Simulating lipophilicity of organic molecules with a back-propagation neural network

From a training set of 7200 chemicals, a back-propagation neural network (BNN) model was developed for calculating the 1-octanol/water partition coefficient (log P) of molecules containing nitrogen, oxygen, halogen, phosphorus, and/or sulfur atoms. Chemicals were described by means of autocorrelation vectors encoding hydrophobicity, molar refractivity, H-bonding acceptor ability, and H-bonding donor ability. A 35/32/1 composite network composed of four configurations was selected as the final model (root-mean-square error (RMS) = 0.37, r = 0.97) because it provided the best simulation results (RMS = 0.39, r = 0.98) on an external testing set of 519 molecules. This final model compared favorably with a recently published BNN model using variables (atoms and bonds) derived from connection matrices.