We created classification models to study how the structure of A/H1N1 flu inhibitors relates to their activity (structure-activity relationship (SAR) models). These models rely on a unique machine learning approach known as a self-consistent extreme classifier. The training sample consisted of expert-selected information regarding the structures and biological test outcomes of 2,255 drug-like compounds. Independent variables included the IC₅₀ (50 % inhibitory concentration), the CC₅₀ (half-maximal cytotoxic concentration), and the SI (selectivity index). The structure was described using molecular descriptors called quantitative atomic neighborhoods (QNA). The predictive accuracy (AUC-ROC) for the IC₅₀, CC₅₀, and SI values was 0.822, 0.875, and 0.875, respectively. Evaluation of the developed models on an independent test set of 16 diverse chemical compounds demonstrated a classification accuracy of 63 % for predicting antiviral activity and cytotoxicity. The study shows that original SAR models can be used to select promising drugs to fight the A/H1N1 flu virus before they are made or tested in a lab.