After the discovery of target binding antibodies, the characteristics of developability such as high levels of aggregation, escalated viscosity and other stability issues assume critical importance. However, the design space for improving antibody stability is enormous because of the large size of proteins and the dependence of physiochemical properties on the formulation conditions. In this work, we implemented a combination of experimental and computational approaches to elucidate the key determinants for aggregation of 21 therapeutic monoclonal antibodies (mAbs). We have applied accelerated aggregation rate assessment, molecular dynamics simulations and a machine learning method for predicting antibody aggregation. The machine learning method combines features from sequence, amino acid properties, molecular descriptors (spatial aggregation propensity and spatial charge map) developed from our group. The contribution of different domains of antibodies is also included. With a rational feature selection, we extracted the most important features to classify and predict therapeutic mAbs with high and low aggregation rate. The cross-validation accuracy is 97% using only three features. The protocol proposed in this study has implications in training and predicting other stability properties with limited experimental data.

Predicting antibody developability from molecular simulations and machine learning

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.