Mechanistic models in medicine development
Research Objectives:
Mechanistic models, by which we mean attempts to model an underlying biological or chemical process, typically via a set of differential equations plus some distributional assumptions, are widely used in medicine development. Examples include the modelling of chemical processes within the drug manufacturing process and the modelling of how concentrations of drugs, drug targets or drug products changes with time in the human body. Such models may also be used to analyse endpoints in clinical trials, for instance measures of lung function in asthma trials or measures of cognitive function in Alzheimer's disease trials. However here standard repeated measures analysis, based on purely descriptive modelling of the data, provides a simpler and widely used alternative. This project aims to understand the properties of mechanistic models in this context, determine under what circumstance the two approaches (mechanistic vs descriptive), might be expected to give different answers and understand how to design studies under mechanistic models generally, with application to the example areas above.
Description of work:
We will work with several real examples from GSK's development pipeline. These will be used to motivate the detailed work-plan, and will be provide a framework to answer a series of important unanswered questions regarding mechanistic models.For instance--- what happens when the model is saturated with many random effects? How best to introduce a treatment effect into the model? Is it sensible to locate a linear part in the model where additional covariates can be slotted in and out? Can these be used safely with partial data collected at interim reviews of the data, where several subject have incomplete data series, to trigger adaptive design decisions? How sensitive are mechanistic models to the modelling assumptions used, and how should this be investigated? In particular, how sensitive are they to partial data, given their reliance on a functional shape and the importance of a decision based on treatment effect at the end of the time sequence? We are particularly interested in Bayesian solutions, because we are increasingly using Bayesian inference to make decisions within the medicine development process. The project will involve both methodological development and applied data analysis; the balance between these can be tailored to some extent to suit the researcher's preferences and skills.
Host Institution: GlaxoSmithKline