Classification of Potent Dengue Inhibitors Using 2-D Molecular Images With Structure-Based Augmentation and Grad-CAM Visualization

Dengue fever is responsible for a significant number of deaths worldwide. Despite ongoing research, there is currently no cure for this mosquito-borne disease. The Non-Structural 3 (NS3) protein is a crucial component of the Dengue virus replication and has emerged as a promising target for novel therapeutic interventions. Traditional drug discovery is often costly and inefficient, making in-silico approaches like Quantitative Structure-Activity Relationship (QSAR) increasingly popular. Although several QSAR methods exist for identifying NS3 inhibitors, they rely on molecular descriptors and may fail to capture the full range of molecular structural properties. Our study proposes a novel application to identify NS3 inhibitors based on 2D images of molecular structures. These images offer a more complete representation of molecular features. To address data scarcity, we applied both image augmentation and Simplified Molecular Input Line Entry System (SMILES) enumeration to identify the most effective one. The model trained on images of enumerated SMILES achieved 96.86% test accuracy on the augmented balanced test set, although its performance dropped on the smaller raw test set. Regardless, our image-based method significantly outperforms existing descriptor-based models. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations revealed that one of the models consistently focused on the hydroxyl group, pointing to it as a potentially relevant compound associated with NS3 inhibition. Future work could focus on applying advanced deep learning and in-vitro techniques to better understand molecular interactions.