(157y) Predicting Antibody Developability from Molecular Simulations and Machine Learning

In the present study, we developed a general machine learning feature selection protocol to predict antibody stabilities. These features only require information from sequences and molecular structures of antibodies. In addition, the physically relevant features are easily interpretable and help to design experiments to understand the underlying mechanisms of antibody stabilities. This feature selection protocol avoids fitting data with hundreds of descriptors and is readily applicable to find predictive models for various antibody stabilities. We applied this protocol to predict the aggregation behaviors of monoclonal antibodies (mAbs). We have measured the accelerated aggregation rate of 21 therapeutic mAbs at high concentration (150 mg/mL). We found that using only three features, the accuracy of predicting the aggregation rate of the 21 mAbs reaches 97%. Models trained using structural features, rather than sequence features, correlated better with aggregation. A predictive equation for screening aggregation behaviors was proposed, facilitating early stage design. This framework can be extended to train other models that predict different physical properties.