Estimation of Hansen solubility parameters using multivariate nonlinear QSPR modeling with COSMO screening charge density moments

[Display omitted] • Nonlinear QSPR models were developed for prediction of Hansen solubility parameters. • Estimations were done with artificial neural networks. • COSMO-RS sigma-moments were used as molecular descriptors. New QSPR multivariate nonlinear models based on artificial neural network (ANN) were developed for the prediction of the components of the three-dimensional Hansen solubility parameters (HSPs) using COSMO-RS sigma-moments as molecular descriptors. The sigma-moments are obtained from high quality quantum chemical calculations using the continuum solvation model COSMO and a subsequent statistical decomposition of the resulting polarization charge densities. The models for HSPs were built on a training/validation data set of 128 compounds having a broad diversity of chemical characters (alkanes, alkenes, aromatics, haloalkanes, nitroalkanes, amines, amides, alcohols, ketones, ethers, esters, acids, ion-pairs: amine/acid associates, ionic liquids). The prediction power of the correlation equation models for HSPs was validated on a test set of 17 compounds with various functional groups and polarity, among them drug-like molecules, organic salts, solvents and ion-pairs. It was established that COSMO sigma-moments can be used as excellent independent variables in nonlinear structure–property relationships using ANN approaches. The resulting optimal multivariate nonlinear QSPR models involve the five basic sigma-moments having defined physical meaning and possess superior predictive ability for dispersion, polar and hydrogen bonding HSPs components, with test set correlation coefficients R 2 d = 0.85, R 2 p = 0.91, R 2 h = 0.92 and mean absolute errors of Δ δ d = 1.37 MPa 1/2, Δ δ p = 1.85 MPa 1/2 and Δ δ h = 2.58 MPa 1/2.