Comparison of Statistical and Physiological Modeling Methods Using Examples in Drug Discovery and Development
C. Friedrich (1), R. Savic (2), R. Baillie (1), J. Bosley (1), R. Beaver (1)
(1) Rosa and Co. LLC, San Carlo, CA, USA; (2) Stanford University, Stanford, CA, USA
Objectives: Both statistical (e.g., NLME) and physiological (systems biology ODE models based on physical laws and physiological knowledge) modeling methods can be used to support decision-making in drug discovery and development. There is a lack of clarity in the field about which method is appropriate under what conditions. This study delineates (1) what questions can be addressed, (2) what data are needed, (3) how hypotheses are used and tested, and (4) how to build confidence in each kind of model.
Methods: The authors systematically reviewed ten examples from their modeling disciplines in similar therapeutic areas to address the research questions above.
Results: (1) Statistical modeling is best suited to quantify pharmacokinetic/pharmacodynamic processes and to separate/quantify different sources of variability. Physiological modeling is ideal for exploring mechanistic connections between pathophysiology, therapeutic pathways and outcomes. (2) Statistical models are fully inferred from clinical or pre-clinical data sets with most model parameters being estimated; the model complexity is determined by the data. Physiological models start with knowledge and hypotheses of biological processes. Many types of data are used to inform and parameterize the models. (3) The statistical modeling process is guided throughout by the hypotheses to be tested: for model building, addition of mechanistic components, covariate relationships etc. In physiological models, scope is guided by the decision to be made, modeling uncovers knowledge gaps, and the models facilitate investigation of the systemic implications of alternative hypotheses. (4) For statistical models, many tools are available to evaluate models internally and externally and assess goodness of fit. For physiological models, matching data is also critical, and additional criteria must be met to ensure that the model is relevant and adequately addresses uncertainty and variability.
Conclusion: Statistical and physiological modeling methods in drug discovery and development share some attributes, which should help make each method conceptually accessible to practitioners of the other. There are also distinguishing features that could inform choice of approach and interpretation of results. Based on our exercise, we conclude that the methods are complementary. Additional work is under way to crisply define hand-off points and methodologies to optimize overall use of modeling in drug discovery and development.