CS-92

A theoretical model has been developed that discriminates between active and nonactive drugs against HIV-1 with four different mechanisms of action for the active drugs. The model was built up using a probabilistic neural network (PNN) algorithm and a database of 2720 compounds. The model showed an overall accuracy of 97.34% in the training series, 85.12% in the selection series, and 84.78% in an external prediction series. The model not only correctly classified a very heterogeneous series of organic compounds but also discriminated between very similar active/nonactive chemicals that belong to the same family of compounds. More specifically, the model recognized 96.02% of nonactive compounds, 94.24% of active compounds that inhibited reverse transcriptase, 97.24% of protease inhibitors, 97.14% of virus uncoating inhibitors, and 90.32% of integrase inhibitors. The results indicate that this approach may represent a powerful tool for modeling large databases in QSAR with applications in medicinal chemistry.

Computational approaches are developed to design or rationally select, from structural databases, pyrimidyl nucleosides with anti-HIV activity. A data set of 141 nucleoside derivatives was selected from literature, and a discriminant function was derived with the use of TOPS-MODE descriptors. The model is able to classify correctly 83% of the compounds in a training set and 88.5% in a cross-validation set. The use of an external prediction set selected from the most recent literature proved that the model has good predictive ability, with a good classification of 85% of the compounds in this set. This model permitted the structural interpretation of the anti-HIV activity of these nucleoside analogues. This interpretation is formulated as several rules concerning the influence of several structural features on the activity/inactivity of such compounds. A QSAR model for the most active compounds was developed with the combined use of 2D and 3D connectivity indices. This model explains 88% of the variance in the activity of these compounds in MT4 assay. The combination of both models will permit the selection of pyrimidyl nucleoside leads and their optimization to improve the potency of the selected ones.