GeoNicheTrans: Spatial Niche Inference and Immune Architecture Characterization in Oral and Oropharyngeal Squamous Cell Carcinoma Validated by CODEX Protein Profiling

Abstract

Background: The tumor microenvironment is spatially organized into distinct ecological niches that shape immune responses and therapeutic outcomes, yet inferring spatial niches from transcriptomic data and validating them at the protein level remains challenging, particularly in oral and oropharyngeal squamous cell carcinoma. Here, we applied GeoNicheTrans — a graph neural network integrating information geometry, hyperbolic geometry, and persistent homology — to infer spatial niches from spatial transcriptomics, and validated them at the protein level using an independent CODEX dataset (792,349 single cells, 31 protein markers, 9 oral and oropharyngeal squamous cell carcinoma samples). Three SPACE-GM CODEX cohorts (~2.93 million cells) and the GSE179730 anti-PD-1 OSCC cohort (12 patients, 23 samples) were used for training and independent external validation. Results: GeoNicheTrans performed robustly across cohorts (Charville AUC=0.9966, UPMC AUC=0.9338, s271 AUC=0.9311; GSE179730 AUC=1.0000, permutation p<0.001). CODEX validation confirmed four key findings: (1) CD8+ T cells exhibited a 97.6% contact rate with tumor cells at a median distance of 12 pixels; (2) PDL1 and CD279 showed significant spatial co-localization (Pearson r=0.55, 120/121 tiles significant); (3) two distinct tertiary lymphoid structure maturation states were identified — TLS1 enriched in plasma cells (39.5%) and TLS2 dominated by M2 macrophages (29.2%) and regulatory T cells (15.0%); (4) immune "cold" tumors contained 89% tumor cells versus 40% in "hot" tumors. Ablation experiments confirmed that the three geometric strategies collectively contributed ~9% performance gain, and feature importance analysis identified FPR2 as the top predictive factor (importance 0.241). Conclusions: This study provides the first systematic protein-level validation of spatial niche inference methods, establishing a complete evidence chain from spatial transcriptomics prediction to CODEX protein orthogonal validation. GeoNicheTrans provides a spatially resolved tool for understanding the immune architecture of oral and oropharyngeal squamous cell carcinoma and offers a theoretical basis for precision immunotherapy strategies.