Protein aggregation is one of the key challenges in the biopharmaceutical industry as its control is crucial in achieving long-term stability and efficacy of biopharmaceuticals. Attempts have been made to develop regression models for predicting the aggregation of monoclonal antibodies in solution using machine learning methods. These efforts have yielded varying levels of success, with current state-of-the-art AI approaches achieving good prediction accuracies ( \(r=0.86\) ). Here, we demonstrate the prediction of aggregation rate in monoclonal antibodies with beyond state-of-the-art reliability using a coupled AI-MD-Molecular surface curvature modelling platform. The scientific novelty of this approach lies in using local geometrical surface curvature of proteins as the core element for protein stability analysis. By combining local surface curvature and hydrophobicity, as derived from time-dependent MD simulations, we are able to construct aggregation predictive features that, when coupled with linear regression machine learning techniques, give a high prediction accuracy ( \(r=0.91\) ) on a dataset of 20 molecules. More generally, this approach shows significant potential for quantitative in silico screening and prediction of protein aggregation, which is of great scientific and industrial relevance, particularly in biopharmaceutics.
