Leveraging Graph Neural Networks for Interpretable Prediction of Pathological Stages in Prostate Can
Updated: Sep 29, 2022
Wenkang Zhan, Department of Computer & Information Science, Temple University Chen Song, Department of Computer & Information Science, Temple University Xinghua Shi, PhD, Department of Computer & Information Science, Temple University
Early-stage cancer diagnosis is an effective approach to predict cancer treatment outcomes and improves survival rates. With the vast amount of transcriptomics data quickly available, it is desirable to develop new computational tools that utilize such data to allow for early detection of cancer and predict cancer stages. Recently, one particular type of machine learning models, namely graph neural networks (GNNs), have gained increasing attention thanks to their superior performance in modeling complex relationships in data to improve prediction tasks. However, GNNs are typically treated as black-box models, and the interpretability of the model's prediction remains challenging. Hence, in this study, we propose an end-to-end GNN framework for predicting cancer stages, that can automatically generate a graph representation using gene expression data and implement a multi-head graph attention network to make the prediction. Furthermore, we provide interpretable GNN models to interpret prediction results from multiple perspectives, including at gene level, sub-network and patient levels. To demonstrate the usage of this newly proposed GNN framework, we applied the framework to predict pathological stages in prostate cancer based on gene expression profiles of patients using The Cancer Genome Atlas (TCGA) data. Experimental results showed that our method not only outperformed competing methods for predicting cancer stages, but also identified marker genes and their interactions that are effective for such prediction. Many of these genes have been previously reported to be associated with prostate cancer and other cancers, while some new genes we identified that are effective for predicting prostate cancer stages deserve further investigation and follow-up studies. Our study suggests the potential of our proposed method for interpretable prediction of diseases taking advantage of complex relationships in clinical and omics data.