Protein expression profiling and molecular classification of gastric cancer by the tissue array method

Abstract

Purpose: Gastric cancer is heterogeneous clinically and histologically, and prognosis prediction by tumor grade or type is difficult. Although previous studies have suggested that frozen tissue-based molecular classifications effectively predict prognosis, prognostic classification on formalin-fixed tissue is needed, especially in early gastric cancer.

Experimental design: We immunostained 659 consecutive gastric cancers using 56 tumor-associated antibodies and the tissue array method. Hierarchical cluster analyses were done before and after feature selection. To optimize classifier number and prediction accuracy for prognosis, a supervised analysis using a support vector machine algorithm was used.

Results: Of 56 gene products, 27 survival-associated proteins were selected (feature selection), and hierarchical clustering identified two clusters: cluster 1 and cluster 2. Cluster 1 cancers were more likely to have intestinal type, earlier stage, and better prognosis than cluster 2 (P<0.05). In 187 early gastric cancers (pT1), cluster 2 was associated with the presence of metastatic lymph nodes (P=0.026). Kaplan-Meier survival curves stratified by pathologic tumor-lymph node metastasis revealed that cluster 2 was associated with poor prognosis in stage I or II cancer (P<0.05). Support vector machines and genetic algorithms selected nine classifiers from the whole data set, another nine classifiers for stage I and II, and eight classifiers for stage III and IV. The prediction accuracies for patient outcome were 73.1%, 88.1%, and 76%, respectively.

Conclusions: Protein expression profiling using the tissue array method provided a useful means for the molecular classification of gastric cancer into survival-predictive subgroups. The molecular classification predicted lymph node metastasis and prognosis in early stage gastric cancer.