VISION TRANSFORMER FOR ACTIVE COMPOUND FUNCTION CLASSIFICATION BASED ON 2D MOLECULAR STRUCTURES

Accurate classification of active compounds based on molecular structure is crucial for accelerating drug discovery while reducing laboratory costs and time. However, existing structure-based classification methods, particularly convolutional neural networks and graph-based models, often struggle to capture long-range dependencies or require large-scale datasets and extensive feature engineering. This study investigates the use of the Vision Transformer (ViT) model to classify 2D molecular structure images of compounds into cancer and cardiovascular therapy categories. A dataset containing 500 images, consisting of 250 per class, was obtained from the PubChem database, processed for consistency, and divided into 72% training, 20% testing, and 8% validation. To address the limited dataset size, careful preprocessing, regularization through weight decay, and systematic hyperparameter tuning were applied to reduce overfitting risks. The ViT model was trained with the Adam optimizer and a linear learning rate scheduler. Hyperparameters were systematically tuned to identify the optimal configuration. Results show that the best settings, with batch size 60, weight decay 0.1, learning rate 3.0×10⁻⁶, and 15 epochs, achieve an accuracy, F1 score, and loss of 80.0%, 79.9%, and 0.597, sequentially. These findings highlight the potential of ViT for small-scale cheminformatics tasks, offering an alternative to conventional methods while maintaining competitive performance.