Colloquium Mathematics, M. Shafiee Kamalabad (PhD)

Title:                        

Advanced non-homogeneous dynamic Bayesian network models for statistical analyses of time series data

Abstract:

Learning network structure of interacting units from time series data is statistically a challenging task in many fields. The aim is to infer the dependencies between variables from temporal data and to represent them in the form of a network. E.g. one of the major goals in the field of systems biology is to learn cellular network, such as gene regularity transcription networks and protein signaling pathways. One class of models that has been widely applied to deal with this challenge, is the class of dynamic Bayesian network (DBN) models which is a powerful statistical tool to learn network structures from temporal data. The traditional homogeneous dynamic Bayesian network model, has certain limitations, so that it is often not able to describe the regulatory relationships correctly, leading to erroneous conclusions. Hence,  these shortcomings call for more advanced models. A lot of model extensions have been proposed in the literature. One of the most popular ideas is to infer regulatory networks with time-varying interaction parameters by combining dynamic Bayesian models, with Bayesian multiple changepoint processes.  The idea is to divide the data into disjoint segments and the network interaction parameters are allowed to vary from segment to segment.

During my PhD program, I have developed the class of non-homogeneous dynamic Bayesian network models further. For typical data situations, arising in systems biology,  I have proposed tailor-made advance dynamic Bayesian network models. My empirical results on synthetic and real network data have shown that the new models outperform the earlier competing models in terms of the network reconstruction accuracy. In my colloquium presentation I will focus on two of those new models corresponding to chapter 4 and chapter 5 of my thesis: First, I will present a new improved edge-wise sequentially coupled non-homogeneous dynamic Bayesian network model and I apply it to reverse-engineer the circadian clock gene network in Arabidopsis thaliana. In the second part of my presentation,  I will present a partially non-homogeneous dynamic Bayesian network model, where only some of the interaction parameters are time-varying, while the other interaction parameters stay constant in time. I use this model to infer the mTORC1 protein signaling pathway.

Colloquium coordinators are Prof.dr. A.J. van der Schaft (a.j.van.der.schaft rug.nl),

Dr. A.V. Kiselev (e-mail: a.v.kiselev rug.nl)

http://www.rug.nl/research/jbi/news/colloquia/mathematics-colloquia/