Machine Learning Models of Antibody-Excipient Preferential Interactions for Use in Computational Formulation Design

Area of Science:

• Biopharmaceutical formulation
• Protein stability
• Computational chemistry

Background:

• Excipient interactions significantly influence protein stability in therapeutic formulations.
• Predicting these interactions experimentally is time-consuming, delaying early drug development.
• Developing computational methods for predicting interactions is crucial for efficient formulation design.

Purpose of the Study:

• To develop machine learning models for predicting local antibody-excipient preferential interactions.
• To identify key antibody surface features driving these interactions.
• To enable early-stage computational formulation design for antibody therapeutics.

Main Methods:

• Developed a feature set to numerically describe local antibody surface regions.
• Trained machine learning models (including elastic net) on antibody-excipient interaction data.
• Quantified the contribution of antibody surface features to interaction coefficients.

Main Results:

• Machine learning models achieved up to 85% accuracy in predicting antibody-excipient interactions.
• Identified distinct feature contributions: carbohydrates and proline share similar features, while arginine·HCl and NaCl interactions are charge-driven.
• Demonstrated model utility in predicting experimental aggregation and viscosity behavior.

Conclusions:

• Computational models can accurately predict antibody-excipient interactions, facilitating early formulation design.
• Understanding feature contributions allows for targeted selection of excipients.
• Proposed a computational framework for designing antibody sequences and selecting excipients concurrently.