Skip to main content
PMC home page
eLife logoLink to eLife
. 2018 Sep 4;7:e36967. doi: 10.7554/eLife.36967

Topography of cancer-associated immune cells in human solid tumors

Jakob Nikolas Kather 1,2,3,4,✉, Meggy Suarez-Carmona 1,3, Pornpimol Charoentong 1,3, Cleo-Aron Weis 5, Daniela Hirsch 5, Peter Bankhead 6, Marcel Horning 1, Dyke Ferber 1,3, Ivan Kel 1, Esther Herpel 7,8, Sarah Schott 9, Inka Zörnig 1,3, Jochen Utikal 10,11, Alexander Marx 5, Timo Gaiser 5, Herrmann Brenner 2,12,13, Jenny Chang-Claude 14,15, Michael Hoffmeister 12, Dirk Jäger 1,2,3, Niels Halama 1,2,3,✉
Editors: Jeffrey Settleman16, Ian Tannock17
PMCID: PMC6133554  PMID: 30179157

Abstract

Lymphoid and myeloid cells are abundant in the tumor microenvironment, can be quantified by immunohistochemistry and shape the disease course of human solid tumors. Yet, there is no comprehensive understanding of spatial immune infiltration patterns (‘topography’) across cancer entities and across various immune cell types. In this study, we systematically measure the topography of multiple immune cell types in 965 histological tissue slides from N = 177 patients in a pan-cancer cohort. We provide a definition of inflamed (‘hot’), non-inflamed (‘cold’) and immune excluded patterns and investigate how these patterns differ between immune cell types and between cancer types. In an independent cohort of N = 287 colorectal cancer patients, we show that hot, cold and excluded topographies for effector lymphocytes (CD8) and tumor-associated macrophages (CD163) alone are not prognostic, but that a bivariate classification system can stratify patients. Our study adds evidence to consider immune topographies as biomarkers for patients with solid tumors.

Research organism: Human

Introduction

Malignant tumors growing in an immunocompetent host elicit an immune response, evident by the presence of various inflammatory/immune cell in tumor tissue (Shalapour and Karin, 2015; Mantovani et al., 2008; Bindea et al., 2013). In order to grow to a clinically relevant size, tumor cells develop specific escape mechanisms against the immune system by manipulating inflammatory cells for their benefit (de Visser et al., 2006; Dunn et al., 2002; Fridman et al., 2013). One of the key strategies is that tumor cells interfere with immune signaling, hijacking immunosuppressive cells and thereby shaping the immune infiltrate, which allows for tumor cell proliferation (Chen and Mellman, 2013; Chen and Mellman, 2017).

These mechanisms have been in the focus of oncology for several years (Kather et al., 2018a). Currently a number of immunotherapeutic drugs are available which interfere with immune cells in the tumor microenvironment in order to facilitate tumor control (Becht et al., 2016a; Galluzzi et al., 2014). However, the complex nature of immune infiltrates impairs the development of more targeted approaches. Specifically, tailored combination treatments are widely proposed as a way to more effective cancer therapy (Sharma and Allison, 2015a; Sharma and Allison, 2015b; Zitvogel et al., 2011). Systematically deciphering tumor-immune phenotypes is key to a better understanding and more effective tailoring of immunotherapies (Greenplate et al., 2016).

Analysis of solid tumor tissue slides by immunohistochemistry (IHC) is the gold standard to assess tumor immune infiltrate because it allows for exact quantification of type, density and localization of immune cells (Fridman et al., 2017; Becht et al., 2016b). For more than a decade, digital pathology has been the method of choice to reliably and reproducibly analyze large cohorts of patient samples and can provide potential biomarkers for immunotherapy (Becht et al., 2016a; Kather et al., 2016; Gurcan et al., 2009). Immune cell quantification in digitized tissue has been used to identify robust and clinically relevant biomarkers in numerous cancer entities, for example in colorectal cancer (CRC) primary tumors and liver metastases (Galon et al., 2006; Halama et al., 2011; Mlecnik et al., 2016). Histological analysis of tumor-infiltrating lymphoid cells has been proven to be a reliable and prognostically relevant marker (Galon et al., 2014; Denkert et al., 2016; International TILs Working Group 2014 et al., 2015). Antitumor immunity arises in a complex ecosystem of various cell types that closely interact with one another, such as effector lymphocytes (Li et al., 2016), macrophages (Halama et al., 2016; Biswas and Mantovani, 2010), dendritic cells (Gardner and Ruffell, 2016), granulocytes (Coffelt et al., 2016), innate lymphoid cells (Crome et al., 2017), regulatory T cells (Nishikawa and Sakaguchi, 2010), natural killer cells (Crome et al., 2013; Barrow and Colonna, 2017), myeloid-derived suppressor cells (Talmadge and Gabrilovich, 2013) and other cell types (Kather et al., 2017). In tumor tissue, T lymphocytes (T cells) and macrophages are among the most abundant immune cells and are closely related to clinical outcome (Fridman et al., 2017; Kather et al., 2017; Fridman et al., 2012; Kather et al., 2018b).

After years of detailed IHC analysis of solid tumor slides, a paradigm for the classification of tumor-immune phenotypes has emerged and three classes of tumors are generally assumed: ‘cold’ tumors (or ‘immune desert’, showing no immune cell infiltration), ‘immune-excluded’ tumors (with immune cells aggregating at the tumor boundaries) and ‘hot’ tumors (or ‘inflamed’ tumors, showing pronounced immune infiltrates in the tumor core (Chen and Mellman, 2017; Lanitis et al., 2017; Joyce and Fearon, 2015). However, although this classification is generally accepted, it is backed by surprisingly little quantitative data. Fundamental biological questions regarding this concept are essentially unanswered, such as: Do these topographies exist in all tumor entities? Is there a difference between immunotherapy-sensitive and immunotherapy-insensitive tumor types? Does the concept of cold/excluded/hot apply to lymphoid and myeloid cells, or only to one of them? Do all tumors use the same strategies for immune escape?

In the last years, several large studies have systematically investigated immune-tumor phenotypes in detail. However, most of these studies were not suitable to distinguish cold, excluded and hot tumors because they did not look at the tumor core and the invasive margin at the same time. Two recent studies of immunophenotypes in colorectal cancer (CRC) have shown that the average immune cell density is higher around the tumor than in the tumor core (Bindea et al., 2013; Mlecnik et al., 2016). Yet, the concept of cold/excluded/hot tumors was not investigated in these studies. Also, other comprehensive studies have not taken spatial patterns of immune cell phenotypes into account (Becht et al., 2016). Recent large-scale studies have looked at high-dimensional phenotypes of tumor-infiltrating immune cells, but have not specifically addressed different topographies (Newell and Davis, 2014; Wong et al., 2016; Newell and Becht, 2018; Kather et al., 2018c). Another previous study investigated spatial patterns of lymphoid and myeloid cells in human solid tumors – however, only tissue microarrays (TMA) were included in that study, precluding any possible differentiation between invasive margin and tumor core (Tsujikawa et al., 2017). Lastly, the correlation of the tumor immunophenotype to its transcription profile was investigated previously, but was lacking a spatially resolved approach (Charoentong et al., 2017).

In the present study, we have attempted to close this knowledge gap by a systematic analysis of a large cohort of various human tumors in a spatially resolved way. In particular, this included a systematic analysis of the following immune cells: CD3+ T-lymphocytes, CD8+ T-lymphocytes, PD1+ T lymphocytes, FOXP3+ T lymphocytes (which we assume to be largely ‘regulatory’, although the role of FOXP3 is more complex (Wang et al., 2007) and CD68+ and CD163+ monocytes/macrophages.

Results

Immune cell densities in major cancer entities

We measured immune cell densities (number of cells per mm², Figure 1A–B) in 965 immunostained histological tissue slides (listed in Supplementary file 3) in the pan-cancer cohort in three spatial compartments per slide (Figure 1C): outer invasive margin (0 – 500 µm outside the tumor invasion front), inner invasive margin (500 µm to the inside) and in the tumor core. In accordance with previous studies, colorectal adenocarcinoma primary tumors and liver metastases had a higher cell count in the outer invasive margin than in the tumor core (Figure 2—figure supplement 2 and Figure 2—figure supplement 3). Melanoma samples had higher cell counts in the inner tumor than outside of the tumor for all analyzed cell types. The other tumor types showed less clear-cut patterns, highlighting the need for a more detailed analysis.

Figure 1. Semiautomatic image analysis defines immune cell topography.

(A) Manual delineation of three compartments: outer 500 µm invasive margin, inner 500 µm invasive margin, tumor core. (B) Example of automatic cell detection in a CD3-stained gastric carcinoma slide. Left: original image, right: after cell detection and classification. (C) Cell counts in all three compartments can be used to create a ‘target plot’ (visualization resembling a shooting target) where the color of each compartment corresponds to the percentile-normalized cell density. Here, two examples of CD3-stained gastric carcinoma tissue slides are shown. The upper sample has an immune-excluded phenotype while the lower sample has an inflamed phenotype. Unit on the color scale: percentile-normalized cell density. Scale bar in B is 100 µm, scale bars in C are 1 mm.

Figure 1.

Figure 1—figure supplement 1. Example images for cell count in all six immunostains.

Figure 1—figure supplement 1.

For all slides with sufficient sample availability, we performed immunostaining for Foxp3, CD68, PD1, CD8, CD163 and CD3. QuPath was used for automatic cell detection and classification. Here, example images are shown for each staining type (20x magnification) along with the result of the cell segmentation.
Open in a new tab

Unsupervised clustering of lymphocyte densities separates hot and cold tumors

As previous studies have assumed the existence of ‘hot’, ‘cold’ and ‘immune excluded’ tumors with regard to lymphocyte infiltration, we assessed our data set for evidence of this clustering. We used cell densities for CD3+ and CD8+ cells in all three spatial compartments and used multiple methods to determine the optimal number of clusters. For CD3 and CD8 separately as well as for both together, the majority of the optimization runs converged on two clusters (Figure 2—figure supplement 4), mostly representing ‘hot’ and ‘cold’ tumors. There was no strong tendency to converge on three clusters, which means that there is no strong inherent grouping into ‘hot’, ‘cold’ and ‘immune excluded’ tumors.

Conceptual definition of hot, cold and immune excluded tumors

We asked whether there is a rationale to define ‘immune excluded tumors’ as a separate group although it does not naturally emerge in unsupervised clustering methods. With three spatial compartments, each of them having either high or low cell density, there are in theory eight possible topographies. We asked whether this number could be reduced and analyzed the statistical correlation between the spatial compartments for each type of staining in all tumor types. We found that in general, there was a high correlation between ‘tumor core’ and ‘inner invasive margin’, but not between either compartment and the ‘outer invasive margin’ (Figure 2—figure supplement 5). ‘Tumor core’ and ‘inner invasive margin’ can be collapsed into one compartment because cell counts in these compartments are highly correlated. This leaves only four possible topographies that are by equivalent to three previously postulated phenotypes (Chen and Mellman, 2017): high density outside of the tumor with a low density inside the tumor can be described as ‘immune excluded’. Low density inside and outside is ‘cold’ and high density inside the tumor is ‘hot’ regardless of cell density outside of the tumor (Figure 2A–F).

Figure 2. Cell densities in the tumor core and in the outer invasive margin in the pan-cancer cohort.

Raw cell densities are plotted for each cell type and both major compartments. Gray lines indicate the median density for this cell type. Split at the median, tumors can be classified as cold, hot or immune excluded for all immune cell types. However, the scatter plot shows that for all immune cell types, there is no natural clustering into these phenotypes – the phenotypes blend into each other.

Figure 2.

Figure 2—figure supplement 1. Replication experiment for the full tissue analysis pipeline.

Figure 2—figure supplement 1.

60 tissue specimens were randomly selected and stained for CD3 or CD163, scanned and analyzed independently and by blinded observers. The resulting cell counts were similar to the original analysis. In this scatter plot, each data point corresponds to one spatial compartment in one patient. There are two spatial compartments per patient. Pearson’s correlation coefficient is 0.74.
Figure 2—figure supplement 2. Normalized immune cell counts for cytotoxic T lymphocytes (CD8) and pro-tumor macrophages (CD163).

Figure 2—figure supplement 2.

For all three compartments outer invasive margin [out], inner invasive margin [in] and tumor core [core], normalized cell density is shown. In accordance with previous reports, lymphocytes and macrophages are on average more abundant around the tumor in colorectal cancer primary tumors (COAD-PRI) and liver metastases (COAD-MET), the opposite was seen in melanoma (MEL). However, these average values might obscure underlying patterns that are independent of tumor type, highlighting the need for more detailed analysis of spatial immunophenotypes.
Figure 2—figure supplement 3. Average cell density percentile score for all compartments in ten tumor entities and six immunostains.

Figure 2—figure supplement 3.

For all tumor types and all immunostains, the average target plot is shown based on the mean value of all tumor samples in each group.
Figure 2—figure supplement 4. Optimal number of lymphocyte topography clusters arising in repeated optimization runs with different methods.

Figure 2—figure supplement 4.

Using lymphocyte densities in three regions in tumors from 168 cancer patients, we looked for the optimal number of clusters (1 to 12) to group the topography phenotypes. We used CD3+ and CD8+ cell densities separately and together. We used three clustering methods with three loss functions each. The vast majority of clustering runs converged on two clusters which mostly corresponded to high and low overall cell densities.
Figure 2—figure supplement 5. Correlations between cell densities in different spatial compartments.

Figure 2—figure supplement 5.

Pearson’s correlation coefficient is shown for all pairwise comparisons between cell densities in three spatial compartments in tumors from N = 144 cancer patients. Cell densities in the ‘inner invasive margin’ (MARG-500-IN) are highly correlated to cell densities in the ‘tumor core’ (TU-CORE) for all six immune cell types. The ‘outer invasive margin’ (MARG-500-OUT) is less well correlated to the other compartments. This provides a rationale to reduce the complexity of the data set and use only ‘MARG-500-OUT’ and ‘TU-CORE’ in further steps of the analysis.
Open in a new tab

Prevalence of immune topographies in different tumor types

Having operationalized these definitions of immune topographies, we next asked how they are distributed in different tumor types in the pan-cancer cohort. As a cutoff value for high versus low cell density we used the median cell density for each cell type (median number of cells per mm² in any tumor type in any compartment, listed in Supplementary file 4). Based on these definitions, different tumor types showed very different distributions of immune topographies (Figure 3A–F). Most strikingly, tumor types that are to some degree sensitive to approved cancer immunotherapies (such as melanoma (MEL), lung adeno (LUAD), lung squamous (LUSC) and head and neck squamous (HNSC) had a large proportion (approximately half) of CD3-hot, CD8-hot and PD1-hot tumors. In accordance with previous studies (Halama et al., 2011), colorectal primary (COAD-PRI) and colorectal metastatic (COAD-MET) had a very high proportion of CD3-excluded tumors (Figure 3A). Differences between tumor types were most pronounced for regulatory T-cells (Foxp3+ cells, Figure 3D) with more than half of lung adeno (LUAD), head and neck squamous (HNSC), stomach adenocarcinoma (STAD) and esophageal (ESCA) cancer samples having Foxp3-hot topographies. Also, while close to half of all analyzed COAD-PRI samples were Foxp3-hot, the vast majority of all COAD-MET samples were Foxp3-cold (Figure 3A).

Figure 3. Distribution of immune topography phenotypes among different tumor types in the pan-cancer cohort.

Figure 3.

Open in a new tab

Analysis for all six immune cell types (A–F) and for all analyzed tumor types (MEL = melanoma, LUAD = lung adeno, LUSC = lung squamous, BLCA = bladder, HNSC = head and neck squamous, STAD = stomach adeno, ESCA = esophageal squamous, COAD-PRI = colorectal primary, COAD-MET = colorectal liver metastasis, OV = ovarian). These data comprise all N = 965 tissue slides from N = 177 patients. MEL through HNSC are to some degree sensitive to approved immunotherapies and predominantly have ‘hot’ phenotypes for most immune cells. However, among these tumor types, different phenotypes for immunosuppressive immune cells (Foxp3+ regulatory T-cells [Tregs]) and CD163+ macrophages prevail.

Bivariate classification of immune topographies

Subsequently, we asked whether the topography for a given immune cell type co-occurs with the same topography for other immune cell types – or whether tumors can be ‘cold’ for one immune cell type and ‘hot’ or ‘excluded’ for another cell type at the same time. This question is important because in the clinic, immune topographies are often used as stratifying biomarkers for immunotherapy trials (Kather et al., 2018c) but it is unclear how many spatial compartments and which histological markers should be looked at.

Indeed, we saw that tumors often had different topographies for a pair of immune cell markers (Figure 4). Especially, regarding lymphocytic and myeloid markers, there was a discordance in a substantial number of cases that were lymphoid-hot and myeloid-cold (Figure 4). We further stratified this by different tumor entities (Figure 5A–E) and found pronounced differences between the tumor types. CD3 and CD8 topographies were mostly concordant (Figure 5A) as can be expected because CD8+ lymphocytes are a subset of CD3+ lymphocytes. Regarding CD8 vs. Tregs (Figure 5B), there was a notable discordance. For example, most colorectal primary tumors (COAD-PRI) that were Foxp3-excluded were CD8-cold. Analyzing tumor-associated macrophages (TAMs) labeled with CD68 (Figure 5D), we found that several melanomas (MEL) that were CD68-hot were CD8-cold which was rarely observed in other tumor types. Colorectal cancer liver metastases (COAD-MET) were mostly CD8-excluded, corresponding to a CD68-excluded or CD68-hot phenotype. This was similarly observed for CD163+ macrophages (Figure 5E). These findings suggest that tumors of different immunophenotypes require different ways of immune escape and that these mechanisms can only be distinguished when considering a two-dimensional myeloid-lymphoid classification system.

Figure 4. Pairwise analysis of immune phenotypes for all immune cell types in the pan-cancer cohort.

Figure 4.

Open in a new tab

For all tissue samples in all tumor types, a pairwise classification into cold-excluded-hot was done for all immune cell types. This analysis was based on the median cutoff for high and low cell densities. Absolute numbers of tumor are given, black marks the most abundant and white the least abundant group. For some pairwise comparisons such as CD3 and CD8 cells (top left corner), there is a high concordance between the phenotypes. For other comparisons such as Foxp3 and CD163 cells, phenotypes are mostly discordant.

Figure 5. Bivariate immune phenotypes for each tumor type in the pan-cancer cohort.

Figure 5.

Open in a new tab

We analyzed the concordance between hot-cold-excluded topographies for CD8+ lymphocytes and all other cell types for each tumor type separately. Absolute numbers of patients assigned to each of nine phenotypes are overlaid on the heat map. For some cell types and some tumor types, immune topographies are concordant. This suggests that in these settings, a detailed analysis of multiple immune cell types in biomarker studies is not necessary. On the other hand, some cell types in other tumor entities (such as CD8+ lymphocytes and CD68+ macrophages in panel (D) show a high discordance in multiple tumor types. This suggests that measuring one of these cell types only may not be informative enough for biomarker studies.

Pan-cancer similarity based on spatial immune phenotype

Using a full panel of immune cell markers (CD3, CD8, PD1, Foxp3, CD68 and CD163), we asked how similar different tumor types were in terms of immune cell spatial layout. For all 144 tumors that were stained for this full panel (Figure 6A–G, six immunostains, three compartments, 144 patients, after percentile normalization for each cell type), we used unsupervised hierarchical clustering to define similarity. We found that ovarian cancer (OV) was an outlier (Figure 6H) due to its low infiltration with almost all immune cell subsets (Figure 2—figure supplement 3). Highly immunogenic tumor types such as melanoma (MEL), stomach cancer (STAD) and non-small cell lung cancer (NSCLC: LUAD and LUSC) were close in hierarchical clustering.

Figure 6. Overall similarity between tumor entities based on full immune topography.

Figure 6.

Open in a new tab

Hierarchical clustering based on all normalized cell densities of (A) PD1+exhausted lymphocytes; (B) Foxp3+regulatory T cells; (C) CD8+cytotoxic T lymphocytes; (D) CD68+monocytes/macrophages; (E) CD3+lymphocytes and (F) CD163+pro tumor macrophages. Unit on the color scale: percentile-normalized cell density. (G) corresponding H & E image of the colorectal cancer sample used in this figure. (H) Hierarchical clustering of tumor types (N = 144 total patient samples).

Clinical utility of immune topography

Having shown that hot, cold and excluded immune patterns exist, and that these patterns vary between tumor types and immune cell types, we investigated the clinical utility of this classification system in an independent patient cohort (DACHS cohort). We analyzed the topography of CD8+ lymphocytes and CD163+ macrophages in colorectal cancer (CRC) primary tumors because these cell types have previously been linked to prognosis (Fridman et al., 2012) and showed discordant topographies in the pan-cancer cohort (COAD_PRI in Figure 5E).

We derived immune topographies for CD8 and CD163 from cell counts in the outer invasive margin and the tumor core for N = 287 patients using the median cell densities from the pan-cancer cohort as cutoff values (Figure 7A and B). As in the pan-cancer cohort, most patients were cold or immune excluded for these two cell types. The bivariate analysis showed that ‘CD8-cold, CD163-cold’ was the most prevalent category (Figure 7D).

Figure 7. Prognostic value of the myeloid-lymphoid topography in primary colorectal cancer (CRC) in the DACHS cohort.

Figure 7.

Open in a new tab

In a validation cohort of N = 287 colorectal cancer patients (N = 286 with follow-up data) from the DACHS study, CD8 and CD163 staining of the primary surgical sample was correlated to clinical outcome (overall survival) using the cutoffs from the pan-cancer cohort. (A) As in the CRC subgroup in the pan-cancer cohort, CD8-cold was the most prevalent phenotype, followed by CD8-excluded. (B) A similar distribution of phenotypes was observed for CD163+ macrophages. (C) In the bivariate analysis for CD8 and CD163, most patients had cold and excluded phenotypes for both antigens. (D) For each of these nine phenotypes, the hazard ratio (HR) for death of any cause (inverse overall survival, OS) was derived from a multivariable Cox proportional hazard model (covariates UICC stage, age and sex). Bold HR indicates statistically significant findings with p<0.05. Bivariate analysis of both antigens is essential for risk stratification as CD8 and CD163 phenotypes show non-trivial interaction. Raw data for panel D are shown in Supplementary file 5. (E) The Kaplan-Meier plots for the reference groups (CD8-cold CD163 cold) and the two significant groups (‘CD8-excl., CD163 excl’ and ‘CD8 excl., CD163 hot’) show significant differences in overall survival. These differences are not captured by univariate, but only by this bivariate stratification system. Log rank p-value=0.019.

We then used multivariable Cox proportional hazard models (including tumor stage, age and sex as potential confounders) to analyze the association between cell counts and cell topographies with survival. Cell densities of CD8 and CD163 in the tumor core and the outer invasive margin were not significantly correlated to overall survival (hazard ratios [HR] for death from any cause were 1.00 for both cell types in both compartments, Supplementary file 5). However, bivariate immune topographies showed significant association to overall survival: With ‘CD8-cold, CD163-cold’ as a reference cohort, the HR was 1.75 for ‘CD8-excluded, CD163 excluded’ (p=0.041) and the HR was 2.71 for ‘CD8-excluded, CD163-hot’ (p=0.025, Supplementary file 6 and Figure 7D).

Discussion

The tumor microenvironment is a highly complex, heterocellular ecosystem (Bindea et al., 2013; Tape, 2016). Multiple immune cell types are involved in this system and ultimately shift it towards a tumor-promoting or tumor-rejecting environment (Chen and Mellman, 2017; Fridman et al., 2017). Also, these immune cells determine whether a tumor will be responsive to immunotherapy (Becht et al., 2016a). However, immunotherapy outcomes vary pronouncedly between different tumor entities and also between different patients within a given entity. The biological basis for these differences has been the subject of various studies but is still not entirely clear. Most importantly, we currently lack a comprehensive classification system for multiple effector parts of the immune system.

In the present study, we have performed a large-scale systematic analysis of lymphoid and myeloid phenotypes of human solid tumors. We provide evidence for a classification of tumor-immune phenotypes into hot, cold and immune excluded spatial patterns. These patterns can be detected not only in lymphocyte infiltrates (lymphoid classification, as previously assumed by most studies) but also in macrophage infiltrates (myeloid classification). Interestingly, lymphoid and myeloid patterns are not always identical and a two-dimensional classification is needed to accommodate all possible lymphoid-myeloid phenotypes.

Addressing the biological differences and the different responses to immunotherapy across tumor entities, we systematically compared different tumor types. We analyzed ten tumor entities in the framework of this classification and showed pronounced differences, but also unexpected cross-entity similarities: We found that lymphoid-hot tumors are more prevalent in immunotherapy-responsive tumor entities such as melanoma and lung adenocarcinoma than in other entities such as colorectal cancer. Immune-excluded tumors, such as lymphoid-excluded and myeloid-excluded tumors are common in head and neck tumors (HNSC). Lastly, we show that characteristic patterns of immune evasion (via T cell exhaustion or infiltration by Tregs) are not exclusive to specific tumor-immune phenotypes but can be detected across the defined categories.

As part of this study, we present a clinical validation of the proposed classification system in the context of colorectal cancer (COAD). It has been previously shown that high density of CD8+ cells in COAD is associated with long survival (Galon et al., 2006) while high density of CD163+ cells is associated with poor survival (Fridman et al., 2012). However, some large studies have shown only a very modest impact of CD8+cell density on overall survival (Glaire et al., 2018). Likewise, in the DACHS cohort which we analyzed as part of this study, CD8+ cell density alone and CD163+ cell density alone were not clearly associated with differences in overall survival. Also, for both cell types, the immune topographies (cold, excluded, hot) were not associated with such differences. However, classifying patients according to bivariate (CD8 combined with CD163) immune topographies identified statistically significant associations to survival in a multivariable statistical model (Figure 7D). This was particularly clear for CD8-excluded tumors: in this case, CD163-cold, excluded and hot tumors had a significantly (log rank p-value 0.019) different prognosis (Figure 7E). These findings suggest that a multivariate analysis of spatial distributions of multiple immune cell types may be superior to merely quantifying a particular type of immune cells in COAD and possibly other human solid tumors.

Our study adds a systematic approach to a hitherto subjective classification of tumor-immune phenotypes. This new classification could constitute a novel framework to investigate immunotherapy responsiveness in clinical trials. Also, this classification sheds light on common immunological aspects by describing a shared immune topography across human solid tumors.

Materials and methods

Key resources table.

Reagent type (species) or resource Designation Source/Reference Identifier Additional information
Software,
algorithm		 QuPath v0.1.2 Bankheadet al. DOI: 10.1038/s41598-017-17204-5 -
Antibody Anti-human
CD3		 Leica
Novocastra		 RRID:AB_563544 Dilution 1:100
Antibody Anti-human
CD8		 Leica
Novocastra		 RRID:AB_442068 Dilution 1:50
Antibody Anti-human
Foxp3		 eBioscience RRID:AB_467555 Dilution 1:100
Antibody Anti-human
CD163		 BioRad RRID:AB_2074540 Dilution 1:500
Antibody Anti-human
CD68		 Thermo Fisher
Scientific		 RRID:AB_720547 Dilution 1:2000
Antibody Anti-human
PD1		 Abcam RRID:AB_881954 Dilution 1:50
Open in a new tab

Ethics statement and tissue samples

All experiments were conducted in accordance with the Declaration of Helsinki, the International Ethical Guidelines for Biomedical Research Involving Human Subjects (CIOMS), the Belmont Report and the U.S. Common Rule. Anonymized archival tissue samples were retrieved from the tissue bank of the National Center for Tumor diseases (NCT, Heidelberg, Germany) in accordance with the regulations of the tissue bank and the approval of the ethics committee of Heidelberg University (tissue bank decision numbers 2152 and 2154, granted to NH and JNK, ovarian cancer tissues granted to SS; informed consent was obtained from all patients as part of the NCT tissue bank protocol, ethics board approval S-207/2005, renewed on 20 Dec 2017). Another set of tissue samples was provided by the pathology archive at UMM (University Medical Center Mannheim, Heidelberg University, Mannheim, Germany) after approval by the institutional ethics board (Ethics Board II at University Medical Center Mannheim, decision number 2017 – 806R-MA, granted to AM and waiving the need for informed consent for this retrospective and fully anonymized analysis of archival samples). Also, a set of melanoma samples was provided by the pathology archive at UMM after approval by the institutional ethics board (Ethics Board II at University Medical Center Mannheim, decision number 2014 – 835R-MA, granted to JU and waiving the need for informed consent for this retrospective and fully anonymized analysis of archival samples).

We analyzed tissue samples of primary esophageal carcinoma (ESCA), primary gastric cancer (STAD), primary colorectal cancer (COAD-PRI), primary lung adenocarcinoma (LUAD), primary lung squamous cell carcinoma (LUSC), primary head and neck squamous cell carcinoma (HNSC), primary bladder cancer (BLAC), ovarian cancer primary tumors (OV) as well as melanoma primary tumors (MEL) and colorectal cancer liver metastases (COAD-MET).

In addition to this pan-cancer cohort, we acquired a set of N = 287 primary surgical specimen of colorectal adenocarcinoma from the DACHS study (Hoffmeister et al., 2015; Brenner et al., 2011) which were provided by the NCT biobank under the same ethics board approval as stated above and including informed consent by all patients. Clinical data for this cohort are listed in Supplementary file 1.

Immunohistochemistry

We performed histological staining for CD3 (dilution 1:100 with Leica antigen retrieval ER1 solution), CD8 (1:50 with ER1), Foxp3 (1:100 with Leica antigen retrieval ER2 solution), CD163 (1:500 with ER2), CD68 (1:2000 with Leica Fast Enzyme digestion) and PD1 (1:50 with ER1) on a Leica Bond automatic staining device using a hematoxylin-diaminobenzidine (DAB) staining protocol as described previously (Halama et al., 2016). For melanoma, FastRed (Leica #DS9390) was used as a chromogen. Stained whole slide tissue sections were digitized as described previously (Halama et al., 2016). Almost all samples were stained for these six immune cell markers. In cases of insufficient tissue availability, only CD3, CD8 and CD163 staining was performed.

Image analysis

Our image analysis pipeline was composed of several steps: First, manual annotation of three regions of interest (ROI) in each histological whole slide image (Figure 1A). The ROIs were ‘tumor core’ (TU_CORE), ‘inner invasive margin’ (MARG_500_IN) and ‘outer invasive margin’ (MARG_500_OUT). Invasive margins were 500 µm wide. Second, we automatically counted all positively stained cells using the open source software QuPath (Figure 1B) (Bankhead et al., 2017). Intensity thresholds and other parameters for cell detection and classification were set manually for each staining type and were identical for all samples in the pan-cancer cohort. All parameters are listed in Supplementary file 2. All cell detection scripts were manually checked for plausibility in all tumor entities. Examples for cell detection are shown in Figure 1—figure supplement 1. For all further analyses, cell density values were normalized by percentile within each staining type and were visualized as ‘target plots’ (Figure 1C). Staining intensity thresholds were slightly adapted for the DACHS cohort as listed in Supplementary file 2.

Reproducibility

To ensure reproducibility of our digital pathology pipeline, we randomly selected 60 tissue specimen and repeated all analysis steps. We used new tissue slides with a distance between 4 µm and 40 µm from the original slice. We stained 30 of these slides for a macrophage marker (CD163) and 30 slides for a lymphocyte marker (CD3) and a blinded observer delineated ROIs for cell quantification as before. There was a high correlation (Pearson’s correlation coefficient was >0.74, p-value<0.001) between these replicates (Figure 2—figure supplement 1).

Furthermore, two slides that served as negative controls for CD3 and CD163 staining, 231/86516 (<<0.1%) and 1/38311 (<<0.1%) cells were false positive.

Automatic determination of optimal cluster number

Analysis of all six immunostains for 177 patients in our collective yielded 965 sets of cell density counts in three spatial compartments each (2895 data points in total, missing values due to tissue availability or quality). For visualization, cell densities were subjected to percentile normalization (quantile normalization with 100 quantiles) for each staining type, across all tumor entities. For clustering and all other analyses, we used absolute cell density (cell number per mm²).

To determine the optimal number of clusters in this data set, we used three different methods for clustering: a gaussian mixture model, k-means and hierarchical clustering. For 1 up to 12 clusters, we computed three loss functions for each approach, using the Davies-Bouldin (Davies and Bouldin, 1979), the Calinski-Harabasz (CalińskiCalinski and Harabasz, 1974) or the silhouette (Rousseeuw, 1987) criteria for quality of clustering. For all optimization methods, we performed 10 technical replicates.

Implementation and data availability

All image analysis steps were implemented in QuPath (see key resources) and all downstream analyses were implemented in MATLAB (Mathworks, Natick, MA, USA) R2017a. All experiments were run on a standard workstation (Intel i7 Processor, 8 cores, 32 GB RAM, Microsoft Windows 10.1). We release all source codes under an open access license (Kather, 2018; copy archived at https://github.com/elifesciences-publications/immuneTopography). Also, we release all raw data from our experiments (Supplementary file 3). All survival analyses were performed in R version 3.5.1 (R-project.org) using the packages survminer, survival, ggfortify and ggplot2.

Acknowledgements

The authors would like to thank Rosa Eurich and Jana Wolf (National Center for Tumor Diseases, Heidelberg, Germany), Katrin Wolk (University Medical Center Mannheim, Mannheim, Germany) and Nina Wilhelm (NCT Biobank, National Center for Tumor diseases, Heidelberg, Germany) for expert technical assistance. The authors are grateful to the participants of the DACHS study, the cooperating clinics which recruited patients for this study, and the Institute of Pathology, University of Heidelberg, for providing tissue samples for this study.

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Contributor Information

Jakob Nikolas Kather, Email: jakob.kather@nct-heidelberg.de.

Niels Halama, Email: niels.halama@nct-heidelberg.de.

Jeffrey Settleman, Calico Life Sciences, United States.

Ian Tannock, Princess Margaret Cancer Centre, Canada.

Funding Information

This paper was supported by the following grants:

  • Heidelberg School of Oncology to Jakob Nikolas Kather.

  • Bundesministerium für Bildung und Forschung to Alexander Marx.

Additional information

Competing interests

No competing interests declared.

Author contributions

Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft.

Conceptualization, Resources, Formal analysis, Writing—review and editing.

Conceptualization, Formal analysis, Writing—review and editing.

Conceptualization, Resources, Methodology, Writing—review and editing.

Resources, Writing—review and editing.

Conceptualization, Software, Methodology, Writing—review and editing.

Data curation, Formal analysis, Methodology, Writing—review and editing.

Conceptualization, Methodology, Writing—review and editing.

Conceptualization, Validation, Visualization, Methodology, Writing—review and editing.

Resources, Methodology, Project administration, Writing—review and editing.

Conceptualization, Resources, Data curation, Writing—review and editing.

Conceptualization, Data curation, Supervision, Project administration, Writing—review and editing.

Resources, Data curation, Investigation, Project administration, Writing—review and editing.

Conceptualization, Resources, Formal analysis, Funding acquisition, Writing—review and editing.

Conceptualization, Resources, Methodology, Writing—review and editing.

Resources, Data curation, Funding acquisition, Project administration, Writing—review and editing.

Resources, Formal analysis, Funding acquisition, Investigation, Methodology, Writing—review and editing.

Resources, Formal analysis, Supervision, Funding acquisition, Writing—review and editing.

Conceptualization, Supervision, Project administration, Writing—review and editing.

Conceptualization, Supervision, Validation, Writing—review and editing.

Ethics

Human subjects: All experiments were conducted in accordance with the Declaration of Helsinki, the International Ethical Guidelines for Biomedical Research Involving Human Subjects (CIOMS), the Belmont Report and the U.S. Common Rule. Anonymized archival tissue samples were retrieved from the tissue bank of the National Center for Tumor diseases (NCT, Heidelberg, Germany) in accordance with the regulations of the tissue bank and the approval of the ethics committee of Heidelberg University (tissue bank decision numbers 2152 and 2154, granted to NH and JNK, ovarian cancer tissues granted to SS; informed consent was obtained from the patients as part of the NCT tissue bank protocol). Another set of tissue samples was provided by the pathology archive at UMM (University Medical Center Mannheim, Heidelberg University, Mannheim, Germany) after approval by the institutional ethics board (Ethics Board II at University Medical Center Mannheim, decision number 2017-806R-MA, granted to AM and waiving the need for informed consent for this retrospective and fully anonymized analysis of archival samples). Also, a set of melanoma samples was provided by the pathology archive at UMM after approval by the institutional ethics board (Ethics Board II at University Medical Center Mannheim, decision number 2014-835R-MA, granted to JU and waiving the need for informed consent for this retrospective and fully anonymized analysis of archival samples). In addition to this pan-cancer cohort, we acquired a set of N=287 primary surgical specimen of colorectal adenocarcinoma from the DACHS study which were provided by the NCT biobank under the same ethics board approval as stated above and including informed consent by all patients.

Additional files

Supplementary file 1. Clinical characterization of the DACHS cohort.

This table lists summary statistics of all relevant clinico-pathological features of the DACHS cohort.

elife-36967-supp1.docx (12.8KB, docx)
DOI: 10.7554/eLife.36967.015
Supplementary file 2. List of all image analysis parameters.

In this table, all parameters for cell detection and classification using the open source software QuPath are listed. Two sets of parameters are distinguished: ‘DAB’ (diaminobenzidine), used for blue-brown staining, and ‘Red’, used for blue-red staining in melanoma. OD = optical density. All parameters were used in the pan-cancer cohort unless labeled as ‘DACHS’, in which case they were used in the DACHS cohort.

elife-36967-supp2.docx (18.8KB, docx)
DOI: 10.7554/eLife.36967.016
Supplementary file 3. List of all samples and all measurement values of the pan-cancer cohort.

In this table, we report all raw measurements for all samples that were used in this study. Column names are: ‘class’ (tumor type as listed above), ‘patient’ (patient pseudonym), ‘antigen’ (antigen for immunostain), ‘TU_CORE_cells_mm2’ (number of positively stained cells per square millimeter in the tumor core), ‘MARG_500_IN_cells_mm2’ (number of positively stained cells per square millimeter in the inner invasive margin, defined as ranging 0–500 µm to the inside from the tumor edge), ‘MARG_500_OUT_cells_mm2’ (number of positively stained cells per square millimeter in the inner invasive margin, defined as ranging 0–500 µm to the outside from the tumor edge).

elife-36967-supp3.xlsx (49.2KB, xlsx)
DOI: 10.7554/eLife.36967.017
Supplementary file 4. List of all cutoff values for all cell types.

On the full data set of N = 965 tissue slides from N = 177 patients in 10 tumor types, we calculated the median cell density for each antigen, taking the compartments ‘outer invasive margin’ and ‘tumor core’ into account. These median values were subsequently used as cutoff values for low and high cell densities which were then used to define hot, cold and excluded phenotypes.

elife-36967-supp4.docx (12.9KB, docx)
DOI: 10.7554/eLife.36967.018
Supplementary file 5. Continuous cell densities of CD8+ and CD163+ cells are not significantly associated with overall survival in colorectal cancer.

A multivariable Cox proportional hazard model was fitted to all variables listed in this table. N = 286 CRC patients in the DACHS cohort, number of events = 108, significance codes (sig): *<0.05, **<0.01, ***<0.001. HR = hazard ratio, UICC = Union internationale contre le cancer.

elife-36967-supp5.docx (13.7KB, docx)
DOI: 10.7554/eLife.36967.019
Supplementary file 6. Bivariate immune phenotype predicts risk of death of any cause.

A multivariable Cox proportional hazard model was fitted to all variables listed in this table. N = 286 CRC patients in the DACHS cohort, number of events = 108, significance codes (sig): *<0.05, **<0.01, ***<0.001. HR = hazard ratio, UICC = Union internationale contre le cancer.

elife-36967-supp6.docx (15.1KB, docx)
DOI: 10.7554/eLife.36967.020
Transparent reporting form
elife-36967-transrepform.docx (246.2KB, docx)
DOI: 10.7554/eLife.36967.021

Data availability

We release all source codes under an open access license (http://dx.doi.org/10.5281/zenodo.1407435; copy archived at https://github.com/elifesciences-publications/immuneTopography). Also, we release all raw data from our experiments (Supplementary File 3).

References

  1. Bankhead P, Loughrey MB, Fernández JA, Dombrowski Y, McArt DG, Dunne PD, McQuaid S, Gray RT, Murray LJ, Coleman HG, James JA, Salto-Tellez M, Hamilton PW. QuPath: open source software for digital pathology image analysis. Scientific Reports. 2017;7:16878. doi: 10.1038/s41598-017-17204-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Barrow AD, Colonna M. Tailoring natural killer cell immunotherapy to the tumour microenvironment. Seminars in Immunology. 2017;31:30–36. doi: 10.1016/j.smim.2017.09.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Becht E, de Reyniès A, Giraldo NA, Pilati C, Buttard B, Lacroix L, Selves J, Sautès-Fridman C, Laurent-Puig P, Fridman WH. Immune and stromal classification of colorectal cancer is associated with molecular subtypes and relevant for precision immunotherapy. Clinical Cancer Research. 2016;22:4057–4066. doi: 10.1158/1078-0432.CCR-15-2879. [DOI] [PubMed] [Google Scholar]
  4. Becht E, Giraldo NA, Dieu-Nosjean MC, Sautès-Fridman C, Fridman WH. Cancer immune contexture and immunotherapy. Current Opinion in Immunology. 2016a;39:7–13. doi: 10.1016/j.coi.2015.11.009. [DOI] [PubMed] [Google Scholar]
  5. Becht E, Giraldo NA, Germain C, de Reyniès A, Laurent-Puig P, Zucman-Rossi J, Dieu-Nosjean MC, Sautès-Fridman C, Fridman WH. Immune contexture, Immunoscore, and malignant cell molecular subgroups for prognostic and theranostic classifications of cancers. Advances in Immunology. 2016b;130:95–190. doi: 10.1016/bs.ai.2015.12.002. [DOI] [PubMed] [Google Scholar]
  6. Bindea G, Mlecnik B, Tosolini M, Kirilovsky A, Waldner M, Obenauf AC, Angell H, Fredriksen T, Lafontaine L, Berger A, Bruneval P, Fridman WH, Becker C, Pagès F, Speicher MR, Trajanoski Z, Galon J. Spatiotemporal dynamics of intratumoral immune cells reveal the immune landscape in human cancer. Immunity. 2013;39:782–795. doi: 10.1016/j.immuni.2013.10.003. [DOI] [PubMed] [Google Scholar]
  7. Biswas SK, Mantovani A. Macrophage plasticity and interaction with lymphocyte subsets: cancer as a paradigm. Nature Immunology. 2010;11:889–896. doi: 10.1038/ni.1937. [DOI] [PubMed] [Google Scholar]
  8. Brenner H, Chang-Claude J, Seiler CM, Hoffmeister M. Long-term risk of colorectal cancer after negative colonoscopy. Journal of Clinical Oncology. 2011;29:3761–3767. doi: 10.1200/JCO.2011.35.9307. [DOI] [PubMed] [Google Scholar]
  9. Calinski T, Harabasz J. A dendrite method for cluster analysis. Communications in Statistics - Theory and Methods. 1974;3:1–27. doi: 10.1080/03610927408827101. [DOI] [Google Scholar]
  10. Charoentong P, Finotello F, Angelova M, Mayer C, Efremova M, Rieder D, Hackl H, Trajanoski Z. Pan-cancer immunogenomic analyses reveal Genotype-Immunophenotype relationships and predictors of response to checkpoint blockade. Cell Reports. 2017;18:248–262. doi: 10.1016/j.celrep.2016.12.019. [DOI] [PubMed] [Google Scholar]
  11. Chen DS, Mellman I. Oncology meets immunology: the cancer-immunity cycle. Immunity. 2013;39:1–10. doi: 10.1016/j.immuni.2013.07.012. [DOI] [PubMed] [Google Scholar]
  12. Chen DS, Mellman I. Elements of cancer immunity and the cancer-immune set point. Nature. 2017;541:321–330. doi: 10.1038/nature21349. [DOI] [PubMed] [Google Scholar]
  13. Coffelt SB, Wellenstein MD, de Visser KE. Neutrophils in cancer: neutral no more. Nature Reviews Cancer. 2016;16:431–446. doi: 10.1038/nrc.2016.52. [DOI] [PubMed] [Google Scholar]
  14. Crome SQ, Lang PA, Lang KS, Ohashi PS. Natural killer cells regulate diverse T cell responses. Trends in Immunology. 2013;34:342–349. doi: 10.1016/j.it.2013.03.002. [DOI] [PubMed] [Google Scholar]
  15. Crome SQ, Nguyen LT, Lopez-Verges S, Yang SY, Martin B, Yam JY, Johnson DJ, Nie J, Pniak M, Yen PH, Milea A, Sowamber R, Katz SR, Bernardini MQ, Clarke BA, Shaw PA, Lang PA, Berman HK, Pugh TJ, Lanier LL, Ohashi PS. A distinct innate lymphoid cell population regulates tumor-associated T cells. Nature Medicine. 2017;23:368–375. doi: 10.1038/nm.4278. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Davies DL, Bouldin DW. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1979;PAMI-1:224–227. doi: 10.1109/TPAMI.1979.4766909. [DOI] [PubMed] [Google Scholar]
  17. de Visser KE, Eichten A, Coussens LM. Paradoxical roles of the immune system during cancer development. Nature Reviews Cancer. 2006;6:24–37. doi: 10.1038/nrc1782. [DOI] [PubMed] [Google Scholar]
  18. Denkert C, Wienert S, Poterie A, Loibl S, Budczies J, Badve S, Bago-Horvath Z, Bane A, Bedri S, Brock J, Chmielik E, Christgen M, Colpaert C, Demaria S, Van den Eynden G, Floris G, Fox SB, Gao D, Ingold Heppner B, Kim SR, Kos Z, Kreipe HH, Lakhani SR, Penault-Llorca F, Pruneri G, Radosevic-Robin N, Rimm DL, Schnitt SJ, Sinn BV, Sinn P, Sirtaine N, O'Toole SA, Viale G, Van de Vijver K, de Wind R, von Minckwitz G, Klauschen F, Untch M, Fasching PA, Reimer T, Willard-Gallo K, Michiels S, Loi S, Salgado R. Standardized evaluation of tumor-infiltrating lymphocytes in breast cancer: results of the ring studies of the international immuno-oncology biomarker working group. Modern Pathology. 2016;29:1155–1164. doi: 10.1038/modpathol.2016.109. [DOI] [PubMed] [Google Scholar]
  19. Dunn GP, Bruce AT, Ikeda H, Old LJ, Schreiber RD. Cancer immunoediting: from immunosurveillance to tumor escape. Nature Immunology. 2002;3:991–998. doi: 10.1038/ni1102-991. [DOI] [PubMed] [Google Scholar]
  20. Fridman WH, Dieu-Nosjean MC, Pagès F, Cremer I, Damotte D, Sautès-Fridman C, Galon J. The immune microenvironment of human tumors: general significance and clinical impact. Cancer Microenvironment. 2013;6:117–122. doi: 10.1007/s12307-012-0124-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Fridman WH, Pagès F, Sautès-Fridman C, Galon J. The immune contexture in human tumours: impact on clinical outcome. Nature Reviews Cancer. 2012;12:298–306. doi: 10.1038/nrc3245. [DOI] [PubMed] [Google Scholar]
  22. Fridman WH, Zitvogel L, Sautès-Fridman C, Kroemer G. The immune contexture in cancer prognosis and treatment. Nature Reviews Clinical Oncology. 2017;14:717–734. doi: 10.1038/nrclinonc.2017.101. [DOI] [PubMed] [Google Scholar]
  23. Galluzzi L, Vacchelli E, Bravo-San Pedro JM, Buqué A, Senovilla L, Baracco EE, Bloy N, Castoldi F, Abastado JP, Agostinis P, Apte RN, Aranda F, Ayyoub M, Beckhove P, Blay JY, Bracci L, Caignard A, Castelli C, Cavallo F, Celis E, Cerundolo V, Clayton A, Colombo MP, Coussens L, Dhodapkar MV, Eggermont AM, Fearon DT, Fridman WH, Fučíková J, Gabrilovich DI, Galon J, Garg A, Ghiringhelli F, Giaccone G, Gilboa E, Gnjatic S, Hoos A, Hosmalin A, Jäger D, Kalinski P, Kärre K, Kepp O, Kiessling R, Kirkwood JM, Klein E, Knuth A, Lewis CE, Liblau R, Lotze MT, Lugli E, Mach JP, Mattei F, Mavilio D, Melero I, Melief CJ, Mittendorf EA, Moretta L, Odunsi A, Okada H, Palucka AK, Peter ME, Pienta KJ, Porgador A, Prendergast GC, Rabinovich GA, Restifo NP, Rizvi N, Sautès-Fridman C, Schreiber H, Seliger B, Shiku H, Silva-Santos B, Smyth MJ, Speiser DE, Spisek R, Srivastava PK, Talmadge JE, Tartour E, Van Der Burg SH, Van Den Eynde BJ, Vile R, Wagner H, Weber JS, Whiteside TL, Wolchok JD, Zitvogel L, Zou W, Kroemer G. Classification of current anticancer immunotherapies. Oncotarget. 2014;5:12472–12508. doi: 10.18632/oncotarget.2998. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Galon J, Costes A, Sanchez-Cabo F, Kirilovsky A, Mlecnik B, Lagorce-Pagès C, Tosolini M, Camus M, Berger A, Wind P, Zinzindohoué F, Bruneval P, Cugnenc PH, Trajanoski Z, Fridman WH, Pagès F. Type, density, and location of immune cells within human colorectal tumors predict clinical outcome. Science. 2006;313:1960–1964. doi: 10.1126/science.1129139. [DOI] [PubMed] [Google Scholar]
  25. Galon J, Mlecnik B, Bindea G, Angell HK, Berger A, Lagorce C, Lugli A, Zlobec I, Hartmann A, Bifulco C, Nagtegaal ID, Palmqvist R, Masucci GV, Botti G, Tatangelo F, Delrio P, Maio M, Laghi L, Grizzi F, Asslaber M, D'Arrigo C, Vidal-Vanaclocha F, Zavadova E, Chouchane L, Ohashi PS, Hafezi-Bakhtiari S, Wouters BG, Roehrl M, Nguyen L, Kawakami Y, Hazama S, Okuno K, Ogino S, Gibbs P, Waring P, Sato N, Torigoe T, Itoh K, Patel PS, Shukla SN, Wang Y, Kopetz S, Sinicrope FA, Scripcariu V, Ascierto PA, Marincola FM, Fox BA, Pagès F. Towards the introduction of the 'Immunoscore' in the classification of malignant tumours. The Journal of Pathology. 2014;232:199–209. doi: 10.1002/path.4287. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Gardner A, Ruffell B. Dendritic cells and cancer immunity. Trends in Immunology. 2016;37:855–865. doi: 10.1016/j.it.2016.09.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Glaire M, Domingo E, Nicholson G, Novelli M, Lawson K, Oukrif D, Kidal W, Danielsen HE, Kerr R, Kerr DJ, Tomlinson I, Church DN. Tumour-infiltrating CD8 + lymphocytes as a prognostic marker in colorectal cancer: A retrospective, pooled analysis of the QUASAR2 and VICTOR trials. Journal of Clinical Oncology : Official Journal of the American Society of Clinical Oncology. 2018;36:3515. doi: 10.1200/JCO.2018.36.15_suppl.3515. [DOI] [Google Scholar]
  28. Greenplate AR, Johnson DB, Ferrell PB, Irish JM. Systems immune monitoring in cancer therapy. European Journal of Cancer. 2016;61:77–84. doi: 10.1016/j.ejca.2016.03.085. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Gurcan MN, Boucheron LE, Can A, Madabhushi A, Rajpoot NM, Yener B. Histopathological image analysis: a review. IEEE Reviews in Biomedical Engineering. 2009;2:147–171. doi: 10.1109/RBME.2009.2034865. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Halama N, Michel S, Kloor M, Zoernig I, Benner A, Spille A, Pommerencke T, von Knebel DM, Folprecht G, Luber B, Feyen N, Martens UM, Beckhove P, Gnjatic S, Schirmacher P, Herpel E, Weitz J, Grabe N, Jaeger D. Localization and density of immune cells in the invasive margin of human colorectal cancer liver metastases are prognostic for response to chemotherapy. Cancer Research. 2011;71:5670–5677. doi: 10.1158/0008-5472.CAN-11-0268. [DOI] [PubMed] [Google Scholar]
  31. Halama N, Zoernig I, Berthel A, Kahlert C, Klupp F, Suarez-Carmona M, Suetterlin T, Brand K, Krauss J, Lasitschka F, Lerchl T, Luckner-Minden C, Ulrich A, Koch M, Weitz J, Schneider M, Buechler MW, Zitvogel L, Herrmann T, Benner A, Kunz C, Luecke S, Springfeld C, Grabe N, Falk CS, Jaeger D. Tumoral immune cell exploitation in colorectal cancer metastases can be targeted effectively by Anti-CCR5 therapy in cancer patients. Cancer Cell. 2016;29:587–601. doi: 10.1016/j.ccell.2016.03.005. [DOI] [PubMed] [Google Scholar]
  32. Hoffmeister M, Jansen L, Rudolph A, Toth C, Kloor M, Roth W, Bläker H, Chang-Claude J, Brenner H. Statin use and survival after colorectal cancer: the importance of comprehensive confounder adjustment. JNCI: Journal of the National Cancer Institute. 2015;107:djv045. doi: 10.1093/jnci/djv045. [DOI] [PubMed] [Google Scholar]
  33. International TILs Working Group 2014. Salgado R, Denkert C, Demaria S, Sirtaine N, Klauschen F, Pruneri G, Wienert S, Van den Eynden G, Baehner FL, Penault-Llorca F, Perez EA, Thompson EA, Symmans WF, Richardson AL, Brock J, Criscitiello C, Bailey H, Ignatiadis M, Floris G, Sparano J, Kos Z, Nielsen T, Rimm DL, Allison KH, Reis-Filho JS, Loibl S, Sotiriou C, Viale G, Badve S, Adams S, Willard-Gallo K, Loi S. The evaluation of tumor-infiltrating lymphocytes (TILs) in breast cancer: recommendations by an international TILs working group 2014. Annals of Oncology. 2015;26:259–271. doi: 10.1093/annonc/mdu450. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Joyce JA, Fearon DT. T cell exclusion, immune privilege, and the tumor microenvironment. Science. 2015;348:74–80. doi: 10.1126/science.aaa6204. [DOI] [PubMed] [Google Scholar]
  35. Kather JN, Berghoff AS, Ferber D, Suarez-Carmona M, Reyes-Aldasoro CC, Valous NA, Rojas-Moraleda R, Jäger D, Halama N. Large-scale database mining reveals hidden trends and future directions for cancer immunotherapy. OncoImmunology. 2018a;7:e1444412. doi: 10.1080/2162402X.2018.1444412. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Kather JN, Charoentong P, Suarez-Carmona M, Herpel E, Klupp F, Ulrich A, Schneider M, Zoernig I, Luedde T, Jaeger D, Poleszczuk J, Halama N. High-throughput screening of combinatorial immunotherapies with patient-specific in silico models of metastatic colorectal cancer. Cancer Research. 2018b:5155–5163. doi: 10.1158/0008-5472.CAN-18-1126. [DOI] [PubMed] [Google Scholar]
  37. Kather JN, Halama N, Jaeger D. Genomics and emerging biomarkers for immunotherapy of colorectal cancer. Seminars in cancer biology. 2018c doi: 10.1016/j.semcancer.2018.02.010. [DOI] [PubMed] [Google Scholar]
  38. Kather JN, Poleszczuk J, Suarez-Carmona M, Krisam J, Charoentong P, Valous NA, Weis CA, Tavernar L, Leiss F, Herpel E, Klupp F, Ulrich A, Schneider M, Marx A, Jäger D, Halama N. In Silico modeling of immunotherapy and Stroma-Targeting therapies in human colorectal cancer. Cancer Research. 2017;77:6442–6452. doi: 10.1158/0008-5472.CAN-17-2006. [DOI] [PubMed] [Google Scholar]
  39. Kather JN, Weis CA, Bianconi F, Melchers SM, Schad LR, Gaiser T, Marx A, Zöllner FG. Multi-class texture analysis in colorectal cancer histology. Scientific Reports. 2016;6 doi: 10.1038/srep27988. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Kather JN. Immune topography: QuPath and Matlab scripts for histological image analysis. v0.1Zenodo. 2018 doi: 10.5281/zenodo.1407435. [DOI]
  41. Lanitis E, Dangaj D, Irving M, Coukos G. Mechanisms regulating T-cell infiltration and activity in solid tumors. Annals of Oncology. 2017;28:xii18–xii32. doi: 10.1093/annonc/mdx238. [DOI] [PubMed] [Google Scholar]
  42. Li B, Li T, Pignon JC, Wang B, Wang J, Shukla SA, Dou R, Chen Q, Hodi FS, Choueiri TK, Wu C, Hacohen N, Signoretti S, Liu JS, Liu XS. Landscape of tumor-infiltrating T cell repertoire of human cancers. Nature Genetics. 2016;48:725–732. doi: 10.1038/ng.3581. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Mantovani A, Allavena P, Sica A, Balkwill F. Cancer-related inflammation. Nature. 2008;454:436–444. doi: 10.1038/nature07205. [DOI] [PubMed] [Google Scholar]
  44. Mlecnik B, Bindea G, Angell HK, Maby P, Angelova M, Tougeron D, Church SE, Lafontaine L, Fischer M, Fredriksen T, Sasso M, Bilocq AM, Kirilovsky A, Obenauf AC, Hamieh M, Berger A, Bruneval P, Tuech JJ, Sabourin JC, Le Pessot F, Mauillon J, Rafii A, Laurent-Puig P, Speicher MR, Trajanoski Z, Michel P, Sesboüe R, Frebourg T, Pagès F, Valge-Archer V, Latouche JB, Galon J. Integrative analyses of colorectal cancer show immunoscore is a stronger predictor of patient survival than microsatellite instability. Immunity. 2016;44:698–711. doi: 10.1016/j.immuni.2016.02.025. [DOI] [PubMed] [Google Scholar]
  45. Newell EW, Becht E. High-Dimensional profiling of Tumor-Specific immune responses: asking T cells about what they "See" in cancer. Cancer Immunology Research. 2018;6:2–9. doi: 10.1158/2326-6066.CIR-17-0519. [DOI] [PubMed] [Google Scholar]
  46. Newell EW, Davis MM. Beyond model antigens: high-dimensional methods for the analysis of antigen-specific T cells. Nature Biotechnology. 2014;32:149–157. doi: 10.1038/nbt.2783. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Nishikawa H, Sakaguchi S. Regulatory T cells in tumor immunity. International Journal of Cancer. 2010;127:759–767. doi: 10.1002/ijc.25429. [DOI] [PubMed] [Google Scholar]
  48. Rousseeuw PJ. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics. 1987;20:53–65. doi: 10.1016/0377-0427(87)90125-7. [DOI] [Google Scholar]
  49. Shalapour S, Karin M. Immunity, inflammation, and cancer: an eternal fight between good and evil. Journal of Clinical Investigation. 2015;125:3347–3355. doi: 10.1172/JCI80007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Sharma P, Allison JP. The future of immune checkpoint therapy. Science. 2015a;348:56–61. doi: 10.1126/science.aaa8172. [DOI] [PubMed] [Google Scholar]
  51. Sharma P, Allison JP. Immune checkpoint targeting in cancer therapy: toward combination strategies with curative potential. Cell. 2015b;161:205–214. doi: 10.1016/j.cell.2015.03.030. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Talmadge JE, Gabrilovich DI. History of myeloid-derived suppressor cells. Nature reviews. Cancer. 2013;13:739–752. doi: 10.1038/nrc3581. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Tape CJ. Systems biology analysis of heterocellular signaling. Trends in Biotechnology. 2016;34:627–637. doi: 10.1016/j.tibtech.2016.02.016. [DOI] [PubMed] [Google Scholar]
  54. Tsujikawa T, Kumar S, Borkar RN, Azimi V, Thibault G, Chang YH, Balter A, Kawashima R, Choe G, Sauer D, El Rassi E, Clayburgh DR, Kulesz-Martin MF, Lutz ER, Zheng L, Jaffee EM, Leyshock P, Margolin AA, Mori M, Gray JW, Flint PW, Coussens LM. Quantitative multiplex immunohistochemistry reveals Myeloid-Inflamed Tumor-Immune complexity associated with poor prognosis. Cell Reports. 2017;19:203–217. doi: 10.1016/j.celrep.2017.03.037. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Wang J, Ioan-Facsinay A, van der Voort EI, Huizinga TW, Toes RE. Transient expression of FOXP3 in human activated nonregulatory CD4+ T cells. European Journal of Immunology. 2007;37:129–138. doi: 10.1002/eji.200636435. [DOI] [PubMed] [Google Scholar]
  56. Wong MT, Ong DE, Lim FS, Teng KW, McGovern N, Narayanan S, Ho WQ, Cerny D, Tan HK, Anicete R, Tan BK, Lim TK, Chan CY, Cheow PC, Lee SY, Takano A, Tan EH, Tam JK, Tan EY, Chan JK, Fink K, Bertoletti A, Ginhoux F, Curotto de Lafaille MA, Newell EW. A High-Dimensional atlas of human T cell diversity reveals Tissue-Specific trafficking and cytokine signatures. Immunity. 2016;45:442–456. doi: 10.1016/j.immuni.2016.07.007. [DOI] [PubMed] [Google Scholar]
  57. Zitvogel L, Kepp O, Kroemer G. Immune parameters affecting the efficacy of chemotherapeutic regimens. Nature Reviews Clinical Oncology. 2011;8:151–160. doi: 10.1038/nrclinonc.2010.223. [DOI] [PubMed] [Google Scholar]

Decision letter

Editor: Ian Tannock1
Reviewed by: Ian Tannock2, Roberto Salgado3, Evan W Newell4

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Topography of cancer-associated immune cells in human solid tumors" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Ian Tannock as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by a Reviewing Editor and Jeffrey Settleman as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Roberto Salgado (Reviewer #2); Evan W Newell (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We are sending the original reviews, but the main points are:

There are major questions about reproducibility of the method. Your team should show:

(i) That using your current algorithm you have reasonable agreement in the results and classification when duplicate sections from the same patient's tumour are analyzed – a sample to show reproducibility is important.

(ii) The effect of specifying different cluster numbers in your algorithm – say 2 or 4 instead of 3. To what extent are the results determined by the algorithm used?

(iii) You should recognize that some of your presentation is highly technical and needs to be written in a way that will be understood by most readers of eLife.

Other more specific comments requiring your response are given in the individual reviews below.

Reviewer #1:

The authors classify 965 tumour samples (obtained from tumour banks) from 177 patients, and use digital analysis of sections stained for immunohistochemistry (IHC) with 6 immune markers to classify them on the basis of infiltration of immune cells (T-lymphocytes and macrophages). The major contribution is looking at immune cell infiltration within the core, at the periphery (and outside) different regions of tumours. Important missing information is any measure of the reproducibility of their IHC, automated machine learning approach and tumour classification when undertaken on duplicate specimens (adjacent sections). Also, the technical language of the text and figure legends will make it very hard for most readers of eLife to have a good understanding of what is being done and plotted.

The authors comment that tumours that are more likely to respond to immunotherapy (melanoma and lung cancer) are more likely to have immune cell infiltration than tumour types that tend not to respond, such as colorectal tumours, but this is a group correlation, not individual. The real test would of course be to compare the topography of immune cell infiltration with outcome of modern immunotherapy (anti-PD1 or – PDL1) in individual patients.

Reviewer #2:

This is an interesting study, with, as major novelty, the dichotomy between lymphoid and myeloid clusters in human cancer, defining herewith the three major entities "cold", "hot" and "excluded".

A major question arises on the clustering. When you cluster data points you need to tell the algorithm how many clusters. The authors start saying that there are three classifications, using this concept to cluster their data by running the clustering algorithm with k=3. Of course, their data will then end up in 3 clusters. The Authors could have used k=2 and would have ended up with two clusters and said only two types exist. Could have used k=5 and gotten 5 different clusters, etc.

What is missing is evidence that three clusters give a classification system that is the most informative of outcome. PAM50, as an example, ended up with 5 classes/clusters based on trying to group patients in 2,3,4,5,6,7,8 groups and seeing if the groupings made any sense biologically based on expression data from 50 genes. They picked 5 and was reasonable but other subclasses definitely exist.

So, how valid is their statement that only three types of immune-cell patterns exist?

Why wasn't clustering analysis done with k=2,3,4,5 and 3 shown to be optimal?

It may be suggested to compare data with number of clusters 2,3,4,5 and then show outcomes comparing the clusters. To improve the paper in its current form a survival analysis based on these defined classes might be important. Is OS data available for the samples that would allow for survival analysis comparing outcomes for each cluster k=2,3,4,5 and that 3 is most informative of outcome with maximum HR or lowest p-value? If there is no survival data available, they should include that in discussion and reason why the survival analysis wasn't done, stating that in future studies a comparison with survival is important so that other groups can make progress on their findings.

Reviewer #3:

In the study by Kather et al., the authors use single color IHC to comprehensively assess the immune infiltration of 10 different types of tumor tissue (9 cancer types). By analyzing each of their 177 patient samples in three compartments (core, inner and outer invasive margins) with 6 different cellular stains, they set out to assess the prevailing notion that there are three main patterns of lymphocytic invasion of tumors: "cold/immune desert", "immune-excluded" and "hot". As described, this paradigm, while widely accepted and discussed is lacking in quantitative evidence, making this an important and timely study. The approach taken here is clearly described and rational, which makes it a useful example of how standard IHC can be used to objectively assess tumor immune cell infiltration in a high dimensional way. Based on their analysis, the authors conclude that their data are consistent with the prevailing view (3 major clusters of lymphocyte infiltration profiles) and suggest a more complex view that also incorporates the profiles of myeloid cell (CD68) infiltration.

Although I agree with the overall conclusions made, and despite the rigor in data collection and preprocessing, I am a little bit disappointed in the rigor of the interpretation of the data. Based on density measurements of CD3 and CD8 cells, a tSNE plot is constructed to display the range T cell infiltration phenotypes and this is useful. However, based on this and k-means clustering (using k=3), the authors conclude that there are 3 clusters of immune infiltrate profiles. While this is reasonable and roughly fits, based on the tSNE representation, some attempt to rigorously assess a best-fit in terms of cluster numbers, or otherwise, would have been much appreciated. For the analysis of CD68 staining, this is even worse. Only k-means clustering (k=3) was used to conclude that 3 clusters exist. This is exactly what this algorithm will produce if 3 clusters are requested. No other evidence is provided that 3 clusters of profiles should be used.

All in all, I think this is an interesting and important study. However, I think the authors should better assess and/or qualify their statements about the validity of the 3-cluster paradigm. Although I agree that their data support this general notion and their analyses provide interesting comparative insights about these tumor types, alternatives are not formally addressed or disproven.

Comments:

Although important and nicely framed, the hypothesis that 3 clusters of lymphocytic infiltration patterns exists is not adequately tested. What would the data need to look like for the authors to conclude otherwise? Perhaps some sort of best-fit analysis could be used to assess the validity of the 3-cluster claim – for both lymphocytes and myeloid cells. tSNE plots should be shown for myeloid cell data.

The tSNE analysis performed is informative but incomplete. Could these plots be shown for myeloid cell parameters as well? It might also be useful to perform this analysis on the entire dataset (similar to that shown in Figure 5H – e.g., one dot per patient, colored by cancer type or colored by cluster definition from Figure 2). This comment highlights the value of the dataset the authors are providing: it is exciting that many different data analysis/visualization strategies could be used to evaluate these data.

It is not clear why the authors have provided the exaggeration of 10 for tSNE and used a non-default parameter for this. Does this value largely influence the outcome of their analysis? Did the authors also experiment with various perplexity values for their analysis?

For the tSNE analysis of this number of events, the stochasticity and perplexity value might be important factors. The authors should show several replicate tSNE runs to assess the consistency of the pattern shown. To justify which of these runs to shown in the main figure, the authors should choose the run with the best fit values provided by the tSNE algorithm. In other words, I am concerned that the conclusion that the tSNE visualization fits well with k-means (k=3) clustering could be a coincidence for this particular run and perplexity setting and the authors should assess the reproducibility of this analysis.

eLife. 2018 Sep 4;7:e36967. doi: 10.7554/eLife.36967.024

Author response

We are sending the original reviews, but the main points are:

There are major questions about reproducibility of the method. Your team should show:

(i) That using your current algorithm you have reasonable agreement in the results and classification when duplicate sections from the same patient's tumour are analyzed – a sample to show reproducibility is important.

We have repeated the full pipeline (staining, scanning and image analysis) for 60 slides (line 311-316, Figure 2—figure supplement 1). The resulting cell counts were highly (rho>0.74) correlated to the original analysis. Also, we included negative controls (subsection “Reproducibility” demonstrating a near-negligible fraction of falsely positively stained cells.

(ii) The effect of specifying different cluster numbers in your algorithm – say 2 or 4 instead of 3. To what extent are the results determined by the algorithm used?

In the revised manuscript, we have added results from multiple unsupervised methods to determine an optimal number of clusters K. These methods converged on K=3 in ~25% of runs, but mostly converged on K=2 (subsection “Unsupervised clustering of lymphocyte densities separates hot and cold tumors”, Figure 2—figure supplement 4). Thus, we conclude that the reviewer’s concerns are correct: the inherent structure of the data does not provide evidence for the existence of three clusters.

However, after dichotomizing cell counts at the median, we saw that the phenotypes “high cell density”, “low cell density” and “high density outside, low density inside” were equally common (Figure 2, subsection “Conceptual definition of hot, cold and immune excluded tumors”). We investigated whether this simple, qualitative definition of topographies translates into a clinically meaningful biomarker. We stained, annotated and analyzed tumor tissue from 287 colorectal cancer patients and saw that indeed, some topographies predict overall survival (more details below).

Thus, we conclude that the definition of hot/cold/excluded immune topographies is useful although there is no fully unbiased, hypothesis-free definition of these phenotypes.

(iii) You should recognize that some of your presentation is highly technical, and needs to be written in a way that will be understood by most readers of eLife.

We have addressed this issue by revising major parts of the manuscript, making it much more understandable. Also, we have fundamentally revised the figures to present the data more clearly.

Other more specific comments requiring your response are given in the individual reviews below.

Reviewer #1:

The authors classify 965 tumour samples (obtained from tumour banks) from 177 patients, and use digital analysis of sections stained for immunohistochemistry (IHC) with 6 immune markers to classify them on the basis of infiltration of immune cells (T-lymphocytes and macrophages). The major contribution is looking at immune cell infiltration within the core, at the periphery (and outside) different regions of tumours. Important missing information is any measure of the reproducibility of their IHC, automated machine learning approach and tumour classification when undertaken on duplicate specimens (adjacent sections).

We have performed these replication experiments as required (subsection “Image analysis”). Comparing the cell densities to the original results, we saw a very high concordance (Figure 2—figure supplement 1). Also, we have added negative controls (subsection “Reproducibility”).

Also, the technical language of the text and figure legends will make it very hard for most readers of eLife to have a good understanding of what is being done and plotted.

We have addressed this by re-phrasing large parts of the manuscript.

The authors comment that tumours that are more likely to respond to immunotherapy (melanoma and lung cancer) are more likely to have immune cell infiltration than tumour types that tend not to respond, such as colorectal tumours, but this is a group correlation, not individual. The real test would of course be to compare the topography of immune cell infiltration with outcome of modern immunotherapy (anti-PD1 or – PDL1) in individual patients.

The reviewer is right: We hypothesize that immune topographies might be predictive of immunotherapy response, but we show this only on a group level, not on the level of individual patients. Unfortunately, whole tumor specimen that include the tumor core and the invasive margin are usually not available for patients prior to immunotherapy.

However, in the revised manuscript, we have analyzed an additional patient cohort of N=287 colorectal cancer patients (subsection “Pan-cancer similarity based on spatial immune phenotype”). In this cohort, we saw that immune topographies carry prognostic information for overall survival.

Reviewer #2:

This is an interesting study, with, as major novelty, the dichotomy between lymphoid and myeloid clusters in human cancer, defining herewith the three major entities "cold", "hot" and "excluded".

Thank you for this positive response.

A major question arises on the clustering. When you cluster data points you need to tell the algorithm how many clusters. The authors start saying that there are three classifications, using this concept to cluster their data by running the clustering algorithm with k=3. Of course, their data will then end up in 3 clusters. The Authors could have used k=2 and would have ended up with two clusters and said only two types exist. Could have used k=5 and gotten 5 different clusters, etc.

What is missing is evidence that three clusters give a classification system that is the most informative of outcome. PAM50, as an example, ended up with 5 classes/clusters based on trying to group patients in 2,3,4,5,6,7,8 groups and seeing if the groupings made any sense biologically based on expression data from 50 genes. They picked 5 and was reasonable but other subclasses definitely exist.

So, how valid is their statement that only three types of immune-cell patterns exist?

Why wasn't clustering analysis done with k=2,3,4,5 and 3 shown to be optimal?

It may be suggested to compare data with number of clusters 2,3,4,5 and then show outcomes comparing the clusters. To improve the paper in its current form a survival analysis based on these defined classes might be important. Is OS data available for the samples that would allow for survival analysis comparing outcomes for each cluster k=2,3,4,5 and that 3 is most informative of outcome with maximum HR or lowest p-value? If there is no survival data available, they should include that in discussion and reason why the survival analysis wasn't done, stating that in future studies a comparison with survival is important so that other groups can make progress on their findings.

Thank you for this detailed critique. We have done exactly as required and performed an unbiased optimization procedure to find the best number of clusters (Figure 2—figure supplement 4). Indeed, the inherent structure of the data does not provide strong evidence for the existence of three clusters (see above).

However, in the revised manuscript we show that the inner invasive margin and the tumor core are highly correlated. Thus, only two compartments have to be considered: inside and outside of the tumor. With this reduced data set, we show that a pragmatic definition of hot/cold/excluded tumors (subsection “Conceptual definition of hot, cold and immune excluded tumors”) can explain patterns observed in our pan-cancer cohort.

As suggested by the reviewer, we went on and asked whether the classification system into hot/cold/excluded tumors is useful in a clinical setting. To this end, we acquired a large cohort of human colorectal cancer patients, more than doubling the number of patients in our study. We found that indeed, this classification system can be used as a biomarker of potential clinical usefulness (subsection “Pan-cancer similarity based on spatial immune phenotype”).

Reviewer #3:

In the study by Kather et al., the authors use single color IHC to comprehensively assess the immune infiltration of 10 different types of tumor tissue (9 cancer types). By analyzing each of their 177 patient samples in three compartments (core, inner and outer invasive margins) with 6 different cellular stains, they set out to assess the prevailing notion that there are three main patterns of lymphocytic invasion of tumors: "cold/immune desert", "immune-excluded" and "hot". As described, this paradigm, while widely accepted and discussed is lacking in quantitative evidence, making this an important and timely study. The approach taken here is clearly described and rational, which makes it a useful example of how standard IHC can be used to objectively assess tumor immune cell infiltration in a high dimensional way. Based on their analysis, the authors conclude that their data are consistent with the prevailing view (3 major clusters of lymphocyte infiltration profiles) and suggest a more complex view that also incorporates the profiles of myeloid cell (CD68) infiltration.

Thank you for this concise summary.

Although I agree with the overall conclusions made, and despite the rigor in data collection and preprocessing, I am a little bit disappointed in the rigor of the interpretation of the data. Based on density measurements of CD3 and CD8 cells, a tSNE plot is constructed to display the range T cell infiltration phenotypes and this is useful. However, based on this and k-means clustering (using k=3), the authors conclude that there are 3 clusters of immune infiltrate profiles. While this is reasonable and roughly fits, based on the tSNE representation, some attempt to rigorously assess a best-fit in terms of cluster numbers, or otherwise, would have been much appreciated. For the analysis of CD68 staining, this is even worse. Only k-means clustering (k=3) was used to conclude that 3 clusters exist. This is exactly what this algorithm will produce if 3 clusters are requested. No other evidence is provided that 3 clusters of profiles should be used.

This comment is in line with the feedback by the other reviewers. As discussed above, we have thoroughly addressed this point in the revised manuscript, removing unclear visualizations.

All in all, I think this is an interesting and important study. However, I think the authors should better assess and/or qualify their statements about the validity of the 3-cluster paradigm. Although I agree that their data support this general notion and their analyses provide interesting comparative insights about these tumor types, alternatives are not formally addressed or disproven.

Thank you for this positive feedback. As discussed above, we have fundamentally revised the manuscript and addressed these points as required.

Comments:

Although important and nicely framed, the hypothesis that 3 clusters of lymphocytic infiltration patterns exists is not adequately tested. What would the data need to look like for the authors to conclude otherwise? Perhaps some sort of best-fit analysis could be used to assess the validity of the 3-cluster claim – for both lymphocytes and myeloid cells. tSNE plots should be shown for myeloid cell data.

We have done exactly as suggested (Results section).

The tSNE analysis performed is informative but incomplete. Could these plots be shown for myeloid cell parameters as well? It might also be useful to perform this analysis on the entire dataset (similar to that shown in Figure 5H – e.g., one dot per patient, colored by cancer type or colored by cluster definition from Figure 2). This comment highlights the value of the dataset the authors are providing: it is exciting that many different data analysis/visualization strategies could be used to evaluate these data.

As discussed above, the tSNE plot was not an optimal choice to present the data. We have revised the figures (new Figure 2—figure supplement 5) to present the data in a more straightforward way.

It is not clear why the authors have provided the exaggeration of 10 for tSNE and used a non-default parameter for this. Does this value largely influence the outcome of their analysis? Did the authors also experiment with various perplexity values for their analysis?

See above, the tSNE plot has been removed.

For the tSNE analysis of this number of events, the stochasticity and perplexity value might be important factors. The authors should show several replicate tSNE runs to assess the consistency of the pattern shown. To justify which of these runs to shown in the main figure, the authors should choose the run with the best fit values provided by the tSNE algorithm. In other words, I am concerned that the conclusion that the tSNE visualization fits well with k-means (k=3) clustering could be a coincidence for this particular run and perplexity setting and the authors should assess the reproducibility of this analysis.

See above, the tSNE plot has been removed.

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    Supplementary file 1. Clinical characterization of the DACHS cohort.

    This table lists summary statistics of all relevant clinico-pathological features of the DACHS cohort.

    elife-36967-supp1.docx (12.8KB, docx)
    DOI: 10.7554/eLife.36967.015
    Supplementary file 2. List of all image analysis parameters.

    In this table, all parameters for cell detection and classification using the open source software QuPath are listed. Two sets of parameters are distinguished: ‘DAB’ (diaminobenzidine), used for blue-brown staining, and ‘Red’, used for blue-red staining in melanoma. OD = optical density. All parameters were used in the pan-cancer cohort unless labeled as ‘DACHS’, in which case they were used in the DACHS cohort.

    elife-36967-supp2.docx (18.8KB, docx)
    DOI: 10.7554/eLife.36967.016
    Supplementary file 3. List of all samples and all measurement values of the pan-cancer cohort.

    In this table, we report all raw measurements for all samples that were used in this study. Column names are: ‘class’ (tumor type as listed above), ‘patient’ (patient pseudonym), ‘antigen’ (antigen for immunostain), ‘TU_CORE_cells_mm2’ (number of positively stained cells per square millimeter in the tumor core), ‘MARG_500_IN_cells_mm2’ (number of positively stained cells per square millimeter in the inner invasive margin, defined as ranging 0–500 µm to the inside from the tumor edge), ‘MARG_500_OUT_cells_mm2’ (number of positively stained cells per square millimeter in the inner invasive margin, defined as ranging 0–500 µm to the outside from the tumor edge).

    elife-36967-supp3.xlsx (49.2KB, xlsx)
    DOI: 10.7554/eLife.36967.017
    Supplementary file 4. List of all cutoff values for all cell types.

    On the full data set of N = 965 tissue slides from N = 177 patients in 10 tumor types, we calculated the median cell density for each antigen, taking the compartments ‘outer invasive margin’ and ‘tumor core’ into account. These median values were subsequently used as cutoff values for low and high cell densities which were then used to define hot, cold and excluded phenotypes.

    elife-36967-supp4.docx (12.9KB, docx)
    DOI: 10.7554/eLife.36967.018
    Supplementary file 5. Continuous cell densities of CD8+ and CD163+ cells are not significantly associated with overall survival in colorectal cancer.

    A multivariable Cox proportional hazard model was fitted to all variables listed in this table. N = 286 CRC patients in the DACHS cohort, number of events = 108, significance codes (sig): *<0.05, **<0.01, ***<0.001. HR = hazard ratio, UICC = Union internationale contre le cancer.

    elife-36967-supp5.docx (13.7KB, docx)
    DOI: 10.7554/eLife.36967.019
    Supplementary file 6. Bivariate immune phenotype predicts risk of death of any cause.

    A multivariable Cox proportional hazard model was fitted to all variables listed in this table. N = 286 CRC patients in the DACHS cohort, number of events = 108, significance codes (sig): *<0.05, **<0.01, ***<0.001. HR = hazard ratio, UICC = Union internationale contre le cancer.

    elife-36967-supp6.docx (15.1KB, docx)
    DOI: 10.7554/eLife.36967.020
    Transparent reporting form
    elife-36967-transrepform.docx (246.2KB, docx)
    DOI: 10.7554/eLife.36967.021

    Data Availability Statement

    We release all source codes under an open access license (http://dx.doi.org/10.5281/zenodo.1407435; copy archived at https://github.com/elifesciences-publications/immuneTopography). Also, we release all raw data from our experiments (Supplementary File 3).

    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES