Lung cancer risk discrimination of prediagnostic proteomics measurements compared with existing prediction tools

Abstract

Background: We sought to develop a proteomics-based risk model for lung cancer and evaluate its risk-discriminatory performance in comparison with a smoking-based risk model (PLCOm2012) and a commercially available autoantibody biomarker test.

Methods: We designed a case-control study nested in 6 prospective cohorts, including 624 lung cancer participants who donated blood samples at most 3 years prior to lung cancer diagnosis and 624 smoking-matched cancer free participants who were assayed for 302 proteins. We used 470 case-control pairs from 4 cohorts to select proteins and train a protein-based risk model. We subsequently used 154 case-control pairs from 2 cohorts to compare the risk-discriminatory performance of the protein-based model with that of the Early Cancer Detection Test (EarlyCDT)-Lung and the PLCOm2012 model using receiver operating characteristics analysis and by estimating models' sensitivity. All tests were 2-sided.

Results: The area under the curve for the protein-based risk model in the validation sample was 0.75 (95% confidence interval [CI] = 0.70 to 0.81) compared with 0.64 (95% CI = 0.57 to 0.70) for the PLCOm2012 model (Pdifference = .001). The EarlyCDT-Lung had a sensitivity of 14% (95% CI = 8.2% to 19%) and a specificity of 86% (95% CI = 81% to 92%) for incident lung cancer. At the same specificity of 86%, the sensitivity for the protein-based risk model was estimated at 49% (95% CI = 41% to 57%) and 30% (95% CI = 23% to 37%) for the PLCOm2012 model.

Conclusion: Circulating proteins showed promise in predicting incident lung cancer and outperformed a standard risk prediction model and the commercialized EarlyCDT-Lung.

Figure 1.

Proportion of proteins selected in 500 training datasets by LASSO logistic regression model. Proteins selected more than 400 times are marked as black. LASSO = least absolute shrinkage and selection operator; CEACAM5 = carcinoembryonic antigen-related cell adhesion molecule 5; MMP12 = macrophage metalloelastase; IL6 = interleukin 6; CDCP1 = CUB domain-containing protein 1; CASP-8 = caspase-8; CXCL13 = C-X-C motif chemokine 13; IGFBP-1 = insulin-like growth factor-binding protein 1; CXL17 = C-X-C motif chemokine 17; MUC-16 = mucin-16; TNFSF13B = tumor necrosis factor ligand superfamily member 13B; CXCL9 = C-X-C motif chemokine 9; S100A11 = Protein S100-A11; GDF-15 = growth/differentiation factor 15; IGFBP-2 = insulin-like growth factor-binding protein 2; OSM = oncostatin-M; SYND1 = syndecan-1; WFDC2 = WAP four-disulfide core domain protein 2; CHI3L1 = chitinase-3-like protein 1; TGF-alpha = transforming growth factor alpha; LAMP3 = lysosome-associated membrane glycoprotein 3; MK = midkine; U-PAR = urokinase plasminogen activator surface receptor.

Figure 2.

Comparison of ROC curves for the PLCOm2012 model and protein-based risk model in the validation set. The ROCs and associated AUC estimates reflect the residual risk-discriminatory performance of the risk models after accounting for age, sex, and smoking status (matching factors in the case-control study). AUC = area under the curve; EarlyCDT = Early Cancer Detection Test; ROC = receiver operating characteristics.