Virtual multiplex immunofluorescence characterization of spatial immune architectures and tertiary lymphoid structures in high-grade serous ovarian cancer.

5585 Background: GigaTIME ( Cell , 2026) is a deep learning–based virtual multiplex immunofluorescence (mIF) model that has been validated using TCGA datasets and enables spatial tumor immune microenvironment (TIME) profiling directly from H&E whole-slide images (WSIs). This study aims to evaluate the feasibility of GigaTIME for characterizing the spatial TIME of high-grade serous ovarian cancer (HGSOC) using public datasets. Methods: We analyzed UBC-OCEAN and CPTAC-OV HGSOC H&E WSIs using GigaTIME to generate virtual 21-channel (e.g., DAPI, CK, CD3, CD8, CD20, and Ki67) mIF pixel-level inferred probability maps. Exploratory analyses of 5×5 stitched regions of interest (ROI, 2560×2560 pixels ≈ 1.28×1.28 mm) were performed to detect tertiary lymphoid structures (TLS) via density-based spatial clustering of applications with noise (DBSCAN), identified as CD20⁺ clusters (major axis ≥100 μm) adjacent to CD3⁺ zones, with maturity defined by intra-cluster Ki67⁺ proliferation. Tumor-immune interaction metrics were analyzed in ROIs using Group Degree Centrality (GDC, fraction of tumor cells contacting immune cells) and Cluster Co-occurrence Ratio (CCR, observed/expected contacts). CPTAC-OV HGSOC H&E WSI–derived virtual mIF features were correlated with bulk proteomics using Spearman’s rank correlation. Integrated K-means clustering was performed on the combined cohorts after Z-score normalization. Results: The UBC-OCEAN cohort (N=195) exhibited an immune-inflamed phenotype with high tumor-immune engagement (mean GDC [mGDC] 0.25) and structural aggregation (mean CCR [mCCR] 1.41), whereas CPTAC-OV (N=222) displayed an immune-excluded phenotype with minimal contact (mGDC 0.05) and near-random distribution (mCCR 1.03). While total TLS density was comparable (UBC-OCEAN 3.92 vs. CPTAC-OV 3.60), UBC-OCEAN exhibited a twofold higher density of mature TLS compared to CPTAC-OV (1.83 vs. 0.77). Integrated clustering (k=2, silhouette=0.71) revealed that High-TLS (N=39) tumors paradoxically exhibited significantly lower mGDC (p=0.016) and mCCR (p=0.001) despite higher TLS density, indicating a spatially segregated and organized immune architecture. In CPTAC-OV validation, virtual CD3 intensity showed a significant positive correlation with bulk proteomics (N=17, ρ =0.547, p=0.023), supporting the biological validity of the T-cell-based spatial metrics. Conclusions: GigaTIME is feasible for evaluating comprehensive spatial TIME including TLS and demonstrates the heterogeneous immune landscape of HGSOC, providing spatial insights that are not captured by bulk proteomics alone. This warrants the utility of virtual mIF-based profiling as a scalable tool to identify distinct immune phenotypes for personalized immunotherapy at minimal cost.