Computer Science ETDs


Sushmita Roy

Publication Date



Condition-specific cellular networks are networks of genes and proteins that describe functional interactions among genes occurring under different environmental conditions. These networks provide a systems-level view of how the parts-list (genes and proteins) interact within the cell as it functions under changing environmental conditions and can provide insight into mechanisms of stress response, cellular differentiation and disease susceptibility. The principle challenge, however, is that cellular networks remain unknown for most conditions and must be inferred from activity levels of genes (mRNA levels) under different conditions. This dissertation aims to develop computational approaches for inferring, analyzing and validating cellular networks of genes from expression data. This dissertation first describes an unsupervised machine learning framework for inferring cellular networks using expression data from a single condition. Here cellular networks are represented as undirected probabilistic graphical models and are learned using a novel, data-driven algorithm. Then several approaches are described that can learn networks using data from multiple conditions. These approaches apply to cases where the condition may or may not be known and, therefore, must be inferred as part of the learning problem. For the latter, the condition variable is allowed to influence expression of genes at different levels of granularity: condition variable per gene to a single condition variable for all genes. Results on simulated data suggest that the algorithm performance depends greatly on the size and number of connected components of the union network of all conditions. These algorithms are also applied to microarray data from two yeast populations, quiescent and non-quiescent, isolated from glucose starved cultures. Our results suggest that by sharing information across multiple conditions, better networks can be learned for both conditions, with many more biologically meaningful dependencies, than if networks were learned for these conditions independently. In particular, processes that were shared among both cell populations were involved in response to glucose starvation, whereas the processes specific to individual populations captured characteristics unique to each population. These algorithms were also applied for learning networks across multiple species: yeast (S. cerevisiae) and fly (D. melanogaster). Preliminary analysis suggests that sharing patterns across species is much more complex than across different populations of the same species and basic metabolic processes are shared across the two species. Finally, this dissertation focuses on validation of cellular networks. This validation framework describes scores for measuring how well network learning algorithms capture higher-order dependencies. This framework also introduces a measure for evaluating the entire inferred network structure based on the extent to which similarly functioning genes are close together on the network.




Machine learning, Computational Biology, Probabilistic graphical models, Gene expression, Condition-specific response

Document Type


Degree Name

Computer Science

Level of Degree


Department Name

Department of Computer Science

First Advisor

Lane, Terran

Second Advisor

Werner-Washburne, Margaret

First Committee Member (Chair)

Moses, Melanie

Second Committee Member

Atlas, Susan

Project Sponsors

National Institute of Health, National Science Foundation, Howard Hughes Medical Institute Interfaces Program, Program in Interdisciplinary Biological & Bio-medical Sciences at UNM.