Prediction of photosensitizers activity in photodynamic therapy using artificial neural networks:: A 3D-QSAR study

Biological activity in photodynamic therapy was predicted from the molecular structure of pyropheophorbide derivatives using artificial neural networks (ANN). First, the structure of molecules was optimized and various descriptors were calculated. ANN architecture was optimized while suitable descriptors were selected applying a novel variable selection method. The reliability of models was tested by cross-validation and randomization of biological activity data. Models are able to predict biological activity from the molecular structure of the phorbide derivatives with a leave-one-out crossvalidation Q(2) of 0.956. The size of the substituents is decisive in the third direction (perpendicular to the main plain of the molecules).