Glioma is a primary tumor, and pathological analysis serves as the gold standard for its diagnosis. Deep learning has emerged as a potent tool for grading gliomas by analyzing whole slide images (WSIs). The phenotypic and topological distribution of constituent histological entities are crucial in glioma diagnosis; however, exploring their spatial and interactive organizations has posed a significant challenge. This study conducted a comparative validation of graph neural networks (GNNs) that characterize histological entities with graphical connections for glioma grading in WSIs. We benchmarked two different graph construction methods: 8-adjacency and K-Nearest Neighbors (KNN). Additionally, we evaluated two types of GNN models, the graph convolutional network (GCN) and the graph attention network (GAT), along with three different pooling methods. Our data cohort, consisting of 1430 glioma slides, was sourced from The Cancer Genome Atlas (TCGA). The results indicate significant performance differences among various graph construction methods and model combinations in the task of glioma grading. Among them, the combination of GAT with SAGPool demonstrates the best performance in in terms of in terms of accuracy (ACC) and Area Under the Curve (AUC), highlighting its significance potential in addressing cancer grading task from WSIs.