Study of Drug Attribute Prediction Model Based on the Neural Network of Divided Attention Mechanism Graph

Abstract

Accurate drug molecular classification is essential for understanding biological activity and guiding clinical applications in drug discovery and development. Traditional methods such as Support Vector Machines (SVM) have demonstrated some success, yet they often struggle with graph-structured data, particularly when confronted with sparse node features and redundant information. Despite significant advances in topology-based drug attribute prediction and machine learning, classification accuracy and efficiency remain critical challenges. Traditional models often fail to capture the intricate relationships within molecular graphs, especially when it comes to feature extraction and information processing. In this study, we focus on enhancing drug molecular classification using graph-based models. This paper begins by preprocessing a graph dataset and manually extracting feature vectors. Dimensionality reduction with t-SNE is then applied to visualize the molecular distributions. A baseline binary classification model using SVM with a Gaussian kernel is implemented, achieving an accuracy of 69.67%. However, this performance is suboptimal. To address this, we introduce an end-to-end Graph Neural Network (GNN) model, incorporating a basic GNN layer, a pooling layer, and a classifier. This model improves accuracy to 89.19%. To further tackle challenges such as node feature sparsity and information redundancy, we augment the model with a frequency division attention mechanism. Additionally, we expand the dataset using Graph Generative Adversarial Networks (GraphGANs), increasing the sample size to 940. The resulting model achieves an accuracy of 92.41%, demonstrating significant performance improvements. Finally, we conduct robustness and sensitivity tests to thoroughly evaluate the models strength's and limitations. Our results reveal the effectiveness of the proposed approach in overcoming key challenges in drug molecular classification and emphasize the potential of GNNs in advancing computational drug discovery.