Background <p>Gastric cancer (GC) shows marked cellular heterogeneity and a complex immune microenvironment, which limits reliable biomarkers and therapeutic targets. We constructed a primary GC single-cell atlas to resolve malignant cell states linked to prognosis and to identify candidate targets.</p> Methods <p>Single-cell transcriptomes from tumor, adjacent/normal, and premalignant gastric tissues were integrated. After quality control and doublet removal, data were batch-corrected, clustered, and annotated. Malignant epithelial cells were inferred from large-scale copy-number variation patterns and re-clustered. A prognostic signature was constructed by using the top marker genes of the low-differentiation ARHGDIB⁺ malignant epithelial state as a biologically informed candidate pool, followed by univariable Cox regression and LASSO Cox modeling in the TCGA stomach adenocarcinoma cohort. Survival was evaluated by Kaplan–Meier analysis with log-rank testing, and discrimination was assessed by time-dependent receiver operating characteristic analysis. Immune context was estimated by computational deconvolution and Tumor Immune Dysfunction and Exclusion scoring. Cell-cell communication was inferred from ligand-receptor expression. EMB expression across GC cell lines was assessed using CCLE/DepMap data, and EMB knockdown efficiency was validated by RT-qPCR. EMB function was tested by shRNA knockdown followed by wound-healing and Matrigel Transwell invasion assays and immunoblotting of epithelial-mesenchymal transition markers.</p> Results <p>Thirteen major cell populations spanning epithelial, immune, and stromal lineages were identified. Copy-number analysis separated malignant from non-malignant epithelium and resolved 8 malignant epithelial clusters. An ARHGDIB⁺ malignant epithelial state showed the lowest differentiation state and highest stem-like potential. A seven-gene signature stratified patients into high- and low-CSL groups; the high-CSL group had worse overall survival and showed bulk-level immune features consistent with an inflammatory but impaired immune context, including an increased CIBERSORT-estimated M2 macrophage fraction and higher dysfunction/exclusion scores. EMB was prioritized as a candidate prognostic hub gene, and EMB-high cells within the ARHGDIB⁺ malignant epithelial subcluster showed inferred ligand-receptor interactions with monocytic/macrophage-related immune populations. EMB knockdown reduced gastric cancer cell migration and invasion and shifted epithelial–mesenchymal transition markers, increasing E-cadherin and decreasing N-cadherin.</p> Conclusions <p>This primary GC single-cell atlas identifies a low-differentiation ARHGDIB⁺ malignant epithelial state and uses its marker genes to derive a seven-gene CSL signature associated with poor prognosis and bulk tumor-level immune features. EMB emerges as a candidate biomarker and functional mediator of GC cell migration and invasion, warranting further mechanistic validation.</p>