Dunn's index for cluster tendency assessment of pharmacological data sets.

Cluster tendency assessment is an important stage in cluster analysis. In this sense, a group of promising techniques named visual assessment of tendency (VAT) has emerged in the literature. The presence of clusters can be detected easily through the direct observation of a dark blocks structure along the main diagonal of the intensity image. Alternatively, if the Dunn's index for a single linkage partition is greater than 1, then it is a good indication of the blocklike structure. In this report, the Dunn's index is applied as a novel measure of tendency on 8 pharmacological data sets, represented by machine-learning-selected molecular descriptors. In all cases, observed values are less than 1, thus indicating a weak tendency for data to form compact clusters. Other results suggest that there is an increasing relationship between the Dunn's index as a measure of cluster separability and the classification accuracy of various cluster algorithms tested on the same data sets.

Comparison of combinatorial clustering methods on pharmacological data sets represented by machine learning-selected real molecular descriptors.

Cluster algorithms play an important role in diversity related tasks of modern chemoinformatics, with the widest applications being in pharmaceutical industry drug discovery programs. The performance of these grouping strategies depends on various factors such as molecular representation, mathematical method, algorithmical technique, and statistical distribution of data. For this reason, introduction and comparison of new methods are necessary in order to find the model that best fits the problem at hand. Earlier comparative studies report on Ward's algorithm using fingerprints for molecular description as generally superior in this field. However, problems still remain, i.e., other types of numerical descriptions have been little exploited, current descriptors selection strategy is trial and error-driven, and no previous comparative studies considering a broader domain of the combinatorial methods in grouping chemoinformatic data sets have been conducted. In this work, a comparison between combinatorial methods is performed,with five of them being novel in cheminformatics. The experiments are carried out using eight data sets that are well established and validated in the medical chemistry literature. Each drug data set was represented by real molecular descriptors selected by machine learning techniques, which are consistent with the neighborhood principle. Statistical analysis of the results demonstrates that pharmacological activities of the eight data sets can be modeled with a few of families with 2D and 3D molecular descriptors, avoiding classification problems associated with the presence of nonrelevant features. Three out of five of the proposed cluster algorithms show superior performance over most classical algorithms and are similar (or slightly superior in the most optimistic sense) to Ward's algorithm. The usefulness of these algorithms is also assessed in a comparative experiment to potent QSAR and machine learning classifiers, where they perform similarly in some cases.