henhaobio
Sign in

Cancer DiscoveryCancer BiologyPeer-reviewed journal article

Single-cell spatial atlas of high-grade serous ovarian cancer unveils MHC class II as a key driver of spatial tumor ecosystems and clinical outcomes

Fernando Pérez‐VillatoroiD, Aleksandra ShabanovaiD, Lilian van WagensveldiD, Ada JunqueraiD, Iga NiemieciD, María Mercedes Hincapié-OteroiD, Ziqi KangiD, Matías Marín FalcoiD, Kürşat BirginiD, Sarah WolfiD, Ella AnttilaiD, Gayani AnandagodaiD, Julia CasadoiD, Eric MarcusiD, Duco GaillardiD, Essi KaheliniD, Foteini ChamchougiaiD, Matilda SalkoiD, Saundarya ShahiD, Salvatore RussoiD, Jacopo ChiaroiD, Mikaela GrönholmiD, Joseph NdikaiD, Otto K. KariiD, iCAN, , Anniina FärkkiläiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

The tumor microenvironment (TME) is a complex network of interactions between malignant and host cells, yet its orchestration in advanced high-grade serous ovarian carcinoma (HGSC) remains poorly understood. We present a comprehensive single-cell spatial atlas of 280 metastatic HGSCs, integrating high-dimensional imaging, genomics, and transcriptomics. Using 929 single-cell maps, we identify distinct spatial domains associated with phenotypically heterogeneous cellular compositions, and demonstrate that immune cell co-infiltration at the tumor-stroma interface significantly influences clinical outcomes. To uncover the key drivers of the tumor ecosystem, we developed CEFIIRA (Cell Feature Importance Identification by RAndom forest), which identified tumor cell-intrinsic MHC-II expression as a critical predictor of prolonged survival, independent of clinicomolecular profiles. Validation with external datasets confirmed that MHC-II-expressing cancer cells drive immune infiltration and orchestrate spatial tumor-immune interactions. Our atlas offers novel insights into immune surveillance mechanisms across HGSC clinicomolecular groups, paving the way for improved therapeutic strategies and patient stratification.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (3 Dec 2024) · CC BY 4.0

Cyclic immunofluorescence of 265 high-grade serous ovarian carcinomas and random-forest feature ranking identify cancer-cell MHC-II as an adjusted predictor of longer survival marking immune-activated tumor-stroma hotspots.

In advanced high-grade serous ovarian carcinoma (HGSC), how malignant and host cells organize spatially, and how that organization relates to homologous recombination deficiency (HRD), BRCA status, neoadjuvant chemotherapy and survival, remains poorly resolved. This atlas ties clinicomolecular groups to single-cell spatial domains and uses an interpretable random-forest screen to nominate cancer-cell MHC-II as a prognostic organizer of tumor-immune interactions.

Key results

  1. ShownCancer cell-intrinsic MHC-II ranked among top CEFIIRA features separating long- from short-overall-survival (OS) patients with HR-proficient (HRP) tumors in primary (PDS) and interval (IDS) debulking samples; a higher MHCII+ cancer cell fraction predicted longer OS cohort-wide after adjustment.

    CEFIIRA random forests with 43 features per cell, fit separately by compartment, HRD-like or HRP-like profile and PDS or IDS over 30 nested cross-validation iterations; Cox model adjusted for age, FIGO stage and molecular profile in survival-eligible tumors (n=234); replicated in high-stage TCGA HGSC using PRISM-deconvolved cancer-cell HLA-DPB1. No numeric hazard ratios are given in the text. · CEFIIRA enables the discovery of clinically…; Cancer cell intrinsic MHCII expression is a key…; Machine learning supervised classification; Survival analysis

  2. ShownNeighborhood clustering (30µm radius) defined 14 recurrent cellular neighborhoods; an immune-infiltrated tumor-stroma interface (TSI, RCN10) associated with better OS, whereas a fibroblast-rich immune-cold TSI (RCN09), more common after neoadjuvant chemotherapy, tracked with worse outcome.

    RCN10 had the highest Shannon diversity (2.3) and more T cells, myeloid cells, macrophages and B cells than RCN09; three RCNs associated with both progression-free survival (PFS) and OS in therapy-stratified adjusted Cox models; RCN09 more abundant in IDS than PDS; similar clusters at 60µm and 90µm radii. · Spatial TME is shaped by molecular profiles and…; Spatial tumor-stroma interface shapes clinical…; Recurrent cellular neighborhoods and spatial…

  3. SuggestedMHCII+ cancer cells concentrated in an immunogenic cancer neighborhood (RCN05) and the immune-infiltrated TSI and sat closer to immune cells, whereas MHCII- cancer cells mostly bordered stroma; an independent whole-slide cohort reproduced these immune hotspots and a PFS link.

    Whole-slide 14-plex tissue cyclic immunofluorescence (tCycIF) of 15 HRP tumors after neoadjuvant chemotherapy (ONCOSYS-OVA, NCT06117384; 10.5 million cells): MHCII+ cancer cells had more CD4+ T, CD8+ T and CD11c+ myeloid neighbors (unadjusted P < 0.05); MHCII+ fraction was higher in long-PFS (>=12 months) patients, unlike CD8+ T-cell abundance. · Cancer cell MHCII expression is spatially…; Whole slide cohort description; Highly multiplexed immunofluorescence and image…

  4. ShownTumors rich in MHCII+ cancer cells expressed immune-activation and interferon-gamma (IFN-γ) programs, and T cells and CD11c+ myeloid cells next to MHCII+ cancer cells at the TSI showed higher phospho-STAT1, consistent with local IFN signaling.

    TCGA DESeq2 (top vs bottom cancer-cell HLA-DPB1 tertile): 14 enriched immune pathways including PD-1, chemokine and IFN-γ signaling, none among downregulated genes; TIGIT, CTLA4, TMIGD2 up; CXCL9, CXCL10, CXCL11 and CXCR3 top hits in TCGA and the 335-gene Nanostring panel; Visium (8 IDS tumors): MHCII and IFN-γ signatures correlated. · Cancer-intrinsic MHCII governs single-cell and…; Spatial transcriptomics analysis; RNA differential expression analysis and pathway…

  5. SuggestedCancer-cell MHCII was enriched in HRD-like tumors, most clearly BRCA-loss, and in single-cell RNA-seq, tumors with many MHCII+ cancer cells carried more antigen-experienced and precursor-exhausted T cells plus IFNG-correlated memory T cells.

    CEFIIRA BRCAloss-vs-HRP cell classifier (PDS samples): median F1=0.77, MHCII most differential feature; scRNA-seq of 51 samples from 29 patients (136,452 cells): 12% of cancer cells MHCII-high; MHCII-high tumors had more type 17 helper CD4+, central memory CD8+ and precursor exhausted CD8+ T cells. · Cancer cell intrinsic MHCII expression is a key…; Cancer MHCII expression shapes immune…; Cancer-intrinsic MHCII governs single-cell and…; sc-RNAseq analysis

Methods, data and reproducibility

Tissue cyclic immunofluorescence (tCycIF, 30 markers) on tissue microarrays was segmented, phenotyped and clustered into 30µm neighborhoods, then related to BRCA/HRD/CCNE1 profiles and survival via adjusted Cox models and CEFIIRA, a nested cross-validated random-forest feature-ranking pipeline, with checks in whole slides, TCGA, scRNA-seq and Visium.

  • Retrospective HGSC cohort (FIGO IIb-IV; NKI-AVL, MUMC, AUMC; 2008-2015): 30-plex tCycIF tissue microarrays, IHC, targeted and shallow WGS, BRCA1 methylation, Nanostring · human · n = 265 patients (PDS n=100, IDS n=158); 4,644,485 cells
  • ONCOSYS-OVA trial whole-slide 14-plex tCycIF validation set: HRP tumors from interval debulking surgery, Helsinki University Hospital · human · n = 15 patients; 10.5 million cells · NCT06117384
  • TCGA high-stage ovarian HGSC bulk RNA-seq with PRISM-deconvolved cancer-cell expression and CIBERSORT immune fractions (GDC-PanCanAtlas) · human
  • Public pre-processed, annotated HGSC scRNA-seq from PDS and IDS samples · human · n = 51 samples from 29 patients; 136,452 cells · GSE266577
  • Public Visium spatial transcriptomics of IDS HGSC with varying chemotherapy response, spots annotated by a pathologist · human · n = 8 samples · GSE211956
  • HGSC antigen genes from prior immunopeptidomics (Schuster et al 2017) · human
Methods & tools
tCycIF (tissue cyclic immunofluorescence) on RareCyte CyteFinder, CEFIIRA (scikit-learn random forest, Bayesian tuning, RFECV, permutation importance), PyCaret, ASHLAR; BaSiC; CyLinter, Ilastik v1.3.2; CellProfiler v3.1.8; StarDist, Mclust; FlowSOM; ConsensusClusterPlus, Tribus, Scimap; YellowBrick, shallowHRD; ovaHRDscar, PRISM; CIBERSORT, DESeq2; PantherDB v19.0 (Reactome), Seurat v4.3.0; UCell v1.3.1; survival/survminer
Software & environment
R 4.2.3; MATLAB 2018b; Ilastik v1.3.2; CellProfiler v3.1.8; Seurat v4.3.0; UCell v1.3.1; PantherDB v19.0; StarDist '2D_versatile_fluo' model; Tribus run with sigma=1, learning_rate=1, clustering_threshold=100.
Reproducibility gaps
tCycIF images and cell feature tables only on reasonable request for non-commercial academic use; sequence data withheld for patient privacy but available on reasonable request; NCR clinical data shared case-by-case after NCR approval; no public deposit of atlas data.
Limitations (6)
  • Retrospective use of archival FFPE tissue may introduce preservation and heterogeneity biases; quality control removed part of the material (e.g., 64 tissue cores with under 1000 good cells, 10 samples with poor cancer-cell clustering). · stated by the authors
  • The 30-protein tCycIF panel cannot represent every relevant cell type in the tumor microenvironment. · stated by the authors
  • Highly intercorrelated immune infiltration limits interpretation of individual hazard ratios, and the random-forest models do not explicitly handle correlated features, which could affect predictive performance. · stated by the authors
  • The proposed mechanism (MHCII antigen presentation triggering IFN-γ-producing T cells) remains a hypothesis untested in functional models, and transcriptomic deconvolution accuracy could be improved. · stated by the authors
  • Support for MHCII driving immune infiltration is correlational (tertile splits, deconvolution, co-localization) with no perturbation experiment; several spatial and feature comparisons rely on unadjusted P < 0.05. · noted in this summary
  • Cohort size is stated inconsistently (280 tumors in the abstract, 265 in the main text), and the whole-slide PFS validation uses 15 preselected HRP patients split at 12 months without covariate adjustment. · noted in this summary

Read this if you work on the ovarian cancer immune microenvironment, multiplexed tissue imaging, or interpretable machine learning that links single-cell spatial features to HRD status and survival.

Citations
10
FWCI
35.12
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLi '25Khan '26Zhang '23Sarkkinen '23Ferri-Borgo '22Kleunen '24
74% semantic matchGenomics
Weishan Li, Laurie Grieshober, Jason Gertz+4 more·bioRxiv·Sep 15, 2025· 1 citations