Supervisor: Prof. Mati Karelson, TÜ Keemia Instituut
Opponent: Prof. Maria Cristina Menziani, University of Modena and Reggio Emilia, Italy
Summary:
Linear and non-linear MLR and ANN QSA(P)R approaches were successfully applied to build accurate models for biological and physico-chemical endpoints of a great importance for the regulatory purposes. An extension of the standard QSA(P)R approach allowed building of predictive models for mixtures, thus demonstrating the applicability of the chemometric methods for modeling of complex systems. Various methods of validation such as cross-validation, hold-out test set validation and Y-scrambling were utilized to build reliable QSA(P)R models adhering to the OECD principles. Meta-analysis of the data for overlapping sets of compounds allowed transferability of the modeling results from one endpoint to another (e.g. Log(1/EC50) and Log(1/LC50)). Due to the specific nature of the utilized molecular descriptors (most originating from the quantum chemistry) the majority of models, except for those based on ANN were accompanied by a mechanistic interpretation pointing to specific molecular characteristics affecting the modeled endpoints. It was demonstrated that QSA(P)R approach can be used for making regulatory decisions solely on the basis of computational models, avoiding thus lengthy and expensive laboratory testing.