Prediction of Pharmacological Classes by Deep Learning Method

Deep learning methods have been applied in the drug design context as an alternative to reach new biologically active compounds¹ ². This study proposes an approach for classifying compounds into drug categories based on the analysis of their 2D structure. Thus, a dataset with 310, 194, and 280 images of Anti-inflammatories, Diuretics, and Corticosteroids, respectively, was built, which was divided into train and test sets, 80%, and 20%, respectively. As a result, a Convolutional Neural Network³ (CNN) had a classification capacity evaluated with an F1-Score of 0.77, 0.92, and 0.68 for Anti-Inflammatories, Diuretics, and Corticosteroids, respectively; and AUC-ROC values ​​of 0.81, 1.00, and 0.78. In conclusion, CNN could classify compounds into these pharmacological classes. These models can be helpful for pharmacological classification of natural compounds, prediction of adverse effects, and drug repositioning.