Validation of gene regulatory network inference based on controllability

There are two distinct issues regarding network validation: (1) Does an inferred network provide good predictions relative to experimental data? (2) Does a network inference algorithm applied within a certain network model framework yield networks that are accurate relative to some criterion of goodness? The first issue concerns scientific validation and the second concerns algorithm validation. In this paper we consider inferential validation relative to controllability; that is, if an inference procedure is applied to data generated from a gene regulatory network and an intervention procedure is designed on the inferred network, how well does it perform on the true network? The reasoning behind such a criterion is that, if our purpose is to use gene regulatory networks to design therapeutic intervention strategies, then we are not concerned with network fidelity, per se, but only with our ability to design effective interventions based on the inferred network. We will consider the problem from the perspectives of stationary control, which involves designing a control policy to be applied over time based on the current state of the network, with the decision procedure itself being time independent. The objective of a control policy is to optimally reduce the total steady-state probability mass of the undesirable states (phenotypes), which is equivalent to optimally increasing the total steady-state mass of the desirable states. Based on this criterion we compare several proposed network inference procedures. We will see that inference procedure ψ may perform poorer than inference procedure ξ relative to inferring the full network structure but perform better than ξ relative to controllability. Hence, when one is aiming at a specific application, it may be wise to use an objective-based measure of inference validity.