A Normalisation Strategy to Optimally Design Experiments in Computational Biology

Abstract

In this work we describe a new methodology to improve predictive capabilities of dynamic models when parameters differ in orders of magnitude. The main idea is to normalise the model unknown parameters before solving the classical problem of optimal experimental design based on the Fisher information matrix. The normalisation improves the relative confidence intervals of the estimated parameters and the conditioning of the Fisher matrix, especially for those criteria aiming to decorrelate the model parameters. Using the so-called core predictions, we show how the new approach improves the final model predictive capabilities in two terms: predictions are closer to the real dynamics and with better confidence intervals.