Article

Single-cell dissection of cellular components and

interactions shaping the tumor immune phenotypes
in ovarian cancer
Graphical abstract Authors
Milena Hornburg, Mélanie Desbois,
Shan Lu, ..., Richard Bourgon,
Anneleen Daemen, Yulei Wang

Correspondence
[email protected]

In brief
Hornburg et al. dissected the
composition of the ovarian tumor
microenvironment by scRNA-seq to
define the underlying biology of the T cell
infiltration pattern, namely the tumor
immune phenotypes. The identified
features, including the enrichment of
immature myeloid cells in desert tumors,
may help inform immunotherapeutic
strategies in ovarian cancer.

Highlights
d OXPHOS and IFN gene sets are enriched in infiltrated and
excluded tumor cells.

d Pre-dysfunctional CD8+ GZMK T cells are enriched in

excluded tumors.

d FCN1 monocytes and immature MARCO macrophages are

enriched in desert tumors.

d CXCL16 is expressed primarily by infiltrated tumor cells and

CXCR6 by T cells.

Hornburg et al., 2021, Cancer Cell 39, 928–944

July 12, 2021 ª 2021 Elsevier Inc.
https://doi.org/10.1016/j.ccell.2021.04.004 ll
 ll

Article
Single-cell dissection of cellular components
and interactions shaping the tumor
immune phenotypes in ovarian cancer
Milena Hornburg,1,9 Mélanie Desbois,2,9 Shan Lu,2 Yinghui Guan,2 Amy A. Lo,3 Susan Kaufman,4 Ashley Elrod,5
Alina Lotstein,5 Teresa M. DesRochers,5 Jose L. Munoz-Rodriguez,6 Xingwei Wang,7 Jennifer Giltnane,3 Oleg Mayba,1
Shannon J. Turley,8 Richard Bourgon,1 Anneleen Daemen,1,10 and Yulei Wang2,10,11,*
1Department of Bioinformatics & Computational Biology, Genentech, Inc., South San Francisco, CA 94080, USA
2Department of Oncology Biomarker Development, Genentech, Inc., South San Francisco, CA 94080, USA
3Department of Research Pathology, Genentech, Inc., South San Francisco, CA 94080, USA
4Department of Biochemical Cellular Pharmacology, Genentech, Inc., South San Francisco, CA 94080, USA
5KIYATEC, Inc., Greenville, SC 29605, USA
6Companion Diagnostics (CD), Roche Tissue Diagnostics, Tucson, AZ 85755, USA
7Department of Digital Pathology, Roche Tissue Diagnostics, Santa Clara, CA 95050, USA
8Department of Cancer Immunology, Genentech, Inc., South San Francisco, CA 94080, USA
9These authors contributed equally
10Senior author
11Lead contact

*Correspondence: [email protected]
https://doi.org/10.1016/j.ccell.2021.04.004

SUMMARY

Distinct T cell infiltration patterns, i.e., immune infiltrated, excluded, and desert, result in different responses
to cancer immunotherapies. However, the key determinants and biology underpinning these tumor immune
phenotypes remain elusive. Here, we provide a high-resolution dissection of the entire tumor ecosystem
through single-cell RNA-sequencing analysis of 15 ovarian tumors. Immune-desert tumors are characterized
by unique tumor cell-intrinsic features, including metabolic pathways and low antigen presentation, and an
enrichment of monocytes and immature macrophages. Immune-infiltrated and -excluded tumors differ mark-
edly in their T cell composition and fibroblast subsets. Furthermore, our study reveals chemokine receptor-
ligand interactions within and across compartments as potential mechanisms mediating immune cell infiltra-
tion, exemplified by the tumor cell-T cell cross talk via CXCL16-CXCR6 and stromal-immune cell cross talk via
CXCL12/14-CXCR4. Our data highlight potential molecular mechanisms that shape the tumor immune phe-
notypes and may inform therapeutic strategies to improve clinical benefit from cancer immunotherapies.

INTRODUCTION studies lack higher-resolution information on cellular heteroge-

neity and spatial distribution, and a better understanding of the
The tumor microenvironment (TME) is a complex ecosystem distribution of T cells in the TME is critical for the success of
comprising tumor cells, infiltrating immune cells, and stromal immunotherapies.
cells intertwined with non-cellular components. The diverse To better capture the spatial distribution of the immune infil-
cellular and functional phenotypes, as well as the dynamic inter- trates in addition to their overall quantity, the concept of the tu-
play within and between these components, shape tumor mor immunity continuum has been introduced (Hegde and
biology and may contribute to different responses to immuno- Chen, 2020; Hegde et al., 2016). This concept proposes three tu-
therapies. However, a high-resolution characterization of these mor immune phenotypes (TIPs) based on the spatial distribution
important cellular heterogeneities and interactions is lacking. of T cells in the TME: (1) the inflamed/infiltrated phenotype in
Most previous studies relied on relatively low-resolution tech- which the T cells infiltrate the tumor epithelium, (2) the im-
niques such as immunohistochemistry (IHC) (Zhang et al., mune-excluded phenotype in which infiltrating T cells accumu-
2003) or bulk RNA sequencing (RNA-seq) deconvolution algo- late in the tumor stroma rather than the tumor epithelium, and
rithms (e.g., CIBERSORT, Newman et al., 2015; xCell, Aran (3) the immune-desert phenotype in which T cells are either pre-
et al., 2017). More recent work integrated multi-omics platforms sent in very low numbers or completely absent. Building upon
and in situ lymphocyte quantifications identifying distinct im- this model, our group has recently developed a machine-
mune phenotypes (Thorsson et al., 2018). However, these learning approach that integrates digital pathology CD8 IHC

928 Cancer Cell 39, 928–944, July 12, 2021 ª 2021 Elsevier Inc.
 ll
Article

with bulk transcriptome analysis to classify ovarian tumors ac- studies (Jerby-Arnon et al., 2018; Puram et al., 2017; Tirosh
cording to their TIP (Desbois et al., 2020). This classification al- et al., 2016) (Figure 1B). In contrast, cells from both the stro-
lowed us to characterize features associated with the different mal and the immune compartments clustered by cell type,
TIPs based on bulk RNA-seq data. However, bulk RNA-seq with secondary clustering by patient to varying extents within
lacks the resolution to interrogate the heterogeneity of the com- the cell-type clusters (Figures 1C, 1D, S2F, and S2G). We
plex TME at the cellular level as well as the sensitivity to capture defined the cell types based on the expression of known
changes in underrepresented cell populations. To this end, sin- cell-type markers: fibroblasts (COL1A1, PDGFRA), endothelial
gle-cell RNA-seq (scRNA-seq) has been widely used over the cells (PECAM1), and pericytes (RGS5) in the stromal compart-
last decade to dissect the composition of the TME in various ment and myeloid cells (CD14), T cells (CD3E), and plasma
cancer indications (Guo et al., 2018; Lambrechts et al., 2018; cells (CD79A, SDC1), as well as tumor-infiltrating B lympho-
Li et al., 2019; Puram et al., 2017; Savas et al., 2018; Tirosh cytes (B-TILs) (CD79A, MS4A1), in the immune compartment
et al., 2016; Wu et al., 2020; Yost et al., 2019; Zhang et al., (Figures 1C and 1D). The relative prevalence of cell types
2018, 2019). However, most scRNA-seq studies focused on within the stromal and immune compartments was highly var-
the characterization of tumor-infiltrating T cells, and a systematic iable between patients and did not show a clear association
single-cell characterization of how other cell types in the TME with the TIP (Figures 1E and 1F, Table S3).
shape the TIP has not been reported to date.
In this study, we dissected the tumor immunity continuum in Tumor-intrinsic features associated with different
human ovarian cancer, characterizing the composition and tumor immune phenotypes
cell-cell interactions in this complex ecosystem with scRNA- We first investigated whether tumor-intrinsic features
seq profiling, to provide additional insights into the biology that contribute to patterns of immune infiltration. Using the
shapes the distinct TIP and informs new therapeutic strategies scRNA-seq data from the tumor cell compartment, we per-
altering the immunity continuum and thereby optimizing the clin- formed a pseudo-bulk differential expression analysis be-
ical benefits of cancer immunotherapies. tween the TIP to identify differential expression patterns.
Notably, there were no significant tumor-cell transcriptional
RESULTS differences between excluded and infiltrated tumors (all
adjusted p > 0.05). This may be due to intertumor heterogene-
Single-cell transcriptional profiling defines the cellular ity, even within a TIP (Figure 1B), which limits statistical po-
ecosystem of patient-derived primary ovarian tumors wer; alternatively, it may suggest that the major drivers of
To dissect the landscape of the tumor immunity continuum in CD8+ T cell exclusion versus infiltration in the tumor epithe-
ovarian cancer, we performed RNA-seq and CD8 IHC analysis lium lie within other cells in the TME, not the tumor cells them-
on tumor samples collected from 42 ovarian cancer patients selves. In contrast, 29 genes were significantly differentially
(Figure 1A). The RNA-seq data were utilized to predict the TIP expressed between desert tumors and the combined set of in-
of each tumor (i.e., infiltrated, excluded, or desert) with a previ- filtrated and excluded tumors. Gene set enrichment analysis
ously developed, ovarian cancer-specific, transcriptional classi- revealed that these genes were mainly associated with prolif-
fier (Desbois et al., 2020) (Figure S1A). In parallel, CD8+ T cell erative pathways, and that their identification was driven by
infiltration into the stroma and tumor epithelium was evaluated three desert tumors with a proliferative molecular subtype.
based on CD8 IHC staining (Figure S1B). We employed both Hence, to identify other biological processes, we accounted
the classifier-based predictions and the CD8 IHC categorization for the G2/M state in the differential expression analysis. Sub-
to assign a TIP to each tumor sample (Table S1). Next, we sequent gene set enrichment analysis revealed that genes
selected 15 ovarian tumors, with 5 tumors representative of involved in the epithelial-mesenchymal transition and angio-
each of the three TIPs, and performed scRNA-seq to further genesis were significantly enriched in desert tumors (adjusted
characterize the cellular composition associated with these phe- p = 0.01 and 0.03, respectively, Figures 2A, S2H, and S2I and
notypes. We used flow cytometry (fluorescence-activated cell Table 1). On the other hand, tumor cells of excluded and infil-
sorting) to sort the tumor cells (EpCAM+ CD45
), the stromal trated tumors were significantly enriched in interferon
cells (EpCAM
CD45
), and the immune cells (EpCAM
response pathways (adjusted p = 0.004), mainly driven by
CD45+) of each tumor sample (Figure S1C). Each of the three genes encoding for major histocompatibility complex (MHC)
compartments (tumor, immune, stromal) was subjected to class I and II processing and presentation (e.g., HLA-A, B,
scRNA-seq as a separate library, with the exception of the desert and C; HLA-DQA1; TAP1; PSMB9) (Figures 2A, 2B, and S2J
stromal and immune cells, which were pooled by patient due to and Table 1). Interestingly, we also observed a significant
the low fraction of CD45+ immune cells in desert tumors (Fig- enrichment of the oxidative phosphorylation pathway in the
ure S1D) and were separated computationally (STAR Methods, excluded and infiltrated tumors (adjusted p = 0.04, Figures
Figures S2A–S2E). In total, we sequenced 44 libraries resulting 2A, 2C, and S2J). Components of this pathway, including sub-
in 40,539 tumor cells, 35,296 stromal cells, 15,049 immune cells, units of NADH dehydrogenase (NDUF variant genes), succi-
and 2,334 stromal/immune cells from the pooled libraries (Ta- nate dehydrogenase (SDHC, SDHA, and SDHB), cytochrome
ble S2). c oxidase (COX variant genes), and V-type ATPase, showed
To better characterize the heterogeneity and cellular a continuum of expression levels, from low (on average) in tu-
composition, we analyzed each compartment separately. mor cells of desert tumors, to intermediate in excluded tu-
The tumor cells clustered primarily by patient, which likely re- mors, to high expression in infiltrated tumors (Figures 2C
flects the strong interpatient heterogeneity shown in previous and S2J).

Cancer Cell 39, 928–944, July 12, 2021 929

 ll
Article

B C

E F

Figure 1. scRNA-seq defines the cellular ecosystem of patient-derived primary ovarian tumors
(A) Overview of the study design and workflow.
(B) Uniform manifold approximation and projection (UMAP) of tumor cells of all 15 patients colored by patient. des, desert; exc, excluded; inf, infiltrated.
(C and D) (C) UMAPs of stromal cells and (D) immune cells of all patients, colored by the identified cell type. Identical UMAPs show the cell-type marker gene
expression levels.
(E) Stromal-cell-type fractions relative to the total stromal cell count per patient. Each stacked bar represents a patient for which the total stromal cell count was
scaled to 1.
(F) Immune-cell-type fractions relative to the total immune cell count per patient.
Norm. expr., normalized expression.
See also Figures S1 and S2 and Tables S1–S3.

930 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

A C

Figure 2. Tumor-intrinsic features associated with different tumor immune phenotypes

(A) Significantly enriched Hallmark gene sets in the tumor cells of desert versus the combined set of excluded/infiltrated tumors (adjusted p < 0.05). Fold change
expression values and adjusted p values were combined to rank the genes as input for the gene set enrichment analysis.
(B and C) Heatmaps of the (B) antigen presentation or (C) oxidative phosphorylation gene set expression by patient. Z-scored gene expression was calculated
based on scran-normalized expression values. Violin plots show the distribution of the per cell signature scores by TIP and report the 25% (lower hinge), 50%, and
75% quantiles (upper hinge) as well as the kernel density estimates as the width. The signature scores were calculated based on the genes shown in the
heatmaps.
See also Figure S2.

Distinct states of CD8+ T cells characterize immune- effector CD8+ T cells (Li et al., 2019). Interestingly, a distinct clus-
infiltrated and -excluded tumor immune phenotypes ter of the Tgd/NK cells shared markers with the CD8+ FGFBP2
To dissect the composition of the TME for each TIP, we investi- subpopulation: FGFBP2, PRF1, GZMB, KLRG1, and GLNY (Fig-
gated each cell type separately (i.e., T cells and myeloid cells) ures 3A and 3C, Table S4). Both CD8+ FGFBP2 and Tgd/NK
and defined cell subpopulations and cellular functions. Cluster FGFBP2 subpopulations were significantly enriched in the desert
analysis on the batch-corrected uniform manifold approximation tumors compared with excluded and infiltrated tumors (p =
and projection (UMAP) representation of all T cells (STAR 0.0003 and 0.07 in excluded, p = 1.2 3 10
9 and 3.5 3 10
5 in
Methods) revealed a separation of CD8+ from CD4+ T cells (Fig- infiltrated, Figures S3B and S3D); however, the absolute cell
ures 3A, 3B, and S3A). In addition, we identified two clusters ex- numbers were relatively small (Table S5). The CD8+ IL7R cells
pressing the Tgd receptor gene, TRDC, and the natural killer (NK) express IL7R and GPR183 (Figure 3C), with a transcriptional pro-
cell marker, NCAM1. Because NK cells and Tgd cells are tran- file resembling central memory CD8+ T cells previously observed
scriptionally similar, we were not able to differentiate these two in non-small cell lung cancer and colorectal cancer (Guo et al.,
cell subpopulations. Within the CD4+ T cells, we distinguished 2018; Zhang et al., 2018) and regrouped under naive-like cells
three distinct functional and phenotypic states according to their based on the nomenclature proposed by Van der Leun et al.
gene expression markers, CD4+ FOXP3 for regulatory CD4+ (van der Leun et al., 2020).
T cells (FOXP3, CTLA4, IL2RA), CD4+ CXCL13 for activated We next focused on the two largest CD8+ T cell clusters, CD8+
CD4+ T cells (CXCL13, CD200, ICOS), and CD4+ IL7R for resting GZMB and CD8+ GZMK T cells (Figures 3C and 3D), and further
CD4+ T cells (IL7R, GPR183, LMNA, ANXA1) (Figures 3A and 3C, characterized their potential functional states and spatial distribu-
Table S4). The resting CD4+ IL7R T cells were significantly en- tion in detail. Although both GZMB and GZMK T cell populations
riched in excluded tumors compared with infiltrated tumors have been previously reported in scRNA-seq studies (Guo et al.,
(p = 0.01), while infiltrated tumors tended to have more activated 2018; Wu et al., 2020; Yost et al., 2019; Zhang et al., 2018,
and regulatory CD4+ T cells (p = 0.1 and 0.3, Figures S3B and 2019), their functions are still poorly understood. Our data revealed
S3C). We identified four CD8+ T cell states and annotated that both subpopulations expressed markers that are suggestive
them based on their featured marker and using the CD8+ T cell of an activated phenotype (ICOS, CD69), and both demonstrated
nomenclature reported by van der Leun and colleagues (van effector functions by expression of other granzymes (GZMA,
der Leun et al., 2020): (1) CD8+ FGFBP2, (2) CD8+ IL7R, (3) GMZH), interferon-g (IFNG), perforin (PRF1), granulysin (GNLY),
CD8+ GZMB, and (4) CD8+ GZMK (Figures 3A and 3C). CD8+ and CCL4 (Figure S3E, Table S4). Yet, they differed markedly in
FGFBP2 cells have been previously reported as cytotoxic their activation and exhaustion-like status. The CD8+ GZMB

Cancer Cell 39, 928–944, July 12, 2021 931

 ll
Article

Table 1. Hallmark gene set enrichment results based on the differentially expressed genes in desert versus infiltrated/excluded
tumor cells
Adjusted
Pathway p NES Gene set leading-edge genes
HALLMARK_ 0.0129 1.611 CDKN2A, CCL4, B2M, TAP1, HLA-DQA1, CTSS, EREG, IL2RG,
ALLOGRAFT_ STAB1, CD247, HLA-DOB, MMP9, TAP2, LTB, GBP2, ICAM1,
REJECTION CCR1, HLA-A, STAT1, PRKCG, ELF4, HLA-G, CSF1, CD47, RPL3L,
NCK1, IL27RA, IL15, CD40, TPD52, TAPBP, JAK2, C2, CSK, CAPG,
LYN, TRAF2
HALLMARK_ 0.0337
1.685 SPP1, SERPINA5, STC1, VCAN, FGFR1, NRP1, FSTL1
ANGIOGENESIS
HALLMARK_E2F_ 0.0129 1.512 CDKN2A, CDKN2C, CDC20, SUV39H1, JPT1, ASF1B, CDCA8,
TARGETS DDX39A, TACC3, KIF2C, TK1, CTPS1, TRIP13, CCNE1, BIRC5,
SPC24, MYC, CCNB2, CKS1B, POLD1, KPNA2, HUS1, PLK1,
TCF19, AURKA, MELK, ORC6, UBE2S, RAD1, DEPDC1, CHEK2,
AK2, CDKN3
HALLMARK_EPITHELIAL_ 0.0129
1.749 COL1A2, SPP1, DAB2, IGFBP4, WNT5A, MMP2, ACTA2, BGN,
MESENCHYMAL_TRANSITION ADAM12, VIM, VCAN, BASP1, COL6A2, BMP1, FBN2, PMP22,
SPARC, SERPINE2, FBLN1, FBN1, LAMA1, CRLF1, LOXL2,
COL4A1, SLIT2, FSTL1, SFRP1, FGF2, FBLN2, RGS4, CDH2,
TGFB1, HTRA1, FOXC2, FMOD, GJA1, PRRX1, COL4A2, VEGFC,
CDH6, SGCD, FN1, COL5A2, PDGFRB, SLIT3, EMP3, TIMP3,
IGFBP2, COL6A3, GAS1, EFEMP2, MATN3, PTX3, DST, SERPINE1,
PDLIM4, CAP2, SFRP4
HALLMARK_INTERFERON_ 0.0035 2.056 EPSTI1, ISG15, CMPK2, CXCL10, IFI44, CXCL11, BATF2, SAMD9,
ALPHA_RESPONSE RTP4, PARP9, MX1, IFIH1, OASL, SAMD9L, B2M, IFIT3, TAP1,
ADAR, PSMB9, PLSCR1, C1S, RSAD2, HLA-C, TDRD7, USP18,
GBP2, IFIT2, UBE2L6, TMEM140, PARP12, HELZ2, CASP1, SP110,
GBP4, PARP14, LAP3, IFI44L, CSF1, PROCR, NMI, IFI27, CD47,
LGALS3BP, PSMB8, PSME2
HALLMARK_INTERFERON_ 0.0035 2.130 ID O 1, EPSTI1, ISG15, CMPK2, IFIT1, NLRC5, CXCL10, ZBP1, IFI44,
GAMMA_RESPONSE CXCL11, BATF2, VCAM1, RTP4, MX1, IFIH1, SAMHD1, CD274,
OASL, IL10RA, OAS3, SAMD9L, AC124319.1, MX2, B2M, IFIT3,
TAP1, SECTM1, ADAR, HLA-DQA1, PDE4B, PSMB9, HLA-B, OAS2,
PLSCR1, C1S, RSAD2, TDRD7, MYD88, TNFSF10, CIITA, USP18,
DDX58, CFB, ICAM1, IFIT2, UBE2L6, PARP12, C1R, HELZ2, CASP1,
SP110, HLA-A, GBP4, PARP14, STAT1, IL18BP, CD38, PSMB2,
LAP3, IFI44L, HLA-G, NMI, IFI27, TNFAIP6, MT2A, BPGM,
LGALS3BP, SOD2, PSMB8, PSME2, XAF1
HALLMARK_OXIDATIVE_ 0.0361 1.446 ATP6V0B, PDP1, COX7B, PDHA1, UQCRFS1, OPA1, MRPS12,
PHOSPHORYLATION MRPS15, NDUFS6, CYC1, SDHB, NDUFA1, NDUFB7, NDUFS2,
SDHA, TIMM17A, MFN2, MDH2, HSD17B10, MRPS11, HCCS,
MRPL35, SLC25A20, PHYH, NQ O 2, NDUFA3, ATP6AP1, BCKDHA,
OGDH, IDH2, NNT, RETSAT, COX6C, FDX1, COX5A, ATP6V1C1,
IDH1, NDUFB2, ETFA, TCIRG1, POR, NDUFA8, IDH3A, ECH1,
AIFM1, ISCA1, COX8A, NDUFC2, NDUFS8, UQCR11, MRPL15,
ATP1B1, COX5B, DLD, NDUFB4, FH, ETFB, ATP6V1F, NDUFA6,
COX17, ATP6V0E1, ECI1, COX6A1, TIMM50, ATP6V1E
1, NDUFS7, NDUFB3, MGST3, SDHC, NDUFS3, CYCS, SURF1,
ATP6V1G1, TIMM8B
See also Figure S2.

subpopulation displayed a profoundly activated and exhausted- late dysfunctional T cells (van der Leun et al., 2020), respectively.
like phenotype as suggested by the expression of CTLA4, TOX, Interestingly, the activation/exhaustion state of CD8+ GZMB
LAG3, and PDCD1 (Figure 3E, Table S4), aligning with the previ- T cells was greater in infiltrated compared with excluded tumors,
ously defined dysfunctional CD8+ T cells (van der Leun et al., with higher expression of activation and exhaustion markers
2020). In addition, the CD8+ GZMB subpopulation was marked such as LAYN, CD69, and TNFRSF9 (Figure 3E). Applying the pre-
by the expression of ENTPD1 (CD39) and CXCL13 (Figure 3F). viously developed dysfunction score (Li et al., 2019) and core T cell
CD39 and CXCL13 have been previously identified as potential exhaustion signature (Yost et al., 2019) to our dataset also
markers of tumor-reactive CD8+ T cells (Simoni et al., 2018) and confirmed that the CD8+ GZMB T cells showed the highest degree

932 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

A B C

D E F

G H I

Figure 3. Distinct states of CD8+ T cells characterize immune-infiltrated and -excluded tumor immune phenotypes
(A) UMAP of all T cells from all patients, colored by identified cell subpopulation.
(B) UMAPs of all T cells, colored by the expression of CD4+ and CD8+ T cell markers.
(C) Heatmap of Z-scored expression of the top 20 significant markers (adjusted p < 0.05) of each T cell subpopulation. Z-scored gene expression was calculated
based on scran-normalized and per cell subpopulation-averaged expression values.
(D) UMAPs of all T cells, colored by the expression of GZMB and GMZK.
(E) Heatmap of the Z-scored gene expression values of T cell activation and exhaustion markers for the CD8+ GZMB and CD8+ GZMK cell states. The Z scores
were calculated based on the average scran-normalized expression of all cells in the respective cell state from the excluded or infiltrated TIP.
(F) UMAPs of all T cells, colored by the expression of selected genes.
(G) RNA-scope assay quantification boxplots (top). Each data point represents the fraction of the CD8A/GZMK+ or CD8A/GZMB+ cells in stroma or tumor relative
to the total number of CD8A/GZMK+ or CD8A/GZMB+ in tumor and stroma. Excluded, n = 8; infiltrated, n = 9. Representative images of CD8A/GZMB (bottom left)
and CD8A/GZMK (bottom right) co-hybridization. Blue, CD8A; pink, GZMB or GZMK. Arrows indicate double-positive cells. T, tumor area; S, stroma area.
(H) Stacked bar plot with the fraction of CD8+ GZMK T cells relative to all CD8+ T cells in green and the fraction of CD8+ GZMB T cells relative to all CD8+ T cells in
blue. Each bar represents one tumor sample.
(legend continued on next page)

Cancer Cell 39, 928–944, July 12, 2021 933

 ll
Article

of dysfunctionality and exhaustion (Figure S3F). To validate the Although not significant, we hypothesized that there is an associ-
activation state of CD8+ GZMB T cells in infiltrated versus excluded ation of the CD8+ GZMK pre-dysfunctional subpopulation with
tumors at the protein level, we subjected 17 ovarian cancer tissues excluded tumors and the CD8+ GZMB dysfunctional subpopula-
(9 infiltrated, 8 excluded) from an independent validation collection tion with infiltrated tumors. To validate this hypothesis, we used
to multiplex immunofluorescence for cytokeratins (panCK), CD3, an RNA-scope assay and a bulk RNA-seq deconvolution
PD-1, and GZMB. Among all GZMB+ T cells (CD3/GZMB double approach. Our RNA-scope assay showed a significantly higher
positive), we observed a trend for higher fraction of GZMB+ proportion of CD8A/GZMK double-positive cells in excluded sam-
T cells positive for PD-1 in the infiltrated tumors compared with ples compared with infiltrated samples (p = 0.03), but no significant
the excluded tumors in the tumor epithelium (p = 0.077, Fig- difference in the proportion of CD8A/GZMB double-positive cells
ure S3G), supporting the higher activation state of GZMB+ T cells (Figure S3I). In addition, we investigated bulk RNA-seq data from
in infiltrated tumors, specifically those in the tumor area. In three ovarian cancer cohorts: TCGA (n = 412) (Cancer Genome
contrast, CD8+ GZMK cells have been previously described as po- Atlas Research, 2011), the ICON7 clinical trial collection (n = 351)
tential effector memory T cells (Wu et al., 2020; Zhang et al., 2018) (Perren et al., 2011), and the ROSiA clinical trial cohort (n = 308)
or pre-dysfunctional T cells (Guo et al., 2018; Li et al., 2019; Zheng (Oza et al., 2017). We first predicted the TIP of these samples
et al., 2017). We found a similar marker profile expression in the with our previously established classifier (Desbois et al., 2020). De-
CD8+ GZMK cell population including EOMES, KLRG1, and convolving bulk RNA-seq data using the CIBERSORT CD8 signa-
CMC1 (Figure 3F), as reported in these studies. Interestingly, we ture confirmed a higher CD8+ T cell content in excluded and infil-
found a small fraction of CD8+ GZMK cells that exhibited signifi- trated tumors in all three cohorts, similar to what we observed in
cantly higher expression of TCF7 compared with the dysfunctional the 15 ovarian cancer samples used in our study (Figure S3J). To
CD8+ GZMB cells (p = 0.028, Figure 3F and S3H). The lower level of validate the TIP-specific enrichment of the pre-dysfunctional
exhaustion markers together with the expression of TCF7 sug- CD8+ GZMK versus dysfunctional GZMB T cell subpopulations,
gests that a subset of these pre-dysfunctional T cells may repre- we used the ratio of GZMK or GZMB expression over the CD8+
sent stem cell-like memory T (Tscm) cells as previously reported T cell signature score for each sample as a surrogate for the frac-
(van der Leun et al., 2020). Altogether, these observations suggest tion of each subpopulation among CD8+ T cells. Consistent with
a more advanced dysfunctional state of CD8+ GZMB cells that our model, in all three clinical cohorts, excluded tumors showed
might be linked to their tumor-reactive potential, while CD8+ a significantly higher GZMK/CD8+ ratio compared with infiltrated
GZMK T cells represent pre-dysfunctional effector memory cells. tumors (Figure 3I). Similarly, the GZMB/CD8+ ratio was higher in in-
Given their different dysfunctional states, we explored whether filtrated than in excluded tumors, although statistical significance
the spatial distribution of these T cell subpopulations in stroma for this difference was achieved only in the larger TCGA dataset.
versus tumor epithelium contributes to the difference in their func- Furthermore, combining the GZMB/CD8+ and GZMK/CD8+ ratios
tional state. We subjected the same 17 ovarian cancer samples in a logistic regression model predicting infiltrated versus excluded
from the validation collection to an RNA-scope assay for in situ phenotype significantly increased the explained variance over
co-hybridization of GZMK/CD8A and GZMB/CD8A as well as either single-ratio model in all three cohorts (Figure 3I), highlighting
localization within the stromal versus tumor area through H&E that the fractions of pre-dysfunctional CD8+ GZMK and dysfunc-
staining. Both CD8A/GZMK and CD8A/GZMB double-positive tional GZMB cell subpopulations are both associated with the dif-
T cells accumulated in peritumoral stroma of excluded tumors ference between excluded and infiltrated tumors, and thus sup-
and in tumor epithelium of infiltrated tumors (Figure 3G). From porting that these TIPs show important qualitative and
this observation we concluded that both CD8A/GZMK and quantitative differences in their CD8+ T cell functional states.
CD8A/GZMB T cells have the capacity to infiltrate the tumor In our previous study we observed reduced survival of ovarian
epithelium at least in infiltrated tumors and that therefore the prox- cancer patients on chemotherapy when their tumors exhibited
imity to the tumor epithelium cannot explain the differences in the an excluded phenotype compared with an infiltrated or desert
dysfunctional states of the GZMB+ and GZMK+ subpopulations. phenotype (Desbois et al., 2020). Because of the differential asso-
We next asked whether the dysfunctional CD8+ GZMB and pre- ciation of dysfunctional and pre-dysfunctional CD8+ T cells
dysfunctional CD8+ GZMK T cells showed a different prevalence with infiltrated and excluded tumors, we asked whether the
in excluded compared with infiltrated tumors based on our CD8+ T cell state also associates with clinical outcome under
scRNA-seq analysis: three of five infiltrated tumors showed a lower chemotherapy treatment. In a cohort of 64 patients from the
CD8+ GZMK fraction compared with all excluded tumors, and the chemo-treatment arm of the ICON7 clinical trial, we observed
reverse for the CD8+ GZMB subpopulation (Figures 3H and S3B). that CD8+ T cell quantity was weakly associated with longer

(I) Boxplots of the per sample GZMK/CD8+ or GZMB/CD8+ score by TIP calculated using the ICON7 (n = 351), ROSiA (n = 308), and TCGA (n = 412) bulk RNA-seq
data. The statistical significance was determined using a Wilcoxon test. Right: bar graphs indicating the explained variance of the GZMB/CD8+ score, the GZMK/
CD8+ score, or their combination in a logistic regression model. Significant differences in the variance explained by each model were tested using a chi-
square test.
(J) Cox proportional hazard model hazard ratio with 95% confidence interval for CD8+ score and GZMK/CD8+ score variables. Model was based on ICON7
chemo-treatment arm patients with infiltrated and excluded tumors only. Bar plot shows explained variance of the single models with CD8+ score or GZMK/CD8+
score only, compared with the additive model with CD8+ score and GZMK/CD8+ score. Statistical difference of the models was tested using a chi-square test.
(G and I) The boxplots report the 25% (lower hinge), 50%, and 75% quantiles (upper hinge). The lower whiskers indicate the smallest observation greater than or
equal to lower hinge
1.5 3 interquartile range; the upper whiskers indicate the largest observation less than or equal to upper hinge + 1.5 3 interquartile range.
Norm. expr., normalized expression.
See also Figure S3 and Tables S4 and S5.

934 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

A B

C D

Figure 4. Fibroblast phenotype association with the localization of T cells

(A and B) (A) UMAPs of fibroblasts from all patients, colored by identified cell subpopulations or (B) by the signature score expression derived from Elyada et al.
(2019) and Dominguez et al. (2020).
(C) Heatmap of Z-scored gene expression of the top 20 markers of each fibroblast subpopulation. Z-scored gene expression was calculated based on scran-
normalized and per cell subpopulation-averaged expression values.
(D) Bar plots with the fraction of TGFB CAF and IL1 CAF cells relative to the total fibroblast count grouped by TIP. Each single dot represents a patient. Sig-
nificance was determined by the t statistic accounting for the patient variability as a random effect. Data shown represent the mean ± SD.
Norm. expr., normalized expression.
See also Figure S4 and Tables S5 and S6.

progression-free survival (PFS) (p = 0.093, Figure 3J), indicating a panCAF signature derived from Elyada et al. (2019), indicating
good prognostic effect for CD8 T cell infiltration in ovarian cancer, the presence of CAFs (Figures 4B and S4A). Previously, CAFs
consistent with previous findings (Hwang et al., 2012; Lo et al., have been described either as inflammatory CAFs (iCAFs) and
2017; Zhang et al., 2003). Interestingly, while the GZMB/CD8+ ra- myofibroblasts (myCAFs) (Elyada et al., 2019) or as IL1-activated
tio was not associated with clinical outcome in this cohort, the (IL1 CAFs) and TGFB-activated CAFs (TGFB CAFs) (Dominguez
GZMK/CD8+ ratio was significantly associated with shorter PFS, et al., 2020). Of the two clusters with the strongest panCAF
particularly when combined with the overall CD8+ T cell score in signal, one cluster displayed high iCAF and IL1 CAF signature
a multivariate model (p = 0.029, Figures 3J and S3K). Thus, the expression, while the other cluster showed high myCAF and
relative proportion of CD8+ T cells that are GZMK positive is asso- TGFB CAF expression (Figures 4B and S4A). We therefore anno-
ciated with reduced survival, and this ratio provides more informa- tated these clusters as IL1 CAF and TGFB CAF, respectively.
tion about patient outcome than overall CD8+ T cell content. To gain further insights into the potential functions of these
distinct CAF subpopulations, we compared the gene expression
Fibroblast phenotypes associate with the localization of profiles of TGFB CAFs and IL1 CAFs. The top 20 markers of the
T cells TGFB CAF subpopulation included genes previously associated
Recent studies have revealed distinct subsets of cancer-associ- with TGF-b-induced reactive stroma, periostin (POSTN), smooth
ated fibroblasts (CAFs) in several tumor types (Avery et al., 2018; muscle actin (ACTA2), and cartilage oligomeric matrix protein
Costa et al., 2018; Dominguez et al., 2020; Ohlund et al., 2017). (COMP), as well as collagen subunits (COL10A1, COL11A1), ma-
However, their association with TIP is poorly understood. We trix metalloproteinases (MMP11), transgelin (TAGLN), and fibro-
identified three main clusters of fibroblasts (Figure 4A). Two of nectin (FN1) (Figure 4C, Table S6) (Ryner et al., 2015). On the
the three fibroblast clusters showed a high expression of the other hand, markers characterizing the IL1 CAF subpopulation

Cancer Cell 39, 928–944, July 12, 2021 935

 ll
Article

A B

D E F

Figure 5. Phenotypically and functionally diverse subsets of myeloid cells linked to different tumor immune phenotypes
(A and B) (A) UMAPs of myeloid cells from all patients, colored by identified cell subpopulations or (B) by the expression of selected cell subpopulation
marker genes.
(C) Diffusion map projection of all monocytes/macrophages from all patients, colored by the identified subpopulation. The diffusion map was computed based on
the bbknn-corrected data projection.
(D) UMAPs of myeloid cells, colored by signature scores of the TAM-like and MDSC-like myeloid subsets from Zhang et al. (2019). Violin plot with per cell signature
scores grouped by identified macrophage/monocyte subpopulations. All violin plots report the 25% (lower hinge), 50%, and 75% quantiles (upper hinge) and the
kernel density estimates.
(legend continued on next page)
936 Cancer Cell 39, 928–944, July 12, 2021
 ll
Article

included cytokine/chemokine signaling features such as MARCO, SIGLEC1 (CD169), and CX3CR1 (Figures 5A and 5B,
CXCL14, CCL2, and suppressor of cytokine signaling 3 Table S7). The FCN1 cluster was characterized by high expres-
(SOCS3) (Dominguez et al., 2020; Elyada et al., 2019; Higashino sion of VCAN and S100A transcripts that were previously asso-
et al., 2019). We previously identified an association between the ciated with monocytes (Villani et al., 2017; Zilionis et al., 2019)
TGF-b-mediated reactive stroma and the T cell-excluded TIP in (Figures 5B and S5E). Moreover, this cluster exhibited the high-
ovarian cancer (Desbois et al., 2020). Hence, we analyzed est expression of the CIBERSORT monocyte signature (Fig-
whether the distinct fibroblast subpopulations might be linked ure S5F). Since the CD169 and CX3CR1 subpopulations highly
to different TIPs: IL1 CAFs show a trend of being enriched in in- expressed complement factor genes, maturation markers
filtrated tumors (p = 0.11); however, the association of TGFB (CD83, HLA-DQA1, HLA-DQB1, HLA-DRB5), and the M2 CI-
CAFs with the excluded tumors was not significant (Figures 4D BERSORT signature, we defined these subpopulations as tu-
and S4B, Table S5). Although these TGFB CAFs express reactive mor-associated macrophage (TAM)-like macrophages (Figures
stroma genes, creating a dense matrix that could contribute to 5B, S5E, and S5F). Finally, the MARCO cluster was character-
the exclusion of tumor cells, and represent R40% of fibroblast ized by lower VCAN expression compared with the FCN1 cluster
cells in four of five excluded tumors and only two of five infiltrated and lower M2 signature and maturation marker expression
tumors, additional investigations are warranted, including the compared with the TAM-like macrophages. We therefore anno-
distribution of these cells in the TME to further dissect their asso- tated these cells as MARCO macrophages, similar to the previ-
ciation with excluded tumors. ously described population (Zilionis et al., 2019). To interrogate
The third cluster of fibroblasts showed appreciable but the differentiation trajectory of the identified cell states, we per-
reduced expression of the panCAF signature (Figure S4A). We formed a pseudotime diffusion map analysis. The trajectory
annotated this cluster as a fibroblast-like population (Figure S4B, along the diffusion map coordinate 2 (DC2 in Figure 5C) follows
Table S6). The top markers for this cluster include genes that a continuum from MARCO immature macrophages to CD169
have been associated with epithelial cells (SLPI, KRT19, KRT8, and CX3CR1 mature macrophages, suggesting a maturation tra-
KRT18, WFDC2) (Shih et al., 2018). Additional investigations jectory, while the diffusion map coordinate 1 (DC1) reflects a dif-
are needed to better understand the nature of these cells. ferentiation trajectory, from FCN1 monocytes to mature CD169/
CX3CR1 macrophages (Figure 5C). To further investigate the
Phenotypically and functionally diverse subsets of myeloid cell states, we derived and applied signatures of TAM-
myeloid cells are linked to different tumor immune like and MDSC-like cells published by Zhang et al. (2019).
phenotypes Indeed, the TAM-like signature was expressed in the CD169
Analysis of the myeloid compartment identified clusters represen- and CX3CR1 macrophages and the MDSC-like signature was
tative of dendritic cells (DCs) marked by CD1C and CLEC10A highly expressed in the FCN1 monocytes and MARCO macro-
expression, plasmacytoid DCs expressing LILRA4, and macro- phages (Figure 5D). TREM2, a member of the Triggering Recep-
phage/monocyte clusters marked by CD14 expression (Figures tor Expressed on Myeloid cells (TREM) family, has previously
5A, S5A, and S5B, Table S7). A subcluster analysis of the DC pop- been associated with the TAM-like myeloid cells (Zhang et al.,
ulation identified five clusters labeled according to the featured 2019). Interestingly, our study not only confirmed this previous
genes, including: (1) SPP1 DCs, (2) APOE DCs, (3) CCR7 DCs, finding, but also identified that TREM1 and TREM2, the two
(4) CD1C DCs, and (5) XCR1 DCs (Figure S5C). Three of these clus- members of the TREM family, showed dichotomized expression
ters can be assigned to previously described DC subsets, with in MDSC-like FCN1 monocytes/MARCO macrophages and
XCR1 DCs representing the cDC1 population (XCR1, CLEC9A, TAM-like CD169/CX3CR1 macrophages (Figure 5E).
CADM1), CDC1 DCs as the cDC2 population (FCER1A, CD1C, Finally, we evaluated whether these different myeloid cell pop-
CLEC10A), and CCR7 DCs representing mature DCs (CCR7, ulations were associated with particular TIPs. Strikingly, the
LAMP3) (Zhang et al., 2019). Based on the expression of MAFB, MDSC-like subset (FCN1 monocytes and MARCO macro-
FCGR3A, CD14, FCGR1A, and CD163, we hypothesized that the phages) was significantly enriched in desert tumors,
APOE DC cluster represents an inflammatory monocyte-derived whereas the TAM-like myeloid subset (CD169 and CX3CR1
DC population (Collin and Bigley, 2018), while the SPP1 DCs macrophages) was enriched in excluded and infiltrated tumors
shared similarities with the CD1c
CD141
DC subset described (p = 0.02, Figures 5F, S5F, and S5G, Table S5), with no difference
by Villani et al. (SERPINA1 and FCGR3A) (Villani et al., 2017) (Fig- between excluded and infiltrated tumors for either population.
ure S5D). Unfortunately, the number of DCs was too low to inves-
tigate the association of these DC subsets with the TIP. Tumor immune phenotypes are shaped by cross-
We further characterized the four macrophage/monocyte compartment interactions
clusters and annotated them according to featured markers Our single-cell characterization of the TME also allowed us to
and genes that have been previously used to describe similar interrogate potential interactions between cell compartments
cell populations (Aran et al., 2019; Azizi et al., 2018; Cassetta that may help to shape the tumor immunity continuum. Although
et al., 2019; Lavin et al., 2017; Zilionis et al., 2019): FCN1, single-cell data cannot provide definitive evidence of cell-to-cell

(E) UMAPs of all myeloid cells, colored by TREM1 and TREM2 expression.
(F) Bar plots with the fraction of the combined CD169/CX3CR1 and MARCO/FCN1 cells relative to the total monocyte and macrophage cell count. Significance
was determined using a Wilcoxon rank-sum test.
Norm. expr., normalized expression.
See also Figure S5, Tables S5 and S7.

Cancer Cell 39, 928–944, July 12, 2021 937

 ll
Article

A C

D E

Figure 6. Tumor immune phenotypes shaped by cross-compartment interactions

(A) Dot plot of CXCL16 and CXCR6 expression by compartment and cell type.
(B) Boxplot of the average CXCL16 scran-normalized expression in tumor cells for each patient, grouped by TIP.
(C) Dot plot of the CXCR3 and corresponding chemokine ligand expression by cell type with heatmap depicting the enrichment of the corresponding cells in the
three TIPs.
(legend continued on next page)

938 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

signaling, various groups have shown how consideration of cell- by the IL1 CAF subpopulation (Figures 6D), and they are known
specific receptor and ligand expression patterns in single-cell to bind to the receptor CXCR4 (Collins et al., 2017; Tanegashima
data can be used to generate hypotheses (Camp et al., 2017; et al., 2013; Vega et al., 2011). Notably, this receptor was found
Costa et al., 2018; Skelly et al., 2018; Vento-Tormo et al., to be almost exclusively expressed by immune cells, and most
2018). Following such approaches, we focused on 23 known strongly by CD8+ T cells (Figures 6D and S6F), suggesting a
chemokine ligand-receptor pairs (Table S8) and took several role for IL1 CAFs in the CD8+ T cell recruitment specifically in in-
filtering steps resulting in six putative chemokine ligand-receptor filtrated tumors, since these show a weak enrichment of IL1
interactions, which we further investigated (STAR Methods). CAFs, while T cells are rare in desert tumors. Endothelial cells
Intriguingly, we found evidence potentially implicating each of and pericytes were the only cell types to appreciably express
the compartments (tumor, immune, and stromal) in the recruit- the main ligands of CX3CR1: CX3CL1 and the recently described
ment and migration of T cells. ligand CCL26 (El-Shazly et al., 2013) (Figure 6E). Since CX3CR1
Tumor cell-immune cell cross talk marked one of the major mature macrophage populations in in-
The chemokine CXCL16, a chemokine known for the recruitment filtrated and excluded tumors (Figure 6F), we hypothesize that a
of T cells (Matsumura et al., 2008), was expressed by tumor cells recruitment of CX3CR1 macrophages through endothelial cells
as well as CD45+ cells, mostly myeloid cells (Figures 6A and and pericytes is specific to these two TIPs (Figure S6G). In addi-
S6A). Further, we observed that CXCL16 expression was signifi- tion to myeloid cells, B-TILs showed evidence of possible che-
cantly higher in the tumor cell compartment of infiltrated and mokine receptor-ligand cross talk with the stromal cell compart-
excluded tumors compared with desert tumors (p = 0.012 ment. Endothelial cells expressed CCL21, an important ligand
and 0.022, respectively, Figure 6B). Interestingly, the CXCL16 enabling the recruitment of CCR7+ cells (Comerford et al.,
receptor, CXCR6 (Wilbanks et al., 2001), was found to be particu- 2013), with a significantly higher expression in endothelial cells
larly high in dysfunctional CD8+ GZMB T cells and CD4+ of excluded tumors (p = 0.03) and consistent but non-significant
FOXP3 regulatory T cells (Tregs) (Figure 6A). This co-expression higher expression in infiltrated tumors (p = 0.06, Figure S6H and
pattern could be confirmed in all individual patients for which we S6I). In our dataset, most CCR7-expressing cells are B-TILs,
had enough cells, and it was particularly apparent in tumor cells suggesting possible recruitment via endothelial cells, specifically
of infiltrated tumors (Table S7, Figure S6B). Together with our in excluded and infiltrated tumors.
observation that dysfunctional CD8+ GZMB T cells and CD4+ To summarize these observations, we postulate a model in
FOXP3 Tregs were enriched in infiltrated tumors, this finding sug- which the different compositions and the potential cross talk within
gested a potential mechanism by which CXCR6+ T cells could be and between the tumor, immune, and stromal compartments
recruited to the tumor epithelium in infiltrated tumors. might shape the distinct TIP (Figure 6G). In our model, tumors
Immune cell-immune cell cross talk with immune presence in the TME, including both infiltrated and
Similar to the potential CXCL16/CXCR6 cross talk between tu- excluded tumors, share many common features that are distinct
mor cells and T cells, we identified possible cross talk between from desert tumors. For example, both infiltrated and excluded tu-
T cells and myeloid cells. The CXCR3 receptor was expressed mors showed an enrichment of T cells and TAM-like cells, while
by CD8+ and CD4+ T cells, and its major ligands, CXCL9, desert tumors show an almost complete absence of T cells, but
CXCL10, and CXCL11 were mainly expressed by DCs and an enrichment of MDSC-like cells. The T cell presence in infiltrated
CD169 macrophages (Figures 6C and S6C). Since the CD169 and excluded tumors could be in part facilitated by a T cell-TAM-
macrophages were enriched in excluded and infiltrated tumors, like cross talk through CXCR3-CXCL9/10/11 signaling. In addition,
this suggests a mechanism by which immune cell infiltration is endothelial cells and pericytes expressing CX3CR1 ligands might
facilitated in those tumors. In addition, the expression of also participate in the recruitment of CX3CR1 TAM-like cells in in-
CXCR5 by B-TILs and of the corresponding chemokine ligand filtrated and excluded tumors. On the other hand, there were also
CXCL13 by activated CD4+ CXCL13 and dysfunctional CD8+ important differences between infiltrated and excluded tumors.
GZMB T cells in infiltrated tumors suggests a potential mecha- The greater extent of T cell infiltration in the infiltrated tumors might
nism of B-TIL recruitment to infiltrated tumors (Kazanietz et al., be influenced by two factors: (1) infiltrated tumor cells showed the
2019) (Figures S6D and S6E). highest CXCL16 expression, which may promote the recruitment
Stromal cell-immune cell cross talk of CXCR6+ T cells to the tumor epithelium, and (2) infiltrated tumors
Stromal cells may also participate in the recruitment of immune are enriched in IL1 CAFs, whose expression of CXCL12/14 may
cells. The chemokines CXCL14 and CXCL12 were expressed further facilitate T cell recruitment via CXCR4.

(D) Dot plot of the average expression of CXCL14, CXCL12 and CXCR4 in immune and stromal subpopulations as well as in the tumor cell compartment with
heatmap depicting the enrichment of the corresponding cell types in the TIP.
(E) Dot plot of CX3CR1 and corresponding chemokine ligand expression by cell type.
(F) Boxplot of the CX3CR1 expression by myeloid cell subpopulation.
(G) Model of the cross-compartment chemokine ligand-receptor interactions in the context of the TME of each TIP. The model for fibroblasts is depicted based on
a trend for higher IL1 CAFs in infiltrated tumors, supported by previous studies from our group and others.
(A, C, D, and E) The color intensity of each dot indicates the average scran-normalized expression across all patients; the size represents the percentage of cells
that express a gene relative to the total number of cells in that group.
(B and F) All boxplots report the 25% (lower hinge), 50%, and 75% quantiles (upper hinge). The lower whiskers indicate the smallest observation greater than or
equal to lower hinge
1.5 3 interquartile range, the upper whiskers indicate the largest observation less than or equal to upper hinge + 1.5 3 interquartile range.
Each dot indicates the average expression level for each patient. Statistical significance was calculated using a Wilcoxon test.
See also Figure S6 and Tables S7 and S8.

Cancer Cell 39, 928–944, July 12, 2021 939

 ll
Article

DISCUSSION with excluded tumors, as one would expect. However, our

spatial analysis also suggests that the pre-dysfunctional CD8+
While cancer immunotherapy is effective in certain indications, it GZMK T cell population has the capability of infiltrating the tumor
is not effective for everyone and the variability in response is not epithelium. Therefore, the spatial localization, i.e., the exclusion
completely understood (Sharma et al., 2017). Developing a bet- of the CD8+ GZMK T cells from the tumor epithelium, cannot fully
ter understanding of the tumor and its microenvironment explain their functional state.
composition might improve and extend the benefit of cancer im- One important question being investigated by many groups
munotherapies to more patients and enable personalized ap- is whether immune checkpoint blockades (ICBs), in particular
proaches to treatment (Hegde and Chen, 2020). Our study not anti-PD-(L)1 antibodies, can reinvigorate the dysfunctional tu-
only provides a high-resolution depiction of the cellular diversity mor-infiltrating CD8+ T cells. While additional investigation is
in the tumor, immune, and stromal compartments, but it also needed, recent studies support a model wherein pre-dysfunc-
highlights how the cross talk within and between the compart- tional rather than late dysfunctional T cells are targeted by
ments may contribute to shaping the biology of the TIP and ulti- ICBs and promote an anti-tumor response. In particular, the
mately could influence the response to immunotherapies. Tscm cells as part of the pre-dysfunctional T cells have been
We first investigated whether intrinsic properties of tumor cells found to expand upon ICB treatment (Sade-Feldman et al.,
themselves contribute to a milieu that favors or hinders immune 2018; Utzschneider et al., 2016) and to associate with a longer
infiltration (Galon and Bruni, 2019). Downregulation of antigen duration of response to ICB treatment (Miller et al., 2019). While
presentation has been observed in our previous bulk RNA-seq our understanding of the role of these cells is increasing in the
and in situ MHC I IHC studies in desert and excluded tumors context of immunotherapies, their role in the response to
(Desbois et al., 2020). Although we did not observe a strong chemotherapy is unclear. Hence, the mere presence of Tscm
downregulation of antigen presentation at the gene level in the cells in ovarian cancer patients under chemotherapy as pre-
excluded tumors in this study, we cannot rule out that the sented in this study might not have an impact on their outcome.
MHC may be downregulated at the protein level. In addition, Moreover, in our previous work (Desbois et al., 2020) and in the
the upregulation of the oxidative phosphorylation pathway in in- present study, we observed that the quantity of CD8+ TILs infil-
filtrated/excluded tumors is in line with previous findings trating the tumor is only weakly associated with PFS. Patients
showing that the metabolism of tumor cells can directly impair with excluded tumors showed shorter PFS, and the amount
T cell infiltration and function through, for example, competition of CD8+ GZMK cells among all CD8+ cells predicted shorter
for metabolites that are essential for T cell function, or accumu- PFS under chemotherapy. Because we found these pre-
lation of waste products like lactate that create an unfavorable dysfunctional cells enriched in excluded tumors it is possible
microenvironment and impair T cell migration (Haas et al., that the observed association with PFS is intimately linked.
2015; Sugiura and Rathmell, 2018). Although the presence of pre-dysfunctional cells is inversely
The tumor immunity continuum has been largely characterized correlated with PFS in a chemo cohort, these observations
in terms of the quantity and location of T cells in the tumor bed do not rule out an optimal clinical efficacy if both activated
(Hegde et al., 2016). Here, we revealed diverse phenotypic and stroma and pre-dysfunctional/Tscm cells are simultaneously
functional T cell and myeloid cell states, as well as their differen- targeted under ICB.
tial enrichment in the TIP. While the CD8+ GZMB and CD8+ Another key finding of this study is the in-depth dissection of
GZMK T cell subpopulations we identified have been previously the heterogeneous myeloid cell population in the context of
described in single-cell studies (Guo et al., 2018; Li et al., 2019; different TIPs. We observed a specific linkage of TREM1 and
Savas et al., 2018; Wu et al., 2020; Yost et al., 2019; Zhang et al., TREM2 with different subsets of myeloid cells associated with
2018, 2019), one of our key findings is that they not only repre- distinct TIPs, which provides additional insights into their poten-
sent different functional states, but also are differentially en- tial roles and may inform therapeutic strategies for targeting
riched in the different TIPs. Our analysis of the association of TREM molecules. High TREM1 expression in macrophages infil-
these CD8+ T cell subpopulations with the TIP provided two in- trating human tumors has been previously shown to be associ-
sights: (1) the dysfunctional/exhausted CD8+ GZMB T cells ated with aggressive tumor behavior and poor patient survival
were less activated/exhausted in excluded tumors and (2) the (Ho et al., 2008). On the other hand, TREM2 has been shown
pre-dysfunctional/effector memory CD8+ GZMK T cells were en- to act as a tumor suppressor in hepatocellular carcinoma
riched in excluded tumors compared with infiltrated tumors. (Tang et al., 2019) and colorectal cancer (Kim et al., 2019). Tar-
Both observations might be explained by a potential link be- geting TREM molecules has recently drawn increased attention
tween the dysfunction/activation states and the T cell spatial dis- as a novel therapeutic opportunity for the treatment of inflamma-
tribution, i.e., infiltration or exclusion of the T cells from the tumor tory disorders and cancer (Nguyen et al., 2015).
epithelium. The immune cell exclusion in excluded tumors might Last, we dissected how the cellular components of the tumor,
result in a lack of sustained antigenic stimulation by tumor cells stromal, and immune compartments may interact through che-
and therefore contribute to the less activated/exhausted CD8+ mokine-receptor signaling and thereby help to shape the tumor
GZMB T cells and the enrichment of pre-dysfunctional CD8+ immunity continuum. Our study revealed a mechanism by which
GZMK T cells. In line with this model, we found an enrichment tumor cells potentially mediate T cell recruitment via the CXCR6-
of resting CD4+ T cell populations in the excluded tumors, and CXCL16 axis. We found CXCL16 to be expressed by myeloid
activated CD4+ T cells and regulatory T cells in the infiltrated tu- cells as previously described (van der Voort et al., 2005), but,
mor. Taken together, these observations point toward a more more importantly, our analysis revealed that CXCL16 is also ex-
antigen-stimulated immune landscape in infiltrated compared pressed on tumor cells, especially in infiltrated and excluded

940 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

ovarian tumor cells. CXCL16 is known to signal through the che- STAR+METHODS
mokine receptor CXCR6 (Wilbanks et al., 2001), and we found
the highest expression of CXCR6 on CD4+ FOXP3 Tregs and Detailed methods are provided in the online version of this paper
dysfunctional CD8+ GZMB T cells. These observations suggest and include the following:
potential recruitment of these T cell subsets by tumor cells in in-
filtrated and excluded tumors. Supporting these findings, it has d KEY RESOURCES TABLE
been previously shown that ionizing radiation can induce the d RESOURCE AVAILABILITY
secretion of CXCL16, which would otherwise recruit CXCR6+ B Lead contact

CD8+ activated T cells to the tumor in a poorly immunogenic B Materials availability

B Data and code availability
breast cancer mouse model (Matsumura et al., 2008). Hence,
the CXCL16-CXCR6 axis could represent an important factor d EXPERIMENTAL MODEL AND SUBJECT
contributing to the tumor immunity continuum in ovarian cancer. B Clinical samples

Nevertheless, the effect of the CXCL16 chemotaxis gradient d METHOD DETAILS

B Bulk RNA sequencing
might be different between excluded and infiltrated tumors,
whereby T cells cannot reach the tumor epithelium in excluded B Single-cell RNA sequencing
B In situ experimental procedures
tumors despite the presence of a chemokine gradient. In fact,
we found that a large fraction of myofibroblasts in the excluded d QUANTIFICATION AND STATISTICAL ANALYSIS
tumors not only express the myofibroblast-specific marker B Survival analysis

ACTA2 (⍺SMA) (Sahai et al., 2020), but also collagen genes B Statistical analysis

(COL10A1, COL11A1, COL6A3, COL1A1) and genes previously

SUPPLEMENTAL INFORMATION
shown to contribute to reactive stroma (POSTN, MMP11, FN1)
(Planche et al., 2011; Ryner et al., 2015). These observations Supplemental information can be found online at https://doi.org/10.1016/j.
further support the hypothesis that specific CAFs create a phys- ccell.2021.04.004.
ical barrier that blocks the access of T cells to the tumor epithe-
lium by producing a dense extracellular matrix (Salmon et al., ACKNOWLEDGMENTS
2012; Zeltz et al., 2020).
Our study has several limitations worth noting. By sorting live The authors would like to thank all the staff members from the Genentech
FACS core facility who helped with the cell sorting of ovarian tumors; Ching-
cells based on their compartment of origin (i.e., immune, stromal,
Wei Chang for analyzing the molecular subtypes of all ovarian tumors; Luciana
or tumor compartment), we were able to enrich for low-abun- Molinero, Patricia Werner, and Regula Deurloo for the clinical datasets; and
dance cell populations and achieve a high-resolution dissection Shilpa Keerthivasan for her help with the staining of samples for cell sorting.
of each compartment. However, such an approach does not We thank Christopher Corless (Knight Diagnostics Labs, Portland, Oregon)
allow one to describe the composition of the microenvironment for analyzing the mutations of ovarian cancer samples in the KIYATEC collec-
as a whole and can report only the proportion of the identified tion. Finally, we thank many scientists for the great discussions enabling a bet-
ter understanding of this work, including Shomyseh Sanjabi, William O’Gor-
cell states relative to their respective individual compartment.
man, Dorothee Nickles, Yijin Li, and Georgia Hatzivassiliou. The graphical
In addition, it has been reported that ovarian tumors and their im- abstract was created with Biorender. This work was funded by Genen-
mune microenvironments are heterogeneous (Jimenez-Sanchez tech/Roche.
et al., 2017; Roberts et al., 2019). Although our study has utilized
different tumor sections from the same primary tumor and em- AUTHOR CONTRIBUTIONS
ployed orthogonal methods (transcriptome classifier-based pre-
Conceptualization, M.H., M.D., A.D., and Y.W.; software, M.H., O.M., and
dictions and the CD8 IHC categorization) to assign a TIP for each X.W.; methodology, M.H. and M.D.; formal analysis and statistics, O.M. and
tumor, it is still possible that such classification may not repre- M.H.; investigation, M.H., M.D., S.L., Y.G., A.A.L., J.G., and J.L.M.R.; re-
sent the whole tumor. Future studies including characterization sources, A.E., A.L., and T.M.D.; writing – original draft, M.H., M.D., A.D., and
of multiple segments of the same tumor as well as metastatic Y.W.; writing – review & editing, M.H., M.D., A.D., and Y.W.; visualization,
tumors of different sites from the same patient may provide addi- M.H. and M.D.; supervision, S.J.T., R.B., A.D., and Y.W.; project administra-
tion, S.K.
tional insights into the heterogeneous TME of ovarian cancer.
Finally, our chemokine receptor-ligand analysis across the DECLARATION OF INTERESTS
different compartments is derived from a transcriptomic analysis
only, and further validation of these potential interactions by M.H., M.D., S.L., Y.G., S.K., A.A.L., S.K., J.L.M.R., X.W., J.G., O.M., S.J.T.,
R.B., A.D., and Y.W. are Genentech/Roche employees. S.L., Y.G., S.K., A.L.,
high-dimensional multiplex in situ analysis and functional assays
S.K., J.L.M.R., X.W., J.G., O.M., S.J.T., R.B., A.D., and Y.W. hold Roche
in future studies is warranted. stocks.
To conclude, our study provides an in-depth dissection of the
diverse cellular and functional phenotypes in the TME and their Received: July 9, 2020
dynamic interplay, enabling a richer characterization of the tumor Revised: November 12, 2020
immunity continuum. Our work also provides additional insights Accepted: April 6, 2021
into the biology that may help to shape the TME and TIP. Finally, Published: May 6, 2021
our findings may also enable identification of therapeutic targets
REFERENCES
and inform novel therapeutic strategies for overcoming immune
suppression and increasing or expanding the response to can- Aran, D., Hu, Z., and Butte, A.J. (2017). xCell: digitally portraying the tissue
cer immunotherapies. cellular heterogeneity landscape. Genome Biol. 18, 220.

Cancer Cell 39, 928–944, July 12, 2021 941

 ll
Article
Aran, D., Looney, A.P., Liu, L., Wu, E., Fong, V., Hsu, A., Chak, S., Naikawadi, Haas, R., Smith, J., Rocher-Ros, V., Nadkarni, S., Montero-Melendez, T.,
R.P., Wolters, P.J., Abate, A.R., et al. (2019). Reference-based analysis of lung D’Acquisto, F., Bland, E.J., Bombardieri, M., Pitzalis, C., Perretti, M., et al.
single-cell sequencing reveals a transitional profibrotic macrophage. Nat. (2015). Lactate regulates metabolic and pro-inflammatory circuits in control
Immunol. 20, 163–172. of T cell migration and effector functions. PLoS Biol. 13, e1002202.
Avery, D., Govindaraju, P., Jacob, M., Todd, L., Monslow, J., and Pure, E. Hegde, P.S., and Chen, D.S. (2020). Top 10 challenges in cancer immuno-
(2018). Extracellular matrix directs phenotypic heterogeneity of activated fibro- therapy. Immunity 52, 17–35.
blasts. Matrix Biol. 67, 90–106. Hegde, P.S., Karanikas, V., and Evers, S. (2016). The where, the when, and the
Azizi, E., Carr, A.J., Plitas, G., Cornish, A.E., Konopacki, C., Prabhakaran, S., how of immune monitoring for cancer immunotherapies in the era of check-
Nainys, J., Wu, K., Kiseliovas, V., Setty, M., et al. (2018). Single-cell map of point inhibition. Clin. Cancer Res. 22, 1865–1874.
diverse immune phenotypes in the breast tumor microenvironment. Cell 174, Higashino, N., Koma, Y.I., Hosono, M., Takase, N., Okamoto, M., Kodaira, H.,
1293–1308.e1236. Nishio, M., Shigeoka, M., Kakeji, Y., and Yokozaki, H. (2019). Fibroblast acti-
Camp, J.G., Sekine, K., Gerber, T., Loeffler-Wirth, H., Binder, H., Gac, M., vation protein-positive fibroblasts promote tumor progression through secre-
Kanton, S., Kageyama, J., Damm, G., Seehofer, D., et al. (2017). tion of CCL2 and interleukin-6 in esophageal squamous cell carcinoma. Lab.
Multilineage communication regulates human liver bud development from plu- Invest. 99, 777–792.
ripotency. Nature 546, 533–538. Ho, C.C., Liao, W.Y., Wang, C.Y., Lu, Y.H., Huang, H.Y., Chen, H.Y., Chan,
Cancer Genome Atlas Research, N. (2011). Integrated genomic analyses of W.K., Chen, H.W., and Yang, P.C. (2008). TREM-1 expression in tumor-asso-
ovarian carcinoma. Nature 474, 609–615. ciated macrophages and clinical outcome in lung cancer. Am. J. Respir. Crit.
Cassetta, L., Fragkogianni, S., Sims, A.H., Swierczak, A., Forrester, L.M., Care Med. 177, 763–770.
Zhang, H., Soong, D.Y.H., Cotechini, T., Anur, P., Lin, E.Y., et al. (2019). Hwang, W.T., Adams, S.F., Tahirovic, E., Hagemann, I.S., and Coukos, G.
Human tumor-associated macrophage and monocyte transcriptional land- (2012). Prognostic significance of tumor-infiltrating T cells in ovarian cancer:
scapes reveal cancer-specific reprogramming, biomarkers, and therapeutic a meta-analysis. Gynecol. Oncol. 124, 192–198.
targets. Cancer Cell 35, 588–602 e510. Jerby-Arnon, L., Shah, P., Cuoco, M.S., Rodman, C., Su, M.J., Melms, J.C.,
Collin, M., and Bigley, V. (2018). Human dendritic cell subsets: an update. Leeson, R., Kanodia, A., Mei, S., Lin, J.R., et al. (2018). A cancer cell program
Immunology 154, 3–20. promotes T cell exclusion and resistance to checkpoint blockade. Cell 175,
Collins, P.J., McCully, M.L., Martinez-Munoz, L., Santiago, C., Wheeldon, J., 984–997.e924.
Caucheteux, S., Thelen, S., Cecchinato, V., Laufer, J.M., Purvanov, V., et al. Jimenez-Sanchez, A., Memon, D., Pourpe, S., Veeraraghavan, H., Li, Y.,
(2017). Epithelial chemokine CXCL14 synergizes with CXCL12 via allosteric Vargas, H.A., Gill, M.B., Park, K.J., Zivanovic, O., Konner, J., et al. (2017).
modulation of CXCR4. FASEB J. 31, 3084–3097. Heterogeneous tumor-immune microenvironments among differentially
Comerford, I., Harata-Lee, Y., Bunting, M.D., Gregor, C., Kara, E.E., and growing metastases in an ovarian cancer patient. Cell 170, 927–938.e920.
McColl, S.R. (2013). A myriad of functions and complex regulation of the Kazanietz, M.G., Durando, M., and Cooke, M. (2019). CXCL13 and its receptor
CCR7/CCL19/CCL21 chemokine axis in the adaptive immune system. CXCR5 in cancer: inflammation, immune response, and beyond. Front.
Cytokine Growth Factor. Rev. 24, 269–283. Endocrinol. (Lausanne) 10, 471.
Costa, A., Kieffer, Y., Scholer-Dahirel, A., Pelon, F., Bourachot, B., Cardon, M., Kim, S.M., Kim, E.M., Ji, K.Y., Lee, H.Y., Yee, S.M., Woo, S.M., Yi, J.W., Yun,
Sirven, P., Magagna, I., Fuhrmann, L., Bernard, C., et al. (2018). Fibroblast het- C.H., Choi, H., and Kang, H.S. (2019). TREM2 acts as a tumor suppressor in
erogeneity and immunosuppressive environment in human breast cancer. colorectal carcinoma through Wnt1/beta-catenin and erk signaling. Cancers
Cancer Cell 33, 463–479.e410. (Basel) 11, 1315.
Desbois, M., Udyavar, A.R., Ryner, L., Kozlowski, C., Guan, Y., Durrbaum, M., Lambrechts, D., Wauters, E., Boeckx, B., Aibar, S., Nittner, D., Burton, O.,
Lu, S., Fortin, J.P., Koeppen, H., Ziai, J., et al. (2020). Integrated digital pathol- Bassez, A., Decaluwe, H., Pircher, A., Van den Eynde, K., et al. (2018).
ogy and transcriptome analysis identifies molecular mediators of T-cell exclu- Phenotype molding of stromal cells in the lung tumor microenvironment.
sion in ovarian cancer. Nat. Commun. 11, 5583. Nat. Med. 24, 1277–1289.
Dominguez, C.X., Muller, S., Keerthivasan, S., Koeppen, H., Hung, J., Gierke, Lavin, Y., Kobayashi, S., Leader, A., Amir, E.D., Elefant, N., Bigenwald, C.,
S., Breart, B., Foreman, O., Bainbridge, T.W., Castiglioni, A., et al. (2020). Remark, R., Sweeney, R., Becker, C.D., Levine, J.H., et al. (2017). Innate im-
Single-cell RNA sequencing reveals stromal evolution into LRRC15(+) myofi- mune landscape in early lung adenocarcinoma by paired single-cell analyses.
broblasts as a determinant of patient response to cancer immunotherapy. Cell 169, 750–765.e717.
Cancer Discov. 10, 232–253. Li, H., van der Leun, A.M., Yofe, I., Lubling, Y., Gelbard-Solodkin, D., van
Efremova, M., Vento-Tormo, M., Teichmann, S.A., and Vento-Tormo, R. Akkooi, A.C.J., van den Braber, M., Rozeman, E.A., Haanen, J., Blank, C.U.,
(2020). CellPhoneDB: inferring cell-cell communication from combined et al. (2019). Dysfunctional CD8 T cells form a proliferative, dynamically regu-
expression of multi-subunit ligand-receptor complexes. Nat. Protoc. 15, lated compartment within human melanoma. Cell 176, 775–789.e718.
1484–1506. Lo, C.S., Sanii, S., Kroeger, D.R., Milne, K., Talhouk, A., Chiu, D.S., Rahimi, K.,
El-Shazly, A.E., Doloriert, H.C., Bisig, B., Lefebvre, P.P., Delvenne, P., and Shaw, P.A., Clarke, B.A., and Nelson, B.H. (2017). Neoadjuvant chemotherapy
Jacobs, N. (2013). Novel cooperation between CX3CL1 and CCL26 inducing of ovarian cancer results in three patterns of tumor-infiltrating lymphocyte
NK cell chemotaxis via CX3CR1: a possible mechanism for NK cell infiltration response with distinct implications for immunotherapy. Clin. Cancer Res. 23,
of the allergic nasal tissue. Clin. Exp. Allergy 43, 322–331. 925–934.
Elyada, E., Bolisetty, M., Laise, P., Flynn, W.F., Courtois, E.T., Burkhart, R.A., Lun, A.T., Bach, K., and Marioni, J.C. (2016). Pooling across cells to normalize
Teinor, J.A., Belleau, P., Biffi, G., Lucito, M.S., et al. (2019). Cross-species sin- single-cell RNA sequencing data with many zero counts. Genome Biol. 17, 75.
gle-cell analysis of pancreatic ductal adenocarcinoma reveals antigen-pre- Matsumura, S., Wang, B., Kawashima, N., Braunstein, S., Badura, M.,
senting cancer-associated fibroblasts. Cancer Discov. 9, 1102–1123. Cameron, T.O., Babb, J.S., Schneider, R.J., Formenti, S.C., Dustin, M.L.,
Galon, J., and Bruni, D. (2019). Approaches to treat immune hot, altered and and Demaria, S. (2008). Radiation-induced CXCL16 release by breast cancer
cold tumours with combination immunotherapies. Nat. Rev. Drug Discov. cells attracts effector T cells. J. Immunol. 181, 3099–3107.
18, 197–218. Miller, B.C., Sen, D.R., Al Abosy, R., Bi, K., Virkud, Y.V., LaFleur, M.W., Yates,
Guo, X., Zhang, Y., Zheng, L., Zheng, C., Song, J., Zhang, Q., Kang, B., Liu, Z., K.B., Lako, A., Felt, K., Naik, G.S., et al. (2019). Subsets of exhausted CD8(+)
Jin, L., Xing, R., et al. (2018). Global characterization of T cells in non-small-cell T cells differentially mediate tumor control and respond to checkpoint
lung cancer by single-cell sequencing. Nat. Med. 24, 978–985. blockade. Nat. Immunol. 20, 326–336.

942 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article
Newman, A.M., Liu, C.L., Green, M.R., Gentles, A.J., Feng, W., Xu, Y., Hoang, veals cellular diversity and intercommunication in the mouse heart. Cell Rep.
C.D., Diehn, M., and Alizadeh, A.A. (2015). Robust enumeration of cell subsets 22, 600–610.
from tissue expression profiles. Nat. Methods 12, 453–457. Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L.,
Nguyen, A.H., Berim, I.G., and Agrawal, D.K. (2015). Chronic inflammation and Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., and
cancer: emerging roles of triggering receptors expressed on myeloid cells. Mesirov, J.P. (2005). Gene set enrichment analysis: a knowledge-based
Expert Rev. Clin. Immunol. 11, 849–857. approach for interpreting genome-wide expression profiles. Proc. Natl.
Ohlund, D., Handly-Santana, A., Biffi, G., Elyada, E., Almeida, A.S., Ponz- Acad. Sci. U S A 102, 15545–15550.
Sarvise, M., Corbo, V., Oni, T.E., Hearn, S.A., Lee, E.J., et al. (2017). Distinct Sugiura, A., and Rathmell, J.C. (2018). Metabolic barriers to T cell function in
populations of inflammatory fibroblasts and myofibroblasts in pancreatic can- tumors. J. Immunol. 200, 400–407.
cer. J. Exp. Med. 214, 579–596.
Tanegashima, K., Suzuki, K., Nakayama, Y., Tsuji, K., Shigenaga, A., Otaka, A.,
Oza, A.M., Selle, F., Davidenko, I., Korach, J., Mendiola, C., Pautier, P., and Hara, T. (2013). CXCL14 is a natural inhibitor of the CXCL12-CXCR4
Chmielowska, E., Bamias, A., DeCensi, A., Zvirbule, Z., et al. (2017). Efficacy signaling axis. FEBS Lett. 587, 1731–1735.
and safety of bevacizumab-containing therapy in newly diagnosed ovarian
Tang, W., Lv, B., Yang, B., Chen, Y., Yuan, F., Ma, L., Chen, S., Zhang, S., and
cancer: ROSiA single-arm phase 3B study. Int. J. Gynecol. Cancer 27, 50–58.
Xia, J. (2019). TREM2 acts as a tumor suppressor in hepatocellular carcinoma
Perren, T.J., Swart, A.M., Pfisterer, J., Ledermann, J.A., Pujade-Lauraine, E., by targeting the PI3K/Akt/beta-catenin pathway. Oncogenesis 8, 9.
Kristensen, G., Carey, M.S., Beale, P., Cervantes, A., Kurzeder, C., et al.
Thorsson, V., Gibbs, D.L., Brown, S.D., Wolf, D., Bortone, D.S., Ou Yang, T.H.,
(2011). A phase 3 trial of bevacizumab in ovarian cancer. N. Engl. J. Med.
Porta-Pardo, E., Gao, G.F., Plaisier, C.L., Eddy, J.A., et al. (2018). The immune
365, 2484–2496.
landscape of cancer. Immunity 48, 812–830.e814.
Planche, A., Bacac, M., Provero, P., Fusco, C., Delorenzi, M., Stehle, J.C., and
Tirosh, I., Izar, B., Prakadan, S.M., Wadsworth, M.H., 2nd, Treacy, D.,
Stamenkovic, I. (2011). Identification of prognostic molecular features in the
Trombetta, J.J., Rotem, A., Rodman, C., Lian, C., Murphy, G., et al. (2016).
reactive stroma of human breast and prostate cancer. PLoS One 6, e18640.
Dissecting the multicellular ecosystem of metastatic melanoma by single-
Polanski, K., Young, M.D., Miao, Z., Meyer, K.B., Teichmann, S.A., and Park, cell RNA-seq. Science 352, 189–196.
J.E. (2020). BBKNN: fast batch alignment of single cell transcriptomes.
Bioinformatics 36, 964–965. Utzschneider, D.T., Charmoy, M., Chennupati, V., Pousse, L., Ferreira, D.P.,
Calderon-Copete, S., Danilo, M., Alfei, F., Hofmann, M., Wieland, D., et al.
Puram, S.V., Tirosh, I., Parikh, A.S., Patel, A.P., Yizhak, K., Gillespie, S.,
(2016). T cell factor 1-expressing memory-like CD8(+) T cells sustain the im-
Rodman, C., Luo, C.L., Mroz, E.A., Emerick, K.S., et al. (2017). Single-cell tran-
mune response to chronic viral infections. Immunity 45, 415–427.
scriptomic analysis of primary and metastatic tumor ecosystems in head and
neck cancer. Cell 171, 1611–1624.e1624. van der Leun, A.M., Thommen, D.S., and Schumacher, T.N. (2020). CD8(+)
T cell states in human cancer: insights from single-cell analysis. Nat. Rev.
Roberts, C.M., Cardenas, C., and Tedja, R. (2019). The role of intra-tumoral
Cancer 20, 218–232.
heterogeneity and its clinical relevance in epithelial ovarian cancer recurrence
and metastasis. Cancers (Basel) 11, 1083. van der Voort, R., van Lieshout, A.W., Toonen, L.W., Sloetjes, A.W., van den
Berg, W.B., Figdor, C.G., Radstake, T.R., and Adema, G.J. (2005). Elevated
Ryner, L., Guan, Y., Firestein, R., Xiao, Y., Choi, Y., Rabe, C., Lu, S., Fuentes,
CXCL16 expression by synovial macrophages recruits memory T cells into
E., Huw, L.Y., Lackner, M.R., et al. (2015). Upregulation of periostin and reac-
rheumatoid joints. Arthritis Rheum. 52, 1381–1391.
tive stroma is associated with primary chemoresistance and predicts clinical
outcomes in epithelial ovarian cancer. Clin. Cancer Res. 21, 2941–2951. Vega, B., Munoz, L.M., Holgado, B.L., Lucas, P., Rodriguez-Frade, J.M., Calle,
A., Rodriguez-Fernandez, J.L., Lechuga, L.M., Rodriguez, J.F., Gutierrez-
Sade-Feldman, M., Yizhak, K., Bjorgaard, S.L., Ray, J.P., de Boer, C.G.,
Gallego, R., and Mellado, M. (2011). Technical advance: surface plasmon
Jenkins, R.W., Lieb, D.J., Chen, J.H., Frederick, D.T., Barzily-Rokni, M.,
resonance-based analysis of CXCL12 binding using immobilized lentiviral par-
et al. (2018). Defining T cell states associated with response to checkpoint
ticles. J. Leukoc. Biol. 90, 399–408.
immunotherapy in melanoma. Cell 175, 998–1013.e1020.
Sahai, E., Astsaturov, I., Cukierman, E., DeNardo, D.G., Egeblad, M., Evans, Vento-Tormo, R., Efremova, M., Botting, R.A., Turco, M.Y., Vento-Tormo, M.,
R.M., Fearon, D., Greten, F.R., Hingorani, S.R., Hunter, T., et al. (2020). A Meyer, K.B., Park, J.E., Stephenson, E., Polanski, K., Goncalves, A., et al.
framework for advancing our understanding of cancer-associated fibroblasts. (2018). Single-cell reconstruction of the early maternal-fetal interface in hu-
Nat. Rev. Cancer 20, 174–186. mans. Nature 563, 347–353.

Salmon, H., Franciszkiewicz, K., Damotte, D., Dieu-Nosjean, M.C., Validire, P., Villani, A.C., Satija, R., Reynolds, G., Sarkizova, S., Shekhar, K., Fletcher, J.,
Trautmann, A., Mami-Chouaib, F., and Donnadieu, E. (2012). Matrix architec- Griesbeck, M., Butler, A., Zheng, S., Lazo, S., et al. (2017). Single-cell RNA-
ture defines the preferential localization and migration of T cells into the stroma seq reveals new types of human blood dendritic cells, monocytes, and pro-
of human lung tumors. J. Clin. Invest. 122, 899–910. genitors. Science 356, eaah4573.

Savas, P., Virassamy, B., Ye, C., Salim, A., Mintoff, C.P., Caramia, F., Salgado, Wilbanks, A., Zondlo, S.C., Murphy, K., Mak, S., Soler, D., Langdon, P.,
R., Byrne, D.J., Teo, Z.L., Dushyanthen, S., et al. (2018). Single-cell profiling of Andrew, D.P., Wu, L., and Briskin, M. (2001). Expression cloning of the
breast cancer T cells reveals a tissue-resident memory subset associated with STRL33/BONZO/TYMSTRligand reveals elements of CC, CXC, and CX3C
improved prognosis. Nat. Med. 24, 986–993. chemokines. J. Immunol. 166, 5145–5154.

Sharma, P., Hu-Lieskovan, S., Wargo, J.A., and Ribas, A. (2017). Primary, Wu, T.D., Madireddi, S., de Almeida, P.E., Banchereau, R., Chen, Y.J., Chitre,
adaptive, and acquired resistance to cancer immunotherapy. Cell 168, A.S., Chiang, E.Y., Iftikhar, H., O’Gorman, W.E., Au-Yeung, A., et al. (2020).
707–723. Peripheral T cell expansion predicts tumour infiltration and clinical response.
Nature 579, 274–278.
Shih, A.J., Menzin, A., Whyte, J., Lovecchio, J., Liew, A., Khalili, H., Bhuiya, T.,
Gregersen, P.K., and Lee, A.T. (2018). Identification of grade and origin spe- Yost, K.E., Satpathy, A.T., Wells, D.K., Qi, Y., Wang, C., Kageyama, R.,
cific cell populations in serous epithelial ovarian cancer by single cell RNA- McNamara, K.L., Granja, J.M., Sarin, K.Y., Brown, R.A., et al. (2019). Clonal
seq. PLoS One 13, e0206785. replacement of tumor-specific T cells following PD-1 blockade. Nat. Med.
Simoni, Y., Becht, E., Fehlings, M., Loh, C.Y., Koo, S.L., Teng, K.W.W., Yeong, 25, 1251–1259.
J.P.S., Nahar, R., Zhang, T., Kared, H., et al. (2018). Bystander CD8(+) T cells Zeltz, C., Primac, I., Erusappan, P., Alam, J., Noel, A., and Gullberg, D. (2020).
are abundant and phenotypically distinct in human tumour infiltrates. Nature Cancer-associated fibroblasts in desmoplastic tumors: emerging role of integ-
557, 575–579. rins. Semin. Cancer Biol. 62, 166–181.
Skelly, D.A., Squiers, G.T., McLellan, M.A., Bolisetty, M.T., Robson, P., Zhang, L., Conejo-Garcia, J.R., Katsaros, D., Gimotty, P.A., Massobrio, M.,
Rosenthal, N.A., and Pinto, A.R. (2018). Single-cell transcriptional profiling re- Regnani, G., Makrigiannakis, A., Gray, H., Schlienger, K., Liebman, M.N.,

Cancer Cell 39, 928–944, July 12, 2021 943

 ll
Article
et al. (2003). Intratumoral T cells, recurrence, and survival in epithelial ovarian 5-plex fluorescent immunohistochemistry with tyramide signal amplification
cancer. N. Engl. J. Med. 348, 203–213. and same species antibodies. Lab. Invest. 97, 873–885.
Zhang, L., Yu, X., Zheng, L., Zhang, Y., Li, Y., Fang, Q., Gao, R., Kang, B., Zheng, C., Zheng, L., Yoo, J.K., Guo, H., Zhang, Y., Guo, X., Kang, B., Hu, R.,
Zhang, Q., Huang, J.Y., et al. (2018). Lineage tracking reveals dynamic rela- Huang, J.Y., Zhang, Q., et al. (2017). Landscape of infiltrating T cells in liver
tionships of T cells in colorectal cancer. Nature 564, 268–272. cancer revealed by single-cell sequencing. Cell 169, 1342–1356 e1316.
Zhang, Q., He, Y., Luo, N., Patel, S.J., Han, Y., Gao, R., Modak, M., Carotta, S., Zilionis, R., Engblom, C., Pfirschke, C., Savova, V., Zemmour, D., Saatcioglu,
Haslinger, C., Kind, D., et al. (2019). Landscape and dynamics of single im- H.D., Krishnan, I., Maroni, G., Meyerovitz, C.V., Kerwin, C.M., et al. (2019).
mune cells in hepatocellular carcinoma. Cell 179, 829–845.e820. Single-cell transcriptomics of human and mouse lung cancers reveals
Zhang, W., Hubbard, A., Jones, T., Racolta, A., Bhaumik, S., Cummins, N., conserved myeloid populations across individuals and species. Immunity
Zhang, L., Garsha, K., Ventura, F., Lefever, M.R., et al. (2017). Fully automated 50, 1317–1334.e1310.

944 Cancer Cell 39, 928–944, July 12, 2021

 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mouse monoclonal anti-CD8 alpha Abcam Cat#ab17147; RRID: AB_443686
Isotype control IgG Abcam Cat#ab91353; RRID: AB_2811128
Rabbit monoclonal anti-CD3 Abcam Cat#ab135372; RRID: AB_2884903
Rat monoclonal anti-Granzyme B ThermoFisher Cat#14-8889-80; RRID: AB_2572909
Mouse monoclonal pan-cytokeratin Roche Cat#760-2595
Mouse monoclonal anti-PD1 Abcam Cat#ab52587; RRID: AB_881954
Rabbit monoclonal anti-PD-L1 Roche Cat#790-4905; RRID: AB_2819099
Secondary antibody GaRt-HRP Roche Cat#760-4457
Secondary antibody GaMs-HRP Roche Cat#760-7060
Secondary antibody GaRb-HRP Roche Cat#760-7058
Mouse monoclonal anti-CD45 APC/Cy7 BioLegend Cat#304014; RRID: AB_314402
Mouse monoclonal anti-EpCAM APC BioLegend Cat#324208; RRID: AB_756082
Biological samples
Ovarian cancer tissue collection (n = 17) Cureline, Inc N/A
(Brisbane, CA, US)
Ovarian cancer samples KIYATEC N/A
Chemicals, peptides, and recombinant proteins
Formalin Fisher Scientific Cat#SF98-4
Parafin Sigma Aldrich P3683-1KG
Tris-EDTA buffer, pH 9.0 Abcam ab93684
40 ,6-diamidino-2-phenylindole (DAPI) Roche Tissue Diagnostics Cat#760-4196
ProLongTM Diamond Antifade ThermoFisher Scientific Cat#P36962
Mountant with DAPI
DMSO Sigma Aldrich Cat#D2650-5X10ML
dual-fluorescence AO/PI Nexelcom Cat#CS2-0106-5mL
7-AAD BD Biosciences Cat#559925
Calcein Blue ThermoFisher Scientific Cat#C1429
RLT buffer Qiagen 79,216
Critical commercial assays
RNAscope 2.5 LS Duplex Reagent Kit ACD Cat#322440
RNAscope 2.5 LS Green Accessory Pack ACD Cat#322550
TruSeq Stranded mRNA kit Illumina Cat#20020595
RNA extraction kit Qiagen Cat#74136
NextSeq500 High output Illumina Cat#20024907
kit v2 or v2.5 (150 cycles)
Chromium Single Cell 30 10X Genomics Cat#120237
Reagent Kit (v2 Chemistry)
Agilent Bioanalyzer High Sensitivity Agilent Cat#5067-4627
KAPA library quantification universal kit Roche Cat#07960140001
HRP/DAB IHC Detection kit Abcam Cat#ab80436
DISCOVERY Rhodamine 6G kit Ventana Cat#760-244
DISCOVERY FAM kit Ventana Cat#760-243
DISCOVERY Cy5 kit Ventana Cat#760-238
DISCOVERY DCC kit Ventana Cat#760-240
(Continued on next page)

Cancer Cell 39, 928–944.e1–e6, July 12, 2021 e1

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
DISCOVERY Red 610 kit Ventana Cat#760-245
Deposited data
Single cell RNA sequencing and bulk RNA European Genome- EGA Archive: EGAS00001004935
sequencing data phenome Archive
RNA-seq from ICON7 collection Desbois et al.2020 EGA Archive: EGAD00001004988
RNA-seq from ROSiA collection Oza et al., 2017 Available upon reasonable request
RNA-seq from TCGA-OV collection Cancer Genome Atlas https://portal.gdc.cancer.gov/
Research, 2011
Software and algorithms
FlowJo BD Biosciences N/A
Prism GraphPad N/A
edgeR package Bioconductor https://bioconductor.org/
packages/release/bioc/html/edgeR.html
multiGSEA package Bioconductor https://bioconductor.org/
packages/release/bioc/html/multiGSEA.html
CellRanger v3.0.2 10x Genomics http://10xgenomics.com
seurat v3.0.0 GitHub https://github.com/satijalab/
seurat/releases/tag/v3.0.0
scran v1.10.2 Bioconductor https://bioconductor.org/
packages/release/bioc/html/scran.html
scanpy v1.4.3 GitHub https://github.com/theislab/scanpy
fgsea package Bioconductor http://bioconductor.org/packages/fgsea
ProgRes CapturePro Jenoptik https://www.jenoptik.com/products/cameras-and-
imaging-modules/microscope-cameras/software-
solutions/image-software-progres-capture-pro
Inform Akoya Biosciences https://www.akoyabio.com/
phenoptics/software/inform-tissue-finder/
HALO image analysis Indica Labs https://www.indicalab.com/halo

RESOURCE AVAILABILITY

Lead contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Yulei
Wang ([email protected]).

Materials availability
This study did not generate new unique reagents.

Data and code availability

The accession number for the scRNA-seq and bulk RNA-seq reported in this paper is European Genome-phenome Archive:
EGAS00001004935. Three ovarian cancer clinical datasets were used in this study to validate our scRNA-seq findings: i) ICON7
collection (n = 351) (Perren et al., 2011), clinical trial registration: NCT00483782; RNA-seq data accession number:
EGAC00001001188; ii) ROSiA collection (n = 308) (Oza et al., 2017), clinical trial registration: NCT01239732; the study is ongoing
and the RNA-seq data have not been published yet, therefore the data are available upon reasonable request; and iii) TCGA collection
(n = 412) (Cancer Genome Atlas Research, 2011); RNA-seq data are available on the TCGA portal (https://portal.gdc.cancer.gov/).
Study protocols were compliant with good clinical practice guidelines and the Declaration of Helsinki. Ethics approval was obtained
in all participating countries and where required in all participating centers. All patients provided written informed consent.

EXPERIMENTAL MODEL AND SUBJECT

Clinical samples
KIYATEC sample collection. After providing written informed consent, patients >18 years of age with suspected or known ovarian
cancer were enrolled onto an Institutional Review Board (IRB) approved biology protocol by Prisma Health, formerly known as

e2 Cancer Cell 39, 928–944.e1–e6, July 12, 2021

 ll
Article

Greenville Health System, Cancer Institute (IRB-Committee C). Tissue acquisition was carried out in accordance with the guidelines
and regulations specified by Prisma Health and informed consent was obtained from all participants. Fresh tissue was collected at
surgical debulking or laparoscopic biopsy of primary tumor sites.
In situ validation collection. An independent ovarian cancer tissue collection (n = 17) was procured from Cureline, Inc (Brisbane, CA,
US) for in situ validation. Procured samples also had an appropriate Institutional Review Board (IRB) approval.

METHOD DETAILS

Bulk RNA sequencing

Sample preparation, libraries and sequencing
Fresh tissue samples were mechanically dissociated with RLT buffer (#79216, Qiagen), followed by RNA extraction (#74136, Qiagen).
Libraries were generated using TruSeq (#20020595, Illumina) following the manufacturer’s instructions, pooled and sequenced on an
Illumina NextSeq500 with the High output kit v2 (#20024907, Illumina).
Data processing and immune phenotype prediction
Raw data processing of the KIYATEC, ICON7, ROSiA and TCGA datasets was performed as described previously (Desbois et al.,
2020). In brief, raw counts were filtered for lowly expressed genes for which the counts per million (CPM) were smaller than 0.25
in at least 10% of samples. CPM was calculated with the cpm function in the edgeR package. Based on the raw counts of the filtered
gene expression matrix, size factors were calculated using CalcNormFactors (edgeR package) and used for subsequent voom-
limma differential expression analysis. The immune phenotype of each sample was predicted using our previously built gene expres-
sion-based classifier applied to housekeeping gene normalized data as described previously (Desbois et al., 2020).
Data deconvolution and signature scoring
To estimate the proportion of CD8+ GZMB and CD8+ GZMK positive cells, the expression level of GZMB or GZMK over the CD8+
CIBERSORT signature score was calculated for each sample (Newman et al., 2015). In detail, the log and voom transformed data
was used to calculate a z-score of the CIBERSORT_LM22_T_cells_CD8 gene set using scoreSingleSamples from the multiGSEA
package. GZMB and GZMK were excluded from the gene set. Then, the CD8+ signature score was subtracted from the log and
voom transformed GZMB or GZMK expression.

Single-cell RNA sequencing

Cell preparation
Within 24 hr of surgery, ovarian tissues were received and immediately enzymatically dissociated into single cells which were cryo-
preserved in media containing 10% DMSO with the exception of exc5 for which the tissue was frozen before dissociation. This pro-
tocol has been validated by KIYATEC. Frozen tumor cell suspensions were thawed and diluted in FACS Stain buffer (1X PBS pH 7.4,
0.2% BSA, 0.09% NaAzide). Cells were counted on Cellometer Auto 2000 with the dual-fluorescence AO/PI (#CS2-0106-5mL,
Nexelcom). Then, the cells were incubated for 30 min with FcR blocking reagent (#130-059-901, Miltenyi) followed by antibodies
for another 30 min on ice. Immediately prior to sorting on FACSaria Fusion, cells were stained with live and dead markers 7-AAD
(1/16, #559925, BD Biosciences) and Calcein Blue (1/500, #C1429, ThermoFisher Scientific). Doublets were excluded and viable
cells identified based on low 7-AAD and high Calcein blue. Antibodies used for sorting cells into a tumor, immune and stromal
compartment per ovarian tissue were anti-CD45-APC-Cy7 (1/100, #304014, BioLegend) and anti-EpCAM-APC (1/20, #324208, Bio-
Legend). After sorting, samples were immediately spun and resuspended in PBS at 600-1,000 cells/uL according to the cell count
provided by the cell sorter. To prepare Mater Mix + cell suspension, we refer to the Cell Suspension Volume Calculator Table in Chro-
mium Single Cell 30 Reagent Kits User Guide (v2 Chemistry) (10x Genomics, California, USA) to add the appropriate volume of
nuclease-free water and corresponding volume of single cell suspension (targeting cell recovery 5000-6000 cells) to Master Mix
for a total of 100 mL in each tube. 90 mL of the above mixture were loaded in Chromium Chip B, subsequently gel beads and other
reagents loaded in the chip according to the protocol (Gel Bead & Multiplex Kit V2 and Chip Kit (#PN-120237, 10x Genomics)). After
running Chromium Controller for Gel Bead-In Emulsions (GEMs) generation and cell barcoding, GEMs were transferred to thermal
cycler for GEMs reverse transcription incubation, followed by post GEMs–RT Cleanup, cDNA Amplification, QC and quantification.
Preparation of scRNA-seq libraries
For each patient, we generated one library per compartment (tumor, immune and stromal) resulting in three libraries per patient, with
the exception of four desert tumors for which immune and stromal cells were not separated and two libraries per compartment (tumor
and immune/stromal) were generated. 3ʹ Gene Expression Library Construction using the Chromium Single Cell 30 Library (v2 chem-
istry) was performed according to manufacturer’s instructions (http://support.10xgenomics.com/single-cell-gene-expression/
library-prep/doc/user-guide-chromium-single-cell-3-reagent-kits-user-guide-v2-chemistry). To reduce technical batch effects, we
randomized the generation of libraries by both compartment and immune phenotype. Post library construction QC was done by Agi-
lent Bioanalyzer High Sensitivity chip (#5067-4627, Agilent Technologies, Santa Clara, California, USA) and libraries were quantified
by KAPA library quantification universal kit (#07960140001, Roche).
Sequencing and raw data processing
All libraries per patient were pooled and sequenced on an Illumina NextSeq500 with the High output kit v2 or v2.5 (150 cycles).
We confirmed that both kit v2 and v2.5 generated similar results and no technical batch effects were detected comparing the
sequencing results from both kits on the same library. Reads were mapped to the human genome (GRCh38) using CellRanger

Cancer Cell 39, 928–944.e1–e6, July 12, 2021 e3

 ll
Article

v3.0.2 (http://10xgenomics.com). First, the CellRanger mkfastq command with the cellranger sample sheet was used to demultiplex
the base call files for each flow cell into fastq files. Second, the cellranger count command was called to generate single cell feature
counts for each library by specifying the library name in the argument. The filtered feature barcode matrix was used for further data
analysis.
Analysis of scRNA sequencing data
The core scRNA-seq analysis was performed using seurat v3.0.0. First, each library was converted into a seurat object using read10x
and makeseuratobject. To perform filtering of the compartment annotations and to separate the desert stromal from desert immune
cells that were sequenced in a pooled library, the data of all 44 libraries were merged (Figures S2A–S2E). The data was filtered to
include genes that were expressed in at least 10 cells and cells that expressed at least 200 genes, not more than 6000 genes and
less than 5% of mitochondrial transcripts. These cutoffs were determined through QC inspection of each library. Subsequently,
the data was log normalized, variable genes detected through a mean-variance inspection, scaled and the principal components
computed. The top principal components were identified using an elbow plot and used for the UMAP dimensionality reduction. Clus-
ter identification was done at a resolution that would best separate the mixed stroma-tumor interface based on the FACS stromal and
tumor annotation. To identify the immune, stromal, tumor and normal epithelial clusters, mean expression of the following gene
markers for each cluster was calculated: 1) immune compartment: PTPRC, CD79A, 2) stromal compartment: VWF, PECAM1,
COL1A1, COL3A1, DCN, 3) tumor compartment: EPCAM, 4) normal epithelial cells: PIFO, CAPS, TMEM190, SNTN. The cluster iden-
tity was determined at a mean expression cutoff greater than 0.8, which resulted in most concordant annotation with the FACS anno-
tation. To note, stromal and immune cells in desert tumors were pooled for the cell sorting due to a low number of these cells. Hence,
immune and stromal identities in desert tumors were assigned computationally. Cells that clustered with compartments other than
their FACS annotated compartment and a cluster of normal epithelial cells (cluster 34) were removed from further analysis.
The raw data of each filtered and newly annotated compartment was processed separately for further downstream analysis as
described above. The number of principal components for the dimensionality reduction was determined for each compartment indi-
vidually. The major cell types in the stromal and immune compartments were defined through per cluster mean expression of the
following gene markers: 1) fibroblasts: DCN, C1R, PDGFRA, OGN, 2) endothelial cells: PECAM1, 3) pericytes: RGS5, 4) B-TILs:
MS4A1, 5) plasma cells: SDC1, 6) T cells: CD2, CD3E, 7) myeloid cells: CD14, CSF1R, LILRA4. For further analysis of the fibroblast,
T cell and myeloid cell populations the data was subset to these cells and processed starting from the raw data as described above.
The analysis of each individual patient’s T cell and myeloid cell population was performed by splitting the subset population by
patient and processing the raw data as described above. Patient data with less than 50 myeloid or T cells were excluded from
the single patient analysis. Positive cluster gene markers were identified using the seurat FindAllMarkers function and the Wilcoxon
test. For the analysis of the fibroblast phenotypes, desert 4 fibroblasts were excluded since they are clustering separately from all
other tumors.
Gene expression values plotted in UMAPs, heatmaps, violin plots, boxplots are scran normalized expression values calculated
based on raw expression values (scran v1.10.2 (Lun et al., 2016)).
Batch effect correction
Since the analysis of the fibroblast, T cell and myeloid populations revealed patient-driven clustering, we corrected for the patient
effects (herein batch effects) using the batch-balanced k-nearest neighbor correction (bbknn) method (Polanski et al., 2020). To
this end, we imported the python modules scanpy v1.4.3 including anndata v0.6.21, numpy v1.17.0, bbknn v1.3.9 into R using retic-
ulate v1.12. In brief, bbknn was applied to the top principal components as computed by seurat and determined by the elbow plot,
clusters were identified using scanpy and all results transferred back to seurat. To validate that bbknn only corrected for technical
differences and not for biological differences, the resulting clusters and their markers were manually compared to the results of a
single library cluster analysis (Figures S4A and S4B).
Diffusion pseudotime analysis
The diffusion pseudotime analysis was performed through the scanpy diffusionmap function on the bbknn-corrected anndata object
and transferred back to seurat.
Identification of subpopulations
The fibroblast cluster identities were determined by calculating gene signature scores using seurat AddModuleScore of previously
identified fibroblast phenotypes: iCAF and myCAF from (Elyada et al., 2019) and IL1-driven and TGFB-driven CAF from (Dominguez
et al., 2020).
Established myeloid and T cell type clusters were annotated using the cluster mean expression of the following gene markers: 1)
dendritic cells: CD1C, CLEC10A, CSF2RA, CCL19, CCR7, 2) plasmacytoid dendritic cells: LILRA4, 3) proliferative cells: MKI67, 4)
Tgd and NK cells: TRDC, NCAM1, 5) CD8 T cells: CD8A, CD8B, 6) CD4 T cells: CD4, CD40LG. A cluster of cells with high expression
of proliferative genes (e.g. MKI67, PCNA and BIRC) was observed in every cell type analyzed. These proliferative cell clusters typically
represented a mixture of different subpopulations which cannot be separated due to the dominant cell cycle gene expression pro-
gram, and we removed these uninformative clusters from all downstream analyses. Cutoffs for the annotation by mean expression
were determined by manual inspection of the clusters and the gene expression distributions. The subpopulation gene markers of the
T cell, myeloid cell and fibroblast populations as plotted on the gene marker heatmaps and reported in the supplementary tables were
identified by testing for significant differential expression in a subpopulation against all other cells using a Wilcoxon test.

e4 Cancer Cell 39, 928–944.e1–e6, July 12, 2021

 ll
Article

Pseudo-bulk differential expression

We used a pseudo-bulk approach to perform differential gene expression (DGE) analysis. For each sample, raw UMI counts for each
gene were summed across cells of a cell population of interest derived from that sample, resulting in sample-level UMI counts. Sam-
ples with fewer than five cells and genes with less than 50 reads across samples were excluded from the analysis. We then calculated
the size factors for each pseudo-bulk sample using calcNormFactors (edgeR) and used voom-limma to perform differential gene
expression analysis on these sample-level pseudo-bulk expression profiles. For the G2/M corrected pseudo-bulk expression anal-
ysis in the tumor compartment, we calculated a G2/M score per cell using the CellCycleScoring function with G2/M cell cycle genes
as previously described (Seurat, https://science.sciencemag.org/content/352/6282/189). The per cell scores were averaged by pa-
tient and added to the voom-limma design matrix as a covariate.
Gene set enrichment analysis
Gene set enrichment analysis was performed on the results of the differential gene expression analysis using the fgsea package (Kor-
otkevich et al. bioRxiv 060012; doi: https://doi.org/10.1101/060012) and the hallmark gene set from the molecular signatures data-
base collection using the msigdbr package (Subramanian et al., 2005). In detail, the differentially expressed gene list was ranked ac-
cording to the combined log fold change and adjusted p value and used as an input for the gene set enrichment analysis.
Chemokine receptor-ligand interaction analysis
A database of known chemokine receptor and ligand pairs was curated using the combined information from cellphoneDB (Efremova
et al., 2020) (https://www.cellphonedb.org/) and resources from the R&D systems website (https://www.rndsystems.com/resources/
technical-information/chemokine-nomenclature). This database was filtered for possible chemokine-receptor interactions using the
following criteria: 1) each receptor-ligand pair should be expressed in at least 10% of cells of a cell population, 2) each pair should be
expressed within the same immune phenotype. This resulted in six possible chemokine receptor-ligand interactions, for which the
expression levels and the number of cells expressing a receptor/ligand was assessed in each individual patient.

In situ experimental procedures

CD8 immunohistochemistry
Fresh tissue samples were fixed in formalin and embedded in paraffin and 10 mm sections were mounted onto glass slides. Rehy-
dration and antigen retrieval were performed using Tris-EDTA buffer, pH 9.0 (#ab93684, Abcam, Cambridge, MA). Slides were
stained with anti-CD8 [144B] (#ab17147, Abcam, Cambridge, MA) or IgG1 [B11/6] Isotype control (#ab91353, Abcam, Cambridge,
MA) both at a dilution of 1:50 for 2 hr at room temperature. Staining was detected using Mouse and Rabbit Specific HRP/DAB IHC
Detection kit (#ab80436, Abcam, Cambridge, MA). Images were taken on an Olympus IX70 microscope with a Jenoptik (Jena, Ger-
many) ProgRes C14plus camera and ProgRes CapturePro software.
Multiplex immunofluorescence (IF)
4mM tissue sections from 17 ovarian cancer samples procured from Cureline, Inc were subject to multiplex IF assays performed on a
VENTANA BenchMark Ultra automated staining instrument (Ventana Medical Systems). The detailed description of epitope retrieval
from FFPE tissue sections, antibody titration, incubation and image acquisition were previously described (Zhang et al., 2017). In
brief, for each target, the corresponding 1 antibody (1 Ab) (human CD3 (#ab135372, Abcam), GZMB (#14-8889-80, ThermoFisher
Scientific), pan-cytokeratin (#760-2595, Roche), PD-1 (#ab52587, Abcam) or PD-L1 (#790-4905, Roche)) was incubated on the slide,
followed with a horseradish peroxidase (HRP) conjugated 2 Ab (GaRt-HRP (# 760-4457) for GZMB, GaMs-HRP (# 760-7060) for PD1
and pan-cytokeratin, and GaRb-HRP (# 760-7058) for CD3 and PDL1); the target was then detected with a tyramide-conjugated flu-
orophore (TSA-FL): Discovery Red 610 kit (#760-245, Ventana), Discovery Rhodamine 6G kit (#760-244, Ventana), Discovery FAM kit
(#760-243, Ventana), Discovery Cy5 kit (#760-238, Ventana), Discovery DCC kit (#760-240, Ventana). The next target detection
followed the same scheme, and so on. To prevent potential cross-reaction of same species 1 antibodies, a heating step was intro-
duced to deactivate the 1 Ab & 2 Ab complex before detecting the next target. Slides were then counterstained with 40 ,6-diamidino-
2-phenylindole (DAPI) (Roche Tissue Diagnostics, Cat# 760-4196). Slides were cover slipped using micro cover glass, 24 3 50mm
no. 1.5 (VWR, Cat#: 48393241) and a ProLongTM Diamond Antifade Mountant with DAPI (ThermoFisher Scientific, Cat#: P36962).
The fluorescent image acquisition was performed in a ZEISS Axio Scan.Z1 (Oberkochen, Germany). Image analysis on counting
of cells with either uniquely stained or concurrently stained markers within regions of interest (ROI), i.e. tumor, panCK+ epitumor
or stroma regions was performed on scanned images as previously described (Racolta, A. et al. Abstract SITC, 2019; https://doi.
org/10.1186/s40425-019-0763-1). These ROIs were computed by the DP algorithms based on pathologists’ manual annotation of
tumor lesion and invasive front of the tumor on the images. Necrotic areas were manually excluded. For each whole slide, individual
fields of view (FOV) are tiled and processed. Digital pathology (DP) algorithm is used to identify phenotypes/regions of interest that are
detected by the markers. In detail, the algorithm includes the following steps: (1) Preprocessing: preprocessing was applied to re-
move a variety of fluorescence artifacts in FOVs. (2) Cell detection: the radial symmetry algorithm was used to detect and vote for
the center of the cells. (3) Feature extraction: morphology, appearance, intensity, gradient, and direction features were extracted.
(4) Cell classification: different machine learning classifiers such as support vector machine, random forest, and logistic regression
algorithms were used. Accuracy of each classification was subsequently assessed. A classifier with the best accuracy was used to
classify the cells. (5) Epitumor area and stroma region segmentation: a method combining region growing and adaptive thresholding
was used to segment epitumor and stroma area. After identifying the phenotypes/regions of interest, the DP algorithm reports sta-
tistical metrics that characterize the density of objects and their spatial interrelationships in automatically computed ROIs. Different
categories of readout analysis were reported: (1) ROI areas; (2) counts of phenotypes within different ROIs; (3) counts of cells with

Cancer Cell 39, 928–944.e1–e6, July 12, 2021 e5

 ll
Article

specific characteristics; and (4) counts of phenotypes at different distances from ROIs. Two samples were excluded from the down-
stream analysis due to lack of any triple positive cells.
RNA situ hybridization (ISH)
RNA in situ hybridization assays for the dual detection of CD8A and GZMB or CD8A and GZMK in 5mm FFPE ovarian tumor tissue
sections were performed using the RNAscope 2.5 LS Duplex Reagent Kit (ACD, Cat # 322440) with the RNAscope 2.5 LS Green
Accessory Pack (ACD, Cat # 322550) on the BOND RX automated stainer (Leica Biosystems) according to the manufacturer’s in-
structions (Advanced Cell Diagnostics, a Bio-Techne brand, Newark, CA). Tissue RNA quality was assessed using positive control
probes Hs-PPIB (ACD, Cat # 313908) for human cyclophilin B (PPIB) and Hs-POLR2A (ACD, Cat # 310458-C2) for human RNA po-
lymerase subunit IIA (POLR2A) and negative control probe dapB (ACD, Cat # 320758) for bacterial dihydrodipicolinate reductase
(dapB). Only samples demonstrating acceptable RNA quality as defined by the presence of an average of R4 dots per cell with
the positive control probe staining and an average of <1 dot per 10 cells with the negative control probe staining were further analyzed
with the target probes for CD8A (ACD, Cat # 560393, NM_001768.6, 971-2342 nt), GZMK (ACD, Cat # 475903-C2, NM_002104.2, 25-
1025 nt) and GZMB (ACD, Cat # 445973-C2, NM_004131.4, 3-912 nt). FFPE HeLa cell control slides were tested in parallel for
POLR2A and dapB as run controls along with the ovarian cancer FFPE tissue slides. The resulting slides were scanned with a 3DHis-
tech Pannoramic SCAN II digital slide scanner (Thermo Fisher Scientific) using a 403 objective. Scanned images were analyzed for
CD8A, GZMB and GZMK staining in the tumor and tumor associated stroma regions using the HALO image analysis software (Indica
Labs). RNAscope signals were counted and binned into 5 categories based on the number of dots per cell (bin 0 = 0 dot/cell, bin 1 = 1–
3 dots/cell, bin 2 = 4–9 dots/cell, bin 3 = 10–15 dots/cell, and bin 4 = >15 dots/cell with >10% of dots in clusters). A composite H-score
was calculated to combine the signal level and the percentage of cells in each bin as follows: H-Score = (0 x % cells in bin 0) + (1 x %
cells in bin 1) + (2 x % cells in bin 2) + (3 x % cells in bin 3) + (4 x % cells in bin 4). The H-scores ranged from 0 to 400. H-scores for tumor
and stroma regions were scored separately.

QUANTIFICATION AND STATISTICAL ANALYSIS

Survival analysis
Analysis of the association of CD8+ GZMK and GZMB T cells with progression free survival in the ICON7 chemotherapy arm was
conducted using the survival package.

Statistical analysis
Statistical analysis was performed in R. The statistical methods used for each analysis are described within the figure legends. Sig-
nificance was defined as a p value smaller than 0.05. All boxplots report the 25% (lower hinge), 50%, and 75% quantiles (upper
hinge). The lower whiskers indicate the smallest observation greater than or equal to lower hinge – 1.5 * interquartile range, the upper
whiskers indicate the largest observation less than or equal to upper hinge +1.5 * interquartile range as default in the geom_boxplot()
function. Error bars in bar plots represent the standard deviation. All violin plots report the 25% (lower hinge), 50%, and 75% quantiles
(upper hinge) and the kernel density estimates as computed by geom_density as the width.

e6 Cancer Cell 39, 928–944.e1–e6, July 12, 2021