A circulating GPNMB-based multimodal model integrates tumor-immune crosstalk to predict immunotherapy response in esophageal cancer

Neoadjuvant immunotherapy improves outcomes in esophageal squamous cell carcinoma (ESCC), yet ~70% of patients fail to respond. Pretreatment biopsies and plasma thus provide critical opportunities for biomarker discovery. Here, we performed plasma proteomic profiling and identified soluble glycoprotein non-metastatic melanoma protein B (sGPNMB) as the most elevated circulating protein in non-responders. Mechanistically, tumor cell-derived sGPNMB suppressed CD8 + T cell receptor (TCR) signaling to induce functional exhaustion, with secretion being required for its immunosuppressive activity. Cancer-associated fibroblast-epithelial (CAF-Epi) niches promoted SOX2 upregulation in tumor cells, transcriptionally activating GPNMB expression. In humanized PDX models, plasma GPNMB levels predicted response to PD-1 blockade, and GPNMB inhibition synergised with therapy. Across retrospective cohorts and a prospective clinical trial, a multimodal model combining plasma GPNMB levels, CAF-Epi niche detection, and clinical-pathological features achieved robust predictive accuracy for immunotherapy response and survival. These findings establish a mechanistically grounded, spatial-circulating biomarker framework for precision immunotherapy in ESCC.