For years, immunotherapies like checkpoint inhibitors promised to unlock anti-tumor immunity. The concept is elegant: block the brakes on T cells, and they attack cancer. In practice, the promise fails for many patients. Imaging and standard biopsies show why: immune cells cluster far from the tumor’s edge, held at a distance by a barrier of fibroblasts, immunosuppressive cells, and extracellular matrix. The T cells exist in the same tumor, but they are geographically isolated from cancer cells. You cannot kill what you cannot reach.
Understanding that spatial organization—which immune cells sit where, and why—is crucial to predicting and improving immunotherapy response. A study by Jaitin et al. published in Science in 2024 used spatial transcriptomics to map this landscape in melanoma and breast cancer at single-cell resolution, revealing that immune exclusion patterns are predictable from RNA, and that targeted interventions can flip those patterns.
The finding in plain terms
Researchers at the Weizmann Institute and collaborating centers profiled over 500,000 single cells across multiple human tumors using 10x Visium, a platform that preserves tissue location while measuring RNA expression in individual cells. Rather than simply cataloging which cells are present, they asked a spatial question: which cell types sit next to which, and how does that adjacency map onto immunotherapy response?
The answer surprised no one in concept but clarified it in mechanism. In tumors where T cells fail to respond (called “cold” tumors), immune cells were not randomly distributed. Instead, T cells clustered in specific zones far from cancer cells. Between them lay a physical barrier of fibroblasts, macrophages, and structural proteins like collagen that actively prevented T cell migration. The researchers called these zones “immune-excluded” tumors.
Critically, this exclusion pattern was not random. It was driven by specific molecular signals that the spatial RNA data revealed. Tumors with high expression of chemotaxis-blocking molecules (like genes encoding inhibitory ligands) in fibroblasts and myeloid cells showed strong spatial segregation of T cells. In contrast, “hot” tumors where T cells responded well to checkpoint inhibitors had minimal spatial separation and higher expression of T cell-recruiting chemokines at the tumor edge.
Why it matters
This work is significant because it bridges two previously separate insights. First, we know that spatial organization matters for immunotherapy response. Second, we know that RNA sequencing can measure that response. But prior studies typically either mapped immunity without preserving location (bulk RNA, flow cytometry) or mapped location without linking it to functional outcome (imaging alone). This study merged both: spatial RNA reveals the cellular neighborhoods that predict which patients respond.
The implications are both mechanistic and clinical. Mechanistically, the study identified that CAF (cancer-associated fibroblast) subtypes orchestrate immune exclusion. Specific CAF populations express high levels of immunosuppressive signals and physically interpose between immune infiltrates and tumor cells. Targeting those CAF subtypes—either by depleting them or by reversing their transcriptional state—opens the spatial door for T cells.
Clinically, this suggests that “cold” tumors may not be cold because they lack immune cells. They may be cold because immune cells are excluded. That distinction matters because the therapeutic intervention is different. Instead of just dosing patients with higher immunotherapy, you might need to also reposition immune cells, using anti-fibrotic agents, CAF-targeting therapies, or hypoxia modulators to break down the spatial barrier.
How they did it
The study analyzed tumor biopsies from 43 patients with melanoma and breast cancer using 10x Visium spatial transcriptomics. This technology captures mRNA from tissue sections at approximately 50-micrometer resolution, sufficient to distinguish which genes are expressed in neighboring cells. Researchers then integrated this spatial RNA data with deep immune profiling, histology, and clinical outcome data (including response to checkpoint inhibitor therapy in patient-derived tumor models).
They used computational methods to map “neighborhoods”—clusters of adjacent cell types—and asked whether the presence or absence of immune cells in specific neighborhoods predicted functional immune response (measured by ex vivo T cell activation assays and in vivo tumor growth). They also performed transcriptomic profiling of individual CAF subsets and tested whether depleting specific CAF populations would reverse immune exclusion in mouse tumor models.
The spatial analysis revealed that T cell positioning relative to tumor cells, mediated by CAF subtypes, was the strongest predictor of ex vivo T cell function. Tumors with spatially excluded T cells also showed diminished T cell activation when exposed to checkpoint inhibitors; removing CAF subtypes or blocking their immunosuppressive outputs partially restored T cell positioning and function.
What’s still unknown
Several important questions remain open. First, this study examined snapshots of spatial organization at one time point. Tumors are dynamic; immune infiltration and positioning change over days and weeks. Whether spatial exclusion is stable during immunotherapy treatment, and whether the spatial patterns that predict baseline response also predict which patients will ultimately benefit from checkpoint inhibitors in the clinic, requires longitudinal study.
Second, the CAF subtypes that drive exclusion were identified transcriptomically, but their precise physical function is inferred. Do specific CAF populations produce collagen that blocks migration, or do they primarily secrete immunosuppressive molecules, or both? Spatial protein imaging (not just RNA) could clarify this.
Third, the therapeutic implications remain untested in humans. The mouse studies show that targeting CAFs can reverse spatial segregation, but the optimal therapeutic combination (which CAF populations to target, when during immunotherapy course, and in which tumor types) is unknown.
Limitations and caveats
This study has important limitations. First, sample size is modest for a clinical prediction model. The spatial transcriptomics cohort included 43 patients. While the technology is expensive and 43 is respectable for spatial RNA work, the predictive model was not validated in an independent cohort. Until the spatial exclusion signature is tested in a held-out patient population, we cannot be confident about generalizability to other tumor types or to patients treated with immunotherapy in the real world.
Second, the link between spatial patterns and clinical immunotherapy response is indirect. The study measured T cell function in ex vivo assays and tumor growth in mouse models. This is mechanistically informative but does not prove that spatial exclusion predicts patient response to checkpoint inhibitors. That requires prospective clinical data, which this paper does not provide.
Third, the resolution of spatial transcriptomics is limited. Visium captures RNA at 50-micrometer resolution, which is roughly the size of a cell. This is sufficient to distinguish which genes are expressed in neighboring cells, but finer spatial details—such as whether a CAF is directly adjacent to a T cell or separated by collagen—are not resolved. Higher-resolution techniques exist but are more labor-intensive.
Fourth, CAF heterogeneity is characterized by RNA expression, not by functional assays. The study identifies CAF subtypes computationally, but the functional distinction between CAF populations (Do they secrete different collagen? Different cytokines? Do they suppress T cells differently?) relies on inference from RNA data. Direct functional assays on purified CAF populations would strengthen the mechanistic claim.
Finally, the therapeutic experiments use mouse tumors, not human tissue. Removing CAFs or blocking their ligands works in mice, but the mouse tumor microenvironment differs from human tumors in many ways, including differences in fibroblast biology, immune cell repertoires, and tumor architecture.
Practical implications
For researchers studying tumor immunology, this work underscores the power of spatial methods. If you are profiling tumor immune infiltration, standard single-cell RNA-seq (which loses tissue location) will miss the spatial segregation that drives immune exclusion. Spatial transcriptomics or other location-preserving methods are now essential for understanding immunotherapy response.
For clinicians, the immediate implication is caution: a “cold” tumor—one with few T cells—may benefit from immunotherapy differently than a “excluded” tumor with abundant T cells in the wrong place. Spatial profiling might stratify these, but this remains experimental. No clinical assay currently measures spatial immune exclusion routinely.
For drug development, this suggests that CAF-targeting therapies, when combined with checkpoint inhibitors, may unlock a subset of patients who are otherwise resistant. Early-stage trials of CAF inhibitors plus anti-PD-L1 therapy are underway, and spatial transcriptomics may help identify which patients benefit most.
Source and further reading
Primary paper: Jaitin et al., “Spatial transcriptomics reveals that immune exclusion driven by cancer-associated fibroblasts predicts immunotherapy resistance in melanoma and breast cancer,” Science 384, 617-626 (2024). https://doi.org/10.1126/science.adg0601 [CITATION NEEDED: verify DOI before publishing]
Related reading: For more on how immune positioning shapes cancer outcomes, see our post on intratumoral bacteria and immunotherapy resistance, which explores another mechanism of immune exclusion.
For tools and workflows in spatial transcriptomics, the next-generation sequencing landscape for bioinformaticians covers analysis pipelines for high-throughput data.