High-Throughput Self-Interaction Chromatography: Applications in Protein Formulation Prediction

Abstract

PURPOSE: Demonstrate the ability of an artificial neural network (ANN), trained on a formulation screen of measured second virial coefficients to predict protein self-interactions for untested formulation conditions. MATERIALS AND METHODS: Protein self-interactions, quantified by the second virial coefficient, B22, were measured by self-interaction chromatography (SIC). The B22 values of lysozyme were measured for an incomplete factorial distribution of 81 formulation conditions of the screen components. The influence of screen parameters (pH, salt and additives) on B22 value was modeled by training an ANN using B22 value measurements. After training, the ANN was asked to predict the B22 value for the complete factorial of parameters screened (12,636 conditions). Twenty of these predicted values (distributed throughout the range of predictions) were experimentally measured for comparison. RESULTS: The ANN was able to predict lysozyme B22 values with a significance of p<0.0001 and RMSE of 2.6x10(-4) mol ml/g2. CONCLUSIONS: The results indicate that an ANN trained on measured B22 values for a small set of formulation conditions can accurately predict B22 values for untested formulation conditions. As a measure of protein-protein interactions correlated with solubility, B22 value predictions based on a small screen may enable rapid determination of high solubility formulations.