ANC Workshop Talk: Frank Dondelinger and Guido Sanguinetti, Chair: Helen Ramsden

Frank Dondelinger

Model Mismatch in an Exponential Prior

During the past couple of years, I have done a lot of research into information sharing priors for reconstructing gene regulatory networks from data (see e.g. Dondelinger et al. 2010, Husmeier et al. 2010). The idea behind an information sharing prior is that we have a collection of networks that are learned from segmented datasets, and the prior reinforces the commonalities among the networks. The strength of this information sharing is regulated by learned hyperparameters. However, for one of the priors we investigated, which has an exponential form, we observed a consistent mismatch between the learned distribution of hyperparameters and the optimal hyperparameter in terms of network reconstruction. I will discuss how we used an idealised analytical model to show that this mismatch is intrinsic to the form of the prior, and what the challenges are for finding a better model.

Guido Sanguinetti

Towards inference in complex regulatory networks

I'll give a brief review of work on modelling transcriptional regulation using latent stochastic processes. In the last 3 years, we have introduced computational tools that have allowed us to perform inference in increasingly complex networks, so that we can now attempt to address some realistic problems of interest to biologists. The aim is to give an overview rather than a technical talk, and assess how far we have gone, and how far we have to go.