Artificial Intelligence Methods for Crystallization

描述(申请人提供)：人们普遍认为结晶是大多数X射线结构测定中的限速步骤。因此，我们一直在开发计算工具来促进这一过程，包括XtalGrow程序套件。在这里，我们建议从两个方面改进这些工具的功能和范围：1.初始筛选，这是一组(迭代)实验，有望产生一个或多个初步“点击”(明显是蛋白质的晶体材料)；2.优化实验，从初始点击开始，以衍射质量的晶体结束。这项提议的一个核心概念是，这种工具建设需要一个以知识为基础的基础。因此，该提案的广泛目标之一是制定一个框架，以便于计算的形式获取和编码知识；具体地说，这种形式将产生更有效的结晶程序。我们感兴趣的是数据如何相互作用，以及如何利用这些数据来改进结晶过程。虽然在可能的情况下将继续使用文献和实验室其他项目中的现有数据，但我们的分析也表明有必要采取主动行动，即收集完成知识库所需的选定数据。我们建议通过以下方式做到这一点：一、深化数据表示是几个领域，包括额外的蛋白质特征，纳入化学添加剂的层次结构，以及获取详细的响应数据。11.提高结晶筛选的效率：通过将归纳推理应用于改进贝叶斯信念网，将改进初始结晶筛选；还将制定程序，通过确定参数空间的未探索区域和使用其他测量，如动态光散射和浊点测定，进一步完善贝叶斯信念网，引导实验朝着更有希望的方向发展，从而处理没有令人满意的结果的问题。通过在这里应用基于案例的方法和贝叶斯方法，以及通过进一步发展自动图像分析，优化筛选将得到改进。提高整个系统的“人性化”、集成化和自动化。

DESCRIPTION (provided by applicant): It is widely believed that crystallization is the rate-limiting step in most X-ray structure determinations. We have therefore been developing computational tools to facilitate this process, including the XtalGrow suite of programs. Here we propose to improve the power and scope of these tools along two fronts: 1. Initial screening, the (iterative) set of experiments that hopefully, yields one or more preliminary "hits" (crystalline material that is demonstrably protein); and 2. Optimization experiments that begin with an initial hit and end with diffraction-quality crystals. A central concept of this proposal is that this tool building requires a knowledge-based foundation. Therefore, one of the broad goals of the proposal is to develop a framework for the acquisition and encoding of knowledge in computationally tractable forms; specifically, forms that will yield more effective crystallization procedures. We are interested in how the data interact and how that can be used to improve the crystallization process. While available data, both in the literature and from other projects in the laboratory will continue to be used wherever possible, our analysis has also demonstrated the need to be pro-active i.e. to gather selected data required to complete the knowledge base. We propose to do this by: I. Deepening the data representations is several areas including additional protein characteristics, incorporating a hierarchy of chemical additives and acquiring detailed response data. 11. Improving the efficiency of crystallization screens: Initial crystallization screens would be improved by applying inductive reasoning to the refinement of Bayesian belief nets; procedures would also be developed for dealing with the absence of promising results by identifying unexplored regions of the parameter space and using additional measurements, such as dynamic light scattering and cloud point determinations to further refine the Bayesian belief nets and steer experimentation in more promising directions. Optimization screens would be improved by applying Case-Based and Bayesian methods here as well as by further developments of automated image analysis. III. Improving the "user friendliness," integration and automation of the entire system.