Plasma proteomic profiles predict chronic obstructive pulmonary disease up to 16 years before onset: a multi-national, machine learning-guided biomarker discovery study

Plasma proteomic profiles predict chronic obstructive pulmonary disease up to 16 years before onset: a multi-national, machine learning-guided biomarker discovery study | Research Square Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Help Center Sign In Submit a Preprint Cite Share Download PDF Article Plasma proteomic profiles predict chronic obstructive pulmonary disease up to 16 years before onset: a multi-national, machine learning-guided biomarker discovery study Hao Chen, Tong Lin, Jing Feng, Yajie Zhang, Kexin Zhang, Zixun Li, and 13 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8467239/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Chronic obstructive pulmonary disease (COPD) remains a major public health burden, yet early risk prediction remains limited. Using Cox regression and multi-machine learning, we analyzed plasma proteomic data from 36,906 UK Biobank participants and identified nine proteins including GDF15, WFDC2, SCGB1A1, CXCL17, CA14, EDA2R, TNR, AGER, and ODAM. The 9-protein model achieved high accuracy for predicting COPD across different time frames (area under the curve [AUC] = 0.83 overall; 0.86 within 5 years; 0.84 within 10 years; 0.77 beyond 10 years) in a geographically defined UKB testing cohort (n = 15,607), and were further validated in the external EPIC-Norfolk cohort (n = 2,944) with similarly high AUCs. Consistent results were observed in the Southern China cohort (n = 100). Incorporating clinical factors further improved the predictive accuracy, achieving maximum AUCs of 0.89 overall, 0.91 for 5-year prediction, 0.89 for 10-year prediction and 0.83 for prediction beyond 10 years. Individuals with higher baseline protein levels had an 7.29-fold increased COPD risk, and proteomic alterations were detectable up to 16 years before diagnosis. All nine proteins showed significant positive genetic correlations with COPD and causal inference analyses further supported roles for CXCL17 and AGER. These findings demonstrate that plasma proteomics enables accurate long-term COPD risk prediction across diverse populations, provides new insights into disease mechanisms, and supports early identification of high-risk individuals for targeted prevention. Health sciences/Biomarkers/Predictive markers Health sciences/Diseases/Respiratory tract diseases Biological sciences/Computational biology and bioinformatics/Machine learning Proteomics COPD Predictive model Biomarker Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Chronic obstructive pulmonary disease (COPD), one of the main causes of respiratory death worldwide, is a progressive respiratory condition marked by persistent airflow restriction and persistent airway inflammation 1 – 4 . With a rising prevalence among adults over 50, COPD imposes a significant public health burden, contributing to increase disability-adjusted life years (DALYs) annually, which was predicted to rank the fourth position in 2050 5–7 . Current diagnostic criteria rely on post-bronchodilator spirometry (FEV 1 /FVC < 0.70), which often leads to delayed diagnosis, limiting the effectiveness of therapeutic interventions to largely palliative care rather than prevention 8 – 11 . Although biological alterations precede clinical onset by many years, existing biomarkers remain insufficient for accurately predicting COPD development or disease progression 12 – 15 . Therefore, the identification of predictive markers for COPD prior to the onset has become a critical priority for enabling early intervention and improving long-term outcomes 16 , 17 . Plasma proteomic profiling has recently gained traction as a promising approach for biomarker discovery in complex diseases, offering a rich substrate for disease prediction and underlying mechanisms 18 – 20 . In the context of COPD, environmental exposure—such as air pollution and smoking has been shown to upregulate inflammatory proteins 21 , 22 , while circulating proteins like haptoglobin are linked to accelerated decline in pulmonary function 23 – 25 . Several studies have also identified protein candidates for predicting COPD or its subtypes. However, most have been limited by small sample sizes, restricted proteomic coverage, and insufficient external validation, leading to reduced generalizability and suboptimal sensitivity and specificity of the reported biomarkers 26 – 28 . Large-scale longitudinal studies integrating high-throughput proteomics with advanced machine learning approaches are crucial for uncovering early molecular signatures and establishing generalizable, stable models for COPD risk prediction. To address these gaps, we leveraged large-scale plasma proteomics integrated with clinical data from multiple cohorts across diverse populations with up to 16 years of follow-up to identify proteins associated with incident COPD. Using machine learning based analyses, we developed and validated effective proteomic models for COPD prediction and risk stratification. By delineating longitudinal protein changes preceding clinical diagnosis and integrating genetic analyses, we further explored potential causal mechanisms, thereby informing early detection and targeted prevention strategies. Results Characteristics of Participants In the UK Biobank (UKB) cohort, after excluding individuals with a baseline diagnosis of COPD and those with missing protein data, a total of 52,513 participants were included in the study with a median age of 57 years and median follow-up of 13.7 years (interquartile range [IQR]: 12.8-14.3). Participants were divided by recruitment region into training (70%) and testing (30%) cohorts. Among participants in the training cohort (n = 36,906), 1,772 (4.8%) developed COPD during follow-up, including 520 cases within 5 years, 1,244 within 10 years and 528 over 10 years. In the testing cohort (n = 15,607), 906 individuals (5.8%) developed COPD, comprising 267 within 5 years, 649 within 10 years and 297 over 10 years. We found marked differences between the training and testing cohorts in several characteristics, including sex, ethnicity, physical activity, BMI, alcohol consumption and comorbidities including diabetes, hypertension and asthma ( P < 0.05), reflecting demographic and lifestyle variability ( Table 1; Supplementary Table 1 ). In the European Prospective Investigation (EPIC)-Norfolk cohort (n = 2,944), 190 (6.4%) developed COPD including 43 cases within 5 years, 117 within 10 years and 73 beyond 10 years.  Cases were older, more frequently male, and had higher smoking prevalence than non-cases (all P < 0.05; Supplementary Table 2 ). In the Southern China cohort (n=100), 36 participants were diagnosed with COPD and cases were older men, either current or former smokers, and were more likely to have hypertension at baseline. (all P < 0.05; Supplementary Table 3 ). Identification of Proteins Associated with Incident COPD We analyzed 2,736 plasma proteins in the training cohort with different Cox models. In Model 1, after adjusting for age, sex, and ethnicity, 873 proteins were significantly associated with COPD incidence following Bonferroni correction. In Model 2, additional adjustments were made for TDI, BMI, alcohol intake frequency, smoking status, dietary habits, physical activity, education level, income, and comorbidities (diabetes, hypertension and asthma). After Bonferroni correction, 348 proteins remained significant, of which 338 overlapped with the results from Model 1. Among these, WFDC2 exhibited the strongest and most significant association with COPD pathogenesis (hazard ratio [HR] = 2.11, P = 6.71 × 10⁻ 71 ). Elevated levels of CXCL17 (HR = 1.86, P = 1.21 × 10⁻ 55 ), GDF15 (HR = 1.70, P = 1.17 × 10⁻ 40 ) and TNFRSF10B (HR = 1.45, P = 1.95 × 10⁻ 42 ) were also significantly associated with an increased risk of COPD ( Fig.1a and Supplementary Table 4 ). Proteins consistently identified in both Model 1 and Model 2 were then thrown into the functional enrichment analysis, from which we identified biological pathways the positive proteins might involved in, including cytokine-cytokine receptor interaction, viral protein interaction with cytokines and cytokine receptors, and inflammatory response ( Fig.1b and Supplementary Table 5 ). Protein Importance Ranking Proteins positive in both Model 1 and Model 2 were ranked according to their importance in prediction performance. Using the sequential forward selection, the top nine proteins were selected as GDF15, WFDC2, SCGB1A1, CXCL17, CA14, EDA2R, TNR, AGER and ODAM. Incrementally adding these proteins to the prediction model steadily improved its performance, with stabilization observed after the inclusion of all nine proteins. Therefore, these nine proteins were selected for subsequent analyses. SHAP plots were used to measure the importance of these 9 proteins. The effect of each protein was visually interpreted using SHAP values and their association of  incident COPD ( Fig.2a ). Bar plot analysis further quantified the relative importance of individual proteins in model construction, highlighting GDF15 as the most influential contributor ( Fig.2b). Protein–Lifestyle and Lung Function Association To further investigate the association between the top 9 proteins and COPD, we examined their associations with COPD-related modifiable lifestyle factors and lung function parameters. After FDR correction, all four lifestyle factors were significantly associated with multiple proteins ( Fig.2c and Supplementary Table 6 ). Notably, WFDC2, GDF15, CXCL17 and EDA2R were positively associated with smoking, while SCGB1A1, TNR, AGER and ODAM showed predominantly negative associations (all P < 0.05). Linear regression analysis further showed that higher levels of GDF15, WFDC2 and CXCL17 were associated with lower FEV₁, FVC, and FEV₁/FVC, indicating poorer lung function ( Fig.2d and Supplementary Table 7 ). In contrast, SCGB1A1 and CA14 were positively associated with these measures (All P < 0.05). Predictive Accuracy of plasma proteins Using internal 5-fold cross-validation, multi-machine learning and bootstrapping, we further explored the COPD prediction model based on the top nine proteins. For all incident COPD cases, the 9-protein panel achieved an AUC of 0.86 (95% CI: 0.85-0.87) in the training cohort and 0.83 (95% CI: 0.81-0.84) in the testing cohort. The model maintained stable predictive performance across different time frames in the testing cohort, with AUCs of 0.86 within 5 years, 0.84 within 10 years, and 0.77 beyond 10 years ( fig.3a-d ). Integration of the 9-protein panel with clinical data significantly improved predictive performance ( Supplementary Table 8 ). The highest AUC was achieved by a model that combined the proteins with nine demographic factors (e.g., age, smoking status), six serum biomarkers (e.g., C-reactive protein), three spirometry measures (e.g., FEV1/FVC), and the polygenic risk score (PRS). Specifically, the combined model achieved an AUC of 0.89 (95% CI: 0.88–0.90) for overall prediction, 0.91 (95% CI: 0.90–0.93) for 5-year prediction, 0.89 (95% CI: 0.88–0.91) for 10-year prediction, and 0.83 (95% CI: 0.81–0.85) for prediction beyond 10 years, significantly outperforming other models (all DeLong test P < 0.05; Supplementary Table 9 ). Consistent performance was observed in external cohorts. In EPIC-Norfolk cohort, the protein model reached an AUC of 0.81 for all-time prediction and 0.87 for 5-year prediction, 0.85 for 10 years prediction and 0.70 for prediction beyond 10 years ( fig.3e and Supplementary Table 10 ). In the Southern China cohort, the protein model achieved a peak AUC of 0.92( fig.3f ). To assess model stability, we conducted sensitivity analyses across several scenarios and the models’ predictive performance remained consistent across all analyses ( Supplementary Table 11 and Supplementary Fig. 2 ). Given that smoking is a major risk factor for COPD and previous studies have reported the prediction and stratification of COPD across smoking behaviors 29 , we conducted subgroup analyses stratified by smoking status. Participants were classified as never-smokers and ever-smokers (former or current), and both the 9-protein panel and the combined model were evaluated separately in each subgroup. As shown in Supplementary Table 12 and Supplementary Fig. 2 , the 9-protein panel demonstrated comparable predictive performance in never-smokers (AUC = 0.80, 95% CI: 0.78–0.82) and ever-smokers (AUC = 0.82, 95% CI: 0.82–0.83). Incorporation of clinical factors further improved discrimination in both groups, yielding AUCs of 0.91 (95% CI: 0.90–0.92) in never-smokers and 0.89 (95% CI: 0.88–0.90) in ever-smokers. These results indicate that the proteomic model maintains robust and stable performance across smoking status. Association with Disease Progression We assessed the ability of baseline protein levels to discriminate COPD risk across individuals. Participants were stratified into high- and low-risk groups using either the Youden index–derived cutoff for the 9-protein panel or the median value of each individual protein from the training cohort. Using a cutoff of 0.45 for the 9-protein panel, high-risk participants had a markedly increased risk of developing COPD compared with those in the low-risk group in both the training cohort (HR = 11.14, P = 3.98 × 10⁻²⁹⁷) and the testing cohort (HR = 7.29, P = 2.56 × 10⁻ 120 ). This association remained strong in the EPIC-Norfolk cohort, where high-risk individuals were 8.58 times more likely to develop COPD. When stratified by individual proteins, higher baseline levels of GDF15, CXCL17, WFDC2, EDA2R were associated with an elevated risk of COPD, whereas higher levels of SCGB1A1, TNR, CA14, ODAM, and AGER were associated with a reduced risk ( Fig.4 and Supplementary Fig. 3,4,5 ). Trajectories of plasma proteins before COPD onset Using the UKB 1:2 matched case-control dataset and the baseline proteomic measurements, we constructed cross-sectional trajectory curves to characterize the temporal dynamics of nine plasma proteins over a 16-year pre-diagnostic period in individuals later diagnosed with COPD. Comparative analysis between cases and controls revealed distinct temporal patterns. Specifically, plasma levels of GDF15, CXCL17, EDA2R and WFDC2 exhibited accelerated upward trajectories in the COPD group compared to matched controls (all P < 0.01, Mann-Kendall trend test), suggesting their potential utility as early biomarkers for identifying subclinical pathological progression. In contrast, CA14 and TNR displayed significantly attenuated concentrations in COPD cases, with trajectories demonstrating a progressive decline over time (all P < 0.01) while AGER and SCGB1A1 exhibited no discernible temporal trends ( Fig.5 and Supplementary Table 13 ). Genetic Evidence for Nine Proteins in COPD Risk Genetic analyses supported the involvement of the nine candidate proteins in COPD risk. As shown in Supplementary Table 14, LDSC revealed significant genetic correlations between each protein and COPD, with the strongest positive correlations observed for CXCL17 (rg = 0.36, P = 1.19 × 10⁻¹²) and GDF15 (rg = 0.42, P = 4.58 × 10⁻¹⁰), and the strongest inverse correlations for SCGB1A1 (rg = −0.28, P = 1.93 × 10⁻¹¹). MR analysis further supported putative causal relationships for two key proteins. Higher genetically predicted AGER levels were associated with a reduced risk of COPD, suggesting a protective effect (OR = 0.83, P = 0.005), whereas increased CXCL17 levels conferred a notable risk-promoting effect (OR = 1.48, P = 0.013). Discussion In this large-scale prospective study spanning multiple cohorts across different countries, we developed and validated a non-invasive nine-protein panel for predicting incident COPD. The model demonstrated robust predictive performance across multiple time horizons and was independently validated in the EPIC-Norfolk cohort, with consistent performance further confirmed in the Southern China cohort, supporting its robustness across populations. Incorporation of clinical variables further enhanced predictive accuracy, yielding performance superior to that of previously reported models. Beyond prediction, the protein panel enabled effective risk stratification. Using a predefined cutoff value of 0.45, individuals classified as high risk exhibited a markedly increased likelihood of developing COPD in both the UKB testing cohort (HR = 7.29) and the external EPIC-Norfolk cohort (HR = 8.58). Notably, six proteins showed detectable concentration changes up to 16 years before clinical diagnosis, and genetic analyses identified two proteins with potential causal roles in COPD development. Together, these findings underscore the utility of this proteomic approach for early detection, long-term risk stratification, and mechanistic insight into COPD pathogenesis. Compared with existing predictive approaches, our model demonstrates clear advantages.  A recent study 17 reported a prediction model based on 5-protein panel derived from the EPIC-Norfolk and reached a best C-index of 0.77, while Zhang et al. reported a protein model using cox regression and LASSO regression with a C-index of 0.81 but lack of robust external validation, limiting its generalizability 28 . Another study 30 identified GlycA as a key predictive biomarker using the population-based data in Netherlands. Integration of  metabolite with demographic factors yielded an AUC of 0.68. In contrast to these approaches, our model demonstrates consistently higher predictive accuracy, provides longitudinal risk estimation across multiple time horizons (AUC in the UKB testing cohort: 0.83 overall), and shows robust performance across geographically (AUC in the EPIC-Norfolk cohort: 0.80 overall) and ethnically diverse populations (AUC in the Southern China cohort: 0.92). These features enable reliable early risk stratification based on proteomic signatures well before clinical onset. Among the identified proteins, ODAM emerged as a novel finding that has not previously been reported in relation to COPD. ODAM, a protein constituent of calcifying epithelial odontogenic tumor-associated amyloids 31 , has been implicated in the progression of several cancers and may contribute to COPD pathogenesis through pathways such as the PI3K/AKT signaling axis and the inflammatory reaction 32,33 . For the first time, we reported that lower circulating ODAM levels are associated with a higher risk of incident COPD. This association is further reinforced by the negative correlation between smoking and ODAM levels, as well as by the observation that declining ODAM concentrations paralleled reductions in FEV₁/FVC. Together, these findings suggest that ODAM may contribute to COPD development and progression and could represent a potential target for early intervention. Other proteins, including GDF15, WFDC2, SCGB1A1, CXLC17, CA14, EDA2R, TNR and AGER, have been linked to COPD onset and progression 28,34–40 via mechanisms involving airway epithelial senescence 41,42 , smoke-induced inflammation 43 , airway remodeling 44–46 and lung ageing 47,48 . Notably, this study highlights for the first time the predictive relevance of EDA2R in COPD. Associations between protein levels and modifiable lifestyle factors, particularly smoking, as well as their correlation with lung function decline, reveal potential roles of CXCL17, CA14 and TNR in previously unrecognized pathogenic mechanisms, underscoring their value for early diagnosis and intervention. Besides, our genetic analyses provided further evidence supporting the involvement of all nine proteins in COPD. Notably, MR analysis indicated a protective causal effect of higher genetically predicted AGER levels, whereas elevated CXCL17 levels were associated with an increased risk of COPD. These genetic findings were concordant with prospective analyses, in which baseline levels of the same proteins were associated with incident COPD and exhibited early divergence 16 years before diagnosis. Together, the alignment between genetic and longitudinal epidemiological evidence reinforces the mechanistic relevance of these proteins and supports their prioritization as early predictive biomarkers and potential therapeutic targets. This study has several strengths. First, the large population-based cohort with long-term prospective follow-up, together with external validation in geographically and ethnically diverse cohorts supports the robustness and generalizability of the predictive models. Second, the protein panel provides a simple yet sensitive tool for risk stratification, with alterations detectable well before clinical diagnosis, substantially extending the temporal window for early identification and prevention. Finally, the combined use of LDSC and MR enabled us to transcend observational associations and identify proteins with potential causal roles in COPD, strengthening the biological and translational relevance of our findings. Nevertheless, several limitations should be noted. First, the proteomic platform does not encompass the entire human proteome, and potentially important biomarkers may have been missed. Second, as we have validated the model across geographically distinct and multiethnic cohorts, providing substantial evidence of its robustness, external validation in larger populations would strengthen the model’s generalizability. Lastly, the genetic analyses were limited to individuals of European ancestry, and cross-ancestry genetic studies will be needed to confirm the genetic associations identified in our study. In conclusion, we identified and validated a 9-protein panel that accurately predicts COPD incidence up to 16 years before clinical onset and uncovered proteins with potential causal relationships to COPD. This simple and non-invasive model provides optimal early risk stratification and may inform future screening strategies. Beyond risk prediction, the identified protein signatures also capture shared molecular pathways while the genetically supported causal proteins offer promising targets for therapeutic development and mechanistic investigation. Together, these advances have the potential to improve clinical management, personalize prevention strategies, and ultimately enhance outcomes for patients with COPD. Online methods Study Population This study utilized cohorts from the UK Biobank (UKB), the EPIC-Norfolk cohort and the Southern China cohort. UKB, a large prospective cohort study initiated in 2006, comprising biological and medical data from approximately 500,000 UK participants. Participants were recruited between 2006 and 2010 at 22 assessment centers across England, Scotland, and Wales, with all individuals providing informed consent. For this analysis, participants diagnosed with chronic obstructive pulmonary disease (COPD) at or before baseline were excluded. Additionally, individuals lacking plasma proteomics data were omitted. The final study population comprised 52,513 participants, who were subsequently stratified into a training cohort (approximately 70% of the total population) and a validation cohort (30% of the total population), based on recruitment centers of different regions ( Supplementary Table 15 ). Specifically, participants from recruitment centers of Stockport, Oxford, Cardiff, Glasgow, Edinburgh, Reading, Bury, Newcastle, Leeds, Bristol, Barts, Nottingham, Middlesborough, Croydon, Swansea, Wrexham were assigned to the training cohort, while validation cohort includes individuals from Manchester, Stoke, Sheffield, Liverpool, Hounslow, Birmingham. Further details are provided in Table 1 . The validation was conducted in an external cohort. The EPIC-Norfolk cohort is a subset of the European Prospective Investigation into Cancer and Nutrition (EPIC) 49 . This prospective population-based cohort recruited participants from Norfolk, UK, between 1993 and 1997. Baseline data were collected through comprehensive questionnaires and laboratory analyses of blood and urine samples. Following the same inclusion criteria as described above, this study included 2,959 individuals. Baseline characteristics are summarized in Supplementary Table 2 . The Southern China cohort was used to further confirm the generalizability of the model. 36 patients and 64 healthy participants were recruited from Guangdong Provincial People's Hospital. COPD was diagnosed according to the GOLD 3 criteria, requiring relevant respiratory symptoms, a history of exposure to risk factors (e.g., smoking), and spirometric evidence of airflow limitation (FEV₁/FVC < 0.7). Controls were individuals without COPD and no clinical signs of respiratory obstruction. Blood samples and clinical data were collected at enrollment. Outcome Ascertainment As reported before 50 , COPD diagnoses were ascertained using hospital admission records, based on the International Classification of Diseases, 10th Revision (ICD-10) codes: J40, J41.0, J41.1, J42, J43.0, J43.1, J43.2, J43.8, J43.9, J44.0, J44.1, J44.8, and J44.9 ( Supplementary Table 16 ). Follow-up extended from the date of baseline assessment center attendance to the earliest occurrence of COPD diagnosis, death, or the end of follow-up (2022/10/31). To match the 16-year follow-up window, we selected participants with follow-up durations of up to 16 years in the EPIC-Norfolk cohort. In the Southern China cohort, all COPD diagnoses were made by physicians who were blinded to the model development. Plasma Proteomics Plasma proteomics data were derived from the UK Biobank Pharma Proteomics Project (UKB-PPP), a collaborative initiative between UKB and thirteen biopharmaceutical companies aimed at characterizing the plasma proteomic profiles of 54,306 UKB participants. Baseline blood samples were collected using EDTA tubes at 22 assessment centers between 2007 and 2010 with additional samples from participants in the COVID- 19 repeat-imaging study, followed by centrifugation at 4°C for 10 minutes to obtain plasma, which was then promptly stored at -80°C for preservation. Samples were shipped on dry ice to Sweden for analysis using the Olink™ Explore 3072 Proximity Extension Assay (PEA). After rigorous quality control measures, 2,736 unique proteins across eight panels (cardiometabolic, inflammation, neurology, oncology) were measured, with protein levels expressed as Normalized Protein Expression (NPX) values on a log2 scale. In the EPIC‑Norfolk cohort, baseline blood samples were collected into plastic straws and immediately preserved in liquid nitrogen at –196 °C to ensure long-term protein stability. After excluding those failed quality control criteria, participants with high-quality serum samples and complete baseline data on age, sex, BMI, and smoking status were selected to undergo proteomic analysis, which was performed using the Olink™ Explore platform—specifically the 1536 and Expansion panels—covering 2,923 unique protein targets across two nested case–control batches. In the Southern China cohort, baseline blood samples were collected into EDTA tubes and centrifuged at 1,000 g for 15 min at 4 °C, after which plasma was separated, aliquoted, and stored at −80 °C. Only high-quality samples were retained for ELISA analysis, which were conducted according to the protocols (see Supplementary Table 17 for kit details). Protein measurements were converted from optical density (OD) values to concentrations and standardized using the same procedures as in the UKB-PPP. Clinical predictors To further investigate the predictive ability of proteins with other clinical predictors, we incorporated various demographic, serum and spirometry data. The candidate demographic features included age, sex, ethnicity, socio-economic indicators (Townsend Deprivation Index, education status, income and BMI), disease presence (diabetes, hypertension and asthma), lifestyle factors (alcohol consumption, smoking status, exercise). The detailed information of 15 serum indicators (including C-reactive protein) is available in Supplementary Table 18 . The included spirometry measures consists of peak expiratory flow (PEF), forced expiratory volume in 1 second (FEV1), forced vital capacity (FVC) and FEV1/FVC ( Supplementary Table 19 ). Additionally, we included the polygenic risk score (PRS), calculated using a Bayesian method based on pooled GWAS meta-analysis data from external studies with no overlap with the UK Biobank population. Statistical Analysis Identification of Associated Predictors and Model Development Two Cox proportional hazards models were constructed within the training cohort to evaluate associations between plasma protein and incident COPD. Hazard ratios (HRs), 95% confidence intervals (CIs), and P-values were reported. Model 1 was adjusted for age, sex, and ethnicity, while Model 2 incorporated additional covariates, including Townsend Deprivation Index (TDI), body mass index (BMI), alcohol consumption frequency, smoking status, dietary habits, physical activity, education level, income, and comorbidities (diabetes, hypertension, and asthma). All of these covariates were obtained from baseline data and observed missing values less than 20%. Multiple imputation by chained equations were employed to further impute the missing data. Statistical significance was determined using Bonferroni correction (P < 0.05), accounting for the number of proteins tested (n = 2,736). Proteins demonstrating significance in both models were considered robust predictors. To mitigate overfitting, further feature selection was conducted within the training cohort. Significant proteins were input into an untrained Light Gradient Boosting Machine (LGBM) learner using the Fast and Lightweight AutoML (FLAML) framework, where their importance was ranked based on SHapley Additive exPlanations (SHAP) values. Demographic predictors were selected via t-tests, whereas predictors from other categories were initially filtered through Cox models using the same covariate adjustments as the protein analyses, after which all clinical factors went through importance ranking for the best combination for model construction. Predictors were sequentially added to a new AutoML LGBM classifier according to their importance, optimizing predictive performance via internal five-fold cross-validation, balanced class weighting, and a maximum of 200 iterations. The top-ranking proteins were visualized using SHAP plots. Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were used to evaluate model performance. Multi machine learning algorithms including LGBM, XGBoost, Random Forest and Extra Tree were applied to get stable model performances and the LGBM models were finally selected to further analysis. Furthermore, the predictive performance of models incorporating alternative predictor categories was also assessed. The final set of predictors was determined based on optimal cumulative AUCs. Models were trained to predict incident COPD across four time frames: overall (all-time), within 5 years, within 10 years, and beyond 10 years. Model performance was validated in the testing cohort. External validation was conducted using the independent EPIC-Norfolk cohort and further confirmed in the Southern China cohort. Bootstrap resampling (1,000 iterations) and DeLong tests were employed to compare AUC values between models. Biological Pathway Analysis and COPD-related Correlations To investigate the biological relevance of COPD-associated proteins, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were conducted. Statistical significance was defined as a false discovery rate (FDR)-adjusted P < 0.05. To further elucidate the relationship between the top 10 proteins and COPD related lifestyles and measurement, we performed multivariable linear regression analyses assessing associations with obesity, physical activity, smoking status, alcohol consumption frequency, dietary habits and lung function parameters (FEV1, FVC and FEV1/FVC). Models were adjusted for age, sex, and ethnicity, and FDR correction was applied to account for multiple comparisons. Risk Stratification Based on Protein Models Kaplan-Meier survival analyses were conducted to evaluate the prognostic implications of baseline protein levels. Participants were categorized into high- and low-level groups based on the Youden index-derived threshold of protein panel induced by analysis in the training cohort and replicated the stratification in the validation cohort. For individual proteins, the cut-off values were based on the median NPX concentrations from the training cohort. Adjusted Cox models (controlling for age, sex, and ethnicity) were utilized to estimate hazard ratios and 95% CIs for dichotomized protein concentrations. Temporal Dynamics of Plasma Proteins Temporal trends in plasma protein concentrations preceding COPD diagnosis were characterized using a nested case-control study design. Incident COPD cases were identified during follow-up were matched to healthy individuals at a 1:2 ratio based on age, sex, ethnicity, and recruitment region. Controls were assigned observation dates corresponding to their matched cases. Mean protein concentrations over time were visualized using locally weighted scatterplot smoothing (LOWESS) curves. Mann-Kendall trend tests were performed to assess monotonic trends in protein levels among cases and controls. Sensitivity Analyses and Subgroup Analyses To assess model generalizability and robustness, the following sensitivity analyses were conducted: (1) evaluating model performance for predicting COPD only diagnosed before baseline; (2) assessing model performance for predicting COPD diagnosed both before and after baseline; and (3) excluding COPD cases occurring within the first year of follow-up to mitigate potential reverse causation. Subgroup analyses were conducted in populations stratified by smoking status to evaluate the performance of the model across different smoking groups. Linkage disequilibrium score regression (LDSC) We used LDSC 51 to estimate SNP-based heritability for COPD with protein quantitative trait loci (pQTLs) from the UK Biobank 52 . SNPs with INFO score <0.9, MAF <0.01, or absent from the 1000 Genomes Phase III LD reference panel were excluded 53 . Bivariate LDSC were conducted to evaluate the genetic correlations between pQTLs and COPD. A genetic correlation with a P value less than 0.05 was considered significant 54,55 . Mendelian Randomization To scrutinize the potential causal links between cis-acting pQTLs (cis-pQTL) and COPD, we performed bidirectional Mendelian randomization (MR) using “TwoSampleMR” R packages 56 . The cis-pQTL was defined as a pQTL located within 1 Mb of the transcription start sites of the corresponding protein-coding gene. We employed genetic variants strongly associated with the exposure as IVs to examine potential causal relationships between the exposure and the outcome 57 . IVs for MR analyses were selected from GWAS summary statistics using three criteria: (1) genome-wide significance (P < 5×10⁻⁸); (2) linkage disequilibrium clumping via PLINK (r² < 0.001, 10,000 kb window) to ensure SNP independence; and (3) an F-statistic ≥10 to mitigate weak instrument bias, where β is the effect size and SE is the standard error of the SNP-exposure association. Additionally, the MR pleiotropy residual sum and outlier (MR-PRESSO) 58 test was utilized to detect pleiotropic outlier variants. Horizontal pleiotropy was assessed via MR-Egger intercept test (P < 0.05 indicates significant bias). All statistical analyses and visualizations were performed using R (version 4.2.1), Python (version 3). Abbreviations AUC area under the curve BMI body mass index BP biological process CC cellular component COPD Chronic obstructive pulmonary disease CIs confidence intervals CRP C-reactive protein DALYs Disability-adjusted life years FDR false discovery rate GO gene ontology HRs hazard ratios KEGG Kyoto Encyclopedia of Genes and Genomes LGBM light gradient boosting machine LDSC LD score regression MF molecular function MR Mendelian randomization NPX normalized protein expression PEA Proximity Extension Assay PRS polygenic risk score ROC receiver operating characteristic RF random forest SHAP Shapley Additive Explanations TDI Townsend deprivation index UKB-PPP The UK Biobank Pharma Proteomics Project. Declarations Acknowledgements The UK Biobank resource is open to all researchers ( https://www.ukbiobank.ac.uk ). The authors thank the UK Biobank for the access of data, and this research has been conducted under Application Number 83339. The EPIC-Norfolk study (DOI: 10.22025/2019.10.105.00004 ) was supported by the Medical Research Council [MR/N003284/1, MC-UU_12015/1, and MC_UU_00006/1] and Cancer Research UK [C864/A14136]. Genetic analyses in the EPIC-Norfolk study were funded by the Medical Research Council [MC_PC_13048]. This research was conducted using data provided by patients and collected by the NHS as part of their routine care and support. We thank all study participants and the numerous team members at the University of Cambridge who made this research possible. Data availability statement Data used in this study was from the UK Biobank and are available to researchers through an access procedure described at ( https://www.ukbiobank.ac.uk/enable-your-research ). Information on the EPIC-Norfolk cohort is available on its official website ( https://www.epic-norfolk.org.uk/ ), and the data used in this study can be requested through the cohort’s data-sharing application. References Christenson SA, Smith BM, Bafadhel M (2022) Putcha, N. Chronic obstructive pulmonary disease. Lancet 399:2227–2242 Stolz D et al (2022) Towards the elimination of chronic obstructive pulmonary disease: a Lancet Commission. Lancet 400:921–972 Agustí A et al (2023) Global Initiative for Chronic Obstructive Lung Disease 2023 Report: GOLD Executive Summary. Eur Respir J 61:2300239 Agusti A, Singh D, Faner R (2025) Treatment of chronic obstructive pulmonary disease: current pipeline and new opportunities. Nat Rev Drug Discov. https://doi.org/10.1038/s41573-025-01290-6 Boers E et al (2023) Global Burden of Chronic Obstructive Pulmonary Disease Through 2050. JAMA Netw Open 6:e2346598 Varmaghani M et al (2019) Global prevalence of chronic obstructive pulmonary disease: systematic review and meta-analysis. East Mediterr Health J 25:47–57 Wang Z et al (2025) Global, regional, and national burden of chronic obstructive pulmonary disease and its attributable risk factors from 1990 to 2021: an analysis for the Global Burden of Disease Study 2021. Respir Res 26:2 Fazleen A, Wilkinson T, Early (2020) COPD: current evidence for diagnosis and management. Ther Adv Respir Dis 14:1753466620942128 Hoesterey D et al (2019) Spirometric Indices of Early Airflow Impairment in Individuals at Risk of Developing COPD: Spirometry Beyond FEV1/FVC. Respir Med 156:58–68 Brandsma C, Van den Berge M, Hackett T, Brusselle G, Timens W (2020) Recent advances in chronic obstructive pulmonary disease pathogenesis: from disease mechanisms to precision medicine. J Pathol 250:624–635 Svigelj R, de Marco A (2024) Biological and technical factors affecting the point-of-care diagnostics in not-oncological chronic diseases. Biosens Bioelectron 264:116669 Fermont JM et al (2019) Biomarkers and clinical outcomes in COPD: a systematic review and meta-analysis. Thorax 74:439–446 Joglekar MM et al (2025) The Calprotectin Fragment, CPa9-HNE, Is a Plasma Biomarker of Mild Chronic Obstructive Pulmonary Disease. Cells 14:1155 Guzman DE et al (2025) Proteomic discovery analysis of quantitatively assessed emphysema in the general population. The MESA Lung Study. Respir Res 26:236 Serban KA, Pratte KA, Bowler RP (2021) Protein Biomarkers for COPD Outcomes. Chest 159:2244–2253 Suryadevara R et al Blood-based Transcriptomic and Proteomic Biomarkers of Emphysema. Am J Respir Crit Care Med 209, 273–287 Carrasco-Zanini J et al (2024) Proteomic prediction of diverse incident diseases: a machine learning-guided biomarker discovery study using data from a prospective cohort study. Lancet Digit Health 6:e470–e479 Deng Y-T et al (2025) Atlas of the plasma proteome in health and disease in 53,026 adults. Cell 188:253–271e7 Guo Y et al (2024) Plasma proteomic profiles predict future dementia in healthy adults. Nat Aging 4:247–260 Gan Y-H et al (2025) Large-scale proteomic analyses of incident Parkinson’s disease reveal new pathophysiological insights and potential biomarkers. Nat Aging 5:642–657 Ryu MH et al (2022) Impact of Exposure to Diesel Exhaust on Inflammation Markers and Proteases in Former Smokers with Chronic Obstructive Pulmonary Disease: A Randomized, Double-blinded, Crossover Study. Am J Respir Crit Care Med 205:1046–1052 Wang T et al (2025) Ambient air pollution exposure, mediating biomarkers and risk of COPD: a cohort study and meta-analysis. Eur Respir Rev 34:250055 Zhai J et al (2024) Circulating biomarkers of airflow limitation across the life span. J Allergy Clin Immunol 153:1692–1703 Ran X et al (2025) Elevated MMP-9 Is Associated With Accelerated Lung Function Decline and COPD Development: A Prospective Cohort Study and Mendelian Randomisation Analysis. Respirology https://doi.org/10.1111/resp.70099 doi:10.1111/resp.70099 Zhang Y-H et al (2025) Proteomic biomarkers of emphysema-predominant and non-emphysema-predominant chronic obstructive pulmonary disease. eBioMedicine 117:105800 Serban KA, Pratte KA, Bowler RP (2021) Protein Biomarkers for COPD Outcomes. Chest 159:2244–2253 Regard L, Roche N, Burgel P-R The Ongoing Quest for Predictive Biomarkers in Chronic Obstructive Pulmonary Disease. Am J Respir Crit Care Med 208, 511–513 Zhang Y et al (2025) Proteomics-based risk prediction and drug targets identification for chronic obstructive pulmonary disease. Thorax thorax-2024-222397. 10.1136/thorax-2024-222397 He Y et al (2023) Prediction and stratification of longitudinal risk for chronic obstructive pulmonary disease across smoking behaviors. Nat Commun 14:8297 Prokić I et al (2020) A cross-omics integrative study of metabolic signatures of chronic obstructive pulmonary disease. BMC Pulm Med 20:193 Zhu S et al (2022) The versatile roles of odontogenic ameloblast-associated protein in odontogenesis, junctional epithelium regeneration and periodontal disease. Front Physiol 13:1003931 Yu M et al (2016) Odontogenic ameloblast-associated protein (ODAM) inhibits human colorectal cancer growth by promoting PTEN elevation and inactivating PI3K/AKT signaling. Biomed Pharmacother 84:601–607 Yamaguchi K et al (2023) Identification of odontogenic ameloblast associated as a novel target gene of the Wnt/β-catenin signaling pathway. Cancer Sci 114:948–960 Sanders JL et al (2021) The Association of Aging Biomarkers, Interstitial Lung Abnormalities, and Mortality. Am J Respir Crit Care Med 203:1149–1157 Husebø GR et al (2017) Growth differentiation factor-15 is a predictor of important disease outcomes in patients with COPD. Eur Respir J 49 Wan Y, Fu J (2024) GDF15 as a key disease target and biomarker: linking chronic lung diseases and ageing. Mol Cell Biochem 479:453–466 Guerra S et al (2015) The relation of circulating CC16 to lung function growth, decline, and development of COPD across the lifespan. Lancet Respir Med 3:613–620 Martinu T, Todd JL, Gelman AE, Guerra S, Palmer SM (2023) Club Cell Secretory Protein in Lung Disease: Emerging Concepts and Potential Therapeutics. Annu Rev Med 74:427–441 Zhai J et al (2019) Club Cell Secretory Protein Deficiency Leads to Altered Lung Function. Am J Respir Crit Care Med 199:302–312 Regan EA et al (2019) Omics and the Search for Blood Biomarkers in Chronic Obstructive Pulmonary Disease. Insights from COPDGene. Am J Respir Cell Mol Biol 61:143–149 Wu Q et al (2016) Cigarette Smoke Induces Human Airway Epithelial Senescence via Growth Differentiation Factor 15 Production. Am J Respir Cell Mol Biol 55:429–438 Strickson S et al (2023) Oxidised IL-33 drives COPD epithelial pathogenesis via ST2-independent RAGE/EGFR signalling complex. Eur Respir J 62:2202210 Verhamme FM et al (2017) Elevated GDF-15 contributes to pulmonary inflammation upon cigarette smoke exposure. Mucosal Immunol 10:1400–1411 Zhan Y et al (2022) Human epididymis protein 4 aggravates airway inflammation and remodeling in chronic obstructive pulmonary disease. Respir Res 23:120 Bingle L et al (2020) Expression and function of murine WFDC2 in the respiratory tract. 05.05.079293 Preprint at https://doi.org/10.1101/2020.05.05.079293 (2020) Allam VSRR et al (2021) RAGE and TLR4 differentially regulate airway hyperresponsiveness: Implications for COPD. Allergy 76:1123–1135 De Vries M et al (2018) Lung tissue gene-expression signature for the ageing lung in COPD. Thorax 73:609–617 Barbera MC et al (2025) Increased ectodysplasin-A2-receptor EDA2R is a ubiquitous hallmark of aging and mediates parainflammatory responses. Nat Commun 16:1898 Riboli E et al (2002) European Prospective Investigation into Cancer and Nutrition (EPIC): study populations and data collection. Public Health Nutr 5:1113–1124 Li J et al (2024) Associations between long-term night shift work and incidence of chronic obstructive pulmonary disease: a prospective cohort study of 277,059 UK Biobank participants. BMC Med 22:16 Sun BB et al (2023) Plasma proteomic associations with genetics and health in the UK Biobank. Nature 622:329–338 Bulik-Sullivan B et al (2015) An Atlas of Genetic Correlations across Human Diseases and Traits. Nat Genet 47:1236–1241 Auton A et al (2015) A global reference for human genetic variation. Nature 526:68–74 Sudlow C et al (2015) UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age. PLoS Med 12:e1001779 Kurki MI et al (2023) FinnGen provides genetic insights from a well-phenotyped isolated population. Nature 613:508–518 Hemani G et al The MR-Base platform supports systematic causal inference across the human phenome. eLife 7, e34408 Burgess S, Foley CN, Zuber V (2018) Inferring Causal Relationships Between Risk Factors and Outcomes from Genome-Wide Association Study Data. Annu Rev Genomics Hum Genet 19:303–327 Verbanck M, Chen C-Y, Neale B, Do R (2018) Publisher Correction: Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases. Nat Genet 50:1196–1196 Tables Table 1 Baseline characteristics of the UKB cohort. Overall Training cohort Testing cohort P value n = 52,513 n = 36,906 n = 15,607 Age, mean (SD), years 56.76 (8.21) 56.78 (8.18) 56.72 (8.29) 0.509 TDindex, mean (SD) -1.20 (3.18) -1.33 (3.19) -0.89 (3.13) <0.001 BMI, kg/m², mean (SD) 27.46 (4.79) 27.42 (4.76) 27.53 (4.85) 0.022 Income, mean (SD) 1.30 (0.56) 1.30 (0.56) 1.28 (0.55) <0.001 Sex <0.001 Female, n(%) 28,358 (54.0) 20,198 (54.7) 8,160 (52.3) Male, n(%) 24,155 (46.0) 16,708 (45.3) 7,447 (47.7) Physical activity 0.185 No regular physical activity, n(%) 9,933 (18.9) 6,926 (18.8) 3,007 (19.3) Regular physical activity, n(%) 42,580 (81.1) 29,980 (81.2) 12,600 (80.7) Ethnicity <0.001 Other, n(%) 3,331 (6.3) 1,956 (5.3) 1,375 (8.8) White, n(%) 49,182 (93.7) 34,950 (94.7) 14,232 (91.2) Smoking status 0.484 Never smoker, n(%) 28,697 (54.6) 20,231 (54.8) 8,466 (54.2) Previous smoker, n(%) 18,302 (34.9) 12,814 (34.7) 5,488 (35.2) Current smoker, n(%) 5,514 (10.5) 3,861 (10.5) 1,653 (10.6) Alcohol consumption <0.001 Daily or almost daily, n(%) 10,619 (20.2) 7,554 (20.5) 3,065 (19.6) Three or four times a week, n(%) 11,848 (22.6) 8,426 (22.8) 3,422 (21.9) Once or twice a week, n(%) 13,640 (26.0) 9,628 (26.1) 4,012 (25.7) One to three times a month, n(%) 5,746 (10.9) 4,065 (11.0) 1,681 (10.8) Special occasions only, n(%) 6,171 (11.8) 4,261 (11.5) 1,910 (12.2) Never, n(%) 4,489 (8.5) 2,972 (8.1) 1,517 (9.7) Education status 0.002 High, n(%) 35,350 (67.3) 24,692 (66.9) 10,658 (68.3) Low, n(%) 17,163 (32.7) 12,214 (33.1) 4,949 (31.7) Daily diet 0.489 Poor diet, n(%) 32,734 (62.3) 23,041 (62.4) 9,693 (62.1) healthy diet , n(%) 19,779 (37.7) 13,865 (37.6) 5,914 (37.9) Diabetes (%) 4,879 (9.3) 3,270 (8.9) 1,609 (10.3) <0.001 Hypertension (%) 17,490 (33.3) 11,985 (32.5) 5,505 (35.3) <0.001 Asthma (%) 5,041 (9.6) 3,467 (9.4) 1,574 (10.1) 0.015 Continuous data are presented as mean (Standard Deviation) and categorical variables as number (percentage). The 52,896 participants were randomly divided into training and testing cohorts according to the UK Biobank recruitment centers in a ratio of 7:3 roughly (See Supplementary Table14 for further information). Differences between the cohorts were compared using Student’s t-test for continuous variables and Pearson’s chi-squared test for discrete variables. BMI: body mass index; TD index: Townsend Deprivation Index. P-values were derived from two-sided tests without multiple comparison adjustments. Additional Declarations There is NO Competing Interest. Supplementary Files supplementaryfigure.pdf Supplementary figure SupplementaryTables2.doc Supplementary table Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information

Related documents

• A Study to Evaluate Biomarkers in Chronic Obstructive Pulmonary Disease#911982
• Explainable Plasma Proteomics-Based Machine Learning for Osteoporosis Diagnosis, Prognosis, and Protein Biomarker Discovery in the UK Biobank.#621048
• Vascular-related proteomic signatures in COPD with suspected pulmonary hypertension as predictors of FEV₁ impairment.#162030
• Metabolomic Analysis of Exhaled Breath Condensate in Asthmatics#1093405
• Prognosis and Treatment of COPD in Primary Care-use of Biomarkers#851049