Response to ‘Modeling is data driven: Use it for successful virtual patient generation’

Abstract

We agree with the concept that Rieger and colleagues1 raise in their response letter to our work2 that if you could check all aspects of every simulated profile then you would be able to detect anomalies that could then be used as an approach to exclusion of their root causes (the values of the causative parameter vector). To explore this further, consider a thought experiment: In this experiment, we have a complicated quantitative systems pharmacology (QSP) model in which there are many states and even more interactions between states. Also assume, in this QSP model, that there are many places in which nonlinearities may occur due to their recursive structures. If all sources of nonlinearity directly influence a state and all states are observable and experimental data are available then it is reasonable to assume that all influences of strange behavior from vectors of parameter values could be identified (as implied in figure 1).1 However, if not all sources are observable or do not yield states for which data exist and not all strange behavior in these states manifests as extreme observations in downstream output states then it remains possible (perhaps plausible) that the originating parameter vector values would not be identified as incompatible with the underlying biology. Initially it might be argued that if the states of interest are not obviously affected, then what does it matter if some component of the underlying biology has been violated? But due to the curse of nonlinearity, it is impossible to know when and under what circumstances a strange behavior may yield an untoward influence. The simple examples we proposed in our work,2 for illustrative purposes, were observable, and data (or prior knowledge of the behaviors) were available, and hence it may appear that our concerns are moot. However, if we take, for instance, the coagulation QSP model,3 we see that there are many states for which data are not available, and these states are part of nonlinear processes (e.g., many of the activated factors). We therefore believe it is a strong assumption that all such nonlinearity will manifest as strange behaviors in output variables and that hidden strange behaviors are inviolate.