ZERO, ONE AND TWO-SWITCH MODELS OF GENE REGULATION

We compare a hierarchy of three stochastic models in gene regulation. In each case, genes produce mRNA molecules which inturn produce protein. The simplest model, as described by Thattai and Van Oudenaarden (Proc. Nat. Acad. Sci., 2001), assumes that a geneisal way sactive, and uses a first-order chemical kinetics framework in the continuous-time, discrete-space Markov jump (Gillespie) setting. The second model, proposed by Raserand O'Shea (Science, 2004), generalizes the first by allowing the gene to switchr and omly between active and inactive states. Our third model accounts for other effects, such as the binding/unbinding of a transcription factor, by using two independenton/offswitches operating in AND mode. We focus first on the noise strength, which has been defined in the biological literature as the ratio of the variance to the mean at steady state. We show that the steady state variance in the mRNA and protein for the three models can either increase or decrease when switches are incorporated, depending on the rate constants and initial conditions. Despite this, we also find that the overall noise streng this always greater when switches are added, in the sense that one or two switches a real ways noisier tha nnone. On the other hand, moving from one to two switches may either increase ordecrease the noise strength. Moreover, the steady state values may not reflect the relative noise levels in the transient phase. We then look at a hybrid version of the two-switch model that uses stochastic differential equations to describe the evolution of mRNA and protein. This is a simple example of a multiscale modelling approach that allows for cheaper numerical simulations. Although the underlying chemical kinetics appears to be second order, we show that it is possible to analyse the first and second moments of the mRNA and protein levels by applying a generalized version of Ito's lemma. We find that the hybrid model matches the moments of underlying Markov jump model for all time. By contrast, further simplifying the model by removing the diffusion in order to obtain an ordinary differential equation driven by a switch causes them RNA and protein variances to be under estimated.