Exploiting Structure Information for Network Dissimilarity Characterization - Application to Disease Network Analysis and Treatment Prediction

Most cancers lack effective early disease markers, prognostic and predictive signatures, primarily due to tumor heterogeneity. As a result, we fail treating cancer heterogeneity due to multiple ways cancer initiates and develops treatment resistance. Models that represent these differences and the underlying molecular mechanism in cancer enhance the possibility in characterizing and in turn treating cancer successfully.

We introduce novel graph-based methods for exploiting network structure information in the comparison between any graphs, and validate them on non-small cell lung cancer datasets. We generate normal and tumor graphs using normal and tumor samples from gene expression datasets, where vertices are genes, and edges connect co-expressed genes. In the first part of this dissertation, we propose a systems approach with an aim to revert disease conditions to healthy ones through treatments. In order to achieve the objective, we propose three novel methods to 1) systematically identify network structure differences between normal and tumor graphs, 2) identify and prioritize drug combinations based on extracted network structure differences, and 3) computationally estimate the potential of the proposed drug combination to "repair" deregulated subgraphs. Biological validation of the predictions highlights that our systems approach is a promising method to provide treatment options to non-small cell lung cancer through the rewiring of disease networks. In the second part of this dissertation, we introduce the notion of differential graphlet community to detect deregulated subgraphs between any graphs such that the network structure information is exploited. We observed a trend that the shortest path lengths are shorter for tumor graphs than for normal graphs between genes that are in differential graphlet communities, suggesting that cancer creates shortcuts between biological processes that may not be present in normal conditions. In the third part of this dissertation, we propose a heuristic, the differential correlation graph approach, that identifies areas that are different between the normal and tumor graph, and perform graphlet enumeration on the identified areas. Results showed that our approach achieves accurate estimation in the difference between normal and tumor states by performing network comparisons in important areas only.