Spatial heterogeneity of the tumor immune microenvironment across ovarian cancer histologic subtypes: an AI-generated virtual multiplex immunofluorescence study
Objective We aimed to decode the spatial tumor immune microenvironment (TIME) across histologic subtypes of ovarian cancer (OC) using artificial intelligence (AI)-generated virtual multiplex immunofluorescence (mIF). Methods GigaTIME, a deep learning model trained on 40 million cells, generated 21-channel virtual mIF from H&E whole-slide images (WSI). The UBC-OCEAN dataset (513 WSIs: high-grade serous carcinoma [HGSC, n = 217], endometrioid carcinoma [EC, n = 119], clear cell carcinoma [CC, n = 94], low-grade serous carcinoma [LGSC, n = 42], mucinous carcinoma [MC, n = 41]) was analyzed for cellularity-normalized biomarker expression and spatial metrics: Immune Infiltration Efficiency (IIE), i.e. the proportion of CD8 + T cells within 20 μm of a cytokeratin + tumor cell, and the Cluster Co-occurrence Ratio (CCR), i.e. the ratio of observed to randomly expected CD8 + –tumor contacts; alongside tertiary lymphoid structure-like lymphoid aggregate (TLA) density. TLAs were graded by size and Ki-67 activity. Proteomic correlation used the CPTAC-OV dataset (35 patients, 113 WSIs). Results After cellularity normalization and false discovery rate adjustment, CD8, CD20, CD4 and CD138 differed across subtypes. Two marker-level failures were identified: predicted Ki-67 did not separate HGSC from LGSC (P = 0.273), and CD3 and CD8 channels were uncorrelated to weakly inversely correlated in every subtype (ρ = −0.29 to −0.02), because each channel is predicted independently. CD3 correlated with proteomics (ρ = 0.544, P = 0.024, n = 17); the five other markers were non-significant at 5–21% achieved power. CC showed the highest mature TLA proportion yet the lowest CCR (P = 0.041), whereas HGSC showed one of the highest CCR values. IIE did not differ across subtypes (P = 0.500). MC showed the lowest TLA density and LGSC the lowest mature TLA proportion. Conclusion Virtual mIF resolves histology-specific spatial TIME architectures in OC without tissue consumption, but its Ki-67 and CD3–CD8 co-expression readouts are unreliable; the findings are hypothesis-generating and require validation against matched tissue-based mIF.
Authors
- Uiree Jo (ORCID: https://orcid.org/0000-0001-6783-4016)
- Byoung‐Gie Kim (ORCID: https://orcid.org/0000-0002-0572-8450)
- Junhwan Kim (ORCID: https://orcid.org/0000-0001-9114-6555)
- Jae Hyun Cho
- Jeong Sig Kim
Institutions
- Soonchunhyang University (KR)
- Soonchunhyang University Hospital Seoul (KR)
- Asan Medical Center (KR)
- University of Ulsan (KR)
- Chung-Ang University Gwangmyeong Hospital (KR)
Publication Details
- Journal
- Frontiers in Immunology
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fimmu.2026.1900803
- Primary Topic
- Ovarian cancer diagnosis and treatment
- Type
- article
- Field-Weighted Citation Impact
- 0.00