• QSAR
• docking
• CYP3A4
• CYP2D6

The work carried out in this thesis is aimed at estimating in detail the predictive power of QSAR models obtained by using only two-dimensional (2D) or both 2D and three-dimensional (3D) molecular descriptors for analyzing interactions between ligands (substrates and inhibitors) and two isoforms of the cytochrome P450. Both metabolism and inhibition were considered in this study, which exploits different approaches based either on the known structure of the enzymes or ligand properties. A classical equation-based method (CODESSA program) and a method based on Machine Learning (WEKA program) were applied. In order to be able of giving a valuable contribution to Drug Discovery, the “in silico” tools must be very efficient, i.e. they must be, at the same time, very fast and very accurate. Handling 3D descriptors implies that molecular structures are optimized so that their biologically active conformations are properly simulated. The protocol commonly followed, which implies a simple search for low-energy conformations, did not appear to lead to reliable results. The problem of optimizing structures of molecules interacting with a given macromolecular target may be solved by docking the molecules of interest within a proper model of the active site of such a target.
The need of handling properly optimized 3D structures of ligands led to focus the second issue afforded in this thesis. As most of the docking software doesn’t allow flexible docking of the target macromolecule (with the exception of some prototypes of computer programs), part of the work carried out in this thesis was aimed at estimating how much different ligands may affect the enzyme active site conformation.
The research showed that, in the case analyzed in detail (CYP2D6), induced-fit phenomena are not negligible and they significantly affect the enzyme active site structure, which has to be exploited for optimizing ligand conformations in docking calculations.
Moreover the comparison between QSAR models only based on 2D molecular descriptors with more accurate models involving 3D descriptors, computed on the basis of optimized conformations of the ligands and based on quantum-chemical calculations, highlighted a very significant improvement in the model performance when quite accurate calculations are carried out. It suggests that the optimal compromise between fastness and reliability of a predictive model is something quite critical and strongly depends upon the system under analysis.