Multivariate methods in tablet formulation suitable for early drug development: Predictive models from a screening design of several linked responses

Multivariate methods were utilized in a screening design with several responses of a tablet formulation. The aim was to get a predictive model by using as few experiments as possible. Six designed factors and one uncontrolled factor were initially screened in a fractional factorial design assuming a linear model. The results were analyzed by partial least squares (PLS). The PLS loading plot showed that the nine responses were clustered into two separate groups. Therefore the design was analyzed using two different models. To get predictive models some refinements were performed. Some important responses showed conflicting factor settings in the two models. In order to reach the target levels of these responses, the two models were linked together in a simulation with a desirability function and a simplex algorithm. It was then possible to choose levels of the factors for additional experiments to reach response targets. A predictive model was obtained by adding only four extra experiments to separate confoundings in order to evaluate significant interaction terms, as well as one identified quadratic term. A tablet formulation of desired tablet strength and a fast drug release profile was obtained with a high drug:filler ratio, and high amount of disintegrant. If these variables were set at proper levels it was possible to avoid addition of a surfactant despite its significant effect on the in-vitro drug release profile.