(701e) Computation of Reliabilities for Extent of Reactant Conversion Incorporating Batch Effects with Applications to ICH Q8 Design Space

A fundamental reaction model was developed to explain rate observations under periods of both kinetic and mass transfer control.  The reaction starting material was a potential genotoxin such that the residual starting material concentration in the product was strictly regulated.  The primary means of controlling the starting material concentration in the product was via the extent of the reaction.  Hence, it was important to be able to reliably forecast the reaction conversion as a function of process and equipment parameters.  This presentation will demonstrate a means of gauging the reliability of the reaction model.      

It has been shown (Peterson, 2008) that a predictive distribution for quality responses was important for process capability and ICH Q8 design space quantification.  Experiments involving mechanistic models for API production often have relatively small measurement error variability.  However, predicting a future quality response involved sources of variation other than just measurement error.  For example, one needed to model the batch-to-batch variation, which can be substantially larger than the measurement error variation. 

Within the Bayesian statistical framework, development of a predictive distribution for future quality responses is, in theory, straightforward.  However, it can be computationally challenging to compute such distributions, and numerically intensive methods are often necessary.  In this presentation, we use the BUGS software application to develop Bayesian predictive models with and without random variation due to batch effects.  We show that failure to model the batch-to-batch variation can result in noticeable bias for process capability and ICH Q8 design space development.