Data-efficient prediction in tableting using word embeddings and empirically-guided neural networks.

Area of Science:

• Pharmaceutical Sciences
• Computational Chemistry
• Materials Science

Background:

• Oral tablet formulation is complex and time-consuming.
• Existing predictive models struggle with categorical data and nonlinear interactions.
• Machine learning offers predictive power but lacks transparency.

Purpose of the Study:

• To develop a novel neural network framework for predicting tablet quality attributes.
• To integrate categorical formulation variables using word embeddings.
• To accelerate pharmaceutical formulation development.

Main Methods:

• Utilized a neural network with word embedding layers for categorical variables like active pharmaceutical ingredients (APIs).
• Integrated embeddings with empirically-guided output functions and a deep ensemble strategy.
• Predicted tablet quality attributes (tensile strength, density, ejection force, dosing height) based on composition, pressure, and weight.

Main Results:

• Achieved predictive accuracy comparable to or exceeding classical regression models.
• Demonstrated avoidance of physically implausible outputs.
• Revealed meaningful clustering of APIs in learned embeddings, enabling transfer learning and robust predictions for data-scarce APIs.
• Showcased that low-concentration formulations enhance predictive accuracy, supporting material-efficient designs.

Conclusions:

• Embedding-based, empirically-guided neural networks are explainable and practical tools.
• This framework can accelerate pharmaceutical formulation development.
• The approach facilitates more efficient experimental designs and material usage.