Computational Labeling, Partitioning, and Balancing of Molecular Networks

Recent advances in high throughput techniques enable large-scale molecular quantification with high accuracy, including mRNAs, proteins and metabolites. Differential expression of these molecules in case and control samples provides a way to select phenotype-associated molecules with statistically significant changes. However, given the significance ranking list of molecular changes, how those molecules work together to drive phenotype formation is still unclear. In particular, the changes in molecular quantities are insufficient to interpret the changes in their functional behavior. My study is aimed at answering this question by integrating molecular network data to systematically model and estimate the changes of molecular functional behaviors. We build three computational models to label, partition, and balance molecular networks using modern machine learning techniques. (1) Due to the incompleteness of protein functional annotation, we develop AptRank, an adaptive PageRank model for protein function prediction on bilayer networks. By integrating Gene Ontology (GO) hierarchy with protein-protein interaction network, our AptRank outperforms four state-of-the-art methods in a comprehensive evaluation using benchmark datasets. (2) We next extend our AptRank into a network partitioning method, BioSweeper, to identify functional network modules in which molecules share similar functions and also densely connect to each other. Compared to traditional network partitioning methods using only network connections, BioSweeper, which integrates the GO hierarchy, can automatically identify functionally enriched network modules. (3) Finally, we conduct a differential interaction analysis, namely difFBA, on protein-protein interaction networks by simulating protein fluxes using flux balance analysis (FBA). We test difFBA using quantitative proteomic data from colon cancer, and demonstrate that difFBA offers more insights into functional changes in molecular behavior than does protein quantity changes alone. We conclude that our integrative network model increases the observational dimensions of complex biological systems, and enables us to more deeply understand the causal relationships between genotypes and phenotypes