Atomic-level-based AI topological descriptors for structure-property correlations.

Abstract

Multiple linear regression (MLR) analysis is used to construct the structure-boiling point models for 71 sulfur-containing organic compounds in terms of the Xu index and atomic level AI indices. The potential of these descriptors is further verified by three high quality QSPR models obtained for two subsets of compounds and a combined set of all compounds. For these subsets, containing respectively 45 sulfides and 26 thiols, the best three-parameter models are obtained, and the best four-variable model is obtained for the whole data set of 71 compounds. The correlation coefficients r are larger than 0.997 in all three final models. The standard errors s are 3.14, 2.48, and 3.48 degrees C for the sulfide subset, the thiol subset, and the whole data set, respectively. Furthermore, the results indicate that the boiling points are dominated by the molecular size, but some atomic types in a molecule are important due to interactions between atomic groups of the molecules. Both the molecular size and atomic types related to different fundamental interactions provide the separate contributions to boiling points. Finally, the three final models are further validated to be statistically significant and reliable by the leave-one-out cross-validation method.