Construction of high-quality structure-property-activity regressions: The boiling points of sulfides

Instead of using the standard molecular descriptors (topological indices) f or regression analysis, which are numerically fully determined once a molec ule is selected, we outline the use of variable molecular descriptors that are modified during the search for the best regression. The approach is ill ustrated using boiling points of sulfides. We have transformed the connecti vity index (1)chi into a function of two variables (x, y) which differentia te carbon and sulfur atoms. The optimal values of the variables (x, y) were determined by minimizing the standard error of the regression. With the va lues x = +0.25 and y = -0.95 for carbon and sulfur, respectively, we have o btained a regression based on a single descriptor and a standard error of 1 .8 degrees C. With elimination of two outliers (having a deviation of about 4 degrees C) the standard error is reduced to a remarkable 1.3 degrees C.