Mining histopathological and genetic data provides a unique avenue to deepen our understanding of cancer biology. However, extensive cancer heterogeneity across image- and molecular-scales poses technical challenges for feature extraction and outcome prediction. In this study, we propose a feature-enhanced graph network (FENet) for genetic mutation prediction using histopathological images in colon cancer. Unlike conventional approaches analyzing patch-based feature alone without considering their spatial connectivity, we seek to link and explore non-isomorphic topological structures in histopathological images. Our FENet incorporates feature enhancement in convolutional graph neural networks to aggregate discriminative features for capturing gene mutation status. Specifically, our approach could identify both local patch feature information and global topological structure in histopathological images simultaneously. Furthermore, we introduced an ensemble strategy by constructing multiple subgraphs to boost the prediction performance. Extensive experiments on the TCGA-COAD and TCGA-READ cohort including both histopathological images and three key genes’ mutation profiles (APC, KRAS, and TP53) demonstrated the superiority of FENet for key mutational outcome prediction in colon cancer.