A visual exploration route: from time-series to network models
Athanasios Vogogias, Jessie Kennedy, Daniel Archambault · 2017
Network models aim to describe relationships between components and provide an abstract view on how a biological system works as a whole. However, learning the structure of the most representative network from sparse and often noisy data sets is a challenging optimisation problem. In this work we propose a visual analytics approach for inferring Bayesian networks from time-series gene expression data. In particular, we seek to provide visual support for exploring the search space of all candidate networks in conjunction with a representation of the original data. As part of the variable selection phase we apply hierarchical clustering with multiple-level cuts to reduce feature redundancy. The purpose is to effectively reduce the number of variables by detecting and aggregating genes that follow a similar expression pattern over time. As part of the network construction phase we use different search algorithms and parameter settings to sample the search space of all possible Bayesian networks. An appropriate scoring metric is used to assess their fitness to the original data. We extend the small multipiles technique to provide visual support for exploring large collections of scored, directed networks, which constitute the solution space. Depending on its shape, either the top scoring network is selected, or a consensus network is constructed from a number of high-scoring candidate networks. Our approach aims at the design and implementation of a visualisation toolbox which would help analysts in inferring Bayesian network models which are not only reproducible, but also representative of the original time-series data.