Three different techniques were used to develop the models: Multiple Linear Regression (MLR), Partial Least Square (PLS) and Group Method of Data Handling (GMDH) Neural Networks. These studies have been done in collaboration with F. Lemke (KnowledgeMiner, Berlin) and G. Gini (Politecnico di Milano). Thus it was possible to develop several models and compare them. We also used Biochemical analysis (Docking and GRID methodology) to compare and increase the information obtained from field-based descriptors.
The role of biochemical analysis is to understand biological processes and try to analyse and model the interaction of the ligand with the binding site and the formation of the ligand-macromolecule complex. 3D models, using CoMFA, GRID and Docking approaches, gave good results, which also indicate important likely molecular toxic metabolism processes. The use of atom based descriptors and GMDH vice versa is more related on automatic extraction of knowledge, independent on mechanism. Even if it gives lower performances, it can be more general and applicable to more heterogeneous structures.