Variable selection in multivariate modeling of drug product formula and manufacturing process

Area of Science:

• Pharmaceutical Sciences
• Chemometrics
• Data Science

Background:

• Multivariate data analysis, including partial least square (PLS) modeling, is increasingly utilized in pharmaceutical product development.
• Initial PLS models often contain irrelevant and collinear predictor variables, necessitating robust variable selection strategies.

Purpose of the Study:

• To apply PLS modeling to pharmaceutical product development data.
• To assess predictor variable importance and implement effective variable selection techniques.
• To optimize pharmaceutical process models through rigorous variable reduction.

Main Methods:

• Analysis of a pharmaceutical product development dataset using PLS modeling.
• Assessment of predictor variable importance using variable importance for projections (VIP) and coefficient values.
• Implementation of a stepwise predictor reduction strategy for variable selection.

Main Results:

• Irrelevant and collinear predictors were identified as common issues in initial PLS models.
• Variable importance rankings (VIP and coefficients) effectively identified redundant predictors.
• Stepwise reduction of predictors based on rankings proved effective in removing collinear variables.

Conclusions:

• Variable selection is a critical step for optimizing pharmaceutical process models developed with PLS.
• The proposed stepwise variable selection procedure enhances the evaluation of variable importance and model optimization.
• This methodology leads to more accurate and efficient drug product process models.