Natural Sciences
Life Sciences
Scientific Computing
Life Science

Fabian Fröhlich, Institute of Computational Biology, German Research Center for Environmental Health, München

BioQuant, Seminar room 042, Im Neuenheimer Feld 267

Roland Eils, DKFZ/BioQuant, Uni Heidelberg, CellNetworks Member

For most cell populations there is variability in the behaviour of individual cells. This variability can either emanate from the intrinsic stochasticity of chemical reactions or from cell-to-cell variability in the concentration of chemical species. In order to build good, predictive models both effects must be accounted for. However, in most applications the computational complexity of a fully stochastic simulation (e.g. via Stochastic Simulation Algorithm) for the behaviour of individuals is prohibitively large. Moreover, the computational complexity increases with the number cells considered. To circumvent this computational complexity, deterministic approximations to the mean and variance of the average individual behaviour (system-size expansion) as well as the population behaviour (sigma-point based) can be used. In order for the models to be predictive, the unknown parameters must be estimated such that the model agrees with experimental data. A bias in parameter estimates can impede the predictive power of models. In this talk we will compare parameter estimation for the individual level using different orders of system-size expansions. At the population level we will employ traditional mixed-effect modelling and a sigma-point based approximation scheme. For all methods we implement efficient gradient-based parameter estimation methods. Using two small-scale models, we evaluate the estimation bias of the approximation schemes for the individual behaviour. We find that there exists an intermediate volume regime where the mean squared estimation bias can be several orders of magnitudes lower when using higher order system-size expansions. For the population behaviour we compare the computation time and estimation bias using traditional mixed-effect modelling and sigma-point based approximation schemes. Moreover, we will study the effect of technical noise on the estimation bias.

Event data:
Import event data into Outlook Calendar