Geraldine Ayral (1), Pauline Traynard (1), Jonathan Chauvin (1)
Lixoft, Antony, France
Objectives: The use of target-mediated drug disposition (TMDD) models is growing at the same time as the number of biologic drugs in development. A large variety of TMDD models have been proposed in the literature, corresponding to different modeling assumptions [1,2]. Yet, it is often difficult to decide which TMDD approximation is the most appropriate for a given data set. In addition, the long runtime of TMDD models imposes an efficient model testing strategy. In this poster, we present guidelines to choose an appropriate model in a minimal number of iterations, using both a priori information and a posteriori diagnostic plots.
Methods: To identify a priori information that is relevant to TMDD model choice, we simulate the most common TMDD models, which are available in the TMDD model library of the MonolixSuite2018R1. The profiles obtained with different parameter values are compared.
To identify a posteriori information, several data sets have been simulated from the different TMDD models, using different parameter values within the physiological range. Several realistic limits of quantification values have also been tried. The fit of wrong and correct models to the simulated data permits to identify the most informative diagnostic plots and how to interpret the observed patterns.
Results: A priori information can be used to reduce the choice of TMDD approximations, such as binding rate and/or dissociation constants obtained from Biacore experiments. We simulate and plot the typical concentration-time curve of typical values of the ligand-receptor binding and identify threshold values for which the full/quasi-equilibrium models and the full/irreversible-binding models are indistinguishable. In those cases, the full model would be not identifiable.
Next, we performed a sensitivity analysis of each of the TMDD approximations. By comparing the range of concentrations recorded in the data set with the concentration-time curve’s zones where the parameters have an influence, non-identifiable parameters can be identified. These parameters can either be fixed to physiologically meaningful values or a model with less parameters can be chosen. The range of doses used in the data set can also influence the choice of a model. Indeed, if the initial ligand concentration is much smaller than the initial receptor concentration, the initial free ligand concentration drop (due to the binding of the ligand to the receptor) will not be observable. As a result, the parameter characterizing the receptor concentration will not be identifiable.
Finally, we fit each TMDD model to the simulated data sets in order to identify key patterns indicating a model misspecification. The following plots have been found especially informative:
- Residuals (individual weighted residuals and NPDE): trends in the residuals are often visible when an underparameterized model is used. Residuals give information about the time and concentration range where the model lacks flexibility. Combined with the parameter sensitivity analysis, the user can easily see which model could be more appropriate.
- Individual parameters versus dose covariate: the individual parameter distributions can be stratified by dose groups. If the distributions show significant trends with respect to increasing doses, the model is misspecified and a more complex model should be tried.
- Condition number and correlation matrix of the estimates: high correlations between parameters indicate that the parameters have the same (or opposite) influence on the predictions. Using an interactive simulation application such as Mlxplore, the information from the correlation matrix can be verified and visually grasped to guide the choice of another model.
Conclusion: With the development of TMDD model libraries, the testing of several TMDD models to model a data set has been simplified. Yet choosing a first model and interpreting the diagnostic plots and results to arrive at a satisfactory model is still challenging. Using a combination of model simulations, parameter sensitivity analysis and fits on simulated data, we have developed guidelines to efficiently choose and diagnose TMDD models.
References:
[1] Mager, D. E., & Jusko, W. J. (2001). Journal of Pharmacokinetics and Pharmacodynamics, 28(6), 507–32.
[2] Dua, P., Hawkins, E., & van der Graaf, P. (2015). CPT, 4(6), 324–337.
Reference: PAGE 27 (2018) Abstr 8627 [www.page-meeting.org/?abstract=8627]
Poster: Methodology - Model Evaluation