Identification of groupings of graph theoretical molecular descriptors using a hybrid cluster analysis approach

There is an abundance of structural molecular descriptors of various forms that have been proposed and tested over the years. Very often different descriptors represent, more or less, the same aspects of molecular structures and, thus, they have diminished discriminating power for the identification of different structural features that might contribute to the molecular property, or activity of interest. Therefore, it is essential that noncorrelated descriptors be employed to ensure the wider and the less inflated possible coverage of the chemical space. The most usual approach for reducing the number of descriptors and employing noncorrelated (or orthogonal) descriptors involves principal component analysis (PCA) or other factor analytical techniques. In this work we present an approach for determining relationships (groupings) among 240 graph-theoretical descriptors, as a means for selecting nonredundant ones, based on the application of cluster analysis (CA). To remove inherent biases and particularities of different CA algorithms, several clustering solutions, using these algorithms, were "hybridized" to obtain a reliable and confident overall solution concerning how the interrelationships within the data are structured. The calculated correlation coefficients between descriptors were used as a reference for a discussion on the different CA methods employed, and the resulted clusters of descriptors were statistically analyzed for deriving the intercorrelations between the different operators, weighting schemes and matrices used for the computation of these descriptors.