Predicting the Effects of Charge Mutations on the Second Osmotic Virial Coefficient for Therapeutic Antibodies via Coarse-Grained Molecular Simulations and Deep Learning Methods.

The impact of various charge mutations on the second osmotic virial coefficient was examined for three model therapeutic monoclonal antibodies (MAbs) at representative formulation pH values by using coarse-grained (CG) molecular modeling. The wild-type of each mAb was characterized experimentally in previous work, showing a range of behaviors spanning from weak protein self-interactions to strong electrostatically driven attractions or repulsions as a function of pH at low ionic strength. The performance and accuracy of the underlying CG model in identifying key residues that contribute strongly to electrostatically driven self-interactions were validated experimentally in prior work with a relatively small number of candidate mutations. The present work focused on computationally exploring a large number of potential mutations (∼10-10) for each mAb as a case study for an algorithm that could provide a means to assess how altering surface charge distributions affects protein self-interactions quantified in terms of the second osmotic virial coefficient. The results for a set exhaustive or near-exhaustive range of single-, double-, and triple-mutations indicate that simple design rules such as changing the total net charge or trying to identify "charge patches" are not robust for providing predictable improvements in protein self-interactions based on electrostatic interactions, and the approach here can provide an efficient way to make predictions based on physics-based force fields. The molecular simulations were also used as a data generator for a deep neural network and explored an extensive number (∼10-10) of mutations for identifying sequences that improve protein self-interactions. Cross-validation of the output of MLP (multilayer perceptron) with the molecular simulations demonstrated high computational efficiency and prediction accuracy, highlighting its utility as an effective tool for accelerating candidate selection in therapeutics design.