A Hybrid Classification Tree and Artificial Neural Network Model for Predicting the In vitro Response of the Human Immunodeficiency Virus (HIV1) to Anti-Viral Drug Therapy

Abstract

Artificial Neural Networks (ANN) are pattern recognition and prediction tools modeled on the biologic nervous system. Chemical structures can be decomposed to critical elements that can be used to teach an ANN to relate structure to function. This study was carried out to evaluate a hybrid ANN system for predicting the in vitro anti-HIV1 activity of potential anti-viral drugs based on their chemical structures.

Structure and activity data were obtained from an AIDS anti-viral screen database for 371 drugs. A proprietary algorithm was used to decompose molecules into critical elements that were used as input variables for a probabilistic artificial neural network (PNN). Initial modeling and input variable selection was carried out using Classification and Regression Tree analysis (CART™, 1999).

Important variables together with the terminal node outputs from CART™ were used to train the hybrid ANN to predict in vitro anti-HIV1 activity of each drug. Cross-validated testing results for the ANN model using only the important variables as inputs were compared with hybrid ANN predictions. The final hybrid ANN model correctly classified 96.8% (96.1-99.2%) of the chemical structures re anti-HIV1 activity. The area under the receiver operator curve (ROC) for the hybrid model was 0.97 (0.95-0.99), for the conventional ANN 0.89. The hybrid ANN model performed better for sensitivity, specificity, positive/negative predictive values (p<0.01). We conclude that elements of chemical structure can be used to train a hybrid ANN to accurately predict in vitro activity of drugs against wild type HIV1.