Incorporating model structure uncertainty in model-based drug discovery
Peter Gennemark
CVGI iMED DMPK AstraZeneca R&D, Sweden
Objectives: Drug discovery is characterized by relatively small pharmacokinetic and pharmacodynamic (PKPD) data sets for lead compounds that are routinely screened in an animal model. The turn-over time of pre-clinical PKPD analysis is usually short and in depth model selection is seldom executed. In addition, selection of the best model structure for one compound is hard due to sparse data. The objective of this study is to improve standard model-based predictions from preclinical data sets by incorporating model structure uncertainty, and not only parameter uncertainty, in an approach that is useful in practice.
Methods: The time constraint of model selection was addressed by automatization of time consuming modeling steps. Uncertainty in model structure was considered by evaluating several models from a model space using a model selection criterion. In this study we have used AIC. Model based predictions were generated from all models in the model space, and weighted using the posterior model probabilities from the calculated Akaike weights. Our approach was developed and tested using compound exposure and compound targeted receptor occupancy data from mice.
Results: Two main obstacles for proper model selection in drug discovery are time constraints and sparse data. Using PKPD data from a drug discovery project, we addressed both issues. To incorporate structure uncertainty we defined a large model space including a reasonable space of both PK and the PD structural models, and weighed the set of feasible models based on their posterior probability. Concerning the time constraint, we accelerated model selection by implementing a user-friendly computational process with input data in form of an Excel file and output in form of a PowerPoint presentation file. Taken together, we could rapidly obtain robust estimation with uncertainty of, e.g., compound potency, by sampling from the inferred model distribution.
Conclusions: Model structure uncertainty, and not only parameter uncertainty for one single model structure, can be incorporated in drug discovery practice. This implies improved robustness in model selection, which is particularly important when data is sparse. A more realistic estimation of model prediction uncertainty can then be expected, which is pivotal in decision making such as compound selection.
