Selection of features from protein-protein interaction network for identifying cancer genes

Cancer is a group of complex diseases, in which a relatively large number of genes are involved. One of the main goals of cancer research is to identify genes that causally relevant to the development and progress of cancer. The increasingly identified cancer genes and availability of genomic and proteomics data provide us opportunities to identify cancer genes by computational methods. In this work, we investigated five predictive topological features, derived from the protein-protein interaction networks, in identifying cancer genes. We used 10-fold cross validation to assess the predictive ability of all the combinations of these features and found the most predictive feature and feature combinations. Two kinds of neural networks, support vector machine (SVM) and multi-layer perceptrons (MLP), were employed to assess the predictive ability of features. We found that the best feature combination for these two algorithms is the same. At the same time, we found SVM performs slightly better than MLP. Using only 2 or 3 features, the best performance of our classification model can get accuracy as high as 73.9%.