[Prediction of binding affinities of protein-ligand complexes using nonlinear models]

A network model for prediction of the free energy changes in protein-ligand complexes has been developed. The 150 complexes of different nature were used as a training set. The computational physics-chemical parameters of these complexes were used as independent variables. Both classical models of multiple linear regression and several network models with one hidden layer were created and the best was chosen. Significant improvement was shown for network model prediction quality in comparison with classical model of multiple linear regression (R2 on training--0.81 and 0.54; R2 on "leave-one-out" procedure--0.74 and 0.52 respectively).