Maximin and bayesian robust experimental design for measurement set selection in modelling biochemical regulatory systems

Experimental design is important in system identification, especially when the models are complex and the measurement data are sparse and noisy, as often occurs in modelling of biochemical regulatory networks. The quality of conventional optimal experimental design largely depends on the accuracy of model parameter estimation, which is often either unavailable or poorly estimated at the stage of design. Robust experimental design (RED) algorithms have thus been proposed when model parametric uncertainties need to be addressed during the design process. In this paper, two robust design strategies are investigated and the comparative study has been made on signal pathway models. The first method is a maximin experimental design approach which is a worst-case design strategy, and the second method is the Bayesian experimental design that 'takes an average' of the parametric uncertainty effects. The limitations of the maximin design which describes the structural uncertainty using a local Taylor representation are quantitatively evaluated. To better quantitatively assess the differences between the maximin and the Bayesian REDs, a concept of effective design parameters is proposed, from which the advantages of the Bayesian design is demonstrated especially in the case of large model uncertainties.