W0075.html

The high-throughput production of diffraction-quality crystals remains a major obstacle in structural proteomics, with success rates rarely exceeding 15% for soluble proteins. The Center for Biophysical Sciences and Engineering, Diversified Scientific, Inc., and ANALIZA, Inc. present a unique and powerful approach for rapidly and efficiently determining optimum protein crystallization conditions. This involves the combination of three key technologies: (1) automated nano-crystallization, (2) incomplete factorial screening, and (3) a specifically designed neural net crystallization prediction program. This approach demonstrates the most promise for optimizing protein crystallization when combined with a thorough sampling of crystallization space. This is accomplished by using an incomplete factorial screen and a uniquely designed nano-crystallization system. Every crystallization trial outcome, including failures, is used to train the neural network. The neural network exhibits enhanced approximation, noise immunity, and classification properties. The self-organizing and predictive nature of the neural network allows for accurate prediction of previously untested crystallization conditions, even in the presence of noise. If properly trained, the neural network can be used to recognize important patterns of crystallization. This information allows the neural net to predict non-sampled complete factorial conditions used for optimization. Thus, it predicts the conditions that produce crystals from the entire “crystallization space” of possible experimental conditions, based upon results from a much smaller number of actual experiments performed. Therefore, a virtual screen can be performed using all possible combinations of components and variables. The top 50 scores for the predictive crystallization conditions are then experimentally prepared to determine/verify optimum crystallization conditions.