High-throughput experimental protocols have revealed thousands of relationships amongst genes and proteins under various conditions. These putative associations are being aggressively mined to decipher the structural and functional architecture of the cell. One useful tool for exploring this data has been computational network analysis. In this thesis, we propose a collection of novel algorithms to explore the structure and evolution of large, noisy, and sparsely annotated biological networks.

We first introduce two information-theoretic algorithms to extract interesting patterns and modules embedded in large graphs. The first, graph summarization, uses the minimum description length principle to find compressible parts of the graph. The second, VI-Cut, uses the variation of information to non-parametrically find groups of topologically cohesive and similarly annotated nodes in the network. We show that both algorithms find structure in biological data that is consistent with known biological processes, protein complexes, genetic diseases, and operational taxonomic units. We also propose several algorithms to systematically generate an ensemble of near-optimal network clusterings and show how these multiple views can be used together to identify clustering dynamics that any single solution approach would miss.

To facilitate the study of ancient networks, we introduce a framework called ``network archaeology'') for reconstructing the node-by-node and edge-by-edge arrival history of a network. Starting with a present-day network, we apply a probabilistic growth model backwards in time to find high-likelihood previous states of the graph. This allows us to explore how interactions and modules may have evolved over time. In experiments with real-world social and biological networks, we find that our algorithms can recover significant features of ancestral networks that have long since disappeared.

Our work is motivated by the need to understand large and complex biological systems that are being revealed to us by imperfect data. As data continues to pour in, we believe that computational network analysis will continue to be an essential tool towards this end.