Marco Grzegorczyk doe research in the topical field of systems biology there is considerable interest in learning networks, such as gene regulatory networks and protein activation pathways, from post-genomic data.
Dynamic Bayesian networks (DBNs) are a flexible modelling tool which can be used to this end. However, the conventional DBN model, known from textbooks, is based on the assumption that the underlying regulatory processes are homogeneous. This assumption is unrealistic for many biological applications, and can thus lead to biased results and erroneous conclusions.
For the last years he has therefore been developing a variety of novel non-homogeneous DBN (NH-DBN) models. Those advanced NH-DBN models combine DBNs e.g. with change point models, mixture models or hidden Markov models, so that non-homogeneous processes can also be adequately approximated. However, since the available data sets are usually sparse (e.g. short gene expression time series) those advanced NH-DBN models tend to be over-flexible and hence have a tendency towards data-overfitting. For models he uses hierarchical Bayesian modelling approaches and concepts to regularize the model flexibility.
Within the DSSC framework, he haves already started a fruitful collaboration with various biological research groups from ERIBA. They supervised a joint DSSC PhD project, and the goal of the project is to analyse various data sets collected at ERIBA as well as to develop a general pipeline for analysing post-genomic data sets with Bayesian networks. The PhD student, Victor Bernal, started his PhD project in August 2016.