A protein risk score for all-cause and respiratory-specific mortality in non-Hispanic white and African American individuals who smoke

Abstract

Protein biomarkers are associated with mortality in cardiovascular disease, but their effect on predicting respiratory and all-cause mortality is not clear. We tested whether a protein risk score (protRS) can improve prediction of all-cause mortality over clinical risk factors in smokers. We utilized smoking-enriched (COPDGene, LSC, SPIROMICS) and general population-based (MESA) cohorts with SomaScan proteomic and mortality data. We split COPDGene into training and testing sets (50:50) and developed a protRS based on respiratory mortality effect size and parsimony. We tested multivariable associations of the protRS with all-cause, respiratory, and cardiovascular mortality, and performed meta-analysis, area-under-the-curve (AUC), and network analyses. We included 2232 participants. In COPDGene, a penalized regression-based protRS was most highly associated with respiratory mortality (OR 9.2) and parsimonious (15 proteins). This protRS was associated with all-cause mortality (random effects HR 1.79 [95% CI 1.31-2.43]). Adding the protRS to clinical covariates improved all-cause mortality prediction in COPDGene (AUC 0.87 vs 0.82) and SPIROMICS (0.74 vs 0.6), but not in LSC and MESA. Protein-protein interaction network analyses implicate cytokine signaling, innate immune responses, and extracellular matrix turnover. A blood-based protein risk score predicts all-cause and respiratory mortality, identifies potential drivers of mortality, and demonstrates heterogeneity in effects amongst cohorts.

Fig. 1

Schematic of study design. COPDGene, Genetic Epidemiology of COPD study. MESA, Multi-Ethnic Study of Atherosclerosis. SPIROMICS, SubPopulations and InteRmediate Outcomes Measures in COPD Study. ProtRS, Protein risk score.

Fig. 2

Forest plot demonstrating the association of the protein risk score (ProtRS) with all-cause mortality in testing cohorts. See Table 1 for abbreviations.

Fig. 3

Receiver-operating-characteristic-curve (ROC) and area-under-the-ROC-curve (AUC) analysis in each cohort. ProtRS, protein risk score. Clinical model includes the reduced clinical model with age, sex, race, and smoking variables (pack-years or ever smoking status, depending on cohort).