Associate Director Bristol Myers Squibb Brisbane, California, United States
Disclosure(s):
Karyn Sutton, PhD: No financial relationships to disclose
Objectives: Large-scale mechanistic models, built from first principles remain the gold standard in quantitative systems pharmacology; however it is difficult to derive well-constrained parameters, and the needed assumptions are often accompanied by uncertainty. Thus, simpler models can often be comparably favorable in terms of implementation, calibration and simulation times, and most importantly in yielding clinically useful results that are easier to interpret. We demonstrate an approach for reducing higher-dimension models without compromising the goodness-of-fit to key outputs or predictive capability as assessed by appropriate metrics, applied to a general model of solid tumor indications treated by a protein degrader as an illustrative example.
Methods: The full model, with all mechanisms represented explicitly, is reduced by implicitly representing several mechanisms that are not directly measurable in the clinical setting. The results of calibrating both models to the same set of simulated data is compared via (i) residual sum of squares, (ii) observed vs predicted plots, and (iii) standard errors of estimated parameters These data are generated by adding noise to simulated clinically relevant outputs of the original model.
The uncertainty of model predictions is assessed by generating simulations with parameter values sampled from their associated distributions from calibration. The values for those parameters that are not identifiable are either fixed or sampled from a range informed by literature.
Results: The full model is reduced by replacing mechanisms such as binding, ubiquitination, drug distribution to tumor and the cells therein, empirically. A simpler model is comparable to that of the full model in the resulting goodness-of-fit to data and uncertainty of model predictions. Other reduced models are seen to not fit the data set as well or result in predictions with increased uncertainty as compared to the full model.
Conclusions: The rationale to judiciously reduce clinically informative models is demonstrated along with an approach to assess the appropriate dimension and mechanistic detail. A suitably reduced model is then more readily extended to incorporate additional features of complexity commonly associated with clinical data, such as inter-patient variability, and the corresponding impact on model predictions more readily discernible.
