Dr Jorge Gonçalves, University of Cambridge (profile)
Title: Biochemical network inference from data
Abstract: One of the fundamental interests in systems biomedicine is to understand the functionality of certain molecules that are involved with a particular disease. We are looking to develop an understanding of biochemical networks regarding (a) their causal molecular interactions and regulations, (b) their dynamical behaviours, and (c) the map from networks to their phenotypic/disease behaviour. With predictable models, we can then simulate hypotheses for subsequent experimental testing, thereby accelerating the development of an understanding of the causes of a disease.
This talk will concentrate on inferring dynamical causal network interactions between measured species. In spite of the intrinsic difficulty of network inference, our research outlines the experimentation process that makes such discovery possible. If nothing is known about the network between measured species, then experiments must be performed as follows: 1) for a network composed of p measured species, the same number of experiments p must be performed; 2) each experiment must independently control a measured specie, i.e., control input i must first affect measured specie i (e.g. experiments involving gene silencing or inducible overexpression). If something is known about the system, then these experiments may be relaxed. The network representation resulting from this process yields a predictive model commensurate with the informativity of the data used to create it: steady-state data yield static network information, while time series data generate a dynamic representation of the system suitable for simulation.
Further, we demonstrate that in the absence of this essential experimental design, the network cannot be reconstructed and every conceivable network structure between species (e.g. a fully decoupled network or a fully connected network) can be equally descriptive of any particular set of input-output data. Unless this correct methodology is followed, any best-fit measure for network reconstruction can yield arbitrarily poor and misleading results.