COPD SUBTYPES AND EARLY PREDICTION USING INTEGRATIVE PROBABILISTIC GRAPHICAL MODELS R01HL157879

COPD SUBTYPING AND EARLY PREDICTION USING INTEGRATIVE PROBABILISTIC GRAPHICAL
MODELS
ABSTRACT
One of the main obstacles in developing efficient personalized therapeutic and disease management strategies
is that most common diseases are typically defined based on symptoms and clinical measurements, although
they are believed to be syndromes, consisting of multiple subtypes with variable etiology. Identifying disease
subtypes has thus become very important, but so far it has been met with limited success for most diseases. In
asthma, a notable exception, it was the clinical characterization that led to successful subtyping; and this is now
incorporated in treatment guidelines. Unsupervised machine learning approaches of single data modalities (e.g.,
omics, radiographic images) have not produced actionable subtypes due to instability across cohorts. Developing
data integrative approaches for multi-scale data, which are becoming available for a number of diseases, is
expected to lead to robust subtyping and provide mechanistic insights of disease onset and progression.
This proposal focuses on developing new computational methods, based on probabilistic graphical models
(PGMs), to address this unmet need; and apply them to investigate three problems of clinical importance in
chronic obstructive pulmonary disease (COPD), which is the fourth leading cause of mortality in USA. Our
underlying hypothesis is that PGMs can integrate and analyze under the same probabilistic framework
heterogeneous biomedical data (omics, chest CT scan, clinical) and identify disease subtypes and their main
determinants. The objectives of our proposal is to build a comprehensive computational framework for disease
subclassification, identify stable COPD subtypes at the baseline and longitudinally, and build interpretable
models of the disease The deliverables of this project are: (1) new integrative computational approaches for
clinical subtyping from multi-scale data; (2) new predictors of COPD progression and severity; (3) new
discoveries of longitudinally stable COPD subtypes; (4) new predictors of future development of COPD; (5) new
omics datasets that will be invaluable to future research in the area (baseline and longitudinal).
To ensure the success of the project we follow a team science approach. This multi-PI proposal builds on the
ongoing efforts of our group in the area of graphical models and their applications in biomedicine; and the
ongoing collaboration of the three PIs that have complementary strengths: Prof. Benos (systems medicine and
machine learning), Dr. Hersh (COPD genetics and genomics) and Dr. Sciurba (clinical aspects of COPD). It is
powered by the access of the investigators to three major COPD cohorts (COPDGene®, SCCOR, ECLIPSE) that
contain multiple parallel deep phenotyping and omics data from thousands of patients and controls. Although in
this project we focus on COPD, our methods are generally applicable to any disease, therefore our project will
have a positive impact beyond the above deliverables. We believe that due to their robust nature and
interpretability, PGMs will soon become the norm for multi-scale biomedical data integration and modeling, when
genetic and genomic data collection will become routine prognostic and diagnostic tools in clinical practice.

Grant Number: 5R01HL157879-04
NIH Institute/Center: NIH
Principal Investigator: PANAGIOTIS BENOS