Predictive QSAR Modeling

描述（由申请人提供）：该项目的主要目标是开发，验证和提供高效的计算工具，用于快速可靠地预测药物样分子的生物活性和/或相关药物特性。我们计划开发具有统计意义且稳健的定量结构-活性关系（QSAR）方法，该方法包含严格的验证程序，并导致具有高预测能力和实用价值的模型。这些方法建立在相似性原则之上，即化学结构的相似性或多样性决定了其生物作用的相似性或多样性。我们认为，化学相似性应该在目标属性的背景下进行评估，并使用客观的相似性和多样性函数来实现化合物在描述符空间中具有生物学意义的聚类。因此，我们的方法采用变量选择程序，旨在识别与目标属性最相关的描述符。对于我们的方法来说，最重要的是严格的外部数据集模型验证，这确保了预测QSAR模型最终应用于筛选化学数据库或生物活性化合物的虚拟库时的最高命中率。本建议与其具体目标相对应的四个主要重点领域包括：发展新颖的非线性QSAR方法，如k最近邻（kNN）和支持向量机（SVM）方法。重点将放在开发新的化学结构描述符，以及基础方法的效率、自动化和统计稳健性，以使其能够应用于大型商业数据集。开发高效、客观的QSAR模型验证方法，确定模型的适用范围，最大限度地提高模型的预测能力。这项研究应该导致建立广泛接受的可靠和广泛验证的QSAR模型的“良好实践”。与实验研究者合作，将经过验证的QSAR建模方法应用于具有药理或药学意义的各种数据集；这部分还包括基于数据压缩和从高维到低维描述符空间映射的高级算法的有效数据分析和模型解释的新方法的发展。在公开访问的UNC QSAR web服务器上实现在此工作过程中开发和验证的所有建模方法。该提案的成功实施有望提供高度自动化，预测性和可访问的QSAR建模工具，这将使在药物设计和发现领域工作的广泛研究团体受益。

DESCRIPTION (provided by applicant): The main objective of this project is to develop, validate, and deliver efficient computational tools for rapid and reliable prediction of biological activity and/or related pharmaceutical properties of drug-like molecules. We plan to develop statistically significant and robust Quantitative Structure-Activity Relationships (QSAR) methodologies, which incorporate rigorous validation procedures and lead to models with a high predictive power and practical utility. The methodologies are built upon the similarity principle, i.e., similarity or diversity of chemical structures determines similarity or diversity of their biological action. We argue that chemical similarity should be evaluated in the context of the target property and employ objective similarity and diversity functions to achieve biologically meaningful clustering of compounds in the descriptor space. Consequently, our methodologies employ variable selection procedures aimed at identifying descriptors most relevant with respect to the target property. Paramount to our approach is rigorous model validation with external datasets which ensures the highest hit rates when predictive QSAR models are ultimately applied to screening chemical databases or virtual libraries for biologically active compounds. Four major areas of concentration in this proposal corresponding to its Specific Aims include: . development of novel, mainly non-linear QSAR methods, such as k nearest neighbor (kNN) and Support Vector Machines (SVM) approaches. The emphasis will be on the development of novel descriptors of chemical structure, and the efficiency, automation and statistical robustness of the underlying methodologies to afford their application to large commercial datasets. . development of efficient and objective QSAR model validation methodologies, which define the domain of model applicability and maximize the predictive ability of the models. This studies should lead to establishing widely accepted "good practices" of reliable and extensively validated QSAR models. . application of validated QSAR modeling methods to various datasets of pharmacological or pharmaceutical significance in collaboration with experimental investigators; this part also includes the development of new approaches for effective data analysis and model interpretation based on advanced algorithms for data compression and mapping from high- to low dimensional descriptor space. . implementation of all modeling methods developed and validated in the course of this work in the publicly accessible UNC QSAR web server. Successful implementation of this proposal is expected to afford highly automated, predictive and accessible QSAR modeling tools, which will benefit a broad research community working in the area of drug design and discovery.