Metabolic and Immune Heterogeneity Defines Distinct Tumor States in Esophageal Squamous Cell Carcinoma

Abstract

Background: Esophageal squamous cell carcinoma (ESCC) is transcriptionally heterogeneous, but the reproducibility of expression-defined molecular states across cohorts and microarray platforms remains uncertain. This study evaluated whether biologically coherent ESCC expression states could be identified in a pooled cross-platform dataset and whether their boundaries remained portable across cohorts. Methods: Publicly available microarray data from three GEO cohorts were analyzed, comprising 77 ESCC tumors and 77 matched normal tissues from GSE23400, GSE77861, and GSE20347. RMA-normalized expression matrices were harmonized to 13,039 common HGNC symbols and adjusted for dataset/platform-associated variation using ComBat. From the 1,200 most variable tumor genes, 264 genes strongly associated with the tumor-normal axis were excluded, leaving 936 features for consensus clustering. Candidate solutions from k=2 to k=6 were evaluated using PAC, silhouette width, CDF behavior, and concordance with non-negative matrix factorization. The resulting states were characterized by limma differential-expression analysis and ssGSEA, followed by two cross-dataset centroid-assignment robustness analyses. Results: A three-state working solution was selected, comprising S1 (n=30), S2 (n=27), and S3 (n=20), although cluster-number diagnostics were not fully concordant and NMF agreement at k=3 was modest (ARI=0.267). State distribution was not detectably associated with dataset (Fisher-Freeman-Halton p=0.335). S1 showed metabolic and detoxification-associated programs, S2 complement/B-cell and innate immune-associated programs, and S3 interferon/inflammatory/antiviral programs. Cross-dataset portability was limited: 13/24 assignments were ambiguous in the first robustness analysis (54.2%; 95% CI, 35.1%-72.1%) and 8/17 in the second (47.1%; 95% CI, 26.2%-69.0%). Conclusion: ESCC exhibits biologically coherent metabolic and immune-associated expression states, but the exact patient-level boundaries of these states are weakly reproducible across cohorts. The three-state model should therefore be considered hypothesis-generating rather than a validated molecular classifier and requires confirmation in larger, independently processed cohorts using a fully locked validation framework.