Parametrization of a specific free energy function for automated docking against RNA targets using neural networks

A set of 8 RNA–drug complexes was extracted from the NDB database and used to determine new parameters of the empirical free energy model implemented in Autodock software. 248 docking experiments were performed with different values for the contributions of van der Waals, electrostatic, hydrogen bonding, torsion and desolvation, respectively. These parameters were correlated with both RMSD and ΔGbind for all docking computations using a layered neural network with back-propagation algorithm (BP-NN). The model obtained from the correlation has allowed us to adjust the parameters. The most important differences between new and the default values were observed for the electrostatic, the torsion angle loss of entropy and desolvation, while the others' terms are comparable with default data. This new set of parameters could be used specifically for virtual screening against RNA targets.