Automated calibration of agent-based immunological simulations

A major challenge in this field, however, lies in parameterization, particularly in agent-based simulations. These simulations contain many parameters (>50 is plausible), many pertaining to aspects of immunology that either have not or cannot be examined with current wet-lab technologies. Curve-fitting (e.g. linear regression) based-calibration is tractable only for relatively simple simulations with few parameters, and will not necessarily lead to biologicallyplausible parameter values (e.g., if the model is a bad representation of the biology). For larger systems the current state of the art is to calibrate by hand/eye, with some values based on wet-lab data or expert opinion, and the rest on trial and error. Furthermore, it is typical to calibrate simulations against data from only a single wet-lab experiment. Although these data may comprise observations of multiple cells/molecules/disease scores (termed responses), given that a simulation is likely to be used to perform multiple novel experiments that have not been attempted in the wet-lab, it still constitutes calibration against a single datapoint (single experiment). Put another way, with so many degrees of freedom there may be multiple points in parameter space for which a simulation re-creates data from a single experiment; a simulation calibrated in such a manner will not necessarily be representative of the biology when used for a different experiment. We propose that to have genuine trust that they reliably capture the biology, immunological simulations should be calibrated against multiple wet-lab experiments.