A Curated Dataset for Drug Class Prediction and Repositioning

This curated dataset offers a valuable resource for deep learning applications in drug discovery and repositioning domains. It contains 5,350 high-resolution images systematically categorized into pharmacological classes and molecular targets. The pharmacological classes encompass antifungals, antivirals, corticosteroids, diuretics, and non-steroidal anti-inflammatory drugs (NSAIDs), while the molecular targets emphasize Alzheimer's disease-related enzymes, including acetylcholinesterase, butyrylcholinesterase, and beta-secretase 1. The dataset was meticulously compiled using data from well-established databases, including DrugBank, ChEMBL, and DUD-E, ensuring diversity and quality in the compounds selected for training. Active compounds (true positives) were sourced from DrugBank and ChEMBL, while decoy compounds (true negatives) were generated using the DUD-E protocol. The decoy compounds are designed to match the physicochemical properties of active compounds while lacking binding affinity, creating a robust benchmark for machine learning evaluation. The balanced structure of the dataset, with equal representation of true positive and decoy compounds, enhances its suitability for binary and multi-class classification tasks. The collection of compounds is diverse and of high quality, thus supporting a wide range of deep learning tasks, including pharmacological class prediction, virtual screening, and molecular target identification. This ultimately advances computational approaches in drug discovery.