QSPR Studies on Physicochemical Properties using CODESSA

초록

Semi-empirical methods, AM1, PM3, PM6, and PM7, are widely utilized to calculate the heats of formation (HOF) of organic molecules.[1,2] When the semi-empirical method is applied to estimate HOF of 54 organic molecules, the correlation coefficient (R2) between experimental and predicted values is greater than 0.85. The slope of the linear correlation is, however, always greater than unity. The semi-empirical method calculates HOF of organic molecules very fast, but there is no way to improve the results. In this work, therefore, a new method is developed to obtain a better statistical prediction of HOF using QSPR method. The method selected in this work is to employ CODESSA package [3]. The input structures for CODESSA package were fully optimized using AM1 and PM3 and verified as local minima without any imaginary frequency. These structures were read into the CODESSA package to calculate more than 400 molecular descriptors per molecule. Statistical correlation was tried with a heuristic method to obtain the final object functions by gradually increasing the number of the descriptor (Table 1). In the conventional CODESSA study, which uses HOF as the target property, the correlation coefficient improves as the number of descriptor increases, and the final object function shows R2 of 0.9623 (AM1) and 0.9745 (PM3). When the results were applied to the external test set, the R2 values are excellent: 0.9552 (AM1) and 0.9377 (PM3). In this work, a new strategy was developed by using the difference between experimental and calculated HOF (Diff.) as a target physical property in the CODESSA study. Compared with the conventional study, the new strategy showed better results for the training set with R2 (AM1: 0.9688 and PM3: 0.9810). The other statistical variables, AAE and RMSD, are better than those from the conventional study.