(636a) Development and Comparison of Computer-Aided Methods for the Optimal Design of Lyophilized Protein Formulations

Authors: 
Roughton, B. C., University of Kansas
Camarda, K., University of Kansas
Topp, E. M., Purdue University



Protein drugs are increasingly important, both therapeutically and commercially. Of the top 100 drugs by U.S. sales in the fourth quarter of 2012, 28 were protein drugs or other biologics. Protein drug products are often lyophilized to ensure stability and prolong shelf-life. Lyophilization, or freeze-drying, aims to produce an amorphous solid with minimal water content. Despite the benefits to chemical stability, the lyophilization process itself can cause  protein aggregation in the final drug product. Protein aggregation is undesirable due to product loss during manufacturing, reduced efficacy and increased potential for adverse immune response in patients. Formulation aims to minimize aggregation through the inclusion of excipients or additives. Excipient selection decisions are often based on prior history and/or generate-and-test methods. Neither approach guarantees that a best or even good solution will be found. Computer-aided molecular design (CAMD) offers a methodology for the rational selection of excipients based on the optimization of target formulation properties of a lyophilized drug product. The work here focuses on the properties and design of carbohydrate excipients, which represent the most common class of excipients in lyophilized protein products.

CAMD requires predictive property models for all target excipient properties. Properties of interest include glass-transition temperatures, water content in the lyophilized solid and contribution to protein stability. A statistical framework is presented for descriptor selection and cross-validation of property models to insure that the final developed model correlates well with known data and provides reasonable prediction accuracy for new data. Both linear and non-linear property models are considered. For linear models, descriptor selection is performed using Mallow’s Cp statistic and the model quality is assessed using R2 and Q2. For non-linear models, descriptor selection is guided by parameter sensitivity analysis and model quality is assessed using %AAD and reduced χ2. Prediction accuracy for a designed excipient is quantified through the use of prediction intervals. For further validation of property models, prediction accuracy is compared to experimental error.

Given a set of predictive property models, CAMD next requires a solution method. Stochastic methods are considered here due to the presence of non-linearities in descriptors and property models. Stochastic methods are also desirable due to the ability to quickly generate locally optimal solutions. Due to error in experimental measurement and property prediction, locally optimal solutions may have statistically similar values and may be statistically similar to the global optimal solution.  Thus, multiple solutions may be equally valid and worthy of further consideration. Prediction intervals  are used to compare solutions. The two stochastic methods used are Tabu search and genetic algorithms. Tabu search is a metaheuristic method that uses memory of previous solutions to enhance the search for a new solution. Genetic algorithms use the principles of evolution to guide an initial population of solutions to optimality. Both methods are tuned for the CAMD application and compared to determine if one is better suited as a stochastic CAMD solution method. Comparison results are presented specifically for the excipient design application and also in general for CAMD methods. After tuning, the stochastic methods are used to design carbohydrate excipients for several example proteins. Results between methods and proteins are compared to assess the performance of the CAMD method. The final CAMD method provides a novel lyophilized formulation design tool for the selection of optimal carbohydrate candidates given a protein drug of interest. The final design candidates can be further tested post-design through molecular modeling and/or experimental analysis.