Diagnostic MicroRNA Signatures to Support Classification of Pulmonary Hypertension

Abstract

Background: Patients with pulmonary hypertension (PH) are classified based on disease pathogenesis and hemodynamic drivers. Classification informs treatment. The heart failure biomarker NT-proBNP (N-terminal pro-B-type natriuretic peptide) is used to help inform risk but is not specific to PH or sub-classification groups. There are currently no other biomarkers in clinical use to help guide diagnosis or risk.

Methods: We profiled a retrospective cohort of 1150 patients from 3 expert centers with PH and 334 non-PH symptomatic controls (disease controls) from the United Kingdom to measure circulating levels of 650 microRNAs (miRNAs) in serum. NT-proBNP (ELISA) and 326 well-detected miRNAs (polymerase chain reaction) were prioritized by feature selection using multiple machine learning models. From the selected miRNAs, generalized linear models were used to describe miRNA signatures to differentiate PH and pulmonary arterial hypertension from the disease controls, and pulmonary arterial hypertension, PH due to left heart disease, PH due to lung disease, and chronic thromboembolic pulmonary hypertension from other forms of PH. These signatures were validated on a UK test cohort and independently validated in the prospective CIPHER study (A Prospective, Multicenter, Noninterventional Study for the Identification of Biomarker Signatures for the Early Detection of Pulmonary Hypertension) comprising 349 patients with PH and 93 disease controls.

Results: NT-proBNP achieved a balanced accuracy of 0.74 and 0.75 at identifying PH and pulmonary arterial hypertension from disease controls with a threshold of 254 and 362 pg/mL, respectively but was unable to sub-categorize PH subgroups. In the UK cohort, miRNA signatures performed similarly to NT-proBNP in distinguishing PH (area under the curve of 0.7 versus 0.78), and pulmonary arterial hypertension (area under the curve of 0.73 versus 0.79) from disease controls. MicroRNA signatures outperformed NT-proBNP in distinguishing PH classification groups. External testing in the CIPHER cohort demonstrated that miRNA signatures, in conjunction with NT-proBNP, age, and sex, performed better than either NT-proBNP or miRNAs alone in sub-classifying PH.

Conclusions: We suggest a threshold for NT-proBNP to identify patients with a high probability of PH, and the subsequent use of circulating miRNA signatures to help differentiate PH subgroups.

Keywords: biomarkers; early diagnosis; machine learning; miRNA; pulmonary hypertension.

Figure 1.

Study design. The combined number of patients with pulmonary hypertension (Combined PH), pulmonary arterial hypertension (PAH), PH due to left heart disease (PH-LHD), PH due to lung disease (PH-Lung), PH caused by pulmonary artery obstruction (PH Obstruction), and PH caused by miscellaneous factors (PH-Misc) subgroups and disease controls with no PH from the retrospective UK and prospective CIPHER (A Prospective, Multicenter, Noninterventional Study for the Identification of Biomarker Signatures for the Early Detection of Pulmonary Hypertension) Independent Test cohort are shown. The segregation of the Retrospective UK cohort into the Discovery/Training cohort and a Hold-out Validation cohort is shown. The boxes labeled 1 to 4 highlight the sequential steps taken for microRNA (miRNA) feature selection (1), miRNA signature model building using a generalized linear model (GLM; 2), internal testing in a held-out validation cohort (3), and independent external validation in the CIPHER study samples (4).

Figure 2.

Utility of NT-proBNP (N-terminal pro-B-type natriuretic peptide) to classify pulmonary hypertension and subgroups. A, Box plots show the distribution of NT-proBNP expression in serum for combined pulmonary hypertension (PH), pulmonary arterial hypertension (PAH), PH due to left heart disease (PH-LHD), PH due to lung disease (PH-Lung), chronic thromboembolic PH (CTEPH) subgroups, and no PH disease controls (DCs) in the Discovery and Test populations from the retrospective UK cohort. B, Receiver operating characteristic (ROC) with area under the curve (AUC) for the use of NT-proBNP to classify PH and PAH from DC, and PAH, PH-LHD, PH-Lung, and CTEPH from a combined population of other PH groups trained on the Discovery and then tested on the internal UK hold-out Test population. C, Table shows the threshold of NT-proBNP derived from the UK Test population for a 75% sensitivity.

Figure 3.

MicroRNA (miRNA) features selected for each signature and their coefficients. Heat map showing the miRNAs included in each pulmonary hypertension (PH), pulmonary arterial hypertension (PAH), PH due to left heart disease (PH-LHD), PH due to lung disease (PH-Lung), chronic thromboembolic PH (CTEPH) subgroups signatures, with their coefficients.

Figure 4.

Pulmonary hypertension microRNA (miRNA) signature feature selection and coefficients from the UK Retrospective study. Histograms show the miRNAs selected for each signature and their coefficients within the generalized linear regression model used to classify (A) pulmonary hypertension (PH) vs no PH disease controls (DCs); (B) pulmonary arterial hypertension (PAH) vs DC; (C) PAH vs PH; (D) PH due to left heart disease (PH-LHD) vs PH; (E) PH due to lung disease (PH-Lung) vs PH; (F) chronic thromboembolic PH (CTEPH) vs PH; and (G) PAH vs CTEPH.

Figure 5.

MicroRNA (miRNA) signature performance in the UK Hold-out Validation cohort. Receiver operating characteristic (ROC) with area under the curve (AUC) for the performance of the miRNA signatures (solid line) to classify (A) pulmonary hypertension (PH) vs no PH disease controls (DCs); (B) pulmonary arterial hypertension (PAH) vs DC; (C) PAH vs PH; (D) PH due to left heart disease (PH-LHD) vs PH; (E) PH due to lung disease (PH-Lung) vs PH; (F) chronic thromboembolic PH (CTEPH) vs PH compared with NT-proBNP (N-terminal pro-B-type natriuretic peptide; dashed line). P values from a DeLong test comparing the performance of the miRNA signature to NT-proBNP are shown for each. The dotted red line represents the 75% sensitivity threshold.

Figure 6.

Flowchart and performance of UK microRNA (miRNA) diagnostic signatures in the CIPHER study. Flowchart summarizing the results obtained and highlighting the utility of miRNA signatures with NT-proBNP (N-terminal pro-B-type natriuretic peptide) to help guide the diagnosis of pulmonary hypertension (PH) and PH subgroups with blood-derived biomarkers from symptomatic patients to pulmonary arterial hypertension (PAH), PH due to left heart disease (PH-LHD), PH due to lung disease (PH-Lung), and chronic thromboembolic PH (CTEPH).