Mechanistic Models for Pharmaceutical Process Development: Accounting for Input, Parameter and Measurement Uncertainties

Abstract

Mechanistic modeling plays a critical role in design and optimization of pharmaceutical manufacturing processes. This thesis focuses on three challenges: 𝑖) model input uncertainty, 𝑖𝑖) design space construction when there are uncertain model parameters, uncertain process conditions, and uncertain final quality measurements, and 𝑖𝑖𝑖) experimental design when data are limited. The thesis proposes an extension of parameter estimability analysis techniques for use with error-in-variables models (EVMs), enabling simultaneous estimation of uncertain inputs and model parameters. The proposed methodology is demonstrated using a batch synthesis step in production of 9-THP-2,6-difluoropurine, an intermediate for the anti-HIV drug Islatravir. The model includes 6 kinetic parameters and uncertain inputs in two experimental runs (i.e., in initial concentration of trimethyl amine reactant). The estimability analysis shows that only 5 of 8 decision variables should be estimated from 144 concentration measurements obtained from two batch reactor experiments. The resulting best estimates indicate that the amounts of trimethyl amine that were charged to the reactor in the two runs are 4.04 % and 0.93 % higher than the targeted amounts. Using the EVM approach results in a better fit to experimental data, especially at long reaction times. Building on this approach, a larger model that considers additional reaction steps is developed. This model contains 39 parameters and uncertain inputs for 26 available experimental runs. All 26 uncertain inputs and 33 of 39 parameters are estimable from 1584 data values provided by Merck & Co., Inc. The resulting model is used to predict an optimal product yield of 92.0 %. In addition, the model is used to compare different design space (DS) construction methods, including probabilistic strategies that account for uncertainties in model parameters, process conditions and final quality measurements. Recommendations are provided for process developers seeking robust DS construction strategies. In the final chapter, a mechanistic model with 12 parameters is developed for a different pharmaceutical process involving a saponification reaction scheme. A sequential Bayesian A-optimal model-based-design-of-experiment (MBDoE) strategy is applied to propose one and two new experiments that will reduce parameter uncertainty. Together, these contributions advance the state of model-based pharmaceutical process development.