In order to improve the recognition accuracy of metabolic abnormal regions in tumor immune microenvironment, a heterogeneous graph structure integrating single-cell transcriptome, spatial location and tissue labeling information is constructed, and cross-cellular modeling of metabolic perturbations and abnormality detection are realized based on graph neural network. The study introduces a negative sampling strategy for structural perturbations to generate training edge sets, constructs node representations through a three-layer GraphSAGE network, and designs an anomaly scoring mechanism by combining the cross entropy of edge classifications with the local density regular term. The graph object is constructed based on GSE158055 spatial transcriptome data and multiple comparison experiments are carried out. The results show that the model outperforms baseline methods such as GCN, GAT and MLP in several indexes, with an AUC of 0.924 and an F1 of 0.881, demonstrating good adaptability to heterogeneous metabolic networks and predictive stability. Different node and edge combinations have significant effects on the detection performance, and the three types of edge structures integrating metabolic dependence, expression coupling and spatial adjacency have stronger discriminative ability in complex organizational environments.