Prediction of Solute Descriptors in LSER Equation Using Quantitative Structure-Property Relationship Methodology

In this study, a quantitative structure-property relationship method based on multiple linear regressions (MLR) and artificial neural network (ANN) techniques were applied for the calculation/prediction of Sigma beta(H)(2) and pi(H)(2) parameters of the linear solvation energy relationship (LSER). The selected descriptors that appear in multiple linear regression models for Sigma beta(H)(2) are: maximal electrotopological positive variation. average connectivity index chi-5, Geary autocorrelation-lag 1/weighted by atomic polarizabilities, radial distribution function-2/unweighted and leverage-weighted autocorrelation-lag 4/unweighted. Also descriptors that appear in MLR model for pi(H)(2) are: Geary autocorrelation-lag2/weighted by atomic Sanderson electronegativites, 2nd component accessibility directional WHIM index/weighted by atomic vander Waals volumes, d COMMA-2 value/weighted by atomic Sanderson electronegativites,. number of H attached to C-1(sp(3))/C-0(sp(2)) and dipole moment. These descriptors were used as inputs for two ANNs. After training and optimization of these ANNs, they were used to prediction of pi(H)(2) and Sigma beta(H)(2) values of the test set compounds, separately. Analysis of the results obtained indicates that the models we proposed can correctly represent the relationship between these LSER solute parameters and theoretically calculated molecular descriptors. Also results showed the superiority of neural networks over regression models.