Selecting high-confidence predictions from ordinary differential equation models of biological networks

Abstract

Many cellular processes are governed by large and highly-complex networks of chemical interactions and are therefore difficult to intuit. Computational modeling provides a means of encapsulating information about these interactions and can serve as a platform for gaining understanding of the biology and making predictions about cellular response to perturbation. In particular, there has been considerable interest in ordinary differential equation (ODE) models, which have several attractive features: ODEs can describe molecular interactions with mechanistic detail, it is relatively straightforward to implement perturbations, and, in theory, they can predict the concentration and activity of every species as a function of time. However, both the topology and parameters in such models are subject to considerable uncertainty. We explore the ramifications of these sources of uncertainty for making accurate predictions and develop methods of selecting high confidence predictions from uncertain models. In particular, we promote a shift in emphasis from model selection to prediction selection, and use consensus among model ensembles to identify the predictions most likely to be accurate. By constructing decision trees, this consensus can also be used to partition the space of potential perturbations into regions of high and low confidence. We apply our methods to the Fas signaling pathway in apoptosis to satisfy two goals: first, to design a therapeutic cocktail to reduce cell death in the presence of high levels of stimulus, and second, to design experiments that may lead to a better understanding of the biological network.