Development of Novel Discrimination Model and Its Application to Predicting P-gp Substrate

随着组合化学和高通量筛选在药物发现中的出现，如何设计化合物文库变得越来越重要。特别是ADME（吸收、分布、代谢和排泄）是药物开发中的关键问题，因为它与药物的安全性和有效性密切相关。基于计算机的ADME特性预测有望降低药物开发后期的损耗率，并通过寻找有前途的化合物来优化药物筛选和测试。为此，应阐明ADME过程的分子结构特征。p -糖蛋白（P-gp）是一种外排转运蛋白，可表达多种器官。转运蛋白负责，例如，抑制外来生物进入肠道和主动排泄在肝脏和肾脏。许多已知的药物被P-gp识别，导致生物利用度不足，治疗效果持续时间短。因此，建立鉴别目标化合物是否为P-gp底物的方法是重要的问题之一。传统的模式识别算法，如判别分析和神经网络，需要所有化合物的分类信息（底物或非底物）进行分析。不幸的是，没有那么多的文献可以清楚地表明这些化合物是非底物。有限的信息使得进行大规模的数据分析变得困难。本文提出了一种基于化学空间概念的新型判别分析方法。化学空间是由各种独立的化学属性（或分子描述符）组成的超维空间。假设P-gp底物在整个化学空间中形成一个簇，我们开发了一种将P-gp在化学空间中的簇可视化到三维的方法。与投射到三维空间相关的信息损失可以通过找到最小化的加载向量来最小化，测试化合物与整个化学物质的变化率。我们认识到这个数学问题是一个广义特征值/特征向量问题。利用这种方法，我们分析了P-gp底物的分子特征。当使用化合物的拓扑描述符作为分子描述符进行分析时，发现约200个P-gp底物仅定位在整个化学空间的1/60中，包含约8,000个生物活性化合物。同样的方法也应用于口服活性药物的定位。760种口服活性药物分布在整个化学空间的1/12左右，由13万种有机化合物组成，列在可用化学目录中。本研究开发的方法通过基于化学空间概念的大规模数据可视化，直观地了解了目标分子的共同特征，从而有助于加速药物的发现和开发。少

With the advent of combinatorial chemistry and high-throughput screening in drug discovery, it is increasingly important how to design a library of compounds. In particular, ADME(absorption, distribution, metabolism, and excretion) is a critical issue in drug development, because it is closely related with safety and efficacy of drugs. Computer-based prediction of ADME properties is expected to reduce the rate of attrition in the late stage of drug development and optimize drug screening and testing by looking at promising compounds. To this end, molecular structural features responsible for ADME processes should be elucidated. P-glycoprotein(P-gp) is an efflux transporter that expresses many organs. The transporter is responsible, for example, for suppression of entry of xenobiotics in the intestine and active excretion in the liver and kidney. Many of drugs are known to be recognized by P-gp, resulting in insufficient bioavailability and short duration of therapeutic effect. Therefor … More e, it is one of the important issues to develop the method of discriminating whether aimed compounds are P-gp substrates or not. Conventionally used pattern recognition algorithms, such as discrimination analysis and neural network, need categorical information (substrate or non-substrate) for all the compounds subjected to the analysis. Unfortunately, not so many literatures are available to clearly show that the compounds are non-substrate. The limited information makes it difficult to perform a large-scale data analysis. In this study, a novel discrimination analysis method has been proposed based on the chemical space concept. Chemical space is hyper-dimensional space consisting various independent chemical attributes (or molecular descriptors). Assuming that P-gp substrates form a cluster in entire chemical space, we developed a method for visualizing the cluster of P-gp in the chemical space downsized to 3-dimension. The loss of information associated with projection into 3-dimensional space can be minimized by finding the loading vectors that minimize, the variation ratio of test compounds to the entire chemicals. We realized that this mathematical problem is one of generalized eigenvalue/eigenvector problems. By using this method, we analyzed molecular features of P-gp substrates. When the analysis was performed using topological descriptors of compounds as molecular descriptors, it was found that 〜200 P-gp substrates localized in only 1/60 of the entire chemical space comprising 〜8,000 bioactive compounds. The same method was applied to mapping of orally active drugs. Seven hundreds sixty orally active drugs distributed approximately 1/12 of the entire chemical space consisting of 130,000 organic compounds listed in available chemical directory. The method developed in this study provides intuitive understanding of common features of target molecules by visualizing a large-scale data based on chemical space concept, and therefore contributes to accelerating drug discovery and development. Less