Drug-drug interaction (DDI) represents an important element complicating the clinical and pharmacological management of the antiretroviral therapy. Due to the high number of possible drug combinations, a limited number of DDI studies are conducted, cannot be ethically studied, and DDIs have an impact on different phases of drug development. Deep learning, a type of artificial intelligence (AI) inspired by the human neural network, offers a powerful tool to predict DDI magnitude and related risk considering large number of drug combinations. The aim of this study was to develop a deep learning algorithm, called DeepHIV, to predict DDI risk between antiretrovirals (ARV) and comedications.
The dataset was provided by the Liverpool HIV Drug Interaction database (https://www.hiv-druginteractions.org/). Molecular structure of each drug was converted to Morgan fingerprints where each atom was represented in numerical format. A mathematical framework was applied to compare the fingerprints, constructing drug similarity profiles. The architecture of DeepHIV was a feed-forward neural network where drug similarity profiles were fed into the input layer. 28,026 drug pairs were split to 80% for training and 20% for testing. DeepHIV was trained to predict 3 classes of DDI: i) Red: drugs should not be co-administered, ii) Amber: potential interaction may require close monitoring and iii) Green: no clinically significant interaction.
The performance of DeepHIV on the test set (5,824 pairs) achieved 85% mean accuracy, 70% mean precision, 69% mean sensitivity, and 84% mean specificity. A confusion matrix, which is a summary of prediction results versus the actual DDIs, was shown in Figure 1A. DeepHIV was utilised to predict DDI magnitude between first-line ARVs and comedications, which were composed of 687 drugs. The overall performance on first-line ARVs was above 80% accuracy and 70% sensitivity (Figure 1B).
DeepHIV predicts clinical and pharmacological relevant DDI between ARVs and comedication with a mean accuracy of 85%. The deep learning approach was developed exclusively integrating molecular data and provides a comprehensive evaluation of ARV DDIs. The algorithm represents a predictive tool to rationalise the risk related to DDI and delineates opportunities for a strategical integration of AI approaches during the development of novel therapies as well the identification and design of clinical studies.