(11d) Multivariate Data Analysis to Drive Process Understanding and Optimization

Chemical and bio processes can be sensitive to a wide range of parameters. Significant amounts of data are generated to find certain Critical Quality Attributes (CQAs) not meeting specifications. In such cases, a retrospective analysis helps identify trends and patterns in the data. Multivariate data analysis (MVDA) is a very powerful tool which uses various algorithms to elucidate inter-parameter interactions; as well as correlations with output variables. This presentation involves four case-studies wherein MVDA has been implemented to enhance process understanding.

• Spectroscopic methods are widely used for tracking reactions and crystallizations in pharmaceutical processes. In many cases, entire spectra (rather than a single peak specific to a molecule) need to be tracked to accurately represent the process, thus creating a complex multivariate data-set. In this example, a Partial Least Squares (PLS) regression was performed on Raman spectra to understand thermodynamic dependence of crystal forms over a range of temperatures and solvent compositions.
• A lipid nanoparticle formulation was manufactured in a recent manufacturing campaign, with several process parameters and output variables. Principal Component Regression (PCR) provided correlations between parameters and each variable, while PLS regression provided overall scores and loadings plots to eliminate outliers amongst batches.
• Process parameters for a cell-culture process were investigated to quantify influence on cell growth and product quality. Trajectories were generated using the scores plot to track evolution of the batches over time and projection models showed how failed batches deviated from ideal runs. The correlation loadings plot highlighted the continuous variables which had the strongest influence at each stage of the process.
• Data was generated from a series of product stability experiments for two monoclonal antibodies. Multivariate Linear Regression (MLVR) was used to correlate independent variables to CQAs. Results were verified using a PLS regression and a predictive mechanistic model was created.

In each of the above examples, MVDA was able to extract useful insights and successful recommendations were made for process improvement.