Protein biomarkers of interstitial lung abnormalities in relatives of patients with pulmonary fibrosis

Abstract

Rationale: First-degree relatives of patients with pulmonary fibrosis (referred to here as relatives) are at high risk for interstitial lung abnormalities (ILA), highlighting the need for biomarkers for risk prediction. We aimed to identify blood proteins associated with and predictive of ILA among relatives of patients with pulmonary fibrosis.

Methods: Relatives enrolled in two independent cohorts had protein levels measured using an aptamer-based proteomic platform. ILA were assessed with computed tomography scans as per Fleischner Society recommendations. Protein associations with ILA were assessed using regression. Significant proteins were used with clinical variables to detect ILA.

Results: Of 237 relatives from two independent cohorts, 26% had ILA. Seven proteins were associated with ILA in the discovery cohort after false discovery rate adjustment, and all remained significant after adjusting for age, gender and smoking status. Six of the seven proteins were significantly associated in the validation cohort, including growth differentiation factor 15, surfactant protein D and surfactant protein B. In a multivariable model, six proteins combined with basic demographics in the discovery cohort had an area under the curve of 0.92 (0.88 in the validation cohort). Least absolute shrinkage and selection operator modelling identified three proteins and age as predictors, with an area under the curve of 0.89 in the validation cohort. When applied to the combined cohorts, this simple model would reduce the need for computed tomography imaging in one of every three relatives screened.

Conclusion: Peripheral blood proteins are associated with ILA in relatives of patients with pulmonary fibrosis and can be used to detect them. Our findings demonstrate the potential use of blood biomarkers in this high-risk group and suggest molecular targets for future investigation.

Overview of the study. AUC: area under the receiver operating characteristic curve; CT: computed tomography; GDF15: growth differentiation factor 15; ILA: interstitial lung abnormalities; SFTPB: surfactant protein B; SFTPD: surfactant protein D.

FIGURE 1

Discovery of single protein associations with interstitial lung abnormalities using univariable logistic regression models. CDCP1: CUB domain-containing protein 1; CLIC3: chloride intracellular channel 3; GALNS: galactosamine (N-acetyl)-6-sulfatase; GCHFR: GTP cyclohydrolase I feedback regulatory protein; GDF15: growth differentiation factor 15; ICAM5: intercellular adhesion molecule 5; MMP12: matrix metallopeptidase 12; NPW: neuropeptide W; PXDN: peroxidasin; RELT: RELT tumour necrosis factor receptor; SFN: stratifin; SFTPB: surfactant protein B; SFTPD: surfactant protein D; TIMP1: TIMP metallopeptidase inhibitor 1; TREM1: triggering receptor expressed on myeloid cells 1; WFDC2: WAP four-disulfide core domain protein 2.

FIGURE 2

Proteins and demographics for the detection of interstitial lung abnormalities (ILA). Receiver operating characteristic (ROC) curves and area under the ROC curves (AUC) for the detection of ILA. a) Logistic regression models were used in the Clinical Genetics and Screening for Pulmonary Fibrosis (CGS-PF) cohort with clinical data and/or the six proteins significantly associated with ILA as the training set. The multiprotein model with the six validated proteins was a significant improvement over the model with demographics alone (p=0.003). The final model with demographics and the six proteins had an AUC of 0.92. b) The models established in CGS-PF were then tested in the independent Colorado cohort. The final model with demographics and the six proteins had an AUC of 0.88.

FIGURE 3

Simplified model for the detection of interstitial lung abnormalities (ILA). Receiver operating characteristic (ROC) curves and area under the ROC curves (AUC) of the simplified model derived from machine learning for the detection of ILA. Variable selection using least absolute shrinkage and selection operator (LASSO) modelling of the discovery cohort gave four variables shown in the table inset with coefficients. These four variables were used in multivariable logistic regression for the detection of ILA in the discovery cohort (red curve). This model was then applied to the validation cohort to generate an estimate of ILA detection (blue curve) and gave an AUC of 0.89. A table of local maximas of the ROC curve from the validation cohort and corresponding test performance characteristics in the combined cohorts are shown in supplementary table E6. GDF15: growth differentiation factor 15; SFTPB: surfactant protein B; SFTPD: surfactant protein D.