A Novel Explainable Intelligent Computational Framework for Identifying Potential Oncogenes

Objective: This study developed a novel explainable artificial intelligent model, deep mining graph convolutional neural network (DMGNN) model, to identify potential oncogenes. Method: RNA transcriptome data from patients with five types of cancers, including stomach adenocarcinoma (STAD), lung squamous cell carcinoma (LUSC), liver hepatocellular carcinoma (LIHC), esophageal carcinoma (ESCA), and bladder urothelial carcinoma (BLCA), were collected in the TCGA database. A novel explainable intelligent model named DMGNN was developed to identify potential oncogenes and important gene pairs based on the trained GNN model and its explanation algorithm. Potential oncogenes of universal cancers (POUC) and potential oncogenes of specific cancers (POSC), as well as the important gene pairs were identified based on DMGNN. To evaluate the DMGNN model, the numbers of cancers-related genes found in identified POUC, as well as the numbers of related genes of each cancer in this study found in identified POSC, were compared with RF and XGBOOST algorithms. Result: Numbers of cancer-related genes found in identified POUC of DMGNN were much more than RF and XGBOOST. Numbers of related genes of each cancer in this study found in identified POSC of DMGNN were also much more than RF and XGBOOST. Conclusion: The DMGNN model may help us to identify potential oncogenes, and open up potential research directions to explore unknown cancer genes.