Prediction of pH-dependent chromatographic behavior in ion-exchange systems.

In this study, predictive quantitative structure-property relationship (QSPR) models that employed a support vector machine regression algorithm and a set of novel pH-dependent descriptors were employed for the prediction of protein chromatographic behavior at any pH. The calculated pH-dependent descriptors were based on protein crystal structures and sequence information and represent charge and electrostatic potential properties on the protein surfaces. With this set of pH-dependent descriptors, proteins at different pH were treated as distinct molecules, thus enabling the generation of integrated QSPR models, which allow the prediction of chromatographic behavior of test set proteins across a wide range of mobile-phase pH conditions. The predictions from these integrated QSPR models in general showed good agreement with the experimental results. For proof of concept, the steric mass action adsorption isotherm parameters of a binary test set of proteins (lysozyme and aprotinin) at a pH not employed in the training set were predicted from the integrated QSPR models. Further, the predicted parameters were used in a macroscopic transport model to simulate the chromatographic separation of this binary protein mixture at this new pH. The simulated column behavior of these proteins showed good agreement with experimental results. The use of pH-dependent descriptors in this multiscale modeling approach now enables the prediction of various modes of protein chromatography at any mobile-phase pH, which may have significant implications for downstream bioprocessing.

A priori prediction of adsorption isotherm parameters and chromatographic behavior in ion-exchange systems.

The a priori prediction of protein adsorption behavior has been a long-standing goal in several fields. In the present work, property-modeling techniques have been used for the prediction of protein adsorption thermodynamics in ion-exchange systems directly from crystal structure. Quantitative structure-property relationship models of protein isotherm parameters and Gibbs free energy changes in ion-exchange systems were generated by using a support vector machine regression technique. The predictive ability of the models was demonstrated for two test-set proteins not included in the model training set. Molecular descriptors selected during model generation were examined to gain insights into the important physicochemical factors influencing stoichiometry, equilibrium, steric effects, and binding affinity in protein ion-exchange systems. The a priori prediction of protein isotherm parameters can have direct implications for various ion-exchange processes. As proof of concept, a multiscale modeling approach was used for predicting the chromatographic separation of a test set of proteins using the isotherm parameters obtained from the quantitative structure-property relationship models. The simulated column separation showed good agreement with the experimental data. The ability to predict chromatographic behavior of proteins directly from their crystal structures may have significant implications for a range of biotechnology processes.