Tailored Similarity Spaces for the Prediction of Physicochemical Properties

Motivation. In the past, molecular similarity spaces have been developed from arbitrary sets of molecular properties or theoretical descriptors and the results of property estimation based on these methods have always been inferior to SAR and QSAR models. Tailored QMSA methods attempt to create similarity spaces specific for a property of interest, rather than being purely arbitrary spaces characterizing the general aspects of all chemicals within the space or intuitively selected structure spaces whose elements are chosen subjectively. To this end, we have created three similarity spaces, two tailored and one non–tailored, for a set of 166 chemicals for which we have both log P and normal boiling point (BP) data. The tailored spaces were each tailored to one of the properties, while the other similarity space was developed using standard QMSA methods. Method. Ridge regression was used to determine which of the available molecular descriptors were most useful in modeling each of the available properties. Fifteen topological descriptors were selected for use as dimensions within each the tailored similarity spaces. The same number of principal components were developed using principal component analysis for the arbitrary similarity space. Results. The log P tailored similarity space was superior to both the arbitrary structure space and the BP tailored space for the estimation of log P. Also, the BP tailored similarity space was superior to the arbitrary structure space for the estimation of BP. Interestingly, the space tailored to model log P performed as well at modeling BP as did the BP tailored space. This unexpected result is explained by the degree of overlap between the indices used in both of the tailored spaces and in the presence of connectivity indices related to BP in the log P model. Conclusions. The tailored similarity method presents a promising approach to creating property specific similarity spaces derived from structural descriptors based on the results of this study and from a previous study. Further work is necessary to determine to true utility of this method with large, diverse data sets.