Prediction of gas to water partition coefficient of some organic compounds using theoretically derived molecular descriptors

Abstract

An artificial neural network (ANN) is constructed and trained for the prediction of gas to water partition coefficients of various organic compounds. The inputs of this neural network are theoretically derived from molecular descriptors that were chosen by the genetic algorithm-partial least squares (GA-PLS) feature selection technique. These descriptors are: area-weighted surface charge of hydrogen bonding donor atoms (HDCA-2), average bond order of a C atom (P C ), Kier flexibility index ()), atomic charge weighted partial positively charged surface area (PPSA-3), and difference between atomic charge weighted partial positive and negative surface areas (DPSA-3). By comparing the results obtained from PLS and ANN models, one can see that statistical parameters (Fisher ratio, correlation coefficient, and standard error) of the ANN model are better than those of the PLS model, which indicates that a nonlinear model can simulate more accurately the relationship between the structural descriptors and the partition coefficients of the investigated molecules.