Two Case Studies on How Study Designs Can Be Made More Informative Using Modeling and Simulation Approaches

    loading  Checking for direct PDF access through Ovid


In developing the high‐affinity anti‐IgE monoclonal antibody (mAb) ligelizumab for chronic spontaneous urticaria, a key point was not simply “the dose” but a complete posology, a combination of dose and dosing frequency, to optimize safe, efficacious, and convenient usage. Although many mAbs are administered monthly as per half‐life, this does not need to be the case. If dosing intervals could safely be extended, medical center visits, associated healthcare costs, and patient burden could be minimized. To visualize posologies to be confirmed in phase III, a model capable of simulating outcomes for different options was required. The phase II study NCT024773321 was therefore designed to deliver not just a dose but a model capable of simulating primary clinical outcomes for minimally to maximally effective posology options.
Unfortunately, pharmacokinetic–pharmacodynamic (PKPD) model‐based analyses for products with a long half‐life or response time can only normally be run at the end of a study. It can be >12 weeks from last patient last treatment, plus sample shipment and bioanalytical laboratory time, before pharmacometric analysts receive data. To minimize “white space” between trials, phase III design proposal and protocol writing, posology included, commenced before the phase II study was complete.
Although most PKPD parameters can be estimated accurately and precisely using on‐treatment period data from a conventional parallel arm multiple dose study design, the timepoint of maximum sensitivity for estimating the concentration associated with half‐maximum response, EC50, occurs during drug washout. The impact of estimation using the individual time course rather than population cross‐sectional analyses is exemplified in the European Medicines Agency assessment report of omalizumab for chronic spontaneous urticaria, which states “The increased quantity of data for the time course analysis increased the precision of the EC50 estimates compared with the Week 12 data analyses (95% confidence interval was 15–16 μg/mL from the time course model compared with 9.8–43 μg/mL for itch).”2 Given that the accuracy and precision of response prediction depends on that for the data analysis model parameters, clearly there was much to be gained by improving both experimental design and analysis methods.
For the ligelizumab phase II study NCT02477332, design options were tested by trial simulation followed by NONMEM nonlinear mixed effect modeling to evaluate the accuracy and precision with which true parameter values could be returned. Design simulations were carried out using a published omalizumab‐IgE‐itch‐hives placebo and drug‐effect model,3 with substitution of an 18‐fold lower apparent binding dissociation constant, KD, reflecting the higher affinity of ligelizumab for IgE (Figure1). The starting point was the dose‐finding for omalizumab urticaria2; three active dose levels plus placebo administered every 4 weeks for 20 weeks was initially proposed. However, with this design the interim analysis at 12 weeks would have had poor precision for EC50 estimation, resulting in a lack of confidence in phase III posology simulations (wide prediction intervals). A fifth arm with a single high‐dose administration followed by blinded washout was therefore included specifically to enable longitudinal modeling of the increasing proportion of patients experiencing recrudescent symptoms as drug concentrations decreased. Simulation‐estimation demonstrated that adding this arm improved parameter precision, achieving median estimation errors of less than 20% for EC50 even if only this arm were analyzed. Precision improved yet further if multiple dose arm data were included to reflect that one could also analyze data from early‐recruited patients who had completed washout by the date of the interim analysis.
The simulation‐estimation and visualization exercise demonstrated that, although crucial for defining a posology, there would not necessarily be sufficient washout data available at interim analysis in a conventional parallel multiple administration multiple dose‐level design.
    loading  Loading Related Articles