This choice was motivated by the fact that sufficiently accurate tracking information on individual cells was not available for these data. It is possible to interpret the ODE model as an approximation of the time evolution of the mean cell numbers of an underlying stochastic thing Markov process in the discrete space of cell state frequen cies, from which it emerges by expansion of the master equation. For the population sizes and transition types and rates of interest here, the approximation holds well, and effects of the discrete or stochastic nature of such a process on the evolution of the means is expected to be negligible compared to the experimental variability of the data. However, if tracking information had been avail able, then using it might have given more direct results, e. g.

on residence time distributions of the cells in the dif ferent states. Due to the presence of cell death and cell division, tracking needs to be integrated with the model fitting of a suitably defined stochastic process. An example of such an approach was presented in the CellCognition methodology. We used a 10 parameter ODE model with 4 cellular states and 4 independent transition rates. We selected this model based on the following criteria, complexity of the model, goodness of fit, parameter identifiability and bio logical significance of the parameters. We were able to fit our model on the vast majority of spot experiments, demonstrating its overall high goodness of fit, despite the broad variety of dynamic phenotypes of the Mitocheck assay, the overall low cell counts, the cell misclassifica tion noise and the presence of untransfected cells.

At the same time, we were able to reliably estimate the 10 model parameters with satisfactory precision, as is indicated by the reproducibility between the control spots, shown in the clear separation of control phenotypes in Figure 4. As part of the model development, we tested simpler and more complex models. The models with fewer parame ters, however, failed to model the complex phenotypes of some of our positive controls, such as siKIF11. Parameter identifiability was a problem in more complex models, e. g. when allowing three separate cell death transition rates, or two different polynucleated states. In these models, some parameters could not be reliably estimated due to low cell counts and cell mis classification noise, and they were often shrunk to zero due to the penalized estimation procedure.

Our model was primarily designed to quantify the phenotypes of a large scale imaging assay with relatively low tempo ral resolution and showing a broad variety of dynamic phenotypes. Depending on the biological question, more targeted models could be envisioned to focus on certain dynamic aspects, such as introducing different modes of cell death or using a finer model of the mitosis Brefeldin_A phase.