Prediction of hydroxyl radical rate constants from molecular structure

Quantitative structure-property relationships are developed using multiple linear regression and computational neural networks (CNNs). Structure-based descriptors are used to numerically encode molecular features that can be used to form models describing reaction rates with hydroxyl radicals. For a set of 57 unsaturated hydrocarbons, a 5-2-1 CNN was developed that produce d a root-mean-square (rms) error of 0.0638 log units for the training set a nd 0.0657 log units for an external prediction set. The residual sum of squ ares for all 57 compounds was 0.234 log units, which compares very favorabl y with existing methodologies. Additionally, a 10-7-1 CNN was built to pred ict hydroxyl radical rate constants for a diverse set of 312 compounds. The training set rms error was 0.229 log units, and the rms error for the exte rnal prediction set was 0.254 log units. This model demonstrates the abilit y to provide accurate predictions over a wide range of functionalities.