Datamining Biological Networks

Abstract

The biological processes that occur within a cell are known to be organized in networks. There are many types of biological networks, each of which perform a specific function and are known to interact. However, as the network structures themselves are complex, the nature and structure of the interactions across networks are difficult to define. In this paper we assume that if a set of networks interact, the dynamics within each of these networks must share a common latent signal. To identify this signal across multiple networks we propose a Gaussian Process Latent Factor Model (GP-LFM). Our proposed model uses a GP-LFM to represent the observed time course at each network node based on the assumption that a diffusion process governs the dynamics within each network. From the assumption that diffusion of a common latent factor generates all observed node time courses we derive a latent force model. This latent force model explicitly predicts each node's time course based only on the neighborhood of that node and the strength of the diffusion process within that node's network. We then extend this idea to multiple networks by assuming that each network can be considered independent once the common latent function which generates all node time courses over all networks is known. Finally, we consider the transductive case, where some node time courses in some networks are unobserved and we wish to infer them. We show that by a re-parameterization of our proposed GP-LFM model we can estimate these missing node time courses without deviating from the standard GP-LFM optimization methods. We evaluate the performance of our GP-LFM on simulated and real data and clearly show that our GP-LFM approach is capable of identifying a common latent signal which is determines known dynamics across multiple networks.