Creation of novel large dataset comprising several granulation methods and the prediction of tablet properties from critical material attributes and critical process parameters using regularized linear regression models including interaction terms

[Display omitted] Our aim was to understand better the causal relationships between material attributes (MAs), process parameters (PPs), and critical quality attributes (CQAs) using an originally created large dataset and regularized linear regression models. In this study, we focused on the following three points: (1) creation of a dataset comprising several tablet production methods, (2) the influence of interaction terms of MAs and/or PPs, and (3) comparison of regularized linear regression models with partial least squares (PLS) regression. First, we prepared 44 kinds of tablets using direct compression and five kinds of granulation methods. We then measured 12 MAs and two model CQAs (tensile strength and disintegration time of tablet). Principal component analysis showed that the constructed dataset comprised a wide variety of particles. We applied regularized linear regression models, such as ridge regression, LASSO and Elastic Net (ENET), and PLS to our dataset to predict CQAs from the MAs and PPs. As a result of external validation, the prediction performance of the models was sufficiently high, although ENET was slightly better than the other methods. Moreover, in almost all cases, the models with interaction terms showed higher predictive ability than those without interaction terms, indicating that the interaction terms of MAs and/or PPs have a strong influence on CQAs. ENET also allowed the selection of critical factors that strongly affect CQAs. The results of this study will help to understand systematically knowledge obtained in pharmaceutical development.