Statistics Seminars: Bayesian calibration of biological simulators
11 March 2008 14:15 in CM221
The ability to infer parameters of gene regulatory networks is a key problem in systems biology. The biochemical data are intrinsically stochastic and tend to be observed by means of discrete-time sampling systems, which are often limited in their completeness.
A naive approach to parameter inference in this context is to work with a deterministic approximation to the stochastic model. Parameter estimates can then be obtained by using standard least squares or maximum likelihood approaches. However, this strategy has been shown not work well in general. The talk will describe an MCMC scheme which makes exact inferences for a partially and discretely observed stochastic kinetic model. The complicating factor here is to allow for uncertainty in the (unknown) sample path between the (partially) observed data points. Unfortunately the algorithm does not scale well to large systems. Other work in the area has sought to find solutions based on stochastic approximations to the true model.
An alternative strategy is to make use of simulators of the true biological model (such as www.basis.ncl.ac.uk) and to calibrate these models by using Bayesian techniques. This is the aim of the CaliBayes project (www.calibayes.ncl.ac.uk). Recent developments in the calibration of model simulators are being used to build a higher level computational GRID facility which enables biological modellers to make inferences using multiple post-genomic data resources. The talk will give an overview of the project and describe some recent work with applications to real experimental data.
Contact email@example.com for more information