Modeling is data driven: Use it for successful virtual patient generation

We thank Duffull and Gulati ¹ for highlighting pitfalls in generating virtual patients for quantitative systems pharmacology (QSP) models. The authors demonstrate how some parameter selections lead to nonphysiological model outcomes and conclude that virtual patients are hence unsuitable in these instances.

Our prior work ² , ³ on virtual populations avoids Duffull and Gulati’s main critique. In our workflow, the application of prior knowledge into the virtual patient generation algorithm specifically “throws‐out” nonphysiological solutions. To demonstrate how to avoid some issues raised by Duffull and Gulati, we modified their first example to illustrate our workflow (Figure 1a).

FIGURE 1.

(a) Workflow for generating plausible and virtual patients. Our method uses a data‐driven approach to retain or reject parameter sets based on comparing model outputs with available data. (b) Simulation of new “acceptable” data profiles for A1 (solid lines [mean ± range of 100 simulations] compared to selected plausible patients and shaded region [range of 2500 plausible patients]). (c) Same simulations as b for variable B1

We generated simulated data representative of the prior information that Duffull and Gulati referred to as an “acceptable profile.” We introduced variability typical of similar clinical data. ⁴ We then generated “plausible patients” by a Metropolis‐Hastings algorithm. ³ Because the plausible population closely matches the “data,” no further selection step was necessary for this exercise (Figure 1b,c). Following this process, we generated 2500 parameter sets with a range of values that give rise to model simulations that are plausible. ⁵

Virtual populations are not a panacea for dealing with the uncertainty in QSP models. For example, this approach does not guarantee the validity of the model itself. There are two main challenges in a virtual population approach: (1) determining what observations to match, and (2) generating a sufficient number of virtual patients to both match the observations and explore parameter variability. Despite these challenges, virtual populations allow us to propagate observed experimental or clinical variability onto the possible parameter values.

Finally, we would like to emphasize that the generation of virtual patients is not the end‐goal of QSP modeling, any more so than putting up a board is the end goal of carpentry. QSP models address key questions at the interface between biology and pharmacology. In drug development, we never collect a complete set of data. QSP models synthesize what data we possess to permit simulations of our best understanding of drug‐physiology interaction. The synthesis of prior biology knowledge brings uncertainty, but also permits assessment of questions that are not amenable to minimal modeling, such as: evaluating novel combination therapies, extrapolation to longer trials, new indications, and identifying key areas of uncertainty for experiments. Ultimately, the analyst should be aware of all the different types of models that can be applied in a drug development program and reach for the hammer that best fits their nail.

CONFLICTS OF INTEREST

The authors of this manuscript have read the journal’s policy and have the following potentially competing interests: T.R.R., R.J.A., and C.J.M. were employees of Pfizer Inc. during the completion and/or analyses of these studies.