Gated Graph Sequence Neural Network-Based Prediction of Drug-Gene Association for Nucleoside Proteins in Oral Cancer

Abstract

Background: Oral cancer poses a significant public health challenge due to its increasing incidence, late-stage diagnosis, and resistance to treatment. Nucleoside metabolism plays a vital role in oral cancer biology, influencing DNA synthesis, repair, and cellular proliferation. Drug-gene associations can help identify targeted therapies by analyzing the interactions between nucleoside proteins and drugs. Graph Neural Networks (GNNs), particularly Gated Graph Sequence Neural Networks (GGSNNs), offer a promising approach to model these complex interactions.
Aim: This study aims to predict drug-gene associations for nucleoside proteins in oral cancer using a GGSNN model, with the goal of identifying potential therapeutic targets and improving drug discovery strategies.
Methodology: A dataset comprising drug-gene interaction data was preprocessed to remove missing values and normalize features. Graph data frames were constructed to represent nodes (drugs/genes) and edges (interactions). A GGSNN architecture with two layers was implemented using the GatedGraphConv layer for message passing. The model was trained for 100 epochs using an 80:20 train-test split, with performance evaluated using metrics such as Mean Absolute Error (MAE), R-squared (R²), precision-recall curves, and F1 scores. Visualization tools such as Cytoscape and dimensionality reduction techniques were used for analysis.
Results: The GGSNN model achieved an average precision score of 76.94%, an F1 score of 72.42%, and a recall rate of 85.95%. The MAE was 0.0685, indicating low prediction error, while the R² value of 26.66% highlighted moderate explanatory power. Visualization techniques revealed insights into drug-gene interactions and model learning patterns. The precision-recall curve indicated robust performance across different recall values, with a balanced threshold of -0.0147 optimizing precision and recall.
Conclusion: The GGSNN model demonstrates strong predictive capabilities in identifying drug-gene associations in oral cancer, providing a valuable tool for computational drug discovery. While the model achieves notable precision and recall, its moderate R² value suggests areas for improvement. Future work should incorporate additional biological features and validation datasets to enhance model robustness and applicability in precision medicine.