Physics-Informed Neural Networks for Ion-Exchange Chromatography-Based Separation of Monoclonal Antibody and Acidic Variants

Separating monoclonal antibody products from acidic charge variants typically requires linear pH or salt gradients or a dual gradient in ion-exchange chromatography, which can be difficult in commercial applications. A method is proposed that combines the Yamamoto model for linear gradient elution with Møllerup’s thermodynamic approach to predict the distribution coefficient of monoclonal antibodies and their acidic variants based on salt concentration and pH. A physics-informed neural network (PINN) model was employed to identify optimal salt conditions for effective separation, achieving an R2 score of 0.999 in just 90 s. The estimated Gibbs free energy values were consistent with existing literature, and the predicted normalized gradient slope (GH) versus salt concentration (I) curves were within ±4.42% of experimental uncertainty. Using the optimized distribution plots from the PINN model, a step salt gradient and a pH-salt dual linear gradient elution strategy is created. The step gradient elution achieved 96.6% mAb purity and 93.1% yield, while the dual gradient elution resulted in 94% purity and 82% yield. PINN modeling helped enhance chromatographic process development, requiring only three experimental data points per elution pH to effectively separate closely related impurities.