(39e) Machine Learning and Deep Learning Models for the Analysis and Prediction of Pharmaceutical Powder Blend Properties

The bulk properties such as flowability of pharmaceutical powder blends and formulations are difficult to predict as the resulting blend flowability is not a linear interpolations of the individual constituent material’s flowability. Flowability is crucial for tableting but evaluating it experimentally can be time-consuming and resource intensive. Additionally, when one of the blend components is very cohesive or cohesive, improving flowability through conventional means of blending with a good flowing material can be challenging. To this end, dry coating has emerged as an innovative technique and can be used to tailor cohesion and enhance flowability in a blend, but it further complicates property prediction as the coated blends exhibit remarkably different flowability compared to their uncoated counterparts. To overcome these challenges, machine learning (ML) and deep learning (DL) can be used owing to their ability to learn non-obvious connections within complex chemistry-structure-property relationships. Combining ML/DL with mechanistic approaches may yield better predictions of powder and powder blend properties.

Here, we applied ML/DL methods to (1) predict flowability of powder blends from properties of their constituent API and excipient powders, (2) understand the features most predictive of flowability, and (3) propose new blend formulations with enhanced properties. First, supervised ML models were trained on a dataset of 400 powder blends consisting of varying combinations of 7 APIs and 6 excipients, both with and without dry coating of nano-silica of different types and amounts. Both regression (direct prediction of the flow function coefficient) and classification (prediction of flowability regime such as free flowing, easy flowing, or cohesive) ML models were trained using size and morphology features of the constituent API and excipients, with the classification algorithms strongly outperforming the regression ones. Random forest and XGBoost models, the best performing, predict the flowability regime of the API/excipients blends with up to 85% accuracy. Further feature importance analysis demonstrated that the dry coating amount was most important in determining flowability, followed by the size and morphology features of the constituent API and excipient.

Second, a variational autoencoder (VAE), a type of generative algorithm, was trained on the experimental data. The trained VAE was used to produce 163 new “synthetic” powder blend formulations and predict the flowability of each. Following this, 12 new powder blends predicted by the ML/DL models were experimentally created and their flowability measured. Comparison of their measured flowability regimes to those predicted by the ML/DL models resulted in an accuracy of 83%. In summary, we demonstrated that supervised ML and DL models can perform well at predicting the relative flowability of powder blends based on their components, give insights into powder feature-property relationships, and be used for the proposal of new blends with improved properties. The integration of these computational approaches with mechanistic understanding can enhance process and product development in the pharmaceutical industry.