Large-scale profiling of solid tumors has converged on three recurrent patterns of CD8⁺ T-cell infiltration. Immune-inflamed (“hot”) tumors display dense CD8⁺ TIL dispersed throughout the cancer nests and stroma, often accompanied by an interferon-γ–rich cytokine milieu (8–10). Immune-excluded lesions harbor abundant CD8⁺ T cells, but these cells are trapped at the invasive margin or within fibrotic septa and seldom penetrate the parenchyma (11). At the opposite extreme, immune-desert (“cold”) tumors show a near-complete paucity of CD8⁺ T cells in both center and periphery (6). This inflamed–excluded–desert continuum is now widely used to stratify the baseline tumor immune contexture and to anticipate responses to immunotherapy ( Figure 1 ).

Beyond binary “hot/cold” labels, quantitative cut-offs refine prognostication: > 500–750 CD8⁺ T cells/mm² typically denotes high infiltration, whereas < 100 CD8⁺ T cells/mm² signifies low. Crucially, where the cells reside matters. The Immunoscore algorithm, validated across multiple cohorts, integrates CD3/CD8 densities in the tumor center (TC) and IM to yield five tiers (I0–I4); high scores (I3–I4) correlate with prolonged disease-free survival and superior benefit from immune-checkpoint blockade (12). New data indicate that CD8 density at the IM alone can approximate full Immunoscore performance, simplifying routine pathology pipelines (13).

However, it is important to highlight that there is currently no universally accepted consensus definition for the objective classification of tumor immune phenotypes, and existing categorizations (e.g., “hot”, “cold”, “immune-excluded”) may vary considerably across studies. While the Immunoscore represents a significant advance in quantifying immune infiltration, it is primarily validated in and routinely applied to colorectal cancer, limiting its generalizability to other tumor types. As reviewed by Tiwari et al., definitions of immune phenotypes remain heterogeneous across the literature, underscoring the need for standardized, pan-cancer frameworks for immune contexture classification to guide both research and clinical decision-making (14).

Spatial-omics technologies have revealed that CD8⁺ T cells are not randomly dispersed but concentrate in discrete niches (15). At the invasive front, TIL interdigitate with tumor cells undergoing epithelial–mesenchymal transition, positioning them at a critical bottleneck for metastatic spread (16, 17). Perivascular immune hubs—rich in dendritic cells and chemokines such as CXCL9/10—act as staging grounds where newly recruited CD8⁺ T cells are primed and expanded before entering the tumor core (18). Mature tertiary lymphoid structures (TLSs), identifiable by spatial transcriptomics and high-parameter imaging, harbor germinal-center-like B cells, follicular helper T cells and stem-like CD8 progenitors (19); their presence consistently associates with higher intratumoral CD8⁺ T-cell densities and improved clinical outcome across carcinomas such as nasopharyngeal, ovarian and lung cancers (19, 20).

CD8⁺ T-cell abundance and spatial positioning collectively inform prognosis and guide treatment selection. Inflamed tumors respond best to PD-1/PD-L1 or CTLA-4 blockade (11); excluded tumors often require barrier-modulating combinations (e.g., TGF-β or VEGF inhibitors) to enable parenchymal entry; desert tumors benefit from priming strategies—vaccines, oncolytic viruses, STING agonists—to initiate de-novo T-cell recruitment before checkpoint inhibition (6). Thus, precise spatial immunophenotyping not only captures tumor biology but also delineates rational avenues to convert “immunotherapy-resistant” landscapes into “immunotherapy-responsive” ones.

Cell-intrinsic mutations, viral integrations or aberrant splicing generate neoantigens that are liberated when tumor cells undergo immunogenic cell death (ICD) triggered by radiotherapy, oxaliplatin, oncolytic viruses or STING/TLR agonists (Figure 2). Cancers with few non-synonymous mutations release scant antigenic material, providing little substrate for downstream immune activation. Large cohort analyses show that a high tumor-mutational burden (TMB) correlates with elevated neoantigen load, dense CD8⁺ T-cell infiltration and superior response to PD-1 blockade in lung cancer and other entities (21, 22). Nevertheless, antigen quantity is not the sole bottleneck: melanoma datasets reveal that “cold” and “hot” tumors can express comparable levels of putative antigens, pointing to additional barriers that arrest the cycle (23).

After uptake, professional antigen-presenting cells (APCs) must cross-present peptides on HLA-I to prime naïve CD8⁺ T cells. Tumor cells frequently thwart this step by β₂-microglobulin/TAP loss, HLA-I locus deletions or epigenetic silencing, while VEGF and TGF-β secreted into the micro-environment impair dendritic-cell maturation (23). Restoring type-I/II interferon signaling or pharmacologically demethylating HLA promoters can up-regulate the antigen-presentation machinery and re-ignite the cycle.

Effective priming in tumor-draining lymph nodes requires Batf3⁺ cDC1s that deliver TCR, co-stimulatory and cytokine “signals 1–3.” Tumors limit this by excluding cDC1s, expanding FOXP3⁺ Tregs that consume CD80/86, and inducing tolerogenic IDO and IL-10 pathways. Complement-component dysregulation or loss of CD40L signaling similarly cripples priming (24). CTLA-4 blockade, FLT3L + poly-ICLC to expand cDC1s, and IDO inhibitors are under clinical exploration to repair this defect.

Activated T cells egress and follow CXCR3/CCR5 gradients (CXCL9/10, CCL5) to the tumor (25). VEGF-driven, chaotic vasculature down-regulates ICAM-1/VCAM-1 and induces “endothelial anergy,” (26, 27) whereas systemic TGF-β suppresses CXCR3 on effector T cells (28). Anti-angiogenic therapy transiently normalizes vessels and restores adhesion-molecule expression, thereby enhancing T-cell recruitment (29, 30).

Recruitment of cytotoxic T cells relies on chemokine gradients dominated by CXCL9, CXCL10 and CCL5 that engage CXCR3 or CCR5 on circulating CD8⁺ T cells (11). “Hot” tumors typically maintain high interferon-γ signaling and a robust CXCL9/10 axis, whereas “cold” tumors frequently lack these chemo-attractants or preferentially secrete chemokines (e.g., CCL17, CXCL12) that draw suppressive myeloid cells and Treg cells instead (11). Pre-clinical work shows that enforced expression of CXCL9/10, STING agonism or oncolytic viruses can re-establish productive gradients and convert immune-desert lesions into inflamed ones (31).

Crossing the endothelial barrier is only half the journey; dense, CAF-produced extracellular matrix (ECM) and CXCL12/TGF-β chemokine sinks trap T cells at the invasive margin. CAF-rich desmoplastic cancers (e.g., pancreas) exemplify “immune-excluded” lesions in which CD8⁺ T cells rarely enter the core (32, 33). Strategies such as FAP-targeted depletion, PEGylated hyaluronidase or low-dose radiotherapy that remodel ECM can open stromal corridors for lymphocyte entry.

Most solid tumors develop a highly disorganized vascular network driven by VEGF and other pro-angiogenic cues (34). Tortuous, poorly pericyte-covered vessels express low levels of adhesion molecules (ICAM-1, VCAM-1) and generate erratic blood flow, thereby limiting CD8⁺ T-cell arrest and transendothelial migration. Endothelial cells can also up-regulate Fas ligand or PD-L1, actively deleting or silencing incoming effector T cells. Vascular “normalization” with anti-VEGF or angiopoietin-2 blockade restores vessel integrity and markedly boosts intratumoural T-cell entry in pre-clinical and early clinical studies (34). Beyond the endothelium, cancer-associated fibroblasts (CAFs) deposit dense, cross-linked extracellular matrix that forms a physical “mesh”, sequestering T cells at the invasive front (35); CAF-rich, desmoplastic tumors such as pancreas or cholangiocarcinoma are quintessential immune-excluded lesions (36). Targeting CAFs themselves (for example, FAP-directed approaches) or enzymes that remodel collagen can soften this barrier and facilitate T-cell penetration (37).

Effector TILs must engage peptide–HLA-I complexes on target cells. Somatic loss of HLA-I, JAK1/2–IFNGR mutations and tumor-intrinsic WNT/β-catenin signaling all diminish antigen visibility or chemokine output (38, 39). Aberrant WNT/β-catenin suppresses CCL4 and blocks dendritic-cell recruitment; PTEN loss, MYC amplification and stem-like transcriptional programs likewise correlate with immune deserts. Successful immunotherapy may therefore require concurrent targeting of these oncogenic drivers to dismantle “do-not-enter” cues encoded by cancer cells.

Tumor cells and stromal elements secrete a spectrum of factors that dampen T-cell trafficking (40). TGF-β is a central gatekeeper: it stiffens the ECM, suppresses endothelial adhesion molecules and directly curtails T-cell motility; dual blockade of TGF-β and PD-1/PD-L1 can shift immune-excluded tumors toward an inflamed phenotype in mouse and human studies. Additional metabolites—adenosine (via CD39/CD73), kynurenines produced by indoleamine-2,3-dioxygenase (IDO), lactic acid and high extracellular potassium—create a hostile biochemical milieu that impairs T-cell viability and chemotaxis (41). Neutralizing these pathways (e.g., A2A receptor antagonists, IDO inhibitors) is under active clinical evaluation to enhance CD8⁺ T-cell ingress.

Tumor-secreted TGF-β stiffens the ECM, down-regulates endothelial adhesion molecules and directly hampers T-cell motility; dual PD-1/TGF-β blockade can shift immune-excluded tumors toward an inflamed phenotype. Adenosine (CD39/CD73), IDO-derived kynurenines, lactate and high extracellular K⁺ further create a biochemical quagmire that blunts CTL chemotaxis and survival (42, 43). A2A-receptor antagonists and IDO inhibitors are under clinical evaluation to reopen these metabolic choke points.

Continuous antigen exposure, combined with inhibitory receptor engagement, drives CD8⁺ T cells toward a hypofunctional “exhausted” state characterized by diminished cytotoxicity and high expression of PD-1, CTLA-4, TIM-3, LAG-3 and TIGIT (44, 45). Tumor cells, tumor-associated macrophages (TAMs) and endothelial cells up-regulate ligands such as PD-L1 or B7-family molecules, directly silencing T-cell receptor signaling (46). Dual blockade of PD-L1 and TGF-β has shown that relieving co-inhibition and simultaneously dismantling stromal barriers allows expansion of stem-like CD8⁺ precursors and their intratumoural accumulation, converting excluded lesions into responsive, inflamed tumors (47, 48).

Solid tumors reshape nutrient supply and waste removal, creating an extracellular milieu with depleted glucose/amino acids/oxygen and elevated metabolic by-products (e.g., lactate, adenosine, K⁺) that collectively blunt CD8⁺ T-cell proliferation, cytokine production and survival. Quantitative metabolomics of tumor interstitial fluid (TIF) confirms altered metabolite concentrations in situ, providing concentration ranges that map onto T-cell dysfunction (49).

TGF-β is a master suppressor that stiffens the extracellular matrix, down-regulates endothelial adhesion molecules and imprints an exhaustion-prone transcriptional program on CD8⁺ T cells (48, 60); high TGF-β signatures typify immune-excluded tumors and predict poor response to monotherapy ICI (61, 62). VEGF not only fuels aberrant angiogenesis but also induces endothelial “anergy,” further impeding T-cell extravasation (26, 63). Together with IL-10 and prostaglandin-E2, these soluble factors create an anti-inflammatory milieu that limits effector-cell recruitment and survival (64). Combination regimens that co-target VEGF or TGF-β with PD-1/PD-L1 blockade are now actively pursued to dismantle these layered defenses (65, 66).

Regulatory T cells (Tregs), myeloid-derived suppressor cells (MDSCs) and TAMs form a cellular triad that orchestrates resistance to CD8⁺ immunity. Tregs accumulate in hypoxic, adenosine-rich niches and secrete IL-10, TGF-β and granzyme-B to dampen local cytotoxic responses; a high Treg/CD8 ratio correlates with shortened survival across many carcinomas (67). MDSCs deplete arginine and cystine, generate nitric oxide and reactive oxygen species, and express checkpoint ligands, thereby stalling CD8⁺ activation and clonal expansion (68). TAMs of an M2-like phenotype produce IL-10 and PD-L1, phagocytose effector cells via FcγR engagement and remodel stroma to favor immune exclusion (69). Therapeutic depletion or functional reprogramming of these suppressor populations — for example, CCR2/CSF1R blockade for TAMs or arginase/iNOS inhibition for MDSCs — synergizes with checkpoint inhibitors by unleashing CD8⁺ TILs (70, 71).

These overlapping inhibitory layers—checkpoint engagement, metabolic starvation, soluble suppressors and suppressive cell populations—conspire to limit both the quantity and quality of intratumoural CD8⁺ T cells. Rational combination approaches that simultaneously relieve co-inhibition, rectify nutrient stress, neutralize soluble cytokines and curtail suppressor cells are therefore essential to transform immune-”immunotherapy-resistant” tumors into T-cell-inflamed, therapy-responsive states.

Upon T-cell receptor recognition of peptide–MHC-I complexes on tumor cells, activated CD8⁺ T cells release perforin to create transient pores and deliver granzyme B and related serine proteases that trigger caspase-dependent apoptosis (72). Parallel death-receptor pathways—Fas (FasL) and TNF-related apoptosis-inducing ligand—provide redundancy, ensuring cytolysis even when one route is impaired (73). High intratumoural expression of perforin–granzyme transcripts correlates with better survival across carcinomas and can be boosted pharmacologically (e.g., mTOR modulation) to accelerate target-cell elimination (74).

Beyond direct lysis, effector CD8⁺ T cells act as “mobile cytokine factories.” Interferon-γ (IFN-γ) up-regulates tumor MHC-I and components of the antigen-processing machinery (75), amplifies CXCL9/10 chemokine gradients that recruit additional CXCR3⁺ T cells, and exerts anti-angiogenic and anti-proliferative effects on neoplastic and stromal cells (76). Tumor-wide, bystander IFN-γ signaling can propagate hundreds of microns, converting immunologically “cold” niches into inflamed ones (77, 78). Conversely, sustained STAT1/IRF1 activation may drive adaptive resistance (e.g., PD-L1 up-regulation), highlighting the need for balanced cytokine tone (76).

Intratumoural CD8⁺ T cells are heterogeneous. Classical short-lived effector cells provide immediate cytotoxicity but decline rapidly. Tissue-resident memory (T_RM) cells, distinguished by CD69 and CD103, lodge long-term within epithelial niches, secrete IFN-γ “on site,” and independently predict favorable prognosis in multiple solid tumors (79–81). In chronically antigenic TMEs, a hierarchical exhaustion program emerges: a TCF1^high stem-like progenitor pool (TPEX) self-renews and seeds terminally exhausted PD-1^hi TIM-3⁺ cells that retain limited killing capacity (82). Checkpoint blockade preferentially expands TPEX cells, explaining why their baseline frequency foreshadows clinical benefit (83).

Tumour-cell death initiated by CD8⁺ T cells releases danger-associated molecular patterns (DAMPs) and neoantigens that fuel dendritic-cell activation and cross-presentation, a process termed epitope spreading (84). IFN-γ further licenses intratumoural dendritic cells, while TNF-α and GM-CSF reshape myeloid composition toward an M1-like phenotype. These events create a feed-forward circuit that recruits fresh waves of cytotoxic T cells, broadens antigenic breadth, and can ultimately overcome tumor heterogeneity. Therapeutic strategies that sustain this circuit—such as local cytokine delivery, STING agonists, or agents that preserve TCF1⁺ precursors—are under active clinical exploration (85).

Collectively, CD8⁺ T cells exert multifaceted antitumor effects—direct lysis, cytokine-mediated remodeling, and adaptive amplification—whose efficacy depends on maintaining a balanced repertoire of effector, memory, and stem-like subsets within a permissive TME.

Single-cell RNA-sequencing (scRNA-seq) has re-defined intratumoural CD8⁺ T-cell taxonomies, resolving short-lived effectors, tissue-resident memory (T_RM), and the TCF1⁺ stem-like exhausted (T_PEX) progenitors that seed durable responses to immunotherapy (81). Historical development. The first proof-of-concept scRNA-seq was reported by Tang et al. in 2009, analyzing mouse blastomeres with manual micromanipulation (86). SMART-Seq/Smart-Seq2 subsequently improved full-length transcript coverage and sensitivity (87, 88). A transformative leap came in 2015 with droplet microfluidic platforms—Drop-seq and inDrop—which enabled barcoding of thousands of cells in nanoliter droplets (89, 90). Commercial Chromium technology (10x Genomics) further standardized droplet scRNA-seq and introduced unique molecular identifiers (UMIs) for digital transcript counting (91). These advances dropped per-cell costs by >100-fold and unlocked routine immune-profiling of complex tissues ( Figure 3 ).

Recent pan-cancer atlases and disease-focused studies—e.g., in pediatric glioma—demonstrate how shifts in these subsets track with clinical outcome and immune-checkpoint blockade efficacy (92). Integrated scRNA/TCR modalities further couple transcriptional states to clonal evolution, pinpointing which clones breach immune exclusion versus those stalled at the invasive margin.

Early attempts to retain positional gene information relied on laser-capture microdissection coupled with microarrays, limiting throughput and resolution. A conceptual breakthrough came in 2016, when Ståhl et al. arrayed spatially bar-coded oligo-dT spots on glass slides, enabling transcript capture directly from intact tissue sections and co-registration with histology (93). The advent of droplet printing and bead-based strategies soon pushed resolution from 100 µm spots to near-single-cell scales: Slide-seq (94) placed 10-µm DNA-bar-coded beads onto adhesive slides. Commercial Visium (10x Genomics) standardized 55-µm bar-coded spots and paired them with user-friendly software, catalyzing widespread adoption across oncology and immunology. Parallel imaging-based platforms—seqFISH+, MERFISH and Xenium—achieved sub-micron resolution by cyclic in-situ hybridization, but at the cost of limited gene panels, making array-based methods preferable for unbiased chemokine mapping (95, 96) ( Figure 3 ).

Next-generation spatial transcriptomics (ST) overlays gene expression on intact tumor sections at near-single-cell resolution. Analytic frameworks such as ReMiTT trace T-cell migratory paths and identify chemokine “highways” (CXCL9/10, CCL5) as well as stromal “cul-de-sacs” enriched for TGF-β or NOS2/COX2 that impede penetration (97). Combining ST with histology or 3D multiscale modelling charted how TLS-rich niches nucleate CD8⁺ T-cell clusters and forecast prolonged survival across carcinomas (98).

Imaging mass cytometry (IMC) and CODEX antibody cycling now quantify 40–60 proteins per 1-μm pixel, preserving spatial context (99–102). These platforms reveal perivascular “immune hubs” where dendritic cells license newly arrived CD8⁺ T cells, as well as CAF-lined stromal corridors that fence them out. In breast, lung and colorectal cancers, IMC-derived interaction maps of CD8⁺ T cells with NOS2⁺/COX2⁺ tumor islands or PD-L1⁺ macrophage cords stratify responders to PD-1 therapy. Such high-parameter imaging feeds directly into computational tissue atlases that nominate barrier-breaking or TLS-inducing combination regimens.

Early repertoire studies used CDR3‐length “spectratyping,” capturing only a crude size distribution. The first high-throughput TCRβ deep-sequencing was reported by Robins et al. in 2009, using multiplex PCR and 454 pyrosequencing to enumerate >200–000 clonotypes in leukemia patients (103). ImmunoSEQ (Adaptive Biotechnologies, 2011) then standardized bulk-repertoire profiling across thousands of samples. A key leap to single-cell resolution came in 2014 when Stubbington et al. paired SMART-Seq cDNA with nested PCR to recover full α/β chains from individual T cells (104). Droplet microfluidics soon enabled scalable capture: 10x Genomics Chromium V(D)J (2017) bar-codes transcripts and links paired TCRs to whole-transcriptome profiles (91). Recent innovations—VIDJIL-airr, TraCeR2 and spatial bar-coding (Slide-TCR-seq, 2022)—now assign clonotypes to precise tissue coordinates, closing the gap between repertoire and geography.

Bulk and single-cell TCR-seq profile the breadth, depth and spatial provenance of tumor-specific clones. Diverse, expanded TCR repertoires associate with inflamed phenotypes, whereas oligoclonal or non-overlapping repertoires typify deserts (105). Longitudinal TCR tracking exposes clonal replacement after checkpoint blockade and reveals whether new infiltrates originate from peripheral reservoirs or in situ expansion (106, 107). When stitched to ST or IMC, clonotype barcodes register where hot-spot clones accumulate and which stromal routes they exploit or avoid.

Single-cell ATAC-seq and joint RNA/ATAC/TCR platforms chart chromatin accessibility underlying exhaustion trajectories; for example, CXCR4 or Id2 disruption reshapes the epigenetic landscape, delaying T_PEX→terminal transition and enhancing infiltration (108). Machine-learning pipelines that fuse scRNA, ST, multiplex imaging and ATAC layers now predict rate-limiting ligands, metabolic sinks and physical barriers, guiding rational multiplexed interventions (99).

Collectively, these synergistic technologies transform static “snapshot” views into dynamic, multi-scale maps of how CD8⁺ T cells navigate, persist and function within solid tumors—informing precision strategies to ignite, channel and sustain antitumor immunity.

High-resolution single-cell, spatial-omics, proteomic, metabolomic and epigenomic platforms are mature enough to be harmonized through machine-learning pipelines. Emerging spatially aware AI tools—DeepST, CellCharter, and related graph/latent-variable models—integrate single-cell states with tissue context to infer cell–cell communication, niche topology and migratory routes. When coupled to longitudinal sampling, these models can assemble predictive immune atlases that capture the dynamic interplay among T-cell clones, stromal niches and metabolic landscapes across treatment time-points. Such atlases should inform in silico trials, accelerate hypothesis testing, and refine patient-selection algorithms for combination therapy.

Rational “three-layer” regimens are emerging: (i) priming agents that ignite de novo T-cell recruitment in immune-desert tumors (e.g., RNA vaccines, STING or TLR agonists); (ii) barrier-modulating drugs that normalize vasculature or remodel CAF-derived matrix to convert immune-excluded lesions into inflamed ones; and (iii) maintenance strategies—checkpoint blockade, metabolic rewiring, or IL-2/IL-7/IL-15 variants—that sustain stem-like precursors and prevent terminal exhaustion. Synthetic-biology approaches such as logic-gated CAR-T cells, mRNA-encoded cytokine factories and conditionally active bispecific antibodies promise unprecedented spatial and temporal control of effector function while minimizing on-target/off-tumor toxicity.

Non-invasive biomarkers (circulating TCR clonotypes, cell-free RNA/DNA, metabolic tracers) should be linked to AI-derived spatial features to track infiltration kinetics and functional states during therapy. Iterating these signals into patient-specific digital twins may enable adaptive dosing and early switching between priming, barrier-modulating, and maintenance layers.

To move AI-enabled spatial immunology into the clinic, several hurdles must be addressed:(a) Model robustness & generalizability: pre-specify training/validation datasets, perform cross-site testing, and quantify batch effects across platforms and staining protocols; (b) Interpretability & actionability: provide saliency on which spatial features (e.g., perivascular hubs, TLS density, CAF corridors) drive predictions, and map them to trial-eligible interventions; (c) Data governance & privacy: adopt harmonized ontologies and secure data standards; consider federated or privacy-preserving learning for multi-center studies; (d) Regulatory pathway: define software-as-a-medical-device requirements, version control, drift monitoring, and prospective performance benchmarks aligned with clinical endpoints; (e) Reproducible pipelines: containerize end-to-end workflows from raw images/sequencing to clinical reports; release validation kits for external laboratories.

By fusing AI-driven multimodal integration with mechanism-guided therapeutic engineering—and by building a clear regulatory path—the field can transform immune-cold tumors into immune-hot, treatment-sensitive diseases. Success will hinge on interoperable data, interpretable models, and prospective trials that treat computational predictions as testable, patient-benefitting hypotheses.