Predicting Subcutaneous Antibody Bioavailability Using Ensemble Protein Language Models

Monoclonal antibodies are pivotal in modern therapeutics,
yet predicting
their subcutaneous bioavailability remains challenging due to the
intricacies of the SC environment and the limitations of traditional
experimental models. In this study, we introduce a novel machine learning
framework that leverages protein language models (PLMs) to derive
high-dimensional embeddings directly from antibody sequences. Using
three distinct PLMsantiBERTy, ABlang, and ESM-2we
extracted numerical representations that were subsequently refined via feature selection and dimensionality reduction. A systematic
evaluation of multiple classifiers using Leave-One-Out cross-validation
led us to develop a robust ensemble model based on a tuned support
vector machine classifier, which achieved a validation accuracy of
89%. This ensemble approach, which aggregates predictions across antibodies,
outperforms prior computational methods. To facilitate broad accessibility,
we deployed the model as a web application, SubQAvail, enabling rapid
bioavailability predictions from input antibody sequences. Our findings
demonstrate the potential of integrating PLM-derived features with
ensemble learning to enhance the predictive accuracy and scalability
of mAb bioavailability assessment, thereby accelerating the therapeutic
development pipeline.