Data driven predictive modeling of formulation and process conditions of pharmaceutical spray dried intermediates
Elise Miller*, Gregory M. Troup, Justin Moser
Merck & Co., Inc., Formulation Sciences, West Point, PA
The pharmaceutical industry has experienced a shift in the physicochemical properties of pre-clinical drug candidates rendering them less soluble and bioavailable. Spray drying is used to enhance solubility and oral bioavailability by stabilizing the amorphous phase of a drug in a polymer matrix. However, the interaction of formulation and process variables is extremely complex. To fully simulate the process from atomization to final particle attributes based on first principles would require multiple computationally heavy models that still do not adequately capture formulation aspects.
In this work, a series of data driven models is developed and operationalized to relate formulation and process inputs to final spray dried particle properties. A user-friendly simulation tool will be showcased that enables a priori prediction of spray dried intermediate (SDI) particle properties as a function of formulation and process parameters. It combines a set of common formulation and process inputs, thermodynamic calculations of outlet process conditions, and SDI particle attributes in partial least squares projection to latent structure (PLS) models. The tool enables simultaneous evaluation of multiple hypothetical formulations and process conditions on the predicted physical attributes of particle size, bulk density, and tapped density. The multi-Y PLS models utilize data from 54 historical spray drying batches. The variables include SDI composition, solution composition, process inputs, process dependent variables, and SDI particle attributes. This exploratory analysis shows that the complexities of the spray drying process are readily captured and that the multivariate model results are consistent with process fundamentals. Then, separate optimized PLS models are developed for the key quality attributes: particle size, bulk density, and tapped density. The resulting PLS models have R2 values greater than 0.8, and Q2 values greater than 0.6, with root mean square estimation errors that are between 8% and 12% of the range. The optimized PLS models allow for prediction of these particle attributes from a limited set of formulation and process inputs, allowing proposed Design of Experiments to be simulated to ensure that a meaningful range of particle attributes can be achieved. Further, the user interface incorporates multivariate diagnostics to access the similarity between the proposed hypothetical batches and the model calibration data. The diagnostics also highlight input parameters that are outside of historical operating ranges. The tool has added utility as a training platform for scientists new to the technology and requires no modeling expertise.
See more of this Group/Topical: Pharmaceutical Discovery, Development and Manufacturing Forum