(762d) Monte Carlo Stepwise Regression for More Accurate Selection of Critical Process Parameters during Process Characterization

Design of experiments (DOE) is commonly used during process characterization (PC) to define control strategies for the manufacture of drug substance. Although the use of sophisticated DOE techniques in establishing the conditions for PC is common, the analysis of the results is often done with a stepwise regression reliant on p-value thresholds. This technique can often produce artificially large models with multiple false positives, leading to unnecessary work to examine spurious variable relationships and potential implementation of overly complex layers of control in the manufacturing process. This work describes the use of Monte Carlo cross validation (MCCV) simulations to aid in the selection of critical process parameters (CPP). Once models were selected defining the relationship between CPPs and critical quality attributes (CQA), MCCV was used to assess the accuracy of those models. Monte Carlo simulations were then used to establish an appropriate control strategy with defined operating ranges (OR) predicted to meet quality standards.

Three recombinant monoclonal antibody cell culture processes were evaluated. The process parameters that were studied included dissolved oxygen, pH, temperature, inoculation density, and feeding strategy. The quality attributes that were evaluated included aggregate levels and charge variants. Each PC was carried out using a D-optimal design performed in 2L bioreactors. The resulting data was analyzed using MATLAB. Stepwise regression with backward elimination was performed to establish baseline models. MCCV stepwise regression with both forward and backward elimination was then performed to identify process parameters which may have been inappropriately included by the baseline method. When available, large-scale data was used to validate the accuracy of the models along with MCCV. ORs were established by Monte Carlo simulation by modifying CPP ranges until appropriate quality thresholds were met.

A comparison of stepwise regression with other model building techniques was also performed. The simplified interpretation of stepwise regression presented here has the advantage of clearly defining CPPs, despite the modest increases in predictive accuracy afforded by more advanced and complex methods, such as regularization (LASSO, ridge) and dimensionality reduction (PCA) . Applying these techniques to a PC program can lead to more robust interpretation of data, higher statistical confidence in CPP selection, and establishment of more reliable ORs.