Predicting Subcutaneous Antibody Bioavailability Using Ensemble Protein Language Models.

Area of Science:

• Biotechnology
• Computational Biology
• Pharmacology

Background:

• Monoclonal antibodies (mAbs) are crucial therapeutics.
• Predicting subcutaneous (SC) bioavailability of mAbs is difficult due to the complex SC environment and limitations of current experimental models.

Purpose of the Study:

• To develop a novel machine learning framework for predicting mAb subcutaneous bioavailability.
• To leverage protein language models (PLMs) for enhanced predictive accuracy and accessibility.

Main Methods:

• Utilized three distinct PLMs (antiBERTy, ABlang, ESM-2) to extract high-dimensional embeddings from antibody sequences.
• Applied feature selection and dimensionality reduction to refine numerical representations.
• Developed an ensemble model using a tuned support vector machine classifier with Leave-One-Out cross-validation.

Main Results:

• Achieved a validation accuracy of 89% for the ensemble model.
• The ensemble approach, aggregating predictions across antibodies, demonstrated superior performance compared to previous computational methods.
• Developed and deployed the SubQAvail web application for accessible bioavailability predictions.

Conclusions:

• Integrating PLM-derived features with ensemble learning significantly improves the accuracy and scalability of mAb bioavailability assessment.
• The SubQAvail application facilitates rapid predictions, accelerating the therapeutic development pipeline for monoclonal antibodies.