A neural network approach for estimating physicochemical properties using quantitative structure-property relationships (QSPRs)

Neural network based quantitative structure-property relationships (QSPRs), were developed for estimating boiling point temperature, critical properties, vapor pressure, aqueous solubility, octanol/water partition coefficient and the Henry's Law constant for a heterogeneous set of organic compounds, that included branched and cyclic aliphatics, aromatics, PAHs, PCBs, pyridines, mercaptans, amines, phenols, esters, ketones, halogens, alcohols, and ethers. Back-propagation and quantitative fuzzy ARTMAP neural networks were derived to correlate quantitative structure-property relationships (QSPRs) among quantum chemical and topological molecular descriptors and the desired physicochemical properties. Estimations of the vapor pressures ranging from 102 to 107 Pa, at temperatures ranging from 200–600 K, for a set of 274 hydrocarbons (aliphatic, aromatic, and PAHs), were achieved from a back-propagation/QSPR that performed with an average absolute error of 0.039 log units or 9.2% (26.8 kPa). The fuzzy ARTMAP/QSPR and 11-13-1 back-propagation/QSPR models predicted aqueous solubilities for a heterogeneous set of 515 organic compounds, ranging from −11.62 to 4.31 logs [mol/l] with average absolute errors of 0.14 and 0.26 logs [mol/l] units, respectively. The logK_ow for a heterogeneous set of 442 organic compounds were predicted with average absolute errors of 0.14 and 0.23 logK_ow using the fuzzy ARTMAP/QSPR and 12-11-1 back-propagation/QSPR models, respectively. Finally, the fuzzy ARTMAP/QSPR correlated the Henry's Law constant [loges] of 495 compounds, ranging from −6.72 to 2.87 loges with average absolute error of 0.13 loges units, while the optimal 7-17-1 back-propagation/QSPR model predicted the Henry's Law constant with average absolute error of 0.27 loges units. The use of fuzzy ARTMAP and back-propagation neural networks was shown to be a feasible approach for developing QSPRs for physicochemical properties for a wide spectrum of organic compounds at accuracies well within the range of expected experimental errors and of better accuracy than other regression based correlations. The major advantages of the neural network based QSPRs is that they can be developed without having to a priori specify the analytical form of a particular correlation model and that a single QSPR can be developed to be applicable to a wide range of chemical groups and even multiple physicochemical properties.