Gastric cancer (GC) is the third leading cause of cancer-related deaths worldwide. Tumor heterogeneity continues to confound researchers’ understanding of tumor growth and the development of an effective therapy. Digital cytometry allows interpretation of heterogeneous bulk tissue transcriptomes at the cellular level. We built a novel signature matrix to dissect epithelium and stroma signals using a scRNA-seq data set (GSE134520) for GC and then applied cell mixture deconvolution to estimate diverse epithelial, stromal, and immune cell proportions from bulk transcriptome data in four independent GC cohorts (GSE62254, GSE15459, GSE84437, and TCGA-STAD) from the GEO and TCGA databases. Robust computational methods were applied to identify strong prognostic factors for GC. We identified an EMEC population whose proportions were significantly higher in patients with stage I cancer than other stages, and it was predominantly present in tumor samples but not typically found in normal samples. We found that the ratio of EMECs to stromal cells and the ratio of adaptive T cells to monocytes were the most significant prognostic factors within the non-immune and immune factors, respectively. The STEM score, which unifies these two prognostic factors, was an independent prognostic factor of overall survival (HR = 0.92, 95% CI = 0.89–0.94, p=2.05×10−9). The entire GC cohort was stratified into three risk groups (high-, moderate-, and low-risk), which yielded incremental survival times (p<, 0.0001). For stage III disease, patients in the moderate- and low-risk groups experienced better survival benefits from radiation therapy ((HR = 0.16, 95% CI = 0.06–0.4, p<, 0.0001), whereas those in the high-risk group did not (HR = 0.49, 95% CI = 0.14–1.72, p=0.25). We concluded that the STEM score is a promising prognostic factor for gastric cancer.