Spatial transcriptomics (ST) profiles gene expression with spatial context, but most platforms capture multicellular spots containing mixed cell types, making accurate deconvolution essential. Existing reference-based methods using scRNA-seq often ignore spatial dependency and gene-level contribution, yielding fragmented maps and limited insight into domain-specific programs. Here, we propose a gene-aware heterogeneous graph attention network called STGnet for spatial transcriptomics deconvolution and functional annotation. Leveraging a hybrid pseudo-spot generation strategy that captures realistic spatially enriched cell-type patterns, STGnet accurately integrates spatial adjacency, transcriptional similarity, and gene-spot associations within a unified heterogeneous network. Attention weights highlight domain-specific genes for interpretable domain annotation. Importantly, STGnet can characterize spatially ordered functional programs across domains that may be associated with disease progression. These insights may facilitate the discovery of spatial disease mechanisms and improve understanding of pathological tissue organization. Experiments on simulated and real datasets show that STGnet achieves the best overall performance compared with state-of-the-art methods.