Space Is the Place: Mapping the Cell–Cell Interactions That Predict Immunotherapy Responses in Melanoma

Melanoma is a tumor that responds favorably to immune checkpoint inhibitor (ICI) therapy. These therapies, including anti-programmed death protein 1 (PD-1), anticytotoxic lymphocyte-associated protein-4 (CTLA4), and anti-lymphocyte–activating gene 3 (LAG3), serve to abrogate the signals that negatively regulate T-cell function and stimulate antitumor immune responses. The most commonly used of these therapies is anti-PD-1 (e.g., nivolumab), which blocks interactions between PD-1 expressed on activated T cells and its ligand PD-L1, which is expressed on tumor cells, macrophages, dendritic cells, and B cells. Blockade of PD-1/PD-L1 interactions reverses T-cell suppression and leads to improved T-cell–mediated antitumor immunity. Anti-PD-1 is frequently used clinically in combination with anti-CTLA4 and, more recently, anti-LAG3.

At this time, single-agent anti-PD-1 therapy leads to durable responses in 30 to 40% of melanoma patients (1), yet nearly one-third of those who respond relapse within 3 years (2). Escape from anti-PD-1 therapy is complex and can arise through T-cell exhaustion, downregulation of interferon and STAT signaling, increased β-catenin signaling, and loss of tumor MHC and antigen expression (3). Identifying who will best respond to ICI therapy is a challenge, with bulk expression levels of PD-1 and PD-L1 in the tumor immune microenvironment (TiME) being poorly predictive of response. It has been well established that the extent of lymphocyte infiltration into a primary melanoma is prognostic, with the widespread “brisk” infiltrate of tumor-infiltrating lymphocytes (TIL) throughout the tumor portending improved survival. Likewise, the baseline levels of CD8⁺ T-cell infiltrate are predictive of ICI response. Beyond this, robust predictive biomarkers of ICI response have been difficult to define. This difficulty arises partly due to the complex nature of the TiME, which involves many interactions between multiple populations of host immune and stromal cells, as well as heterogeneous populations of melanoma cells. Deconvoluting this host-tumor communication to determine whether the overall balance of the TiME is favorable or unfavorable remains a major challenge.

Recent years have seen the development of multiple novel technologies to interrogate the TiME in a high-dimensional manner. Among these, single-cell RNA-seq (scRNA-seq) has provided valuable insights into immune landscapes, allowing cell type compositions of clinical melanoma specimens to be accurately mapped (4, 5). Of note, these analyses have identified the potential role of a T-cell exclusion program in ICI response and have defined a novel population of dendritic cells (DC3) that correlate with an improved survival (4, 5). Although the initial scRNA-seq analyses were primarily used to describe cell landscapes and transcriptional programs within tumor samples, attempts are now being made to infer cell–cell interactions from single-cell data using methods such as CellPhoneDB and SingleCellSignalR. Although useful for discovery and hypothesis generation, the identified cell–cell interactions from scRNA-seq datasets lack spatial information, and key findings must be confirmed through multiplexed IHC. As it is likely that ICI responses are determined by the sum of interactions between multiple cell types within the TiME, there is a clear need to develop methods to map cell–cell interactions in a high-dimensional manner.

In this issue of Cancer Research, Antoranz and colleagues identify some of the first spatial immune cell–cell interactions that predict responses to PD-1 therapy in patients with melanoma (6). In their studies, Antoranz and colleagues performed a bulk transcriptomic analysis of 16 baseline tumor specimens from melanoma patients prior to anti–PD-1 therapy (8 responders, 8 nonresponders). It was found that responders had a gene expression profile associated with immune response (e.g., T-cell, B-cell, and T-cell receptor signaling pathways). However, when the authors performed a CIBERSORT analysis to deconvolute the immune composition of the samples, no differences in TiME composition could be observed between the responders and nonresponders (6). To address this in more detail, the investigators undertook highly multiplexed immunohistochemical staining for 77 markers, allowing 18 cell types and various states of activation/exhaustion to be identified. Spatial analysis of the landscape uncovered three important habitats: tumor area, tumor-stroma interface, and nontumor area. The most common cell type within the tumor area was melanoma cells, with approximately 24% of the total cells being immune cells. Among these, M1-type macrophages were the dominant immune cell type. The majority of the cytotoxic T cells were found at the tumor-stroma interface. Further profiling of the cytotoxic T cells identified 9 subsets, the largest group of which (accounting for 37% of the total) did not express any of the markers included on the panel. This lack of a clear phenotype of the most abundant T-cell population was a limitation of the study, illustrating the need for future analyses that are not constrained by small numbers of markers. Analysis of the activation state of the cytotoxic T cells demonstrated increased activation in the nontumor areas and more exhaustion within the tumor (Fig. 1), as has been previously reported by others (7, 8).

PD-1/PD-L1 interactions are a major determinant of T-cell exhaustion, which can occur in part through inhibition of the CD28 signaling (9). There is already evidence that myeloid cells, including macrophages and dendritic cells, express high levels of PD-L1 (9). Although PD-L1 is also expressed on tumor cells, the accumulating evidence suggests that PD-L1/PD-1 interactions with myeloid cells may be more critical for T-cell exhaustion (9). One possible explanation for this is that myeloid cells (unlike tumor cells) also express the CD28 ligands B7–1 and B7–2 and upregulate PD-L1 to balance CD28 activation in T cells during antigen presentation (9).

An analysis of the spatial distribution of PD-L1 expression on macrophages identified the lowest levels to be in the tumor area, with peak PD-L1 expression being seen at the tumor-stromal interface in macrophages that were close to T cells (Fig. 1; ref. 6). The proximity of macrophages with the highest expression of PD-L1 to T cells is significant, as expression of PD-L1 in macrophages is regulated by IFNγ released by cytotoxic T cells (Fig. 1). The spatial analyses of Antoranz and colleagues further showed that for responding patients, the cytotoxic T cells closest to M1-macrophages at the tumor/stromal interface showed the highest levels of exhaustion. These findings suggest that in pretreatment samples from responders, PD-L1/PD-1 interactions between macrophages and T cells proceed in a canonical manner, with PD-1 engagement leading to T-cell inactivation (6). In the nonresponders, a spatial relationship between cytotoxic T-cell exhaustion and macrophage proximity was not identified, hinting at a broader dysregulation of normal immune cell-immune cell interactions. The lack of functional PD-1/PD-L1 interactions between macrophages and T cells in nonresponding patients at baseline could explain the lack of subsequent response to anti-PD-1 therapy.

The identification of M1-type macrophages as the potential driver of T-cell exhaustion in responders was unexpected. Typically, M1-macrophages, which are characterized as expressing type-1 inflammatory cytokines such as TNFα and IL6, are thought to have antitumor activity, whereas M2-type macrophages, which are characterized by the release of type-2 anti-inflammatory cytokines, such as IL4 and IL3, are considered to be more protumorigenic. One potential explanation for the identification of M1-type macrophages as the drivers of T-cell exhaustion is that the M1/M2 macrophage distinction may not be an accurate reflection of the diversity of macrophage states, which are difficult to define with the small number of markers used in this study. In support of this idea, another recent spatial analysis of melanoma samples showed that macrophages with high PD-L1 expression in close proximity to cytotoxic T cells expressed high levels of TIM-3 (7), suggestive of a more immune suppressive macrophage phenotype. Previous scRNA-seq studies have highlighted the diversity of myeloid cell transcriptional profiles and are continuing to refine our understanding of macrophage phenotypes (5, 10).

Limitations of the study of Antoranz and colleagues include the low sample numbers in the training set and the relatively small list of markers included. Despite these drawbacks, the authors were able to identify spatial interactions between PD-L1⁺ M1-macrophages and cytotoxic T cells at baseline as being correlated with anti-PD-1 response, an essential step in redefining PD-L1 expression as a predictive biomarker. As technology improves and high-dimensional single-cell spatial analysis can be undertaken, further insights will be gained into the nature of baseline TiMEs that predict ICI response, allowing therapy to be further personalized.

K.S. Smalley reports personal fees from Elsevier, grants from NIH, and grants from Revolution Medicine outside the submitted work. I. Smalley reports grants from NIH, grants from Melanoma Research Alliance, grants from Florida Breast Cancer Foundation, and grants from State of Florida outside the submitted work.