Cancer Distant Metastasis Identification Based On Optimized Graph Representation Of Gene Interaction Patterns

Metastasis is a major cause of high cancer morbidity and mortality,and most cancer deaths are caused by cancer metastasis rather than by the primary tumor.The prediction of metastasis based on computational methods has not been explored much in the previous research.In this paper,we proposed a graph convolutional network embedded with a graph learning module,named glm GCN,to predict the distant metastasis of cancer.Through the graph learning module,we obtained the best graph representation of gene interaction suitable for network structure and data set,and then predicted tumor metastasis under the GCN framework.The framework extracted informative high-level features from the constructed irregular graph structure and obtained highly scalable feature information,which greatly improved the prediction accuracy.We compared the proposed method with other methods,trained based on two data sets,and further verified with the data sets of the other two cancer types.A series of experiments show the effectiveness of this method.The main innovations are as follows:(1)At present,most methods for directly predicting cancer metastasis were based on one expression data,which can not provide enough genetic information,resulting in low algorithm performance.We used two kinds of RNA-seq data(m RNA and lnc RNA),whose expression can provide more genetic information than m RNA alone.We used them to construct best graph representation to consider the influence of genetic interaction.(2)Aiming at the problem that most of the graphs analyzed in the graph convolution layer are existing or artificially constructed in this field,which does not necessarily meet the needs of the network structure,resulting in poor network prediction accuracy,we embedded a graph learning module in the proposed glm GCN to learn the best graph representation of gene interaction,and integrated the graph learning layer and the graph convolution layer into the same network structure.The embedded graph learning layer learns the gene interaction relationship through a single-layer network,and the graph convolution layer extracts gene features from the learned graph representation,so as to better predict cancer metastasis.(3)Aiming at the problem that the graph learning module mainly predicts through the distance between genes,which is easy to ignore the existing interaction relationship in this field,our graph learning layer pays more attention to the gene-gene relationship of the domain graph through power value.Therefore,this method can obtain more accurate prediction performance than previous methods.In this paper,we firstly constructed the protein-protein interaction(PPI)network to represent the initial gene(node)relationship graph.Then through the graph learning module,a new graph representation was built which optimally learned the gene interaction strength.Finally,the GCN was adopted to identify the distant metastasis cases.