Modeling Core

点击翻译按钮获取中文摘要

Project Summary/Abstract – Modeling Core The Modeling Core, as part of SCRIPT, aimed to apply machine learning approaches to clinical and -omics data generated by the SCRIPT projects and cores to develop a models of severe pneumonia and identify novel biomarkers and therapeutic targets. Using an iterative systems biology approach, we generated a detailed model, published in Nature, of how severe SARS-CoV-2 pneumonia, in contrast with severe pneumonia due to other pathogens, possesses a peculiar host response pathobiology that explains its propensity to cause prolonged critical illness. Importantly, SCRIPT’s model predicted the efficacy of an experimental pharmacologic intervention in SARS-CoV-2 pneumonia – the CRAC channel inhibitor Auxora. In this renewal, Super-SCRIPT (SCRIPT2) will continue to leverage serial sampling of biological materials (bronchoalveolar lavage fluid, nasal epithelium, blood) paired with cutting-edge multi-omics technologies and deep clinical phenotyping to develop models of pneumonia pathogenesis which could augment clinical decision making. We used clinical and -omics data collected and generated during the first cycle of this award to generate preliminary data for the renewal. We discretized time in the ICU and related physiological measures on a per-day basis, similar to how physicians view and treat patients with severe pneumonia in the ICU. Our novel approach overcomes a critical limitation in the application of machine learning approaches to clinical data, which often do not take into account interventions that can change the course of the disease and typically focus only on clinical state at presentation and ultimate outcome, analogous to drawing a line between two points. We generated a low-dimensional interpretable latent space model of clinical states in patients with severe pneumonia. We show that transitions between these clinical states are different in patients with SARS-CoV-2 pneumonia and other types of pneumonia. By projecting results of -omics assays onto this clinical latent space, we propose to identify biomarkers associated with favorable and unfavorable clinical transitions. We will use this latent space model of severe pneumonia to test the hypothesis that machine learning approaches can identify interpretable cellular and molecular biomarkers of favorable and unfavorable clinical transitions during the clinical course of severe pneumonia. We will test this hypothesis in three interrelated Specific Aims: Aim 1: To generate an interpretable latent space model of clinical states and transitions (disease trajectories) in patients with severe pneumonia using data collected within SCRIPT2. Aim 2: To identify cellular and molecular biomarkers and clinical interventions predictive of transitions between unfavorable and favorable clinical states in patients with severe pneumonia using data collected within SCRIPT2. Aim 3: To generalize models generated using SCRIPT2 to external datasets.