PREDICTION OF CHROMATOGRAPHIC RETENTION VALUES (R(M)) AND PARTITION-COEFFICIENTS (LOG P-OCT) USING A COMBINATION OF SEMIEMPIRICAL SELF-CONSISTENT REACTION FIELD CALCULATIONS AND NEURAL NETWORKS

A combination of semiempirical solvent effect calculations and artificial neural networks was used to predict the (reversed phase) retention values (R(M)) of steroids and the partition coefficients (log P-oct) of a diverse set of organic compounds. Eleven selected physical parameters from AM1 self-consistent reaction field calculations were used as input for neural networks of the back-propagation type. The performance (standard error 0.08 R(M) units) in predicting retention values of a set of steroids is close to experimental accuracy (0.03 R(M) units). The partition coefficients of a heterogeneous set of organic compounds are predicted with a standard error of 0.29 log P-oct units. Systematic leave-n-out experiments revealed that the solvation energy obtained by simple solvent calculations (spherical model) is the most important parameter in our 11 parameter set and considerably improves the performance in predicting hydrophobicity at little additional computational cost.