algorithmic models ranged from 0.86 to 0.91 on the training set and from 0.90 to 0.95 on the test set. Figure 4D shows the confusion matrix of the LightGBM model, Fig. 4E displays the expression heatmap of the six feature proteins across all samples, and Fig. 4 F presents the PCA, demonstrating that these six proteins effectively differentiated among the pneumoconiosis stages. Detailed data for the three groups are provided in Supplementary Table 4.

Potential of PRSS3 as a circulating biomarker for the early detection of pneumoconiosis in dust-exposed individuals

Accurately distinguishing between DEW and CWP-I is of significant clinical importance in practical applications. Therefore, this study aimed to develop an early screening model to differentiate DEW from CWP-I. Following a procedure similar to that described previously, we employed LASSO regression, RFE, and multiple machine learning algorithms for model construction. Notably, we found that a single protein biomarker, PRSS3, was sufficient to effectively discriminate between DEW and CWP-I (Fig. 5A-B). PRSS3 (serine protease 3), a member of the serine protease family, has recently garnered increasing attention for its role and regulatory mechanisms in various cancers and diseases; however, its function in pneumoconiosis warrants further investigation. The results demonstrated that PRSS3 exhibited exceptional discriminatory power in the training set, with an AUC of 1.00 (Fig. 5 C). In the independent test set, various machine learning algorithms consistently performed strongly, with accuracies, AUCs, F1 scores, sensitivities, and specificities ranging between 0.88 and 1.00. The confusion matrix and ROC curves further confirmed the robustness of the model (Fig. 5D-F). The expression heatmaps revealed that PRSS3 was downregulated in the DEW group but upregulated in the CWP-I group (Fig. 5G). PRSS3 expression showed dynamic changes during pneumoconiosis progression, peaking at CWP-I stage and gradually declining in CWP-II and CWP-III stages (Supplemental Figure S5A-B). Significant differences were observed between healthy controls and all pneumoconiosis stages, as well as between DEW and various pneumoconiosis stages (Supplemental Figure S5B, p < 0.05). Trend analysis confirmed the progression-associated expression pattern of PRSS3 (Supplemental Figure S5C, p < 0.01), and fold-change analysis revealed the most pronounced change in CWP-I (Supplemental Figure S5D).

Fig. 5.

Machine learning-based development and validation of a screening model discriminating between DEW and CWP-I. (A) Eighteen feature proteins selected by the least absolute shrinkage and selection operator (LASSO) regression from DEW vs. CWP-I differentially expressed proteins in the training set. (B) Recursive feature elimination (RFE) showed that a single protein (PRSS3) achieved high performance. (C) Area under the curves (AUC) of multiple machine learning models using PRSS3 in the training set. (D) Evaluation of the discrimination performance of various models in the test set. (E) Confusion matrices of multiple machine learning models in the test set. (F) Receiver operating characteristic (ROC) curves of multiple machine learning models in the test set. (G) Heatmap of PRSS3 expression in DEW and CWP-I samples. DEW: dust-exposed workers, HC: healthy controls, CWP: coal workers’ pneumoconiosis

Furthermore, we applied the same strategy to construct a tri-group screening model involving HC, DEW, and CWP-I. Ultimately, five feature proteins were identified (PRSS3, LMTK2, BNC2, CSF1R, and EIF4A2). The model demonstrated favorable performance in both the training and test sets, as detailed in Supplementary Fig. 4.

Discussion

Pneumoconiosis is one of the most severe occupational diseases worldwide. However, sensitive and specific biomarkers are lacking due to an incomplete understanding of its pathogenesis. In recent years, proteomics integrated with machine learning has increasingly been applied in disease research, providing a powerful approach to diagnosis, treatment and predicting prognosis. Therefore, the aim of this study was to explore biomarkers for early diagnosis, with a particular focus on non-invasive screening methods that are highly accurate and reproducible. Multiple machine learning algorithms were employed to develop and validate predictive models.

A total of 158 subjects were enrolled, including HCs, DEWs and CWP patients at different disease stages. Plasma proteomic profiling identified 1,239 proteins, of which 646 were classified as high-confidence proteins. Subsequent differential expression and functional enrichment analyses revealed significant associations between disease severity and several biological processes, including the PPAR signaling pathway, cholesterol metabolism, EBV infection, and the pentose phosphate pathway. WGCNA was then applied to identify key proteins. Feature selection was performed using LASSO regression and RFE to construct predictive models. To our knowledge, this is the first study to establish machine learning–based grading and early screening models for pneumoconiosis using plasma proteomics. These models provide valuable insights for diagnosing, staging, and detecting this disease early, as well as for its treatment.

This study used differential expression and enrichment analyses to reveal common features across all stages of pneumoconiosis. These analyses emphasized systematic alterations in metabolic pathways during disease progression. Pairwise comparisons (CWP-all vs. HC, CWP-all vs. DEW and CWP-I vs. DEW) revealed that differential proteins were consistently enriched in the PPAR signaling pathway, cholesterol metabolism, EBV infection and the pentose phosphate pathway. PPAR, a master regulator of lipid metabolism, is functionally linked to the pathogenesis of pneumoconiosis [19–21]. Specific activation of the PPAR signaling pathway has been shown to suppress fibroblast activation, attenuate collagen deposition, and reduce the decline in lung function in silica-induced pulmonary fibrosis [22]. Disturbances in cholesterol metabolism are associated with an increased risk of comorbidities [23], and a recent lipidomic study of CWP identified altered lipids closely related to cholesterol metabolism, corroborating our proteomic findings [24]. The pentose phosphate pathway, an important glucose metabolic shunt, is reportedly dysregulated in murine models of silica-induced pulmonary fibrosis [25].

Additionally, DEWs displayed distinct pathway activities, including Rap1 signaling, cytokine–cytokine receptor interaction and cell adhesion molecules. Rap1 governs cellular adhesion, junction formation and polarity. An integrated methylome-transcriptome analysis identified Rap1 signaling as a key discriminator between IPF and normal lung tissue [26, 27]. Cytokine–cytokine receptor interaction is fundamental to autocrine/paracrine immune communication: cytokines are released by multiple cell types in response to stimulatory cues [28]. In silica-challenged mice, the novel anti-fibrotic agent mefenidone markedly modulated this interaction, concurrently suppressing pulmonary inflammation and fibrosis [29]. Cell adhesion molecules are essential for immunity and inflammation [30]. Consistent with our results, combined transcriptomic and proteomic analyses have revealed the enrichment of differentially expressed genes and proteins in adhesion-related pathways [31]. Functional studies demonstrate that silica down-regulates endothelial adhesion molecules, thereby promoting local myofibroblast accumulation and amplifying systemic inflammatory signals [32]. Population-based investigations have also reported an inverse correlation between adhesion molecules and lung function [33]. Taken together, these pathway alterations suggest that individuals with DEW exhibit pronounced innate and adaptive inflammatory responses, reflecting an active injury–repair interplay.

Our initial observation was that plasma PRSS3 levels were significantly higher in DEWs, suggesting its potential as a biomarker for the pre-clinical stage of pneumoconiosis. PRSS3 is a serine protease that is increasingly recognized as a regulator of inflammation and fibrosis in multiple disorders. It is overexpressed in hepatocellular carcinoma and lung cancer, and its expression is controlled by DNA methylation, non-coding RNAs, and proteolytic activation, which correlates with aggressive progression and poor prognosis [34–39]. Notably, distinct PRSS3 splice variants exert oncogenic or tumor-suppressive effects in hepatocellular carcinoma [35, 36]. In inflammatory bowel disease, PRSS3 interacts with innate immune pathways and may destabilize proteins that trigger disease onset [28]. Conversely, in diabetic kidney disease, PRSS3 overexpression ameliorates glucose- and microbiota metabolite-induced renal inflammation and fibrosis [40]. Machine-learning screening has identified aberrant plasma PRSS3 in pneumoconiosis patients, highlighting its potential for early detection. While the exact function of PRSS3 in silicotic lung disease remains unclear, we hypothesize that coal-dust-activated alveolar macrophages release PRSS3, which then amplifies fibroblast activation and collagen deposition via a positive feedback loop, pending functional validation. Rigorous expression profiling and mechanistic studies are underway to clarify this pathway and consolidate PRSS3 as a diagnostic and therapeutic target for pneumoconiosis.

In recent years, multi-omics strategies have been widely used to map the molecular landscape of pneumoconiosis. For instance, Zhang et al. combined proteomic, acetylomic and succinylomic analyses to show that post-translational modifications of STAT1 drive the inflammatory response in silicotic mice [41]. A radiomics-transcriptomics study suggested that the TFCP2/USP22 axis promotes fibrosis through ECM-receptor and AGE-RAGE signaling (AUC = 0.799) [42], while a genomics-proteomics-metabolomics triad identified ERCC1 as an additional potential biomarker [31]. Notably, these investigations were mostly conducted in advanced animal models or cross-sectional cohorts, and the resulting signatures exhibited the greatest differences between CWP stages II/III and healthy controls. Consequently, they are better suited to discriminating between diseases than to providing early warnings. We therefore excluded CWP-II and CWP-III samples (Supplemental Figure S4) and restricted our analysis to HCs, DEWs and CWP-/I subjects, in order to capture molecules specifically relevant to early diagnosis and differential classification. The resulting five-protein model achieved accuracies of 0.94 and 0.80 in the training and test sets, respectively, providing an initial stratification of the three groups. Together, our plasma signature and existing omics data form a temporal continuum: the latter elucidate terminal fibrotic pathways, while the former provides a practical tool for early detection. Together, they offer the prospect of covering the entire pneumoconiosis management spectrum.

To our knowledge, this is the first report of a non-invasive, plasma proteomics-driven grading strategy for pneumoconiosis which uses machine learning to provide a reproducible candidate marker: PRSS3. Our findings lay the groundwork for translational studies focused on early screening. However, the exploratory nature of this design introduces several limitations. Firstly, all participants were recruited from a single occupational health examination central, which ensured standardized sample collection and clinical annotation, but which may have introduced regional and industrial bias. We are now collaborating with pneumoconiosis hospitals in different provinces to enroll subjects exposed to various dust types and mining operations, thereby evaluating the robustness of the model across heterogeneous settings. Secondly, the current plasma proteome covered approximately 1,200 proteins, but did not detect low-abundance inflammatory mediators or post-translational modifications. We plan to apply targeted phospho-/acetyl-enrichment strategies together with metabolomics to obtain a comprehensive view of the molecular drivers of disease progression. Thirdly, the cross-sectional nature of this study and the use of retrospective samples limits causal inference. A large-scale prospective cohort with longitudinal follow-up is underway to validate the early-warning performance of the identified markers. Finally, the discovery phase relied on high-resolution mass spectrometry, which is not readily accessible to primary laboratories. We are therefore developing a high-throughput enzyme-linked immunosorbent assay (ELISA) kit for PRSS3 to standardize large-scale measurements. We anticipate that, through these multi-central, multi-omics and prospective efforts, we will be able to translate our model and biomarkers into a scalable, non-invasive screening protocol for pneumoconiosis.

Conclusion

This study established a plasma proteomics dataset comprising patients with pneumoconiosis (stages I, II, and III), dust-exposed workers, and healthy controls. Quantitative analysis, differential expression analysis, functional enrichment analysis, and WGCNA were performed to construct a comprehensive proteomic profile. Key pathways and proteins (including PRSS3) were identified. Machine learning approaches were subsequently applied to discover biomarkers for early screening and disease staging.

Abbreviations

DEPs

Differentially expressed proteins

HC

Healthy controls

DEW

Dust-exposed workers

CWP

Coal workers’ pneumoconiosis

CWP-I

Coal workers’ pneumoconiosis in stage I

CWP-II

Coal workers’ pneumoconiosis in stage II

CWP-III

Coal workers’ pneumoconiosis in stage III

DDA

Data-dependent acquisition

DIA

Data-independent acquisition

LASSO

Least absolute shrinkage and selection operator

GO

Gene ontology

KEGG

Kyoto encyclopedia of genes and genomes

KNN

k-nearest neighbor complementation algorithm

AUC

Area under the ROC curve

LOESS

Locally estimated scatterplot smoothing

EBV

Epstein-Barr virus

TAK1

Transforming growth factor beta-activated kinase 1

TMT

Tandem mass tag-based quantitative proteomics

PRM

Parallel reaction monitoring

IL-6

Interleukin-6

ROC

Receiver operating characteristic

MFAP4

Microfibril-associated glycoprotein 4

CHBP

Calcium hippocampal binding protein

ILO

International Labor Organization

ATS/ERS

American Thoracic Society/European Respiratory Society

FEV₁

forced expiratory volume in the first second

FVC

forced vital capacity

FEV₁% predicted

Percentage of predicted FEV₁

FVC% predicted

Percentage of predicted FVC

FEV₁/FVC ratio

Expressed as a percentage

HPLC

High-performance liquid chromatography

MS

Mass spectrometry

iRT

Internal standard peptide

LC–MS/MS

liquid chromatography-tandem mass spectrometry analysis

AGC

Automatic gain control

HCD

Higher-energy collisional dissociation

NCE

Normalized collision energy

FC

Fold change

WGCNA

Weighted co-expression network analysis

FDR

False discovery rate

MAD

Median absolute deviation

TOM

Topological overlap matrix

kME

Intramodular connectivity

RFE

Recursive feature elimination

SVM

Support vector machine

LDA

Linear discriminant analysis

ANOVA

Analysis of variance

BMI

Body mass index

PCA

Principal component analysis

IBD

Inflammatory bowel disease

DKD

Diabetic kidney disease

TMAO

The gut microbiota metabolite trimethylamine N-oxide

LMTK2

Lemur tyrosine kinase 2

PRSS3

Serine protease 3

BRF1

BRF1 RNA polymerase III transcription initiation factor

HYI

Hydroxypyruvate isomerase

GLIPR2

GLI pathogenesis-related 2

FKBP4

FK506 binding protein 4

ELISA

Enzyme-linked immunosorbent assay

Author contributions

D.H., J.Y., and J.L. designed and supervised the entire study. Z.H., J.G., and M.W. collected patients’ electronic medical records and serum samples. Y.T., and M.W. conducted the analyses, interpreted the data and wrote the article. Y.T., W.Z., Z.H., J.G., X.Z., and K.M. carefully revised the manuscript. All of the authors have read and approved the final manuscript. D.H., J.Y., and J.L. had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Funding

This project was funded by the National Natural Science Foundation of China Young Student Basic Research Program (823B2001), the National Key Research and Development Program of China Grants (2021YFC2500700, 2023YFC2507102), the National Natural Science Foundation of China (82090010, 82090011), the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences, China (CIFMS2021-I2M-1-049), the Nonprofit Central Research Institute Fund of the Chinese Academy of Medical Sciences (2021RC31002, 2018RC31001), and the National Undergraduate Training Program for Innovation and Entrepreneurship of China (2023zglc06011).

Data availability

The data and materials that support the findings of this study are available within the paper and its Supplementary Materials files, and the mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium (http://proteomecentral.proteomexchange.org) via the iProX partner repository68 with the dataset identifier PXD051382 [43, 44]. All of the data in this article are available upon reasonable request from the corresponding authors.

Declarations

Ethics approval and consent to participate

Baseline questionnaires and clinical data were collected after obtaining voluntary informed consent from all participants. The study protocol was approved by the Institutional Ethics Committee of Sinopharm Tongmei General Hospital (Approval No.: 2017-25-1; 2019-02) and conducted in accordance with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.