Longitudinal machine learning of molecular and phenotypic trajectories of pulmonary hypertension.

Our vision is to fundamentally redefine the diagnosis and treatment of Pulmonary Hypertension (PH), a multifaceted disease that carries significant diagnostic and prognostic difficulties. Even though there are numerous cross-sectional studies on PH, they cannot be combined because the patients are all at different stages of the disease along with different comorbidities. By integrating high-dimensional molecular profiling into traditional longitudinal cohort studies, we aim to leverage machine learning, genomics, and vascular biology to predict future molecular and clinical measures at any time point in a patient's journey. This approach can potentially discover biological mechanisms that drive disease progression and identify biomarkers in patients even when they cannot visit the clinic to provide data.

Aims and Outcomes:

1. Develop machine learning models to predict whole transcriptomes from blood biopsies at any point in a PH patient's journey. This could enable a new form of molecular classification for PH at previously unexamined time points, offering more precision than current clinical classifications.

2. Create an efficient computational system for real-time analysis of high-dimensional data, such as thousands of genes, thus overcoming the current methodological gap for genomic data captured at multiple times.

3. Extrapolate molecular changes observed in blood biopsies to changes in the pulmonary vasculature, providing a non-invasive method for investigating disease mechanisms.

4. Determine the optimal sample size and design for longitudinal molecular studies, increasing these studies' cost-efficiency and statistical power for other diseases.

Our interdisciplinary approach begins with computational modelling of gene expression and symptom trajectories across multiple years following a patient's diagnosis. We will begin by combining molecular trajectories with baseline factors, including gender, ethnicity, and socio-economic status. Time-dependent changes in blood transcriptome and methylome will be associated with changes in patient symptoms, such as mean arterial pressure and 6 minute walk distance. Historically measured gene expression profiles will be used directly as longitudinal inputs in our machine learning models to predict future molecular profiles associated with pulmonary vascular remodelling and patient outcome.

To ensure scalability to many patient measures and time points, we will evaluate and extend a machine learning technique called Gaussian Processes, which has successfully modelled lower-dimensional longitudinal data such as body weights and CO2 emissions. The algorithm will be trained through an iterative process, by gradually increasing the amount of longitudinal data from patients' molecular profiles and electronic health records. Our multidisciplinary team of computer scientists, epidemiologists, biologists and clinicians, will enable the algorithm to use biologically relevant features and be scalable for clinical applications.

The data and methods from this project will not only provide PH patients with better information about their disease progression, but can also be used to understand other complex diseases and age-related disorders.