Logo: to the web site of Uppsala University

Target-mediated drug disposition (TMDD) models are used to model a nonlinear pharmacokinetic (PK) due to one drug binding to its pharmacological target with high affinity, affecting the pharmacokinetic characteristics. TMDD approximation models emerged because of the difficulty of full TMDD model implementation with limited data to solve the overparameterization of a complicated model. Conventional population TMDD model development is time-consuming and subjective, requiring the modeler's experience. This thesis presents one strategy of TMDD model development and ranking that could be implemented to achieve automatic TMDD model development. The current work aims to establish a TMDD model development strategy that can be automated to extend to the Pharmpy/Pharmr package to enable the Automatic Model Development (AMD) tool with a more sophisticated description for nonlinear PK. Simulated data from the published TMDD models for five compounds were used to develop and test the TMDD model development strategy. First, appropriate estimation method was selected based on the literature and practical considerations for the automatic model development procedure for modeling efficiency. Second, an algorithm for setting initial estimates for new parameters during model development was proposed and tested on two potentially representative TMDD approximation models to enable easy estimation convergence. The likelihood-ratio test (LRT) and Bayesian information criterion (BIC) were tested as model selection criteria. Finally, the complete TMDD model development strategy was proposed and tested with five simulated data. The Quasi-steady-state model (QSS) rather than Michaelis-Menten approximation model (MMAPP) was selected as the representative TMDD approximation model after structure model search and found to be sufficient to identify the correct structure model. Other TMDD models updated initial estimates from QSS model with different gradients of initial estimates of target degradation rate constant (KDEG) and baseline target concentration (R0) also provided reasonable objective function value (OFV). Given the ranking criteria of BIC and model development strategy, the best model for each data was at least as complicated as the simulated model. Besides, 4/5 data gave accurate estimates for those non-target related parameters and not significantly worse OFV than the model with "true" parameters as initial estimates. In conclusion, the proposed TMDD model development strategy easy the TMDD model development and selection and potentially can be implemented in AMD to achieve automatic TMDD model development.