Determining the Applicability of QSPR Models to Property Prediction for Query Compounds

宾夕法尼亚州立大学的Peter Jurs教授得到了分析和表面化学项目探索性研究的小额资助，以使用神经网络学习方法来预测化学性质。第一轨道包括通过使用其属性值已知的化合物的训练集来生成定量结构-属性关系（QSPR）模型。 第二个轨道，这是新的这个项目，包括生成一个二元分类器或相似性评估器训练，以区分化合物的结构是类似的训练集与那些结构不类似的训练集。 定量构效关系（quantitative structural/activity relationships，QSAR）是分子设计的一种重要方法，其目标是通过计算来评估分子的性质，从而筛选出具有分离、传感、生物活性等功能的化合物。 这项工作旨在提高使用化学数据库预测重要化学性质（如毒性和医疗功效）的准确性。

Professor Peter Jurs of Pennsylvania State University is supported by a Small Grant for Exploratory Research from the Analytical and Surface Chemistry Program to use neural network learning methods to predict chemical properties. The first track consists of generating the Quantitative Structure-Property Relationship (QSPR) model by using a training set of compounds whose property values are known. The second track, which is novel to this project, consists of generating a binary classifier or similarity assessor trained to distinguish betweeen compounds whose structures are similar to those of the training set versus those whose structures are not similar to the training set. The goal is to be able to assess properties of molecules computationally so that large numbers of compounds can be screened for effectiveness in separation, sensing, biological activity, and so on. The use of computational quantitative structure/activity relationships (QSAR) is an important approach to molecular design in industry. This work seeks to improve the accuracy of prediction of important chemical properties such as toxicity and medical efficacy using chemical databases.