Recent advancements in computer-aided drug discovery revolutionize healthcare by integrating virtual screening (VS) and artificial intelligence (AI). Virtual screening enables the efficient screening of vast chemical libraries in silico, reducing the number of compounds requiring physical testing in the lab before drug synthesis or repurposing. An essential aspect of successful virtual screening is the representation of chemical compounds. While traditionally represented as feature vectors, leveraging convolutional neural networks (CNNs) to interpret chemical structures as images has emerged as a promising approach, harnessing the learning capabilities of CNNs. One potential application of CNNs is in creating classifiers capable of accurately distinguishing between drugs and decoys. These classifiers could serve as a foundation for developing generative adversarial neural networks (GANs), facilitating the synthetic generation of potential non-toxic drugs. This study, which attempts to serve as a basis for future work in the field of smart health, assesses a selection of pre-trained CNNs for their efficacy in classifying drugs associated with diabetes, cancer, and malaria. To enhance model training, a data augmentation phase has been incorporated, introducing variations to the initial images to impart rotational invariance to the learning process. Results indicate that DenseNet201 exhibits superior accuracy, albeit with considerable computational time requirements. Surprisingly, excluding data augmentation significantly improves predictive performance across all models, challenging the initial assumptions. Consequently, applying pre-trained CNNs for drug classification is contingent upon specific conditions, necessitating carefully considering augmentation strategies for optimal outcomes.