Prediction of anti-HIV-1 activity of a series of tetrapyrrole molecules.

Anti-HIV-1 activities of 20 tetrapyrroles (hematoporphyrin derivatives, meso-tetraphenylporphyrins, a chlorin, and a phthalocyanine) were predicted based on their molecular structures using artificial neural networks. The molecular structures were optimized by HyperChem program using MM+ molecular mechanics and conformational search for the global minimum conformer. Eighty-seven theoretical descriptors were calculated for characterization of molecular structures. The network architecture was optimized, and suitable descriptors were selected applying a novel variable selection method. The 3DNET program was used for the calculation of descriptors and for neural network computations. The reliability of models was tested by randomization of biological activity data, leave-one-out, leave-n-out cross-validation, and external validation process. The predictive ability of the artificial neural network was compared to other model building methods, like multiple linear regressions and partial least squares projection to latent structures. For prediction of anti-HIV-1 activity, the artificial neural network gave the best results at cross-validation processes and at external validation as well. We built four nonlinear models with good predictive ability in all validation steps, which can be applied to predict the anti-HIV-1 activity of tetrapyrrole-type compounds in a much better way than with any other three-dimensional quantitative structure-activity relationship methods published to date.