Prediction of methyl radical addition rate constants from molecular structure

Multiple linear regression and computational neural networks (CNNs) are used to develop quantitative structure-property relationships for methyl radical addition rate constants. Structure based descriptors are used to numerically encode substrate information for 191 compounds. Descriptors can be classified as topological, geometric, electronic, or combination. A six-descriptor CNN was developed that produced training set rms error = 0.381 log units and rms error = 0.496 log units for an external prediction set. A seven-descriptor CNN was used to build a model for a subset of 172 of the compounds. Training set rms error was 0.424 log units and prediction set rms error reduced to 0.409 log units. Model predictions were on the order of experimental error.