Anti-cancer immunotherapy is encountering its own checkpoint. Responses are dramatic and long lasting but occur in a subset of tumors and are largely dependent upon the pre-existing immune contexture of individual cancers [1]. Current research is trying to determine why some cancers respond to CIT more than others and the reasons for individuals’ variability within each indication [2, 3].

Several morphological observations based on immune histochemical analyses suggest that three immune landscapes best define distinct varieties of the cancer microenvironment: an immune-active , an opposite immune-deserted and an intermediate immune-excluded [4]. Across cancers, and among subtypes, the prevalence of each landscape may differ. Nevertheless, this trichotomy is observable across most solid tumors suggesting that convergent evolutionary adaptations determine the survival and growth of cancer in the immune competent host leading to predictable patterns determined by uniform immunological principles independent of the biology pertinent to distinct tumor tissue of origin. It is therefore reasonable to postulate that the mechanisms leading to cancer resistance to checkpoint blockade are similar across cancers deriving from different tissues. Functional characterization based on transcriptional analyses cannot distinguish structural differences. Thus a reductionist argument could be made that at the functional level cancers can simply be aggregated into immune-active or immune-silent clusters. Current work from our group suggests that most immune excluded cancer resemble functionally immune active tumors suggesting that the periphery immune cells interact with cancer cells (unpublished observation).

We will refer to the mechanisms allowing persistence of cancer in the immune-active cluster as Compensatory Immune Resistance ( CIRes ) based on the assumption that lack of CIRes would prevent tumor survival against the host’s immune response. Conversely, we refer to survival of cancer in the immune-deserted environment as Primary Immune Resistance ( PIRes ). In 2002, we proposed that human cancer immune responsiveness to antigen-specific vaccination administered in combination with systemic interleukin-2 is predetermined by a tumor microenvironment conducive to immune recognition [5] Likewise, recent observations suggest that CIT is most effective for the treatment of immune active tumors, where a tenuous balance between immune-effector and immune-suppressive mechanisms determines outcomes [6–8].

To explain CIRes and PIRes, several phenomenologies have been described and models proposed that largely outnumber the fewer immune landscapes (Table 1). Such discrepancy can be explained in three ways: a) some models do not translate broadly across the majority of human cancers, b) there are subtler immune landscapes than those discernable by current approaches, or c) some models are redundant and describe different facets of the same pathophysiology. To solve this discrepancy, we surveyed human cancers through readily-available open-access information.Tab1

Principal models related to immune responsiveness

Marincola et al. [9] have previously described a transcriptional signature comprising the concordant activation of innate and adaptive immune effector mechanisms that is required for the occurrence of immune-mediated tissue-specific destruction . This represents a conserved mechanism determining destructive autoimmunity, clearance of pathogen-bearing cells during acute infection, acute allograft rejection, graft-versus-host disease and rejection of cancer. Thus, the signature was termed: the Immunologic Constant of Rejection (ICR) [9]. The ICR signature was derived from bulk tumor transcriptome data sets, as they offer the most readily-available sample/data type and the easiest to apply in the clinic due to the ease of collection. The ICR signature was further trained to be representative of the broader signature as previously described [10] and is currently represented by twenty transcripts and four functional categories: CXCR3/CCR5 chemokines (including CXCL9, CXCL10, CCL5), Th1 signaling (including IFNG, IL12B, TBX21, CD8A, STAT1, IRF1, CD8B), effector (including GNLY, PRF1, GZMA, GZMB, GZMH) and immune regulatory (including CD274, CTLA4, FOXP3, IDO1, PDCD1) functions. The expression of these twenty representative genes is highly correlated with the extended ICR signature that includes approximately five-hundred transcripts and is representative of its main functional orientation as previously described [11, 12]. Importantly, the specific cell types in the tumor microenvironment expressing these genes will ultimately be relevant in elucidating the mechanistic link between the ICR and the immune responsiveness of cancer. It was subsequently observed that the ICR serves both as a positive predictor of responsiveness to immunotherapy and as a favorable prognostic marker for various tumor types [6, 10, 13, 14]. This observation suggests that these related phenomena represent facets of a spectrum within the continuum of anti-cancer immune surveillance. Such continuity leads to the fair, though unproven, assumption that signatures predictive of prolonged survival may mark an immune-favorable cancer phenotype and serve as surrogate predictors of responsiveness to anti-cancer immunotherapies [10, 15]. This assumption is also corroborated by recent reports suggesting that similar gene expression patterns predict response to CIT [6–8]. Specifically, Ayers et al. [6] using RNA from pre-treatment tumor samples of pembrolizumab-treated patients and the nCounter platform identified and validated a pan-tumor T-cell–inflamed gene signature correlating with clinical benefit. This tumor inflammation signature ( TIS ) contains IFN-γ–responsive genes (CD27, STAT1, IDO1, HLA-E, NKG7) related to antigen presentation (HLA-DQA1, HLA-DRB1, PSMB10, CMKLR1) chemokine expression (CCL5, CXCL9, CXCR6), cytotoxic activity (CD8A), and adaptive immune resistance (TIGIT, LAG3, CD274, CD276, PDCD1LG2) and as such, is highly correlated to the ICR signature: composite scores for each signature calculated with ssGSEA software and compared according the expression values in the 999 breast cancer samples from TCGA were highly correlated (r = 0.98). The TIS has been developed into a clinical grade assay running on the nCounter platform currently being evaluated in ongoing pembrolizumab trials (3). We, therefore, developed a strategy to build a navigational map of cancer immunity with the primary purpose of assigning distinct immune responsive and resistant models to their respective immune-landscapes using the expression of twenty transcripts that are representative of the extended ICR signature.

Using the ICR signature [10], we queried the prognostic accuracy of a transcriptional data set of breast cancers from The Cancer Genome Atlas ( TCGA ) as a discovery platform and validated the findings on a second transcriptional set of breast cancers from the Gene Expression Omnibus ( GEO ) repository at the National Center for Biotechnology Information. The TCGA set encompasses RNA-seq-based transcriptional characterization of 999 breast cancer cases while the compilation of 10 GEO studies included 1728 cases of breast cancer (compiled in [16]) that were transcriptionally characterized utilizing a uniform Affymetrix platform. Both datasets were classified according to the coordinated expression of ICR transcripts [10].

ICR groups were ranked 1–4, according to the level of expression of the 20 representative ICR genes (Fig. 1). At the transcriptional level a dichotomy between Immune-active (ICR3–4) and immune-silent (ICR1–2) clusters was apparent [10]. Kaplan-Meier applied to the four ICR classes confirmed that ICR gene expression correlates with survival in breast cancer [10].Fig1Fig. 1

Distribution of sRes gene expression according to distinct models (Table 1) within immune landscapes as defined by ICR gene expression. Four immune landscapes were identified ranked according to the level of expression of ICR genes with purple, green, blue and red representing respectively ICR 1, 2, 3 and 4. Because of similarities in patterns of gene expression, for the purpose of discussion the landscapes will be referred to as immune-silent (ICR1–2) or Immune-active (ICR3–4). Genes were assigned to distinct landscapes according to significant difference in expression between ICR4 and ICR1 (p-value < 0.05 and false discovery rate < 0.1). Genes signatures associated with a particular immune responsiveness model as per Table 1 were assigned to distinct landscapes according to gene enrichment analysis and ubiquitous were considered signatures that did not reach significance (one-tailed Fisher Test p-value < 0.01). *Cluster of ubiquitous genes that segregate with the immune active signatures but did not reach significance and, therefore, were considered ubiquitous

Several models have been proposed to explain proclivity or resistance of cancer in response to immunotherapy (Table 1). Effector T cell exhaustion is broadly observed in the tumor microenvironment manifesting through the expression of a cluster of immune checkpoints often concomitantly expressed in response to chronic interferon stimulation [22, 23]. In addition, it is well established that regulatory T cells balance immune effector mechanisms [24–26]. Other models propose blockade of immune cell homing to cancer tissue by barrier molecules, chemo-inhibitory mechanisms, and by epigenetic silencing of chemokines (CCL5, CXCL9, and CXCL10), Th1 signaling molecules and antigen processing machinery components [12, 27–37].

Other immune resistance models point to alterations of cancer cell signaling that result in secondary dysregulation of myeloid cell function. Cancer-intrinsic β-catenin signaling defects disrupt chemo-attraction of dendritic cells (DCs) and, consequently, antigen presentation in the context of immunogenic cell death [21, 38–40]. In addition, polarization of DCs toward a tolerogenic, IL23 producing phenotype leading to Th17 polarization was described in experimental animal models and in human samples [26, 41–46]. Suppression of anti-cancer immunity has also been attributed to the TAM receptor tyrosine kinase family members that mediate efferocytosis and negative regulation of DC activity [47]. Similarly, hypoxia can drive immune suppression by inducing tolerogenic myeloid DC polarization [48, 49]. Finally, myeloid cell biology is responsible for the immune regulation of the cancer microenvironment through the upregulation of metabolizing enzymes such as arginase and indoleamine 2,3-dioxygenase, which can negatively impact T cell function [50, 51].

The phenotype of suppressive myeloid cells in the microenvironment is often attributed to activation of the PI3Kγ/SFK/pGSK3/β-catenin axis (Fig. 2). Phosphoinositide3-kinase-gamma (PI3Kγ) can act as a molecular switch that triggers immune suppressive mechanisms in myeloid DCs [52, 53]. At the same time, alteration of PI3K functional components plays a widespread role in tumorigenesis [54]. Downstream phosphorylation of serum and glucocorticoid kinase 1 (SGK1) by the PI3K/PDK1 cascade leads to activation of glycogen synthase kinase 3 beta (GSK3β) and subsequently β-catenin [55–57]. Interestingly, most studies describing dysregulation of the PI3Kγ/SFK/pGSK3/β-catenin axis refer to abnormalities intrinsic to tumor cells, although the same pathway can play an important role in myeloid suppressor DC induction and immune suppression downstream of the TAM receptor tyrosine kinases [58–60]. Converging on the same pathway, hypoxia inducible factors (HIF1α) signal through the SGK3β/β-catenin axis promoting cancer cell stemness and immune suppression [48, 49, 61] (Fig. 2).Fig2Fig. 2

Dichotomy in the Myeloid-Centric Hypothesis of immune resistance: the same pathway is relevant to myeloid cell differentiation as well as intrinsic oncogenic activation (in red boxes are included models included in Table 1). It is currently unclear how the two interpretations diverge vs relate to each other and further characterization of the single cell level will need to be entertained to clarify this point

All data download, processing and analyses were done in R programming environment and as described in Hendrickx et al. [11]. For the unsupervised clustering of the TOE genes (Additional file 1), modified distance and hierarchical clustering functions were used. Specifically the distance between 2 genes was defined as 1-“Correlation Coefficient (Spearman)” and for the hierarchical clustering function “Ward.D2” method was used.

Composite correlation between the ICR and the TIS signature was assessed by calculating a cumulative score for each gene included in the respective signature using ssGSEA method form GSVA package and correlating the scores in the breast cancer TCGA data set according to Spearman Correlation.

The metrics used when assigning genes to silent, active and ubiquitous groups are derived from differential expression statistics between ICR1 and ICR4 samples. The genes are assigned to the active cluster if they have significantly higher expression levels in ICR4 samples (p-value < 0.05 and FDR < 0.1). Similarly the genes are assigned to the silent cluster if they have significantly higher expression levels in ICR1 samples (p-value < 0.05 and FDR < 0.1). If the genes do not pass these cutoffs they are grouped as “Ubiquitous”. Geneset enrichment for each signature belonging to individual models of immune resistance (Table 1) against ICR1 and ICR4 clusters was assessed using one-tailed Fisher’s exact test.

In the analyses and corresponding heatmaps, the genes that were identified in multiple signatures were plotted as one, so each gene in the heatmap is unique. When the ICR direction is inferred for each signature, the repeated genes contributed to each signature with the same statistics.