Topological data analysis identifies molecular phenotypes of idiopathic pulmonary fibrosis.

BACKGROUND: Idiopathic pulmonary fibrosis (IPF) is a debilitating, progressive disease with a median survival time of 3-5 years. Diagnosis remains challenging and disease progression varies greatly, suggesting the possibility of distinct subphenotypes.

METHODS AND RESULTS: We analysed publicly available peripheral blood mononuclear cell expression datasets for 219 IPF, 411 asthma, 362 tuberculosis, 151 healthy, 92 HIV and 83 other disease samples, totalling 1318 patients. We integrated the datasets and split them into train (n=871) and test (n=477) cohorts to investigate the utility of a machine learning model (support vector machine) for predicting IPF. A panel of 44 genes predicted IPF in a background of healthy, tuberculosis, HIV and asthma with an area under the curve of 0.9464, corresponding to a sensitivity of 0.865 and a specificity of 0.89. We then applied topological data analysis to investigate the possibility of subphenotypes within IPF. We identified five molecular subphenotypes of IPF, one of which corresponded to a phenotype enriched for death/transplant. The subphenotypes were molecularly characterised using bioinformatic and pathway analysis tools identifying distinct subphenotype features including one which suggests an extrapulmonary or systemic fibrotic disease.

CONCLUSIONS: Integration of multiple datasets, from the same tissue, enabled the development of a model to accurately predict IPF using a panel of 44 genes. Furthermore, topological data analysis identified distinct subphenotypes of patients with IPF which were defined by differences in molecular pathobiology and clinical characteristics.