Molecular similarity based estimation of properties: a comparison of structure spaces and property spaces

Molecular similarity methods have emerged as powerful tools in analog selection, chemical classification based on toxic modes of action, and property estimation. The basic assumption of structure-activity relationships (SAR) is that similar structures usually have similar properties. Therefore, similarity methods can be used for the selection of analogs and estimation of properties of chemicals from their structural analogs in property spaces. Each similarity method is user defined. Its efficacy depends on the set of descriptors used to define the intermolecular similarity of chemicals as well as on the mathematical function used to quantify similarity. Also, similarity methods can be based on experimental data or computed molecular descriptors. We have carried out a comparative study of similarity spaces derived from experimental data vis-a-vis computed structural parameters for two sets of chemicals: (a) a diverse set of 76 chemicals derived from the TSCA Inventory and (b) the 166 structurally distinct constituents of JP-8 identified by GC/MS. Property spaces for these two sets of chemicals were created using experimental and calculated physicochemical properties. Atom pairs (APs) and topological indices calculated by POLLY v2.3 were used to create theoretical structure spaces. These spaces were used for the KNN-based estimation of properties with K=1--10, 15, 20, 25. The results will be presented with a comparative analysis of the effectiveness of property spaces and structure spaces in analog selection and property estimation.