| Cancers, a group of multifactorical complex diseases, are generally caused by mutation of multiple genes or dysregulation of pathways. Identifying dysregulated pathways and biomarkers that can characterize cancers would help to understand and diagnose cancers. They can also serve as potential drug targets. In recent years, complex diseases have a rapid development in the pathogenesis study from a molecular perspective. But for the study of gastric cancer which is the most fatal in the malignant tumor, the bottleneck from periodic characteristics has not been broken through. Traditional methods have been developed to detect differentially expressed genes only between cancer and normal samples, and these genes are considered to be closely related with diseases. Unfortunately, most of the differentially expressed genes detected in one dataset do not work effectively in another dataset for the same disease, especially for complex diseases. This may result from the interactions between genes.Based on this observation, recent studies propose to identify dysregulated pathways as a new direction. Compared with the analysis of individual genes, dysregulated pathways can serve as a better biomarker for the discrimination of the disease states. However, different pathways often have crosstalk with each other, and the deregulation of one pathway may affect the activities of multiple related pathways. Therefore, the detected biomarkers will be more reliable by taking into account the functional dependencies or interactions between pathways. In this paper, we construct pathway interaction network based on protein-protein interactions, cellular pathways and gene expression. The identification of dysregulated pathways and biomarkers is thus formulated as a feature selection problem in the machine learning framework. Firstly, we identify the pathways as seeds that can optimally classify the disease states by SVM. Then, the biomarkers are extended from the seeds to other pathways in the PIN based on a heuristic algorithm. Note that each extension should maximize the performance of classification. The algorithm will be ended until the accuracy of classification no longer increases. Finally, the dysregulated module with the minimal size is detected, which can optimally discriminate the stages of gastric cancer.Experimental results show that the distribution of t-score in identified top 10% seed pathways on gastric cancer three stages by our method is consistent with clinical situations. It indicates that the seed pathways are closely related with gastric cancer. Compared with other three methods, our method is averagely higher than others about 2.0 in t-score, which also verifies that the method is significantly correlated with gastric cancer. When evaluating the identified modules using test set, the average value of corresponding AUC is about 0.7, thus indicating that the identified dysregulated modules is of great robustness in test set. In addition, comparing the other three methods with our approach, our AUC is taller than other three methods at least 0.1 on average. By the analysis of genes in identified dysregulated modules, we find that these genes are of significance and reliability in statistics. By further enrichment analysis of these modules, we find that most of these modules have functional association with gastric cancer, including phosphorylation, apoptosis, cell cycle and cell proliferation et al. Through the comprehensive analysis of the dysregulated modules, we find out the common and specific biomarkers, and these genes have a potential as drug targets. Finally, when applying this method to the PIN and ordinary network, the result using PIN is remarkably superior to that using an ordinary one. Furthermore, the construction of PIN is greater than ordinary network in time complexity. |
PDF Full Text Request