Tumor architecture shapes polarized malignant states that predict outcome and treatment response in high-grade serous ovarian cancer
In brief
Across genetically diverse high-grade serous ovarian cancers (HGSC), malignant cells adopt recurring gene-expression states. These states are linked to a cell's location in the tumour, to patient survival and to response to bevacizumab, a therapy that blocks VEGF signalling.
One axis
Malignant secretory cells vary along the same axis across tumours, from SecA to SecB
We combined nearly 2 million single-cell transcriptomes from 13 published studies, spanning 362 samples from 158 patients before and after chemotherapy. Samples covered 12 anatomical sites, including primary tumours, metastases and ascites, the fluid that accumulates in the abdominal cavity.
Within the epithelial compartment, we identified two opposing gene-expression programs in secretory cancer cells. Secretory A (SecA) expresses developmental and progenitor-associated genes, including MECOM and SOX17. Secretory B (SecB) expresses keratins and genes associated with mucosal injury, including KRT7, KRT19 and LCN2.
SecA and SecB mark the ends of a continuum, with intermediate cells between them. We call a cell's position along this continuum its polarization: higher scores indicate a stronger SecB program relative to SecA.
Polarization was the leading pattern of malignant-cell gene expression in 54% of tumour samples and one of the three leading patterns in 82%. It was distinct from programs associated with the cell cycle, stress, interferon and epithelial–mesenchymal transition (EMT).
Polarization is the leading axis in most samples
Principal component of malignant-cell gene expression most aligned with polarization; one square per sample (n = 250)
Each point is a cell; nearby points have similar gene-expression profiles. A random 200,000 of 575,366 epithelial cells are shown. Scroll or pinch to zoom; hover to see a cell's state. Figure 1
The full cell atlas, shown as a UMAP of gene-expression similarity. A random 400,000 of 1,980,703 cells are shown.
The same states in every study
The atlas resolves 13 cell lineages and 51 cell types and states. Cells from different studies, sites and treatment histories mix within each lineage, so the epithelial states are not an artefact of one cohort.
Colour the atlas by study, site or treatment to see how cells from different samples overlap.
Both states coexist within clones
Secretory state composition of copy-number clones in three tumours
Clones were inferred from copy-number profiles using CopyKAT. SecA, intermediate and SecB cells coexist within individual clones, so these states do not map neatly onto genetic subclones. Figure 1l,m
Two programs
SecA cells proliferate; SecB cells are quiescent and stress-adapted
Signalling and the cell cycle
SecA cells show stronger growth-associated programs, including MYC, E2F and MAPK; 37.8% are in S or G2/M phase. SecB cells show stronger hypoxia and stress-response programs and a greater inferred reliance on glycolysis; 80.6% are classified as G1.
SecB dominates ascites
Before treatment, SecB made up 68% of epithelial cells in ascites, against 14% in primary tumours and 17% in solid metastases. In 14 patients sampled at several sites, the median was 60.7% in ascites and 10.2% in the primary tumour.
SecB accounted for 24% of epithelial cells in post-chemotherapy primary tumours, compared with 14% before treatment. In the separate analysis of 18 patients with matched pre- and post-treatment samples, its share increased in 12, although the paired difference was not statistically significant (p = 0.11).
Growth signalling in SecA, stress signalling in SecB
The three largest differences in each type of inferred activity (scaled), mean per state
PROGENy, Hallmark, DoRothEA and scFEA scores. Figure 2a
Most SecB cells are classified as G1
Cell-cycle phase, % of cells
Epithelial state by site
Treatment-naive samples, % of epithelium
Matched patients
SecB, % of epithelium (n = 14)
Architecture
SecB cells are concentrated in hypoxic regions farther from blood vessels
We mapped where these states occur within tissue using a 477-gene Xenium panel tailored to ovarian cancer. We profiled eight whole-tumour sections, 503 tumour tissue-microarray cores from 378 patients, and 15 healthy fallopian-tube tissue cores: about 3.4 million cells in all. Hypoxia and pathway activity are estimated from gene-expression signatures. Scroll to explore one tumour section.
One tumour section, 181,413 cells
Each mapped cell in this treatment-naive tumour section (WT01), coloured by cell type. Secretory cancer cells are split into SecA, intermediate and SecB by their polarization score.
SecA and SecB cells segregate
Colour shows each secretory cell's polarization. SecA and SecB cells segregate in all eight tumour sections and in 299 of 334 evaluable tissue-microarray patients. Of the SecB cells assigned to a neighbourhood, 89% lie in SecB-enriched neighbourhoods.
SecB cells lie far from vessels
Distance from each cell to the nearest endothelial cell or pericyte. SecA cells lie nearer the blood vessels at the centre of papillary structures; SecB cells lie toward the open space, or lumen.
…where hypoxia is highest
Mean hypoxia score of the cells within 50 µm (ENO1, HIF1A, VEGFA and nine other genes). Hypoxia rises with distance from vessels, and with it the share of SecB cells.
See polarization up close
The gradient across tumour sections
Across more than 1.2 million secretory cells in the eight tumour sections, higher SecB polarization was associated with greater distance from blood vessels, lower local cell density, and higher hypoxia and matrix-remodelling gene-expression scores.
The same holds in the tissue microarrays: across 361 patients, SecB cells were farther from the nearest vessel than SecA cells (p < 0.0001).
Local environment along the polarization axis
Secretory cells of 8 tumour sections; 50 µm neighbourhoods, z-scored · Figure 4f
Distance to nearest vessel
Median per tumour section, µm (n = 8)
Distance to vessel by patient
Median, µm (n = 361)
A shared niche
SecB cells and nearby macrophages share hypoxia-associated programs in regions with fewer B cells
Toward the SecB end of the continuum, secretory cells show lower expression of proliferation, growth and DNA-repair programs, alongside higher glycolysis and TNF/NF-κB signalling.
The cells around them differ too. Across the gradient from lower to higher local hypoxia, macrophages make up a larger share of non-epithelial cells (13.4% at the low end, 32.6% at the high end) and B cells a smaller one (3.8% and 0.7%), in all eight tumour sections. T cells and plasma cells trend lower.
Secretory programs along the axis
Cell-intrinsic UCell scores (z-scored), 8 tumour sections · Figure 5a
Who shares the neighbourhood as hypoxia rises
% of non-epithelial cells within 50 µm, against local epithelial hypoxia (z). Thick line: all sections; thin lines: each tumour section
Macrophages show a similar program
Macrophages in SecB regions show higher expression of glycolysis and NF-κB genes and lower expression of tissue-resident markers. Visium HD profiling of WT01, the tumour shown above, also identified a hypoxia-associated SPP1+ MMP9+ macrophage state beside SecB regions.
Potential signals within the niche
Gene expression points to possible interactions within this niche. Macrophages express VEGFA, while nearby SecB cells show higher expression of the NF-κB target BIRC3. Other genes suggest potential roles in immune protection, lactate export, adhesion and matrix remodelling. Together, these patterns support a model in which SecB cells and macrophages form cohesive groups that could detach into ascites.
Macrophage programs along the axis
Macrophage scores against the polarization of the epithelium around them (z-scored), 8 tumour sections
Signals in the SecB niche
Genes with higher expression in SecB cells and nearby macrophages, grouped by role. Hover a gene for the evidence across platforms.
Missing in models
Patient-derived organoids favour SecA, and none of the conditions we tested fully restored SecB
We profiled eight patient-derived organoid lines by single-nucleus RNA-seq. Fewer than 1% of organoid cells reached the SecB threshold of the atlas. Lines grown from ascites kept higher SecB scores than lines from primary tumours (p = 0.029), and in one line SecB scores were lower after 6–12 days in culture than at day 2.
Of four microenvironmental signals tested, only IFN-γ increased SecB gene expression, and none of the culture media restored the full state. Published organoid and spheroid datasets captured only parts of SecB: hypoxic cultures reproduced the HIF response but retained high proliferation. The KPCA.B mouse model also captured some features, including a Krt7+ luminal population, without a gene-expression program matching human SecB.
SecB score in eight organoid lines
SecB UCell score per cell; dashed lines: atlas medians for intermediate and SecB cells
Outcome
Greater SecB polarization is associated with poorer survival and greater bevacizumab benefit
Among 104 treatment-naive patients represented by two tumour cores, those in the most SecB-enriched third had worse 5-year overall survival than those in the most SecA-dominant third (log-rank p = 0.011). Polarization remained prognostic after adjusting for stage, residual disease and age (HR 1.51 per standard deviation).
The survival association also held in 314 TCGA-OV tumours scored by bulk RNA-seq. Higher polarization was associated with worse overall survival and progression-free survival after adjustment for epithelial content, stage and age (adjusted HRs 1.84 and 1.65, respectively).
Lowest versus highest third of tumour polarization scores; the middle third is not shown. Figure 7a,b,g,h
Bevacizumab in ICON7
Bevacizumab benefit was greater in SecB-polarized tumours
SecB niches are distant from blood vessels and contain macrophages expressing VEGFA. We therefore asked whether polarization was associated with benefit from bevacizumab. We scored pretreatment tumours from 191 patients with stage III/IV disease in ICON7 and divided them at the median polarization score. Within each group, we compared carboplatin–paclitaxel alone with carboplatin–paclitaxel plus bevacizumab.
In SecB-polarized tumours, median progression-free survival was 18.7 months with bevacizumab versus 11.1 months with chemotherapy alone (HR 0.36). The treatment effect differed between the polarization groups (interaction p = 0.025); no statistically significant benefit was detected in the SecA group. In the analysis split at 15 months, the estimated benefit in SecB tumours was strongest during the first 15 months (HR 0.23), with no clear benefit detected afterwards (HR 0.87).
Treatment effect before and after 15 months
Hazard ratio of bevacizumab for progression-free survival, with the follow-up split at 15 months
Implications
A recurring cell state offers a way to investigate relapse, test treatment selection and develop practical biomarkers
SecB cells show lower proliferation and occupy regions distant from blood vessels. Their share increased after chemotherapy in 12 of 18 patients with matched samples, although the paired difference was not statistically significant. These observations raise the possibility that SecB cells persist through treatment and contribute to recurrence, a hypothesis for longitudinal studies.
Our retrospective ICON7 analysis links SecB polarization to greater progression-free survival benefit from bevacizumab. Prospective validation is needed to establish whether it can guide treatment selection. Whether polarization can also inform treatment duration remains a separate question.
KRT7 and KRT19 protein levels correlated with the polarization score measured by Xenium in the same tissue cores; E-cadherin, KRT18 and vimentin did not. These markers provide a starting point for a practical tissue-based test. Serum CYFRA 21-1, a KRT19 fragment, is another candidate to investigate as a blood-based readout.