Integrative Statistical Learning with Applications to Predicting Features of Diseases and Health.

Abstract

This dissertation develops methods of integrative statistical learning to studies of two human diseases - respiratory infectious diseases and leukemia. It concerns integrating statistically principled approaches to connect data with knowledge for improved understanding of diseases. A wide spectrum of temporal and high-dimensional biological and medical datasets were considered.

The first question studied in this thesis examined host responses to viral insult. In a human challenge study, eight transcriptional response patterns were identified in hosts experimentally challenged with influenza H3N2/Wisconsin viruses. These patterns are highly correlated with and predictive of symptoms. A non-passive asymptomatic state was revealed and associated with subclinical infections. The findings were validated and extended to three additional viral pathogens (influenza H1N1, Rhinovirus, and RSV). Their differences and similarities were compared and contrasted. Statistical models were developed for exposure detection and risk stratification. Experimental validations have been performed by collaborators at the Duke University.

The second question studied in this thesis investigated the regulatory roles of Hoxa9 and Meis1 in hematopoiesis and leukemia. Methods were developed to characterize their global in vivo binding patterns and to identify their functional cofactors and collaborators. The combinatorial effects of these factors were modeled and related to specific epigenetic signatures. A new biological model was proposed to explain their synergistic functions in leukemic transformation. Experimental validations have been performed by members of the Hess laboratory.

Motivated by problems encountered in these studies, two algorithms were developed to identify spatial and temporal patterns from high-throughput data. The first method determines temporal relationships between gene pathways during disease progression. It performs spectral analysis on graph Laplacian-embedded significance measures of pathway activity. The second algorithm proposes probabilistic modeling of protein binding events. Based on information geometry theory, it applies hypothesis testing coupled with jackknife-bias correction to characterize protein-protein interactions. Experimental validations were shown for both algorithms.

In conclusion, this dissertation addressed issues in the design of statistical methods to identify characteristic and predictive features of human diseases. It demonstrated the effectiveness of integrating simple techniques in bioinformatics analysis. Several bioinformatics tools were developed to facilitate the analysis of high-dimensional time-series datasets.