Supporting data for "Identification of Antiviral Drug Combinations Using Machine Learning Approaches"

In this study, we acquire approximately 14,000 variables per compound. The individual drug compound features and corresponding average cell survival rates are randomly split into a training set and a test set, making up 70% and 30%, respectively. We employ algorithms such as RF, Enet, GBM to train the ML models. Additionally, we use the Pearson correlation coefficient to filter variables for training the machine learning models and conduct area under the AUC assessments to evaluate model performance. This enables us to select the top five models with the highest R values to establish predictive machine learning models against H1N1. This is a crucial first step in building our single drug machine learning models.To progress towards dual drug predictive machine learning models, we use eight anti-H1N1 drugs as inputs. These drugs are clinically used to treat influenza viruses. Subsequently, we consolidate the final valuable data from the features of these eight anti-H1N1 drugs and 1,612 compounds and their related cell survival rates to feed into the top five models. This allows us to derive the predictive values of each drug combination and ultimately find 148 combinations among the top 200 ranked by efficacy across all models.To validate the synergistic anti-H1N1 effects of these predictive drug combinations, we first conduct wet experiments of IC50 detection for the eight anti-H1N1 drugs and procure the cell survival rates of 1,184 drug combinations (148 x 8 drugs) via the CellTiter-Glo® luminescent assay. Based on the results of the wet experiments, we use the experimental average values as the true values for evaluating the efficacy of the drug combinations against H1N1. Then, we assess the AUC between the true values and the predictive values of the top five models. The Enet models demonstrated the best performance.To examine drug synergism, we use the Bliss and HSA methods to calculate the CI values.