Sushmita Roy, Terran Lane, Margaret Werner-Washburne
To understand how a cell responds and adapts itself to changing environmental conditions, we must build context-specific networks — networks of genes, proteins and metabolites that are re-wired according to particular environmental conditions. Existing machine learning algorithms for biological networks either infer statistical correlation with no physical interpretation of the edges, or infer physical attributes of the network assuming the structure to be fixed. Moreover, these algorithms do not account for condition-specific properties of the networks. We propose a multi-stage, data-driven approach, which combines classification, probabilistic network structure and parameter learning, and graph-theoretic algorithms to construct high-resolution, condition-specific networks. We use classification to predict parts of the physical network that are absent in the physical interaction databases. The physical network is incorporated as prior knowledge in the network structure learning algorithms to construct biologically plausible networks. The structure learning algorithms construct coarse-grained networks from condition-specific measurements of the genes, proteins and metabolites. This is followed by hidden variable inference and structure refinement of the coarse condition-specific network by annotating the edges with values of physical attributes, such as directionality and type of interaction. The high resolution, condition-specific networks are then compared using graph-theoretic concepts to reveal the range of network-level behavior under different environmental conditions. In this abstract, we show results from the preliminary stages of our work, (a) classification-based prediction of physical protein interactions and (b) description of a novel pathwise score that can be used to evaluate how well structure learning algorithms capture long range dependencies.
Subjects: 12. Machine Learning and Discovery; 12.2 Scientific Discovery
Submitted: Apr 10, 2007