Mapping networks of anti-HIV drug cocktails vs. AIDS epidemiology in the US counties

Abstract The implementation of the highly active antiretroviral therapy (HAART) and the combination of anti-HIV drugs have resulted in longer survival and a better quality of life for the people infected with the virus. In this work, a method is proposed to map complex networks of AIDS prevalence in the US counties, incorporating information about the chemical structure, molecular target, organism, and results in preclinical protocols of assay for all drugs in the cocktail. Different machine learning methods were trained and validated to select the best model. The Shannon information invariants of molecular graphs for drugs, and social networks of income inequality were used as input. The nodes in molecular graphs represent atoms weighed by Pauling electronegativity values, and the links correspond to the chemical bonds. On the other hand, the nodes in the social network represent the US counties and have Gini coefficients as weights. We obtained the data about anti-HIV drugs from the ChEMBL database and the data about AIDS prevalence and Gini coefficient from the AIDSVu database of Emory University. Box–Jenkins operators were used to measure the shift with respect to average behavior of drugs from reference compounds assayed with/in a given protocol, target, or organism. To train/validate the model and predict the complex network, we needed to analyze 152,628 data points including values of AIDS prevalence in 2310 counties in the US vs. ChEMBL results for 21,582 unique drugs, 9 viral or human protein targets, 4856 protocols, and 10 possible experimental measures. The best model found was a linear discriminant analysis (LDA) with accuracy, specificity, and sensitivity above 0.80 in training and external validation series.