Exploring Cancer Genomics with Graph Convolutional Networks: A Comparative Explainability Study with Integrated Gradients and SHAP

In the rapidly advancing field of cancer genomics, identifying new cancer genes and understanding their molecular mechanisms are essential for advancing targeted therapies and improving patient outcomes. This study explores the capability of Graph Convolutional Networks (GCNs) for integrating complex multiomics data to uncover intricate biological relationships. However, the inherent complexity of GCNs often limits their interpretability, posing challenges for practical applications in clinical settings. To enhance explainability, we systematically compare two state-of-the-art interpretability methods: Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP). We quantify model performance through various metrics, achieving an accuracy of 76% and an Area Under the ROC curve is 0.78, indicating the model’s effective identification of both overall predictions and positive instances. We analyze and compare explanations provided by IG and SHAP to gain more knowledge in the decision-making processes of GCNs. Our framework interpret the contributions of various omics features in GCN models, with the highest SHAP score observed for feature MF:UCEC and the highest IG score for KIF11. This approach identifies novel cancer genes and clarifies their molecular mechanisms, enhancing GCN interpretability. The study improves GCN accessibility in personalized medicine and contributes to understanding cancer biology.