Spatial transcriptomics technologies capture gene expression profiles with spatial context, enabling tissue architecture analysis. Current spatial domain identification methods rely on gene expression patterns and spatial information, but do not incorporate valuable biological prior knowledge accumulated in large-scale genomics databases. In this study, we present PriorST, a framework that enhances spatial domain identification by integrating foundation model-derived knowledge with spatial transcriptomics data. Our approach combines highly variable gene features with pre-trained scGPT embeddings through feature concatenation, then employs graph convolutional networks for joint representation learning that preserves both biological prior knowledge and spatial context. We evaluate PriorST on human dorsolateral prefrontal cortex (DLPFC) datasets and demonstrate improved clustering performance compared to existing methods. This approach provides a practical framework for incorporating foundation model knowledge into spatial transcriptomics analysis.