A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training.

Area of Science:

• Biopharmaceutical development
• Computational biology
• Machine learning in drug discovery

Background:

• Therapeutic antibodies need high affinity, specificity, and good developability for manufacturing and in vivo performance.
• Predicting antibody developability is challenging due to limited training data, often relying on empirical testing.
• Computational models for antibody binding are improving, but developability prediction lags behind.

Purpose of the Study:

• To establish a high-throughput platform for generating large antibody developability datasets.
• To develop and validate machine learning models for predicting antibody developability.
• To improve the prediction of antibody developability compared to existing methods.

Main Methods:

• Optimized and automated existing developability assays.
• Developed an integrated data analytics pipeline for antibody data.
• Generated data on 246 antibodies across 10 developability assays.
• Trained an XGBoost machine learning model using a tidy data format.

Main Results:

• Collected developability data for 246 antibodies (approved, clinical-stage, and withdrawn).
• Developed an XGBoost model that predicts similarity to approved antibodies more effectively than traditional thresholds.
• Demonstrated that predictive model performance improves with increased training data.

Conclusions:

• The PROPHET-Ab platform enables scalable data generation for antibody developability.
• Improved machine learning models can enhance the prediction of antibody developability.
• This work facilitates the development of more successful therapeutic antibodies.