Workshop on Parameter Estimation in Systems Biology, Manchester 2007

Systems Biology models often have numerous parameters, such as kinetic constants, decay rates and drift/diffusion terms, which are unknown or only weakly constrained by existing experimental knowledge. A crucial problem for Systems Biology is that these parameters are often very difficult to measure directly. Furthermore, they may vary greatly according to their in vivo context. As a result, methods for the estimation of these parameters are of great interest. Standard approaches include maximum likelihood or least squares methods, using various optimisation heuristics such as simulated annealing and evolutionary algorithms. Although these approaches have had some success, it is very difficult to estimate the parameters when there are many interactions in the system under consideration. In this case the likelihood surface may have many local optima, the parameters may be poorly determined because there is not enough data, or there may be ambiguities brought about by symmetries or redundancy in the system.

These same issues arise in machine learning. However, the problems in Systems Biology have an additional facet. In machine learning the models of interest are typically general function approximators, whereas in Systems Biology models are intended to provide a mechanistic description of the system, often using ordinary or stochastic differential equations.

Recently, attention in machine learning and statistical inference has turned to parameter estimation in these models. The main goal of this workshop is to bridge the divide between the fields by bringing together experts in machine learning and statistics with systems biologists and bioinformaticians.