# Graph attention networks for predicting drug-gene association of glucocorticoid in oral squamous cell carcinoma: A comparison with GraphSAGE

## TL;DR

This study compares two graph-based models, GAT and GraphSAGE, for predicting drug-gene interactions in oral squamous cell carcinoma, finding that GAT performs better in some metrics.

## Contribution

The study introduces a novel comparison of GAT and GraphSAGE for drug-gene association prediction in the context of glucocorticoids and oral squamous cell carcinoma.

## Key findings

- GraphSAGE outperformed GAT in accuracy, macro-averaged F1 score, and AUC-ROC metrics.
- GAT achieved higher accuracy and F1 scores in identifying glucocorticoid interactions in OSCC.
- Both models demonstrated effectiveness in predicting drug-gene associations, with GAT showing promise for personalized therapeutic approaches.

## Abstract

The present study evaluates the effectiveness of Graph Attention Networks (GAT) and GraphSAGE in predicting drug-gene interactions for glucocorticoids in oral squamous cell carcinoma, thereby aiding in developing better treatment strategies.

We utilized a curated dataset containing known drug-gene interactions and corresponding molecular profiles. Both GAT and GraphSAGE were implemented to model the biological networks of drug-gene relationships. Experiments were conducted to evaluate each model’s performance using accuracy, precision, recall, and F1-score metrics.

The network analysis details 174 nodes and 409 edges with a sparse structure, moderate connectivity, and low clustering, indicating a diverse node connection. The analysis confirms a fully connected network with efficient computation time. In comparing models, GraphSAGE outperforms GAT with higher accuracy (0.949 vs. 0.947), better macro-averaged F1 score (0.275 vs. 0.195), and higher AUC-ROC (0.780 vs. 0.514), suggesting stronger class-distinction capabilities. Both models achieve high accuracy, but GraphSAGE’s superior scores in F1 and AUC-ROC indicate a more effective balance in precision and recall. The results demonstrated that both GAT and GraphSAGE effectively predicted drug-gene associations. However, GAT outperformed GraphSAGE, achieving higher accuracy and F1 scores in identifying relevant glucocorticoid interactions in the context of OSCC.

Our findings highlight the efficacy of advanced graph-based methodologies in elucidating drug interactions in OSCC. GAT, in particular, shows promise for accurately predicting drug-gene associations, which may facilitate personalized therapeutic approaches. Future research will focus on enhancing these models and exploring additional drug compounds to understand their applicability in OSCC treatment.