Zoubin Ghahramani (Department of Engineering, Cambridge):
Probabilistic Models of Networks
Network data encoding the pairwise relations between objects appear in many fields. For instance in biology, a protein network connects interacting partners, while in a social network, links among people indicate social relationships. The problems of analysing, understanding and modelling such networks has attracted interest from many research communities. I will review probabilistic approaches to modelling networks, including models such as the stochastic blockmodel and various forms of infinite relational models. The key idea behind these models is that each object has certain latent features, and that observed links in the network depend on these latent features. Probabilistic inference allows one to discover the potentially unbounded number of latent features (including discovering communities as a special case), predict missing links, and generally learn about the statistical properties of the networks. Existing latent feature models have certain limitations, which we address by proposing a more flexible Infinite Latent Attribute (ILA) model for network data. We demonstrate the performance of the ILA model on link prediction problems from biological and social networks.
Joint work with Konstantina Palla and David A. Knolwes.
Julio Saez-Rodriguez (European Institute of Bioinformatics, Cambridge):
Logic modelling as a means to link protein signalling networks with functional analysis of signal transduction
Thanks to recent experimental techniques and literature-mining efforts, we are able to construct comprehensive networks describing the interactions among proteins. These networks are useful for exploring complex biochemical pathways but are rarely cell-type specific and do not encode the input-output relationships required for analyzing receptor-mediated signaling cascades and the drugs that target them. Conversely, traditional approaches to studying cell signaling do not make use of the wealth of information that is now encoded in protein networks. I will describe a hybrid method to compress protein networks and convert them into logical models (from discrete Boolean circuits to quantitative Fuzzy-logic and dynamic differential equations) that can be trained against data in which cells are exposed to-combinations of ligands and drugs followed by biochemical measurement of intracellular responses (Mol. Sys. Biol., 5:331, 2009). Our method is implemented in the toolbox CellNOpt (http://www.ebi.ac.uk/saezrodriguez/software.html). I will illustrate its application with a study to distinguishing the topologies of immediate early signaling networks in primary human hepatocytes and four hepatocellular carcinoma (HCC) cell lines (Cancer Research, 71(16):5400-11, 2011).