Edinburgh Napier University

  Network models are often created to describe interactions between components in biological systems. At a molecular level the involved components are genes or other biomarkers such as proteins and metabolites. Building network models can help in understanding the functional role of biomarkers and can also give an idea of how the system works as a whole. However, learning the interactions in complex systems is problematic, even for the most advanced computational inference methods, mainly because of the large number of variables involved. Visualisation combines the ability of the human visual system to detect patterns and anomalies in the data, with the power of modern computers for storing, processing and displaying information. This research focuses on the application of visualisation methods for gaining insights from biological data in order to create better network models. This research is done in collaboration with a group of biologists who are interested in constructing Bayesian network (BN) models from experimental Microarray time-series data. However, this work is applicable and could be extended to ecological, neural and other types of data.

Microarrays measure expression levels of thousands of genes in several time points. The first challenge is to reduce the number of variables involved, so that only the most representative would be used for the construction of the final network model. A common way of achieving that distinction is by applying a clustering algorithm in order to identify genes that follow a similar temporal profile and then merge or reduce them to create one variable for each profile. The problem is in allocating genes to the right clusters. This is not clear and the resulting visualisation often seems cluttered. A second challenge involves the selection of a consensus network from collections of candidate networks. Candidate networks are produced and scored by a heuristic network-inference algorithm. The final network model should not only be informative, but also reproducible. Achieving reproducible results is a challenge when applying probabilistic methods, such as Bayesian networks. From the perspective of visualisation it is difficult to summarise and compare potentially large collections of networks in order to find a reproducible final network. The direction of this research is towards the fulfilment of those research objectives.