To orient cross-disciplinary readers, we briefly define the central clinical states referenced throughout this progression-based review and highlight representative biological and immunological features. Localized hormone-sensitive PCa is organ-confined and typically remains androgen receptor (AR)-driven; across cohorts, the tumor immune microenvironment is often immune-cold on average, with substantial spatial heterogeneity in immune infiltration and functional states (9–11). mCSPC denotes systemic dissemination while disease remains responsive to androgen deprivation therapy (ADT), and metastatic niches—particularly bone—commonly display myeloid-enriched, T-cell-excluded, or dysfunctional immune contexts (8–10). mCRPC is defined by disease progression despite castrate testosterone levels and may manifest as rising prostate-specific antigen (PSA), radiographic progression, and/or new metastases; under the selective pressure of ADT and AR-targeted agents, mCRPC frequently exhibits immunosuppressive remodeling characterized by expanded suppressive myeloid compartments and impaired T-cell effector function (10, 12, 13). Finally, NEPC, including treatment-emergent NEPC, represents lineage plasticity toward an AR-indifferent phenotype and is increasingly recognized in contemporary treatment settings; this phenotype can be associated with altered target expression and additional constraints on effective antitumor immunity, with important implications for immunotherapy responsiveness and trial stratification (14, 15). A concise summary table (Table 1) is provided to serve as a roadmap for the stage-specific sections below.

NEPC represents a clinically aggressive phenotype that can arise de novo or, more commonly in contemporary practice, as treatment-emergent NEPC (t-NEPC) under selective pressure from androgen deprivation and potent AR-targeted therapies (14). Biologically, NEPC reflects lineage plasticity toward an AR-indifferent state, often accompanied by neuroendocrine differentiation programs and distinct transcriptional/epigenetic regulation; clinically, this transition can coincide with rapid progression, atypical metastatic patterns, and reduced reliance on PSA kinetics (14, 16). Significantly, NEPC evolution may reshape therapeutic vulnerabilities and biomarker performance, including reduced prostate-specific membrane antigen (PSMA) expression in subsets, which can complicate imaging- and target-based strategies (15). From an immunotherapy perspective, NEPC is often considered an “immune-cold” context, and lineage transition may further constrain effective antitumor immunity by altering antigen presentation and tumor–immune interactions, contributing to limited responsiveness to immune checkpoint blockade in many patients (14–16). These features underscore the need for state-aware stratification and rational combinations (e.g., approaches that bypass impaired priming via T-cell redirection, and/or strategies that modulate suppressive myeloid niches) when considering immunotherapy across the PCa continuum (15, 16).

Limited PCa is often considered to have few immune cells, but the integrated single-cell and spatial transcriptomic maps show significant intratumoral variation underlying the “low overall invasion”. In the same prostate sample, the immune rejection gland region coexisted with a stromal interface rich in T cells and myeloid cells. The ligand–receptor network suggests that organized epithelial–fibroblast–myeloid cell crosstalk maintains local immune stagnation (17–19). These spatial neighborhoods are clinically significant because sampling can miss a small range of immunoreactive foci, and the dominant barriers in different regions may differ (physical rejection vs. functional inhibition). Therefore, even in early PCa, it should be regarded as a puzzle of multiple immune niches rather than a heterogeneous cold tumor.

A considerable number of primary tumors contain B-cell clusters and TLS. Recent spatial studies have emphasized that TLS is also heterogeneous in location and maturity. Immature TLS showed loose B/T-cell aggregation, while mature TLS showed a structure similar to the germinal center, accompanied by stronger antigen presentation and effector T-cell characteristics; they can coexist in the same tumor, suggesting different local developmental trajectories (20–22). Clinically, the TLS-high/mature TLS region is similar to the immune-inflammatory microenvironment in ICI-responsive tumors, supporting the use of neoadjuvant or local immunotherapy in carefully selected patients for a limited period, rather than for all patients.

When PCa enters metastasis but is still sensitive to hormone, the immune diversity expands along a new axis—the differences between different lesions are significant. Different metastatic sites within the same patient often exhibit distinct immune density and composition, suggesting that diffusion is accompanied by site-specific microenvironment imprinting and a “Founding” immune state (17, 18). This has practical consequences: a single biopsy lesion cannot represent the overall immune landscape of patients, and treatment response may also be inconsistent across lesions. Therefore, multisite heterogeneity should be considered in the early joint scheme, especially when selecting monitoring or irradiation lesions.

Standard treatment will actively reshape the immune state, rather than only select drug-resistant clones. ADT can rapidly transform some lesions into a more inflammatory environment, increase the number of activated CD8⁺ T cells, and expand inhibitory regulatory T (Tregs) and myeloid cells (23, 24). In the metastatic castration-sensitive stage, the combination of programmed cell death protein 1 (PD-1) blockade and ADT can induce intense immune infiltration and interferon-related transcription programs in some patients, demonstrating the available “heating effect” (23). The treatment sequence is also important: a sequential study of ADT and vaccine shows that different sequences can alter the balance between the initiation of effects and compensatory inhibition (25). This evidence indicates that the immune heterogeneity of PCa is dynamic and treatment-dependent, and that the inflammatory window in the early posttreatment period is a key opportunity for rational combination therapy.

After long-term androgen axis inhibition, the immune landscape usually evolves toward more potent inhibition and greater heterogeneity. Multigroup studies showed an inflammatory monocyte/MDSC-like status and an amplification of the inhibitory tumor-associated macrophage(s) (TAM) program, while CD8⁺ T cells were depleted and spatially excluded from the gland (18). Importantly, mCRPC is not just “colder” but is differentiated into multiple myeloid-dominated inhibition niches, each with distinct reversibility and specific inhibition circuits. In clinical transformation, this means that T-cell-oriented therapies (ICI, vaccines, T-cell connector (TCE), etc.) often need supporting strategies to dismantle the lesion-specific myeloid barrier.

CRPC generates treatment-related NEPC through lineage plasticity, which is usually characterized by sparse adaptive immune infiltration and reduced antigen presentation. However, residual congenital inflammatory changes can still be observed across different lesions (26). Bone metastasis is the leading lethal site, forming a more highly specialized immune organ. Single-cell analysis of bone metastases revealed an inhibitory myeloid ecology of bone marrow imprinting, and depleted T cells were embedded within a chemokine field that promoted dysfunction (8). The recurrent and operable pathway is the CCL20–CCR6 axis: the excessive production of CCL20 by myeloid cells in bone marrow drives CCR6⁺ T cells into a tolerant state; blocking this axis can restore T-cell activity and improve survival in the model (8, 27, 28). In addition, heterogeneity in HLA class I expression across lesions can regulate immune visibility, which may explain why, even in a population selected for biomarkers, some metastatic lesions remain treatment-resistant (29). Clinically, these findings support bone-specific immune stratification and suggest that effective systemic immunotherapy must be tailored to the biology of metastatic sites. Together, these stage- and site-dependent immune ecosystems can be viewed as an integrated continuum; an overview is provided in Figure 1.

PCa significantly dictates its tumor immune microenvironment. Conventional acinar adenocarcinoma (AA) typically represents an “immunologically cold” phenotype, characterized by limited T-cell infiltration and a “lymphocyte-excluded” pattern, where immune cells are physically sequestered in the periglandular stroma (30). In contrast, ductal adenocarcinoma (DA) and intraductal carcinoma of the prostate (IDC-P) exhibit a more aggressive and immunosuppressive niche. Recent spatial profiling has revealed that DA, despite higher immune cell densities in some instances, often harbors a higher proportion of dysfunctional CD8⁺ T cells and M2-type macrophages, driven by genomic instabilities such as PTEN loss and BRCA2 alterations (31). Furthermore, IDC-P has been identified as a distinct “hot” yet highly suppressed niche, characterized by significantly increased recruitment of Tregs and myeloid-derived suppressor cells (MDSCs) compared to adjacent acinar components (32, 33). This lineage-specific heterogeneity suggests that, while AA may require strategies to “heat up” the tumor, aggressive variants such as DA and IDC-P may necessitate combination therapies targeting myeloid-driven suppression or T-cell exhaustion (34).

Myeloid cells are the most abundant and transcriptionally diverse immune components in PCa, with limited metastatic potential and therapeutic tolerance. Single-cell and spatial analysis distinguished a variety of TAM states (including SPP1^Hi, inflammatory, interferon-responsive, etc.), MDSC-like monocytes, and neutrophil/tan programs. These subgroups occupied distinct spatial neighborhoods and exerted distinct immune-regulatory effects (18, 35, 36). Functionally, they synergistically inhibit antigen presentation, suppress T cells, and produce inhibitory metabolites in a lesion-specific manner, which explains why “myeloid-high PCa” itself is also a heterogeneous clinical entity. Space studies further showed that the myeloid-rich areas were often aligned with fibroblasts and ECM-rich, high-density matrix areas, forming a physical and biochemical barrier that captured T cells at the edge of the tumor rather than allowing them to enter the gland (17–19, 36). The diversity of multiple myeloid systems determines the reversibility of different lesions to targeting strategies. It also explains why some lesions in the same patient respond to immunotherapy, while others remain tolerant.

One of the key findings in recent years is that androgen axis destruction and tumor stress can amplify the inflammatory cytokines/chemokines program and actively recruit inhibitory myeloid cells. Interleukin (IL)-8 (CXCL8) and its receptors, CXCR1/2, serve as core mediators linking inflammation and immune escape: IL-8 promotes neutrophil/MDSC chemotaxis, supports tumor cell survival and invasion, and is associated with androgen-independent progression (37–39). In the transformation queue, castration or AR blockade can enhance IL-8 signaling, leading to increased myeloid infiltration, decreased antigen-presenting tone, and increased immunosuppression (38, 40). This establishes a clinically operable logic. Although some lesions are “more inflamed” after treatment, the inflammation tends to be myeloid rather than T-cell-mediated, thus turning the potential “heat” into adaptive resistance. Moreover, IL-8/C-X-C chemokine receptor 2 (CXCR2) upregulation was not consistent across all lesions, further indicating that inflammatory reprogramming was an essential source of heterogeneity between metastases.

In addition to classical cytokines, treatment-induced cellular senescence is increasingly regarded as the driving factor of mCRPC immune heterogeneity and drug resistance. Aging tumor or stromal cells release an aging-related secretory phenotype (SASP) rich in proinflammatory factors, which recruits and polarizes myeloid cells into an immunosuppressive state (41–43). A landmark clinical transformation study showed that age-related myeloid inflammation promoted the progression of metastatic CRPC and blocked resistance to AR inhibition of myeloid chemotaxis, which can reduce inhibitory infiltration and bring lasting benefits in some patients (37). Conceptually, this supports the “inflammation switch” model of PCa resistance: treatment triggers senescence/SASP → SASP recruits inhibitory myeloid → myeloid, in turn, enhances tumor survival and immune escape. It is worth noting that aging is reversible in some situations, meaning that blocking the SASP–myeloid circuit may reopen the immune control window and restore the sensitivity of subsequent immune activation (41, 43).

The bone marrow niche will amplify the process described above. In bone metastases, macrophages and inflammatory monocytes are dominant, and depleted T cells are embedded within a chemokine field that renders them dysfunctional (8, 26, 27). The recurrent pathway is CCL20–CCR6 signaling: intramedullary myeloid cells produce CCL20, while CCR6⁺ T cells are tolerant; genetic or pharmacological blockade can restore T-cell activity and improve the survival of the model (8, 26). The clinical inference is that bone marrow-dominant lesions naturally tend to fail immunotherapy unless bone-specific myeloid circuits are neutralized at the same time. This also provides a mechanism explanation for the inconsistent response between lesions in systematic treatment, highlighting the necessity of using the metastasis site as a stratified variable in the experiment.

Even in the presence of T cells, many PCa lesions still showed decreased HLA class I expression, limiting antigen presentation and facilitating immune escape. This phenomenon differs significantly across lesions and may be reversible via the epigenetic pathway (29, 44–46). Simultaneously, hypoxia and changes in lipid/adenosine metabolism form a spatially limited “functional cold zone”, which inhibits T-cell killing and cooperates with myeloid suppression (36). These immune stealth and metabolic binding layers explain why it is not enough to improve immune infiltration alone, unless the dominant inhibitory circuit (usually the medullary center, maintained by inflammation) is disassembled synchronously. Clinically, this strengthens the “two-step” strategy: immune visibility/effect fitness is first restored, and immune effect immunity is then amplified, rather than assuming that deep heterogeneity can be overcome simply by releasing checkpoints.

The circuits highlighted in this section are readily translatable to combination-trial design and correlative biomarker strategies. For instance, IL-8/CXCR2-driven neutrophil and MDSC recruitment can be monitored in peripheral blood (e.g., circulating IL-8 levels and the neutrophil-to-lymphocyte ratio), and elevated systemic or tumor-associated IL-8 has been linked to reduced clinical benefit from PD-(L)1 blockade, supporting its use as a clinically relevant biomarker/readout (40). These blood- and tissue-based readouts also support pharmacodynamic endpoints for CXCR1/2 blockade or IL-8-axis targeting, including rational combinations with ICIs (47, 48). In parallel, senescence/SASP-associated inflammatory programs can reinforce suppressive myeloid states and impair antitumor T-cell function, providing additional rationale for senescence-targeting (senomorphic/senolytic) components in combination regimens (49).

Likewise, lesion-level heterogeneity in antigen presentation—particularly human leukocyte antigen class I (HLA-I)/B2M loss—has been implicated in acquired resistance to PD-1 blockade, underscoring the need to assess baseline antigen presentation capacity when selecting immunotherapy strategies (50). Such profiling can help guide the choice between priming-dependent approaches (e.g., ICI or cancer vaccines) versus priming-bypassing strategies (e.g., bispecific T-cell engagers or chimeric antigen receptor T cell(s) (CAR-T) therapies) (51). Finally, hypoxia- and adenosine-associated immunometabolic suppression and TGF-β-driven immune exclusion provide a translational rationale for incorporating adenosine/CD73-axis or TGF-β-targeting add-ons into rational combinations to overcome “functional cold” barriers (52, 53).