Designing genetic perturbation experiments for model selection under uncertainty

Abstract

Deterministic dynamic models play a crucial role in elucidating the function of biological networks. However, the underlying biological mechanisms are often only partially known, and different biological hypotheses on the unknown molecular mechanisms lead to multiple potential network topologies for the model. Limitations in generating comprehensive quantitative data often prevent identification of the correct model topology and additionally leave substantial uncertainty about a model's parameter values. Here, we introduce an experiment design method for model discrimination under parameter uncertainty. We focus on genetic perturbations, such as gene deletions, as our possible experimental interventions. We start from an initial dataset and a single model whose topology includes all different hypotheses. We obtain the set of models compatible with the initial dataset, their posterior probabilities, and the distribution of compatible parameter values using our previously published topological filtering approach. We then employ a fully Bayesian approach to identify the genetic perturbation that yields the maximal expected information gain in a subsequent experiment. This approach explicitly accounts for parameter uncertainty; it also naturally allows comparing an arbitrary number of candidate models simultaneously. In contrast to previous approaches, our intervention alters the topology of the dynamic system rather than selecting optimal inputs, observables, or time-points for measurements. We demonstrate its applicability with an in-silico study based on a published real-world biological example.