Finding Key Members in Compound Libraries by Analyzing Networks of Molecules Assembled by Structural Similarity

Characterization of chemical libraries is an essential task in everyday chemoinformatics practice. This study describes some potential uses of network visualization and analysis methods to identify distinguished members of compound libraries. Molecules were ordered into networks by their structural similarity defined by molecular fingerprints. Various properties of such networks were examined. It was shown, that the correlation methods used to calculate the similarity between two structures radically determined the topology of networks. From the same set of molecules, the Russel−Rao and the Baroni−Urbani methods created sparser and denser networks, respectively, than using the Tanimoto method. Central nodes, corresponding to central compounds in the libraries, were determined for some example data sets. It was shown by the case of adenosine A1, A2, and dual antagonists that the methods used to identify central nodes could be divided into two groups: (1) centrality methods, exemplified by the centroid centrality, which could pick up structures that were the most similar to the largest number of other molecules and (2) group, exemplified by betweenness centrality, that could identify molecules that had intermediate structures between some homogeneous subsets of the library. The latter method gave significantly higher ranks to dual adenosine antagonists, hinting the suitability of this measure to identify molecules with multiple activities. Some practical applications of the method for clustering of and sample selection from chemical libraries are presented. In the frame of the study, a Jchem plug-in has been developed to the Cytoscape network visualization software, which makes the visual observation of molecular networks more convenient. The plug-in is included in the Supporting Information of the article for free usage.

化学库的表征是日常化学信息学实践中的基本任务。本研究阐述了网络可视化和分析方法在识别化合物库中杰出成员方面的潜在应用。通过分子指纹定义的结构相似性，将分子有序地组织成网络。研究了此类网络的多种属性。研究表明，用于计算两个结构之间相似性的相关性方法极大地决定了网络的拓扑结构。通过对某些示例数据集的分析，发现从同一组分子中，Russel−Rao 和 Baroni−Urbani 方法相较于 Tanimoto 方法，分别创建了更为稀疏和密集的网络。对于一些示例数据集，确定了与库中中心化合物相对应的中心节点。通过腺苷A1、A2和双拮抗剂的案例，表明用于识别中心节点的方法可以分为两类：（1）中心性方法，以质心中心性为例，能够识别出与最大数量其他分子最相似的分子结构；（2）群组方法，以介数中心性为例，能够识别出在库中某些同质子集之间的中间结构的分子。后者显著提高了双腺苷拮抗剂的排名，暗示了该度量标准在识别具有多重活性的分子方面的适用性。展示了该方法在化学库聚类和样本选择中的实际应用。在研究框架内，针对Cytoscape网络可视化软件，开发了一个Jchem插件，使得分子网络的视觉观察更为便捷。该插件作为文章的补充信息部分，可供免费使用。