Precision Medicine in Pulmonary Arterial Hypertension: A First Step

It is increasingly recognized that there are complexities within endophenotypes that contribute to clinical heterogeneity observed among patients with the same disease, including pulmonary arterial hypertension (PAH). These hidden features likely underlie the observed variability in responses to therapeutics as well as differences in short and longer-term clinical outcomes within this patient population. This heterogeneity has catalyzed a movement to utilize precision medicine for the discovery of novel PAH sub-endophenotypes through deep clinical and panomics (genomics, transcriptomics, metabolomics, and proteomics) phenotyping to redefine patient subgroups.

Precision medicine is a strategy that focuses on individual differences at both the molecular level, through state-of-the-art panomics profiling, in concert with high level data derived from medical testing, lifestyle, and environmental exposures to determine a person’s health or disease phenotype. Precision medicine differs substantially from our current clinical practice of reductionism in medicine. Reductionism is our assumption that all patients with similar clinical symptoms and physical signs have the same or similar disease phenotype(s), and should receive the same medical therapies resulting in equivalent clinical responses and outcomes¹. Precision medicine as a discipline overcomes these limitations by focusing on a unique patient profile created through deep phenotyping, rather than relying on that of the average disease group population, to select an appropriate and optimal therapy². This approach also differs somewhat from the concept of personalized medicine, which may utilize a patient’s phenotypic data to create an individualized treatment plan that is based on a known biomarker for response to therapy.

The appeal of precision medicine as a method towards improved disease classification, treatment selection, and prognostication lies in the integration and analysis of complex clinical and molecular datasets, which are then used to construct an individual’s disease pathophenotype. Large scale panomics (gene expression, transcriptomics, epigenomics, metabolomics, proteomics, and microbiomics) profiling is now possible on a scale that is more amenable to routine use in population-based studies¹. This high-level molecular characterization coupled with environmental exposure history and integrated with clinical, laboratory, imaging, and historical electronic health record data form a composite profile of an individual at a granular level. Once compiled, these datasets are amenable to unique analytical methodologies, such as network analysis, or can inform artificial intelligence algorithms, for discovery of novel relationships between previously unknown or unrecognized factors in an unbiased manner.

In the current issue of Circulation Research, Sweatt et al³ use precision medicine to define subgroups of patients with PAH based on immune profiles, an area that is understudied in the disease. By utilizing machine learning (consensus clustering) algorithms, immune sub-phenotypes were discovered within the cohort of PAH patients in an unbiased manner. Although the selection of immune markers was limited to a panel of 48 cytokines, chemokines, and growth factors on the basis of the platform selected, the ability of these factors to define 4 different clusters is intriguing. Further examination of the cohort revealed that patients do not segregate within clusters on the basis of standard clinical phenotyping metrics. This is especially interesting given that patients treated with immune-modulating therapies were included in the study, and they did not segregate to a unique cluster either. The identified clusters were associated with distinctive clinical risk profiles and assignment to a cluster predicted long-term outcomes. The strength of the clusters and the overall findings, despite the small sample size and restricted immune profiling, is supported fully by replication in a second validation cohort. This is similar to what was observed in other studies when patients at-risk for pulmonary hypertension were assigned to clusters on the basis of clinical exercise parameters⁴. Within each of the immune clusters, network analysis was employed to discover a central core group of proteins based on network centrality scores. Central core groups were identified for 3 of the 4 clusters; however, it remains to be determined if the core group immune markers could serve as a panel that would discriminate and segregate patients into the defined clusters or identify relevant drug targets.

This study advances the role of precision medicine in pulmonary hypertension although future investigations with larger sample sizes and expression profiling may test the durability of the findings. As noted, the panel used to define the clusters was limited to 48 proteins using blood samples taken at a single timepoint. While there is no optimal platform to construct a panel that includes every possible immune-related protein, use of a restricted panel has the potential to limit unbiased discovery. Similarly, cross-sectional data from a single timepoint raises several questions that are not limited to this study and remain unanswered to date. First, it’s unknown how immune profiles will change over time and if the observed trends are more significant than the initial measures. Second, patients may “jump” clusters as their immune profiles undergo modification, indicating the potential for loss of fidelity of the clinical and predictive power of the clusters. Nonetheless, the investigators did identify actionable immune markers from a single blood sample and used these successfully to derive the clusters. Whether these unique clusters may be expanded (or contracted) on the basis of adding immune data from other profiling platforms is an unknown. The immune markers were then used to develop cluster-specific networks based on correlation; however the unique relevance of these networks to PAH, relationship to comorbidities, or overlap with other disease states is not yet discovered as these networks were not been mapped to the consolidated human interactome, which would provide important information on functional or physical relationships between immune markers⁵.

The application of machine learning to define and elucidate the immune clusters is novel in the field and should be viewed as a strength of the study. Machine learning is not entirely new to the field of pulmonary hypertension but has been used predominantly with imaging studies to define parameters that identify disease or predict outcomes^{6, 7}. More recently, deep neuronal networks and machine learning were used to estimate prognosis in patients with adult congenital heart disease or pulmonary hypertension. In this study, neural networks were trained to with clinical data and selected laboratory values from 10, 019 patients and found to have 91–97% accuracy⁸. Forthcoming analyses from the National Heart Lung, and Blood Institute initiative “Redefining Pulmonary Hypertension through Pulmonary Vascular Disease Phenomics (PVDOMICS)” study will broaden the datasets subjected to deep machine learning by incorporating clinical phenotyping with dense panomics results to create a wide-ranging molecules-to-man profile. As the overarching goal of this study is discovery of shared biological features for precision endophenotyping, regardless of World Symposium of Pulmonary Hypertension classification, this unique cohort is poised to enable and inform unbiased network analyses and machine learning to derive unique subgroups and clusters from the prevalent endophenotypes⁹.

The widespread use of precision medicine principles and technologies to characterize pulmonary hypertension endophenotypes will be evident in the near future. The work of Sweatt et al³ demonstrates the feasibility of applying unsupervised machine learning to define immune phenotypes and the novel immune identities of a cohort of patients with PAH. The PVDOMICS study aims to follow suit on a broader scale with comprehensive and integrated data from extensive deep clinical and panomics phenotyping. As shown in this study, the application of machine learning algorithms to these types of datasets will allow them to be exemplars of an optimal system that is facile enough to be utilized for prevention, diagnostics, or therapeutics. This evolving use of precision medicine to understand and define subphenotypes of patients with pulmonary hypertension, therefore, has the potential to improve health, support prevention strategies, and select interventions with a high probability of successful outcome.