Cross-validated spatial architecture of tissue-resident memory T-cells predicts recurrence pattern in nasopharyngeal carcinoma

Background: Tumor-immune spatial architecture influences clinical outcomes, yet the role of tissue-resident memory T cells (TRM) in this organization and recurrence remains unclear. TRM are long-lived, non-circulating T cells that maintain local immune surveillance. How TRM patterns reorganize at recurrence-and whether initial patterns predict recurrence type-has not been resolved. Prior studies focused on TRM counts rather than spatial organization. We introduce validated spatial metrics to capture TRM architecture and its predictive value for recurrence pattern, addressing a critical gap for guiding surveillance and immunotherapy strategies. Methods: Multiplexed imaging was applied to paired initial-recurrent nasopharyngeal carcinoma tissues (NPC; n = 19) to compare TRM abundance. Spatial architecture was assessed in pooled (unmatched) samples using proximity-based cellular neighborhoods and a mixing score for heterogeneity. TRM aggregates were identified by DBSCAN (≥80 TRM within 20 µm) and aggregate structure were classified as random vs concrete via the Average Nearest Neighbor Index (ANNI; p < 5 × 10⁻⁶). Aggregate-level features included density (count/tissue area), relative area (aggregate area/tissue area), and TRM per aggregate. TRM subsets (including IFN⁺ TRM, TRM1) were mapped across tumor core, invasive margin, and stroma in initial, local recurrence (LR), and distant recurrence (DR) groups. An interpretable decision tree evaluated spatial/functional predictors of recurrence pattern. Results: Recurrent tumors showed higher TRM density than initial tumors (Wilcoxon p = 0.01), with enrichment confined to PanCK⁺ tumor regions. Median fold-change was 2.2 (IQR: 0.762-7.416), and the median paired difference was 228.1 cells/mm². TRM exhibited lower mixing scores and more homotypic organization than CD8⁺ cells. TRM aggregates were detected in 11/19 initial and 9/19 recurrent samples; aggregate-level metrics were therefore derived from pooled analyses to define recurrence-associated thresholds. Concrete TRM1 aggregates were enriched at the invasive margin, and with LR samples showing the highest TRM1 proportion near the margin (0-20 µm; p = 0.0365). A decision tree identified TRM1 enrichment >76.6% in concrete aggregates as the strongest predictor of LR (gini = 0). ROC analysis of concrete TRM1 count showed robust discriminative performance for recurrence type (AUC = 84%), supporting its biomarker potential. Conclusions: Recurrent NPC tumors harbor more TRM within tumor regions, and concrete TRM aggregates enriched for TRM1 at the invasive margin characterize LR, whereas reduced TRM1 enrichment and low region-specific aggregate density mark DR. These spatial/functional TRM metrics could complement existing pathological imaging or biopsy-based surveillance strategies.