Neural network-based fingerprinting of monoclonal antibody aggregation using biolayer interferometry.

Abstract

Aggregates are widely accepted to be a critical quality attribute (CQA) for biotherapeutics and believed to impact product immunogenicity. Monitoring of aggregates is typically performed using multiple orthogonal tools as any single tool is unable to offer comprehensive characterization of aggregate species over the entire size and morphology range that they can exist in. Researchers have attempted to categorize monoclonal antibody (mAb) aggregates into six classes based on their respective physicochemical properties. In this study, we have developed model based on artificial neural network (ANN) to predict the stress history of mAb contributing to the aggregate formation, based on binding sensogram profiles obtained with biolayer interferometry (BLI). It was observed that each class of mAb aggregates exhibited unique binding profiles that were characteristic fingerprint of that class. The proposed model uses principal components extracted from the mAb-Fcγ receptor binding sensogram (106 profiles from 9 stressed mAb samples) as inputs while the unique identification codes in the form of binary coded numbers are used as model outputs. The latter served as a fingerprint for each class of mAb aggregates generated by subjecting to specific stress conditions. The ANN was trained using Levenberg-Marquardt algorithm with Bayesian regularization, using 86 sensogram profiles, in the ratio of 80:10:10 for internal training, validation, and testing. The trained ANN accurately identified each single stress condition that the samples were subjected to based on their binding sensogram profiles. The model was also able to predict stress history for samples that had been subjected to more than one kind of stress with reasonable accuracy. The proposed approach therefore can be effectively employed for control of product quality in biopharmaceutical industry as well as for prediction of stress history of a sample.