A conceptual blueprint for “turning cold to hot” in Osteosarcoma: from TME stratification hypotheses to adaptive therapeutic prospects

Osteosarcoma (OS) is a quintessential “cold tumor,” and outcomes for patients with metastatic or recurrent disease have remained poor for decades. The failure of immune checkpoint inhibitors (ICIs) in OS reflects a multilayered immunosuppressive architecture rather than a single dominant lesion. This review deconstructs three principal barriers within that architecture: (1) physical T-cell exclusion driven by a dense fibrotic stroma and aberrant vasculature; (2) a myeloid-dominant suppressive network enriched for Tumor-Associated Macrophages (TAMs) and Myeloid-Derived Suppressor Cells (MDSCs); and (3) Antigen Presentation Machinery (APM) defects, most commonly loss of MHC-I/B2M. In parallel, the primary tumor actively engineers the pulmonary environment through exosomes and Neutrophil Extracellular Traps (NETs), establishing a Pre-metastatic Niche (PMN) that facilitates lung metastasis.

Integrating evidence from single-cell and spatial omics, multi-modal imaging (radiomics, digital pathology), and liquid biopsy (ctDNA-minimal residual disease [MRD]), this review translates biological “decoding” of the OS microenvironment into a hypothesis-generating operational framework. We propose an “OS-TME Subtyping V1.0” model in which subtypes are treated as dynamic, dominant-barrier system states rather than fixed biological classes. In its current conceptual form, state assignment is envisioned as a semi-structured, rule-based process using concordant signals from pathology/spatial readouts, imaging surrogates, and ctDNA/immune context; mixed or discordant cases are intentionally retained as indeterminate states for reassessment rather than forcibly classified. On this basis, we outline a sequential “De-suppression → Priming → Checkpoint” logic tailored to different barrier-dominant states. For myeloid-dominant states, we prioritize myeloid reprogramming (e.g., CSF1R/CCR2-axis targeting) combined with immunogenic priming. For dense fibrotic stroma/angio-abnormal states, we emphasize up-front vessel/stroma remodeling before checkpoint therapy. For APM-defective states, we discuss MHC-independent approaches targeting B7-H3 or GD2 (e.g., CAR-T/NK cells, antibody-drug conjugates), while explicitly acknowledging target heterogeneity, trafficking barriers, and on-target/off-tumor risk.

To narrow the translational gap, we further outline a perioperative “Window of Opportunity” (WoO) trial prototype and a conceptual “Cold-to-Hot Readiness Index (RI)” that integrates dynamic imaging, pathology, and MRD monitoring. The RI is presented only as an illustrative, hypothesis-generating summary variable intended for retrospective stratification, simulation modeling, or biomarker-guided early-phase trial design, rather than near-term routine clinical decision-making. Together, these elements define a theoretical blueprint for iterative state assessment and adaptive therapeutic sequencing in osteosarcoma.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12964-026-02883-3.

Highlights

ICI resistance in Osteosarcoma (OS) originates from a triad of barriers—“dense fibrotic stroma/vasculature,” “myeloid suppression,” and “APM defects”—and is compounded by active shaping of the pulmonary pre-metastatic niche.

We propose a conceptual “OS-TME V1.0” framework that treats subtypes as dynamic dominant-barrier states and uses a minimal viable panel of imaging, pathology/spatial readouts, and ctDNA to support hypothesis-generating stratification.

A sequential “De-suppression → Priming → Checkpoint” logic is discussed in an evidence-tiered manner, distinguishing OS-supported pathways from cross-tumor extrapolations and matching APM-defective tumors to MHC-independent strategies only with appropriate caution about heterogeneity and toxicity.

We outline a perioperative “Window of Opportunity” (WoO) prototype and a conceptual “Cold-to-Hot Readiness Index (RI)” for short-interval reassessment and adaptive sequencing; both are proposed as exploratory research tools rather than validated clinical algorithms.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12964-026-02883-3.

Methods

Data sources and search databases

A systematic search was conducted across multiple major international academic databases (PubMed/MEDLINE, Web of Science (WOS) Core Collection, Scopus, Embase, IEEE Xplore, ClinicalTrials.gov) for the period from January 1, 2015, to August 1, 2025. To ensure the completeness of key concepts, a small number of foundational articles published before 2020 were also considered. This study is a narrative review, supplemented by a systematic search methodology to enhance coverage and transparency. A unified definition of terms and abbreviations used in this paper is provided in Supplementary Table S13.

Search strategy and keywords

The search strategy employed Boolean operators (AND, OR) to combine multidimensional keywords. The keywords were primarily divided into five modules: (P) Population, (I) Intervention/Focus, (C) Components & Mechanisms, (T) Technology, and (S) Strategies. All searches were limited to English-language literature.

• (P) Population: (“Osteosarcoma” OR “OS” OR “Osteogenic Sarcoma” OR “Bone Sarcoma” OR “Bone Neoplasms”)

• (I) Intervention/Focus: AND (“Tumor Microenvironment” OR “TME” OR “Immune Microenvironment” OR “Cold Tumor” OR “Immunosuppressive Microenvironment” OR “Immune Evasion” OR “Immunotherapy Resistance” OR “Immune Checkpoint Inhibitor” OR “ICI”).

• (C) Components & Mechanisms: AND (“Myeloid-Derived Suppressor Cells” OR “MDSC” OR “Tumor-Associated Macrophages” OR “TAMs” OR “Myeloid Cells” OR “Neutrophils” OR “Treg” OR “T cell exhaustion” OR “Antigen Presentation” OR “Antigen Processing” OR “MHC” OR “B2M” OR “HLA” OR “Stroma” OR “Cancer-Associated Fibroblasts” OR “CAF” OR “Extracellular Matrix” OR “ECM” OR “Collagen” OR “Mechanobiology” OR “Stiffness” OR “DDR1” OR “YAP” OR “Metabolism” OR “Metabolic Reprogramming” OR “Hypoxia” OR “HIF-1alpha” OR “Lactate” OR “Kynurenine” OR “AhR” OR “Adenosine” OR “CD39” OR “CD73” OR “Ferroptosis” OR “RANKL” OR “RANK” OR “Pre-metastatic Niche” OR “PMN” OR “Exosomes” OR “Structural Variation” OR “Chromothripsis”).

• (T) Technology: AND (“Single-Cell” OR “scRNA-seq” OR “Spatial Omics” OR “Spatial Transcriptomics” OR “Multi-omics” OR “Radiomics” OR “Digital Pathology” OR “Artificial Intelligence” OR “Machine Learning” OR “ctDNA” OR “Liquid Biopsy”).

• (S) Strategies: AND (“Combination Therapy” OR “Turn Cold to Hot” OR “Radiotherapy” OR “Immunogenic Cell Death” OR “ICD” OR “Anti-angiogenic” OR “Targeted Therapy” OR “CAR-T” OR “NK cells” OR “B7-H3” OR “GD2” OR “Antibody-Drug Conjugate” OR “ADC” OR “Bispecific Antibody” OR “Tumor Vaccine”).

Inclusion and exclusion criteria

Inclusion Criteria

• Article Types: Original research articles (including preclinical and clinical studies), Systematic reviews, Meta-analyses, and in-depth Narrative reviews published in high-impact journals.

• Content: Studies related to the osteosarcoma immune microenvironment, immune resistance mechanisms, biomarker discovery, multi-omics analysis, or novel immune/combination therapy strategies.

Exclusion Criteria:

• Non-English language literature.

• Article Types: Case reports, short communications (unless containing critical data), letters to the editor, editorials, conference abstracts (except for clinical trial data), or protocols-only papers.

• Duplicate publications or literature with overlapping data.

Introduction

Osteosarcoma (OS) is the most prevalent primary malignant bone tumor among children, adolescents, and young adults (AYA) [1–4]. Originating from mesenchymal stem cells, its histological hallmark is the direct formation of immature osteoid matrix by neoplastic cells [5]. Despite standardized multimodal treatment regimens—comprising multi-agent neoadjuvant chemotherapy, radical surgical resection, and adjuvant chemotherapy—clinical outcomes for OS patients, particularly those with metastatic or recurrent disease, have stagnated for nearly four decades [6]. While the 5-year Overall Survival for localized OS is approximately 60% to 70%, this rate plummets to below 20% to 30% upon the onset of pulmonary metastasis (the most common site) or disease relapse [7, 8]. This clinical reality underscores the limitations of current paradigms and highlights a profound unmet clinical need.

Over the past decade, Immune Checkpoint Inhibitors (ICIs), particularly antibodies targeting PD-1/PD-L1 and CTLA-4, have revolutionized the standard of care for multiple advanced malignancies, including melanoma, non-small cell lung cancer, and renal cell carcinoma, delivering durable clinical benefits [9, 10]. This success, however, has not been replicated in OS. Multiple high-profile clinical trials, including the SARC028 study (evaluating Pembrolizumab) and investigations of Nivolumab combined with Ipilimumab, uniformly demonstrated exceptionally low objective response rates (ORR) in unselected OS patients, typically below 5% (Supplementary Table S1-2) [11, 12]. This profound primary resistance unequivocally classifies OS as a canonical “cold tumor,” also termed a “non-T-cell-inflamed” phenotype, whose Tumor Microenvironment (TME) is characterized by a deeply immunosuppressive and immune-excluded state at structural, cellular, and functional levels [2, 13, 14].

The “cold” phenotype of OS is not attributable to a single mechanism but rather a multidimensional biological architecture that collectively erects a formidable barrier against immune surveillance and therapeutic intervention. First, at the cellular level, the OS TME is characterized by ‘T-cell exclusion’ or a ‘T-cell desert’ phenotype, exhibiting extremely sparse infiltration of effector T cells. Conversely, the TME is densely populated by profoundly immunosuppressive myeloid cells, particularly M2-like Tumor-Associated Macrophages (TAMs) and granulocytic-Myeloid-Derived Suppressor Cells (g-MDSCs) [2, 15–17]. Second, OS possesses a unique genomic landscape. It is not characterized by a high tumor mutational burden (TMB) but rather a low-to-moderate TMB accompanied by a high frequency of complex Structural Variations (SVs), Copy Number Variations (CNVs), and chromothripsis [18–20]. While this genomic instability can generate neoantigens, it more frequently results in defects in the Antigen Presentation Machinery (APM), such as the downregulation or functional loss of MHC class I (MHC-I) molecules (including B2M), rendering tumor cells “invisible” to CD8 + T cells [21, 22]. Third, as a bone-derived tumor, the OS TME features a unique “dense fibrotic matrix” barrier, rich in a dense collagen network, an abnormal vascular system, and extensive calcification [23, 24]. This high-stiffness physical matrix not only physically impedes T-cell infiltration and migration via mechanobiology signals (e.g., the DDR1/YAP pathway) but also actively participates in immunosuppression through its unique bone-remodeling microenvironment (e.g., the RANKL-RANK signaling axis) [25–28]. Finally, the OS TME constitutes a metabolically hostile niche. Profound hypoxia (HIF-1α activation), high glycolysis (lactate accumulation), and amino acid catabolism (e.g., the Tryptophan-Kynurenine-AhR pathway) converge to suppress T-cell effector functions while simultaneously nurturing suppressive myeloid cells and Tregs [29, 30] (Fig. 1).

Fig. 1.

The clinical dilemma and the “Tri-Barrier” hypothesis of immune resistance in Osteosarcoma (OS)

Understanding the intricate interplay of these mechanisms is a prerequisite for developing effective immunotherapies. Monotherapy with ICIs is clearly insufficient to overcome the multifactorial immune barriers of OS. Consequently, academia and industry are actively exploring more complex “cold-to-hot” combination strategies. Leveraging cutting-edge technologies such as single-cell RNA-sequencing (scRNA-seq), spatial multi-omics, and Artificial Intelligence (AI) to decode the OS TME at high resolution has become a major path toward identifying novel therapeutic targets and precision biomarkers [31–35]. These technologies are helping map the immune-stromal-metabolic interactome of OS with unprecedented resolution and define clinically relevant barrier states that may inform individualized combination therapies [36]. This review aims to systematically elucidate the biological underpinnings of the OS “cold tumor” microenvironment, from myeloid-dominant suppression and APM defects to stromal/mechanical and metabolic barriers. We focus on TME stratification, translational sequencing logic, and biomarker-enabled monitoring, while emphasizing that the OS-TME states, RI, and WoO blueprint proposed here are dynamic, operational hypotheses derived from current biological evidence. They are intended to sharpen discussion, support biomarker-guided trial design, and provide a conceptual starting point for future validation—not to serve as a current clinical standard or autonomous treatment algorithm. For ease of reading, we provide a list of acronyms and term explanations in the supplementary materials (Supplementary Table S13).

Main

The interwoven immune-stromal-vascular network of OS and the “cold tumor” etiology

The Osteosarcoma (OS) Tumor Microenvironment (TME) is a highly dynamic, complex, and spatially heterogeneous ecosystem. Its “cold tumor” phenotype is the ultimate manifestation of a multidimensional dialogue among immune cells, stromal components, vascular networks, and the tumor cells themselves [15, 37, 38]. Unlike “hot tumors,” which are highly infiltrated by T cells, the OS TME is structurally characterized as “immune-excluded” or “immune-desert” [38–40]. In the “immune-excluded” TME, T cells are present in the tumor periphery but are blocked from the tumor parenchyma by a physical and chemical barrier; in the “immune-desert” TME, T cells are entirely absent. The formation of this spatial architecture is the collective result of OS’s unique tissue of origin, its highly dense fibrotic stroma, and its abnormal vascular network (Fig. 2).

Fig. 2.

Barrier I (Physical): “Immune Exclusion” - stroma, vasculature, and mechanobiology

The OS stroma is not merely a physical scaffold but an active architect of immunosuppression [15, 41]. The OS TME is typically enriched with highly cross-linked type I collagen, forming a dense and stiff Extracellular Matrix (ECM). This high matrix stiffness profoundly remodels the TME through mechanobiology signaling pathways. On one hand, the stiff matrix physically impedes T-cell migration and infiltration by activating integrins and their downstream RhoA/ROCK pathways [42–46]. On the other hand, mechanical signals (like stiffness) are sensed by tumor cells and stromal cells, such as Cancer-Associated Fibroblasts (CAFs), particularly via the YAP/TAZ transcriptional co-activators [47–50]. YAP/TAZ activation not only promotes tumor proliferation and drug resistance but also induces the secretion of a suite of immunosuppressive cytokines (e.g., CXCL5, CCL2), which further recruit suppressive myeloid cells [51–53]. Furthermore, collagen receptors highly expressed in the TME, such as Discoidin Domain Receptor 1 (DDR1), activate non-canonical NF-κB signaling upon sensing collagen, promoting the expression of immunosuppressive ligands (like PD-L1) and restricting T-cell infiltration [25, 54]. This “mechano-immune” coupling, driven by matrix stiffness, is one of the core reasons the OS TME is so recalcitrant to T-cell attack.

As a primary bone tumor, the OS TME is deeply influenced by the bone remodeling microenvironment, where the RANKL-RANK-OPG axis serves as a critical bridge linking bone destruction and immunosuppression [55, 56]. OS tumor cells and TME stromal cells (e.g., osteoblasts) highly express Receptor Activator of Nuclear Factor-κB Ligand (RANKL). RANKL binds to its receptor RANK (expressed on osteoclast precursors and certain immune cells), promoting osteoclast differentiation and bone lysis [56–58]. More importantly, the RANKL-RANK signaling axis is proven to have potent immunomodulatory functions. RANKL can directly act on regulatory T cells (Tregs), promoting their proliferation and suppressive function [55, 59]; simultaneously, it induces a tolerogenic phenotype in dendritic cells (DCs), impairing their antigen presentation and T-cell activation capabilities [59–61]. By secreting RANKL, tumor cells not only dissolve bone to create space for their own growth but also actively reshape the TME toward immune tolerance.

The abnormal vascular network within the TME is another critical factor in OS immune exclusion [62–64]. OS tumor vessels are typically structurally chaotic and functionally defective, characterized by high permeability, low perfusion, and extensive hypoxic regions [65–67]. This disorganized vascular system cannot efficiently deliver circulating effector T cells to the tumor parenchyma. Furthermore, abnormal tumor endothelial cells (TECs) actively form a “vascular immune barrier” by highly expressing molecules like PD-L1 and FasL, which induce T-cell apoptosis or exhaustion [68, 69]. Concurrently, the hypoxic environment (HIF-1α activation) and aberrant angiogenic signals (e.g., VEGF) synergize to drive the robust recruitment and polarization of suppressive myeloid cells, especially TAMs and MDSCs [70], further exacerbating the TME’s immunosuppressive state. Notably, the microcalcifications commonly found in the OS TME are not just a pathological feature of osteoid formation; recent studies suggest these calcific deposits may also exert complex control over the immune microenvironment by activating macrophage inflammasomes (e.g., NLRP3) and promoting the release of pro-inflammatory (yet often pro-tumorigenic) cytokines like IL-1β [71–74]. In summary, the unique bone matrix and vascular abnormalities of OS collectively erect a physical barrier that is nearly insurmountable for T cells. More importantly, this physical and chemical barrier—especially the resulting widespread hypoxia (activating HIF-1α) and high matrix stiffness (activating YAP/TAZ)—serves as a critical upstream driver for the subsequent myeloid suppression.

This immunosuppressive niche at the primary site does not exist in isolation. It actively shapes the distant pulmonary Pre-metastatic Niche (PMN) through systemic communication, particularly via exosomes [75, 76]. Exosomes secreted by OS cells carry specific proteins, lipids, and non-coding RNAs (e.g., miRNAs), enabling them to “educate” lung stromal cells (like fibroblasts) and immune cells (like alveolar macrophages), reprogramming them into phenotypes favorable for tumor colonization [77–80]. For example, OS-derived exosomes can induce pulmonary fibroblasts to secrete pro-fibrotic factors and recruit MDSCs and M2 macrophages, establishing an immunosuppressive “soil” in the lungs [2, 81]. This exosome-mediated PMN formation, driven by the stromal-vascular-immune network of the primary TME, explains OS’s high propensity for early lung metastasis and implies that effective disease control must target both the primary tumor and the distant niche simultaneously.

Direct evidence for spontaneous or therapeutically inducible Tertiary Lymphoid Structures (TLS) in primary OS remains limited. Therefore, TLS induction is discussed here only as an exploratory future direction. In the current framework, spatial pathology or multiplex IHC assessment of TLS-like aggregates may serve, at most, as a correlative readout in WoO studies rather than an expected near-term therapeutic endpoint (see Supplementary Table S11 (Table 2) and Supplementary Table S10).

Myeloid dominance and APM defects: from T-Cell exclusion to functional failure

As previously mentioned, the physical and chemical barriers described in Sect. (The interwoven immune-stromal-vascular network of OS and the"cold tumor" etiology) (especially hypoxia and matrix stiffness) actively shape the next layer of TME suppression: an immunosuppressive network that is dominant in both number and function, and a molecular defect that renders T cells ‘ineffective’ [2, 82–84]. This “cold tumor” microenvironment’s core characteristic stems from the synergy of two major mechanisms: first, the TME is dominated by suppressive myeloid cells that actively engineer a niche for T-cell functional failure; second, OS tumor cells frequently harbor defects in the Antigen Presentation Machinery (APM), rendering them ‘invisible’ to T cells [21, 85, 86].

One of the most striking features of the OS TME’s cellular composition is its T-cell-poor and myeloid-rich nature [87, 88]. Among these myeloid cells, M2-phenotype Tumor-Associated Macrophages (TAMs) and granulocytic-Myeloid-Derived Suppressor Cells (g-MDSCs) are predominant [89]. The accumulation of these populations is not coincidental; it is driven by a chemokine network actively secreted by OS tumor cells and stromal cells. It must be emphasized that this myeloid accumulation is largely a passive response to the TME stress signals mentioned in (The interwoven immune-stromal-vascular network of OS and the"cold tumor" etiology). For instance, hypoxia is one of the strongest activators of HIF-1α, which not only drives VEGF (worsening vascular abnormality) but also potently drives the recruitment and M2 polarization of TAMs and MDSCs. Similarly, the high matrix stiffness mentioned in (The interwoven immune-stromal-vascular network of OS and the"cold tumor" etiology) actively induces the secretion of chemokines like CCL2 and CXCL8 via the YAP/TAZ pathway, directly providing the ‘recruitment signal source’ for the myeloid cells described in this section. Among these, the Colony-Stimulating Factor 1 (CSF1) and its receptor CSF1R axis is a critical “master switch” for TAM survival, differentiation, and M2 polarization [90–92]. OS cells highly express CSF1, continuously recruiting circulating monocytes and “educating” them into suppressive M2-TAMs. These TAMs, in turn, promote tumor growth by secreting factors like EGF and IL-6, forming a vicious cycle [93]. In parallel, the CCL2-CCR2 signaling axis is another core pathway mediating the recruitment of monocytes (TAM precursors) and MDSCs to the tumor site [94]. Furthermore, the CXCL8 (IL-8) and its receptors (CXCR1/2) axis is proven to be key for the specific recruitment of g-MDSCs and Tumor-Associated Neutrophils (TANs) [95]. These recruited myeloid cells form the “first line of defense” in the TME, efficiently suppressing T-cell anti-tumor responses through multiple mechanisms.

Myeloid-mediated immunosuppression is multidimensional. First, both TAMs and MDSCs highly express immune checkpoint ligands, such as PD-L1 and PD-L2, which directly bind to PD-1 on infiltrating T cells to deliver inhibitory signals [85, 96]. Second, myeloid cells are key executors of metabolic reprogramming in the TME. MDSCs, particularly g-MDSCs, highly express Arginase-1 (ARG1) and inducible Nitric Oxide Synthase (iNOS) [97]. ARG1 depletes L-arginine, an amino acid essential for T-cell proliferation, leading to T-cell receptor (TCR) ζ-chain downregulation and cell cycle arrest. Meanwhile, iNOS produces reactive nitrogen species (RNS) that nitrate TCRs and cytokine receptors, inactivating their function [98]. Concurrently, myeloid cells and tumor cells in the TME collaboratively overexpress ectonucleotidases CD39 and CD73, which progressively degrade pro-inflammatory extracellular ATP into adenosine [99]. Adenosine, by activating the A2A receptor (A2AR) on the T-cell surface, elevates intracellular cAMP levels, thereby potently inhibiting TCR signaling and effector function—one of the most central metabolic suppressive pathways in the TME [100]. Finally, TAMs and MDSCs are major sources of potent inhibitory cytokines, such as IL-10 and TGF-β [90]. TGF-β not only directly inhibits the cytotoxicity of CD8 + T cells but is also a key factor in inducing and maintaining the phenotype of regulatory T cells (Tregs), further intensifying the TME’s suppressive state [101].

Under such a powerful suppressive network, the few T cells that manage to infiltrate are often in a state of functional failure or “exhaustion” [102]. This exhausted state is characterized by a loss of proliferative capacity, reduced cytokine secretion (e.g., IFN-γ, TNF-α), and the sustained high expression of multiple inhibitory receptors (i.e., “checkpoints”) [103]. In OS, beyond PD-1, other key checkpoints such as T-cell Immunoglobulin and ITIM domain (TIGIT) and Lymphocyte-Activation Gene-3 (LAG-3) are also highly expressed on exhausted T cells [87]. TIGIT binds to CD155 (PVR) on tumor or myeloid cells, while LAG-3 binds to ligands like MHC class II molecules (expressed on TAMs/DCs), co-delivering deeper inhibitory signals that cannot be reversed by PD-1 blockade alone [104] (Fig. 3).

Fig. 3.

Barrier II (Cellular): “immunosuppression” - the myeloid-dominant network and T-Cell exhaustion

The final, critical factor that cripples T-cell immune surveillance stems from the OS tumor cells’ intrinsic antigen presentation defects [21]. As mentioned, OS is characterized by a low-to-moderate TMB and high genomic instability, including complex SVs, CNVs, and chromothripsis. While this instability could theoretically generate neoantigens, it more frequently leads to the inactivation of genes related to the Antigen Presentation Machinery (APM) [105]. The most critical event among these is the mutation or loss of heterozygosity (LOH) of the β2-microglobulin (B2M) gene. B2M is an essential subunit for the correct folding, assembly, and transport of MHC class I (MHC-I, also often referred to as HLA-I in OS) molecules to the cell surface [106]. Loss of B2M results in a complete absence of MHC-I expression on the tumor cell surface, making the cell incapable of presenting any tumor antigens (neither shared antigens nor neoantigens) to CD8 + T cells [22]. Therefore, even if active, tumor-specific T cells are present in the TME, they cannot “recognize” or “kill” these MHC-I-deficient tumor cells, constituting a “hard barrier” to primary immune resistance [107]. Furthermore, the cGAS-STING pathway—a key pathway that senses cytosolic DNA generated by genomic instability and initiates a Type I Interferon (IFN-I) response—is often dysfunctional or suppressed in OS [108]. IFN-I signaling is a core driver for upregulating MHC-I expression and promoting DC maturation; its absence further exacerbates APM defects and the “cold” TME phenotype [109]. In summary, the myeloid-dominant TME actively “extinguishes” T-cell activity via metabolic and checkpoint signals, while the tumor cells’ APM defects render them completely “invisible.” This dual-lock mechanism orchestrates the extreme resistance of OS to immunotherapy (Supplementary Table S3) (Fig. 4). It is noteworthy that, as suggested by combination therapies in other cancers, overcoming severe T-cell exhaustion in OS may require co-blockade of multiple inhibitory receptors (e.g., combining anti-PD-1 with anti-TIGIT or anti-LAG-3 drugs), although this approach remains to be validated in OS.

Fig. 4.

Barrier III (Molecular): “Immune Invisibility” - Antigen Presentation Machinery (APM) Defects

TME subtypes and cellular interactions from a single-cell and spatial multi-omics perspective

Traditional bulk-omics analyses, due to their “averaging” effect, fail to resolve the key cell subpopulations, functional states, and precise spatial interactions within the Osteosarcoma (OS) Tumor Microenvironment (TME) that drive immune evasion [34, 110, 111]. The “cold” phenotype of the OS TME arises from localized suppressive niches organized within specific micro-regions [31, 33, 34, 112]. High-resolution technologies therefore add complementary—not interchangeable—information. Single-cell RNA-sequencing (scRNA-seq) primarily resolves cell identity, activation state, and lineage trajectory, whereas spatial transcriptomics and spatial pathology define neighborhood architecture, cell–cell proximity, stromal rings, vascular niches, and the physical topology of immune exclusion [112–115]. In the framework proposed here, cell-state information is used to identify the dominant suppressive program, while spatial information is used to determine whether that program is organized into a barrier-forming structure with therapeutic consequences.

Single-cell sequencing has profoundly deepened our understanding of cellular heterogeneity within the OS TME. Significant heterogeneity has been revealed within the malignant OS cells. Zhou et al. classified them into two major lineages based on gene expression: osteoblastic-like and chondroblastic-like, further subdividing them into multiple subpopulations. Through inferred Copy Number Variation (CNV) analysis and cell trajectory analysis, they provided evidence that chondroblastic-like OS cells might transdifferentiate towards osteoblastic-like OS cells [116]. Zheng et al. categorized OS cells into seven subpopulations, including an OSc3TAGLN subpopulation with fibroblast-like features and an OSc5TOP2A subpopulation with a proliferative phenotype [112]. Analysis by Liu et al. further revealed that OS cells with a higher CNV burden tend to downregulate Major Histocompatibility Complex class I (MHC-I) molecules and suppress the IFN-γ signaling pathway, suggesting that their reduced immunogenicity may be an immune evasion mechanism driven by genomic instability [117].

Stromal cells, particularly Mesenchymal Stem Cells (MSCs) and Cancer-Associated Fibroblasts (CAFs), also exhibit high diversity. Zhou et al. identified three MSC subpopulations (NT5E+, WISP2+, CLEC11A+) and three CAF subpopulations (characterized by COL14A1, DES, and MYL9, respectively). Zheng et al., based on an integrated cohort, further classified MSCs/CAFs into functional subpopulations such as matrix CAFs (mCAF), inflammatory CAFs (iCAF), myogenic CAFs (myoCAF), vascular CAFs (vCAF), and Pericytes. Liu et al. also annotated iCAFs and myofibroblast-like CAFs.

Myeloid cells are a critical component of the OS TME. Regarding macrophages, the study by Liu et al. identified C1QC+ TAMs and SPP1 + TAMs, subpopulations with M2-like features (e.g., downregulated MHC-II genes). The study by Zhou et al. distinguished M1-TAMs, M2-TAMs, and a class of FABP4 + TAMs enriched in lung metastases, which may possess pro-inflammatory characteristics. The heterogeneity of Dendritic Cells (DCs) has also been revealed. Liu et al. focused on DCs, identifying conventional DC1 (cDC1, XCR1 + CLEC9A+), conventional DC2 (cDC2, CD1C+CLEC10A+), and a class of mature regulatory DCs (mregDCs, CD83 + CCR7+LAMP3+) that are enriched in OS tissues but rare in peripheral blood. Trajectory analysis suggested that mregDCs might differentiate from cDC1s and form an immunosuppressive microenvironment by secreting chemokines like CCL17/19/22 to recruit regulatory T cells (Tregs). Furthermore, Zhou et al. also conducted subpopulation analysis and trajectory inference on osteoclasts (OCs), classifying them into OCprogenitor, OCimmature, and OCmature. They found that as maturation progresses, the cells’ antigen presentation function weakens while bone resorption activity increases.

Tumor-Infiltrating Lymphocytes (TILs), although relatively sparse in the OS TME, also display functional heterogeneity [116]. Zhou et al. identified subpopulations including CD4 + T cells, CD8 + T cells, Tregs, NKT cells, NK cells, and B cells. They specifically noted that infiltrating CD8 + T cells and Tregs highly express the immune checkpoint molecule TIGIT, and CD8 + T cells also express LAG3, indicating a T-cell exhaustion state. Importantly, Zhou et al. demonstrated via in vitro experiments that using an anti-TIGIT antibody could enhance the killing activity of primary CD3 + T cells from patients with high TIGIT infiltration against OS cell lines.

To resolve the spatial organization of these cell subpopulations, Zheng et al. performed the first spatial transcriptomics analysis on a post-chemotherapy OS sample (BS3). Using the Cell2location deconvolution algorithm, they mapped the spatial distribution of major cell types (OS cells, MSCs, endothelial cells, myeloid cells, OCs, and T/NK cells) on the tissue slide. Further spatial clustering with Scanpy identified a unique niche (cluster6) at the front of the tumor necrosis area. This region highly expressed genes such as COL4A1, suggesting a possible association with chemotherapy resistance. Importantly, this type of spatial analysis does more than restate single-cell composition: it reveals how suppressive or resistant programs are physically arranged within stromal rims, necrotic fronts, and vascular interfaces, thereby providing non-redundant evidence for the structural basis of immune exclusion in OS.

Integrating single-cell and spatial data, especially through Ligand-Receptor (L-R) interaction network analysis (e.g., using CellPhoneDB), allows for the systematic inference of intercellular communication within the TME. These analyses not only confirm known key pathways in the OS TME (such as the myeloid recruitment-related CSF1-CSF1R, CCL2-CCR2, and CXCL8-CXCR1/2 axes) but also reveal novel interaction patterns. For instance, Zheng et al. discovered key communication between mCAF/Pericytes and endothelial cells via FN1-(ITGA5 + ITGB1). The analysis by Liu et al. highlighted interactions between cancer cells and macrophages via the novel macrophage checkpoint CD24-SIGLEC10, as well as communication between CAFs/SPP1 + TAMs and endothelial cells mediated by the ACKR family (e.g., CXCL12–CXCR4, CXCL8–ACKR1 axes), potentially involved in angiogenesis. Furthermore, Liu et al. also observed potential interactions between mregDCs and Tregs via checkpoint molecules like PVR-TIGIT.

Based on this high-dimensional omics data, researchers are working to construct more refined OS-TME classification frameworks that move beyond the traditional “cold/hot” dichotomy. For example, Jiang et al., by integrating bulk genomic, epigenomic, and transcriptomic data from 121 OS patients, classified OS into four subtypes with significant prognostic differences: Immune-Activated (S-IA), Immune-Suppressed (S-IS), Homologous Recombination Defect-dominant (S-HRD), and MYC-Driven (S-MD) [36]. Building on this direction, we treat OS-TME subtypes not as immutable taxonomic classes but as dynamic, clinically oriented system states defined by the barrier that is currently most actionable. In the present V1.0 framework, state assignment is deliberately semi-structured and hypothesis-generating rather than algorithmically fixed [36, 112–117]. A dominant state is assigned only when one barrier receives concordant support from at least two evidence layers—for example, tissue/pathology or spatial readouts plus either imaging/radiomics or circulating/immune context—or when a direct tissue-level hallmark is sufficiently strong (e.g., absent B2M/HLA-I in an otherwise immune-desert tumor). The current model is therefore rule-based and operational, whereas future versions may evolve toward probabilistic weighting. Importantly, mixed, co-dominant, or discordant cases are not forced into a single label; they are retained as “Mixed/Indeterminate” states and prioritized for short-interval reassessment after one treatment cycle. On this basis, the conceptual OS-TME V1.0 framework integrates a minimal viable detection panel (imaging-pathology-ctDNA; detailed in Supplementary Table S11 (Table 1)) and maps each state to an initial therapeutic logic: (1) Myeloid-Dominant / mregDC-enriched, in which suppressive myeloid tone dominates and de-suppression should precede checkpoint therapy; (2) Dense Fibrotic Stroma / Angio-abnormal, in which vessel/stroma remodeling is needed to “open the gate” before ICI-forward sequencing; and (3) APM-Defective / MHC-loss, in which recognition failure motivates MHC-independent or APM-restorative approaches. In this formulation, the subtype functions as a temporary decision state that conditions the next therapeutic step, not as a permanent biological identity (Supplementary Table S4).

The primary tumor—pulmonary Pre-metastatic Niche (PMN) and exosome axis

The primary reason for clinical treatment failure in Osteosarcoma (OS) is not the uncontrolled primary tumor, but rather its extremely high and almost exclusively lung-tropic tendency for early metastasis [5, 118–120]. This metastasis is not a passive, random dissemination process, but rather the result of active “orchestration” and “preparation” by the primary tumor [121–125]. Before Circulating Tumor Cells (CTCs) are shed from the primary site into the circulation, the primary TME has already “cultivated” a receptive “soil” in the distant target organ—the “Pre-metastatic Niche” (PMN)—by releasing a series of systemic signals [126]. In this “Primary-PMN” coupling axis, exosomes are considered the core vectors for transmitting these “educating” signals and reprogramming the distant microenvironment [75, 127, 128] (Fig. 5).

Fig. 5.

The systemic barrier - active engineering of the Pulmonary Pre-metastatic Niche (PMN)

Exosomes secreted by OS cells are nano-sized extracellular vesicles, enriched with meticulously packaged bioactive “cargo,” including specific proteins, lipids, metabolites, and non-coding RNAs (like miRNAs and lncRNAs) [129–131]. These exosomes act as “messengers,” entering the circulation and being preferentially taken up by specific cells in the lungs, such as pulmonary fibroblasts, endothelial cells, and alveolar macrophages [132–135]. Once inside the target cells, this cargo is released, systematically reprogramming and remodeling these resident cells at multiple levels, including transcription, translation, and metabolism [76]. For example, OS-derived exosomes have been shown to carry molecules like miR-148a-3p [135], miR-21-5p [136, 137], or linc00881 [138]. By targeting different messenger RNAs, they induce lung fibroblasts (LFs) to activate into a Cancer-Associated Fibroblast (CAF)-like phenotype. These “activated” LFs begin to secrete large amounts of Extracellular Matrix (ECM) proteins (like fibronectin and type I collagen), leading to pulmonary matrix fibrosis and increased stiffness, providing “anchoring sites” for subsequently arriving CTCs [123, 129]. Furthermore, OS exosomes can also carry proteins such as integrins or Annexin A2 (ANXA2). These proteins not only mediate the specific adhesion of exosomes to lung endothelial cells but also disrupt the integrity of the pulmonary vascular barrier, increasing vascular permeability and paving the way for CTC extravasation (Supplementary Table S5) [75, 139].

PMN formation involves not only physical remodeling of the stroma but, more critically, the systemic reprogramming of the immune microenvironment—namely, the pre-establishment of an immunosuppressive niche in the distant organ [121]. This process is driven by a combination of cytokines and exosomes secreted from the primary site. Among these, the CCL2-CCR2 axis plays a crucial role [94, 140]. Signals from the primary tumor (or the lung cells it induces) upregulate CCL2 expression in the lungs, which recruits a large number of CCR2-expressing circulating monocytes and MDSCs. In the lung microenvironment, these cells are “educated” into suppressive M2-like macrophages (TAMs) and MDSCs. They construct a powerful immunosuppressive network in the lungs before the tumor cells even arrive. Similarly, the CXCL family of chemokines (e.g., CXCL1, CXCL12) and their receptor (CXCR2, CXCR4) axes are also involved in the recruitment of neutrophils and MDSCs to the pulmonary PMN [121, 141]. Moreover, systemic inflammation driven by primary tumor-derived circulating cytokines (such as G-CSF or IL-6) also plays a significant role in PMN formation, suggesting potential intervention targets (e.g., targeting IL-6R) that are distinct from the exosome and chemokine axes.

In recent years, the role of neutrophils in PMN formation has received considerable attention, especially through the formation of Neutrophil Extracellular Traps (NETs) [142]. Factors released by the primary tumor (or its TME), such as G-CSF or IL-8, can systemically mobilize neutrophils. Upon arriving in the lungs, these neutrophils are further activated and release NETs—a web-like structure composed of a decondensed DNA fiber backbone and granular proteins (like neutrophil elastase, NE) [143–146]. Studies indicate that NETs play a dual role in the PMN: on one hand, they act like a “spider web” to directly capture circulating OS tumor cells (CTCs), greatly increasing their retention and colonization efficiency in the lungs [147]. On the other hand, proteases within the NETs (such as NE and MMP9) can degrade the basement membrane, promote CTC extravasation, and potentially “awaken” dormant disseminated tumor cells by activating specific signals (e.g., TLR9), driving their proliferation [142, 144].

The formation of the PMN and its coupling with the primary TME have profound implications for the clinical management of OS. It explains why many patients still develop lung metastases shortly after complete surgical resection of the primary tumor—the “seeds” of metastasis were already sown, and the “soil” was already fertile [134, 148]. It also highlights why perioperative intervention must be tested as a biomarker-rich platform rather than assumed to be automatically beneficial. While conventional chemotherapy can reduce primary tumor burden, treatment-related stress may also promote the release of pro-metastatic exosomes or mobilize suppressive myeloid cells, potentially accelerating PMN formation [149]. Accordingly, this review proposes a conceptual Window-of-Opportunity (WoO) design (Supplementary Table S11 (Table 2)) in which a provisional baseline state is assigned before surgery, short preoperative de-suppression/priming is administered, and paired postoperative pharmacodynamic readouts are used to test whether the expected state transition actually occurs. The aim is not to prescribe routine perioperative management, but to define an early-phase experimental platform in which changes in circulating MDSCs, NET-associated markers, ctDNA-MRD, and spatial pathology can be linked to mechanism.

Diagnostics and stratification: multi-modal biomarkers, radiomics, and ctDNA-MRD

Given the high degree of heterogeneity within the Osteosarcoma (OS) Tumor Microenvironment (TME) (as described in TME subtypes and cellular interactions from a single-cell and spatial multi-omics perspective), developing biomarkers that can accurately, dynamically, and minimally invasively stratify and monitor a patient’s TME phenotype is a prerequisite for realizing precision immunotherapy [2, 15]. Relying on a single, invasive tissue biopsy to guide treatment decisions faces significant challenges, including biopsy bias (failure to capture spatial heterogeneity [31, 150, 151]), poor patient compliance, and the difficulty of repeated sampling to monitor the dynamic evolution of the TME [152, 153]. Consequently, research is shifting toward multi-modal biomarkers that integrate imaging, pathology, liquid biopsy, and multi-omics data, leveraging Artificial Intelligence (AI) to extract deep information from them [154–156].

Radiomics and digital pathology offer the potential to “visualize” the TME non-invasively or minimally invasively [157–160]. Radiomics involves the high-throughput extraction of a vast number of quantitative features—imperceptible to the human eye (e.g., texture, shape, intensity distribution)—from standard clinical images (such as MRI, CT, PET), and correlating them with underlying biological phenotypes (e.g., gene mutations, TME infiltration). MRI-based radiomic features have been shown to significantly correlate with histological response, genomic instability, and the infiltration levels of immune cells (such as CD8 + T-cell and macrophage density) [161]. Similarly, applying Deep Learning algorithms to H&E-stained pathological slides (digital pathology) can not only automate the identification of tumor regions and quantify stroma-lymphocyte ratios but may even be able to predict spatial transcriptomics-defined TME subtypes or key ligand-receptor interactions directly from morphology [162–166]. This “imaging-omics” correlation is critical; it implies a future possibility of non-invasively inferring whether a patient belongs to a “myeloid-dominant” or “stroma-excluded” subtype using only preoperative imaging or routine pathology slides, thereby enabling preliminary screening of populations suitable for specific combination immunotherapies (e.g., anti-CSF1R or anti-LOX) (Fig. 6) (Supplementary Table S6) [157, 167].

Fig. 6.

OS-TME subtyping V1.0 as dynamic barrier states: defining readouts, conceptual decision logic, and initial therapeutic mapping

Simultaneously, liquid biopsy, especially the analysis of circulating tumor DNA (ctDNA), is emerging as the most promising tool for monitoring OS disease burden and therapeutic response [168, 169]. OS is a tumor of high genomic disarray; its characteristic Structural Variations (SVs) and Copy Number Variations (CNVs) (as described in Myeloid dominance and APM defects: from T-Cell exclusion to functional failure) provide abundant “tumor fingerprints” for ctDNA detection. Through ultra-deep sequencing of patient-specific genomic alterations (e.g., bTMB [170], or blood tumor mutational burden), ctDNA analysis can sensitively detect Minimal Residual Disease (MRD) post-surgery [171]. Multiple studies have confirmed that postoperative ctDNA-MRD positivity is the most powerful independent predictor of relapse and poor prognosis in OS patients, with sensitivity and specificity for prognostic judgment significantly exceeding traditional imaging surveillance [172, 173]. Incorporating MRD status into clinical trial designs (e.g., as a basis for prognostic stratification or treatment escalation) has become a trend. Furthermore, ctDNA can provide not only “quantitative” but also “qualitative” information. By analyzing variations in ctDNA, it is possible to track the clonal evolution of the tumor under therapeutic pressure, such as monitoring for resistance mutations in antigen presentation pathway genes (like B2M) that emerge during immunotherapy [174].

The scope of liquid biopsy is continually expanding beyond ctDNA. High-throughput sequencing of the circulating T-cell receptor (TCR) repertoire can dynamically assess the breadth and clonality of the systemic immune response [175–178]. In successful immunotherapy, one expects to see an expansion of tumor-specific TCR clones in the peripheral blood; conversely, a persistently “cold” or contracting TCR repertoire may signal primary resistance [179–181]. Additionally, analyzing circulating immune cell subpopulations, such as circulating Myeloid-Derived Suppressor Cells (cMDSCs) or specific monocyte phenotypes, provides a systemic window into the TME’s “myeloid bias” [182]. High levels of cMDSCs have been confirmed to correlate with poor prognosis and immunotherapy resistance in OS [183].

The immediate translational role of AI and Machine Learning (ML) in OS is therefore likely to be integrative decision support rather than autonomous clinical control. By combining radiomics, digital pathology, ctDNA-MRD, TCR repertoire data, circulating immune subsets, and, where available, spatial or single-cell readouts, computational models may help generate composite state summaries, prioritize biomarker-defined trial cohorts, and test retrospective hypotheses more efficiently than single-modality analyses alone. However, current multi-modal markers remain primarily prognostic rather than treatment-directive, and any AI-enabled state classifier will require prospective calibration, uncertainty quantification, external validation, and safeguards against dataset shift or model drift before clinical deployment.

“Turning cold to hot”: multidimensional combination strategies and AI-driven precision intervention

The multidimensional barriers of the Osteosarcoma (OS) “cold tumor” microenvironment—spanning stromal-vascular exclusion (Sect.  The interwoven immune-stromal-vascular network of OS and the"cold tumor" etiology), myeloid-dominantfunctional suppression (Sect.  Myeloid dominance and APM defects: from T-Cell exclusion to functional failure), and molecular defects in antigen presentation (Sect.  Myeloid dominance and APM defects: from T-Cell exclusion to functional failure)—collectively explain why monotherapy with PD-1/PD-L1 blockade has failed [2, 11, 184]. Merely “releasing the brakes” is ineffective if effector T cells are absent, if antigen presentation is broken, or if T cells remain physically trapped by the stromal barrier [185]. Accordingly, any “turning cold to hot” strategy in OS must be interpreted as state-conditioned and evidence-tiered rather than as a validated universal algorithm. At present, direct OS clinical signals are strongest for anti-angiogenic/ICI combinations, whereas many other state-action pairings—including durable myeloid reprogramming, robust RT-driven immunogenic conversion, and MHC-independent cell/ADC strategies—remain supported mainly by preclinical OS data, early-phase studies, or extrapolation from other tumor types [11, 12, 186–188]. Supplementary Tables S10-S12 therefore present these pathways as a conceptual decision framework designed for hypothesis generation and trial design, not as routine clinical guidance. Representative osteosarcoma combination immunotherapy strategies and their current supporting evidence are summarized in Supplementary Table S7.

The first route to immunogenic priming is to induce Immunogenic Cell Death (ICD), thereby releasing tumor antigens, Damage-Associated Molecular Patterns (DAMPs), and activating Type I Interferon signaling [189–192]. Radiotherapy (RT) is the classic means to achieve this. High-dose Stereotactic Body Radiotherapy (SBRT) or hypofractionation can, in principle, activate cGAS-STING signaling, upregulate MHC-I, and generate abscopal immune effects [193–197]. However, in OS, durable RT-driven conversion from a barrier-dominant state to a clinically meaningful ICI-responsive state has not yet been demonstrated; current support remains strongest at the preclinical or extrapolative level. RT is also a double-edged sword, as it may induce PD-L1, recruit Tregs, or deplete lymphocytes depending on dose and schedule. Thus, RT in OS is best framed as a rational priming lever to be tested within biomarker-defined sequences rather than assumed to be uniformly immunogenic. Besides RT, direct STING agonists or certain chemotherapeutic agents (e.g., doxorubicin) may also contribute to ICD-based priming [198, 199].

Antigen release alone, however, is insufficient to overcome the TME’s physical barriers. To target the abnormal vascular network and dense fibrotic stroma described in Sect. (The interwoven immune-stromal-vascular network of OS and the"cold tumor" etiology), anti-angiogenic and stromal-remodeling strategies have emerged [200–202]. Anti-angiogenic TKIs such as apatinib or lenvatinib may normalize chaotic tumor vasculature, improve T-cell delivery, and attenuate VEGF-driven myeloid suppression [203–206]. Among the candidate state-action pairings discussed in this review, anti-angiogenic TKI plus PD-1 blockade currently has the clearest direct OS-specific clinical signal, although the available evidence remains non-randomized and the magnitude of benefit is still uncertain [207, 208]. Stromal interventions targeting LOX, DDR1, or related mechanosignaling pathways are conceptually attractive for the dense-fibrotic-stroma state, but in OS they remain largely preclinical and their ability to remodel heavily mineralized, collagen-rich tissue in patients is unproven [209, 210].

The central battle in remodeling the TME lies in reversing the myeloid-dominant suppressive state described in Sect. (Myeloid dominance and APM defects: from T-Cell exclusion to functional failure). Inhibitors of the CSF1-CSF1R axis, CCL2-CCR2 axis, and CXCL8-CXCR1/2 axis aim to reduce or redirect suppressive macrophage and MDSC influx [140, 211–215]. This is a biologically compelling strategy for myeloid-dominant OS states, but the evidence base remains more limited than for anti-angiogenic/ICI combinations. Most support in OS is preclinical, and experience from other cancers suggests that durable myeloid reprogramming is difficult to maintain because compensatory pathways, rapid rebound after drug withdrawal, and tissue-specific toxicities may blunt benefit. Likewise, metabolic modulators targeting the adenosine, TGF-β, lactate, or kynurenine–AhR axes may help dismantle functional suppression, but in OS they should currently be considered exploratory combination partners rather than validated sequence-defining interventions [216–219].

Faced with the hard barrier of widespread MHC-I/B2M antigen-presentation defects in OS, conventional T-cell-directed therapies alone are unlikely to be sufficient [21, 22]. Two broad strategies are therefore considered: partial repair of the defect (e.g., STING agonists or epigenetic approaches to restore MHC-I expression) and functional bypass through MHC-independent targeting [220, 221]. NK-cell and CAR-NK approaches are attractive because their killing does not require tumor-cell MHC-I expression [222–224]. B7-H3- and GD2-directed CAR-T cells, antibody-drug conjugates, and bispecific antibodies similarly redirect recognition from TCR-MHC interactions to tumor surface antigens [225–237]. Nevertheless, for OS these approaches remain constrained by antigen heterogeneity, poor trafficking and persistence in a stromal/myeloid-suppressive environment, manufacturing complexity, and on-target/off-tumor toxicity. Accordingly, we position MHC-independent strategies as rational options for APM-defective states, but not as universally mature first-line solutions. Their use is most defensible when direct tissue evidence of APM loss coexists with demonstrable target expression and a plan to address parallel stromal or myeloid barriers.

How, then, should one select the most appropriate sequence for an individual patient? In the near term, the most realistic role of multi-modal biomarkers is to support structured state reassessment within biomarker-guided trials or highly specialized translational programs, rather than to automate routine care. For example, a patient provisionally classified as Myeloid-Dominant with RI < 0.35 and positive ctDNA-MRD could enter a de-suppression/priming track first (Fig. 7; Supplementary Table S12). Reassessment at 2–4 weeks would include ctDNA-MRD kinetics, an imaging-derived barrier score, and repeat tissue or blood myeloid readouts when feasible. A falling MRD signal together with reduced barrier intensity would support transition toward checkpoint therapy, whereas persistent MRD, worsening stromal exclusion, or newly dominant APM-loss features would trigger switching or intensification toward an alternative barrier track. In this review, such adaptivity is framed primarily as a design principle for biomarker-rich early-phase trials and only secondarily as a future template for individualized management within specialized centers. Any AI-enabled model in this context is envisioned as a decision-support and hypothesis-generating tool, not an autonomous clinical system.

Fig. 7.

Conceptual “Cold-to-Hot” Readiness Index (RI): Illustrative Multimodal Inputs, Reassessment Triggers, and Adaptive Sequencing

Discussion

This review has systematically decoded how Osteosarcoma (OS), as a quintessential “cold tumor,” exhibits immunotherapy resistance driven by multidimensional, interwoven biological mechanisms. From the clinical predicament of stagnant survival rates over nearly four decades (recounted in the Introduction) to the in-depth analysis of four core barriers in the main text—namely, the “immune-stromal-vascular” physical barrier characterized by dense collagen and abnormal vessels; the overwhelming “myeloid suppressive network” dominated by TAMs and MDSCs; the “antigen presentation defects” represented by MHC-I/B2M loss; and the “pulmonary pre-metastatic niche” (PMN) actively “engineered” by the primary TME in distant organs—we have constructed a panoramic map of immune evasion in OS (Supplementary Table S8). The critical insight from this map is that immune resistance in OS is not caused by a single factor (like low PD-L1 expression) but is a profoundly immunosuppressive state “locked” at the genomic (high SVs, APM defects), cellular (myeloid-dominant), and tissue (stromal stiffness, bone remodeling) levels. This realization itself is the fundamental explanation for the dismal failures of past ICI monotherapy trials. However, this breakthrough also brings a clear limitation: the profound heterogeneity of the OS TME means no “one-size-fits-all” cold-to-hot strategy exists; rather, each patient’s TME may be dominated by a different combination of barriers.

Based on this framework, the multidimensional combination strategies outlined in Sect. ("Turning cold to hot": multidimensional combination strategies and AI-driven precision intervention) can be understood as attempts to match a specific “combination key” to a dominant barrier state. Importantly, these state-action relationships are not supported at the same evidentiary level. In OS, the clearest direct clinical signal currently comes from anti-angiogenic/PD-1 combinations; by contrast, the anticipated benefits of SBRT/STING priming, durable myeloid reprogramming, and many MHC-independent cellular or antibody platforms remain supported mainly by preclinical OS studies, early-phase trials, or extrapolation from other tumor types. This distinction matters because the negative or very limited activity observed in unselected OS trials such as SARC028 and nivolumab/ipilimumab illustrates that a favorable state transition does not occur automatically [11, 12]. Priming strategies (e.g., SBRT plus ICI) are attractive because they leverage existing infrastructure to induce ICD and Type I IFN signaling, but the optimal dose, fractionation, and sequencing remain unresolved and the clinical ability of RT to durably “heat” OS is unproven [195–197]. Stromal-vascular remodeling strategies, particularly anti-angiogenic TKIs combined with ICIs, are among the fastest translating pathways in OS because they may simultaneously normalize vessels and dampen myeloid recruitment, although toxicity and resistance remain major liabilities [207, 238–240]. Myeloid reprogramming strategies are biologically central, yet the plasticity of TAMs and MDSCs means that compensatory pathways and rebound after treatment withdrawal are likely to limit durability [241–244]. Finally, MHC-independent killing strategies—CAR-T, CAR-NK, ADCs, and bispecific antibodies directed against B7-H3 or GD2—offer a genuine conceptual solution to APM loss, but their clinical maturation in OS is constrained by antigen heterogeneity, trafficking barriers, manufacturing complexity, and on-target/off-tumor toxicity (Supplementary Table S9) [245–250].

These complex combination strategies also bring severe clinical translation challenges. First is regulatory and toxicity management. As we move from “dual” to “triple” or even “quadruple” combinations (e.g., RT + anti-CSF1R + anti-PD-1 + anti-TIGIT), the incidence, severity, and unpredictability of immune-related adverse events (irAEs) will increase exponentially [251–254]. The toxicity profiles may be completely different from traditional ICI irAEs, especially when targeting myeloid cells (which may harm homeostatic macrophages), metabolic pathways (which may affect normal T-cell function), or introducing cell therapies. This demands the establishment of entirely new toxicity monitoring and management standards. Second, biomarker development and commercialization are the critical bottlenecks determining the success or failure of these strategies. As mentioned in Diagnostics and stratification: multi-modal biomarkers, radiomics, and ctDNA-MRD, we must move from “prognostic” markers (like ctDNA-MRD) to “predictive” biomarkers. We desperately need clinically viable assays (e.g., multiplex IHC, radiomic AI models, or reproducible spatial-omics-based scores) that can answer “Which patient needs anti-stroma, and which needs anti-myeloid?” Without precise stratification, OS immunotherapy trials are likely to repeat the failure of SARC028, failing overall due to the inclusion of too many “cold” TME subtypes. For example, we urgently need to differentiate among ctDNA-MRD-positive (poor prognosis) patients, which ones are dominated by APM defects (should receive CAR-T) and which are dominated by myeloid suppression (should receive anti-CSF1R). Currently, ctDNA provides only ‘prognostic’ information, lacking this ‘predictive’ guidance, which is a core challenge in the field.

Looking ahead, several translational priorities appear most realistic. First, biomarker-defined state assignment must be improved using multi-modal data integration, with AI/ML serving as a human-supervised decision-support layer rather than an autonomous treatment controller. In this context, a future “digital twin” should be understood as a research model for simulation and hypothesis generation. The RI proposed here is accordingly presented only as an illustrative prototype: its weights and cutoffs are not data-derived, and any future implementation would require prospective calibration, external validation, uncertainty quantification, and monitoring for dataset shift or model drift. Second, adaptivity should initially be tested at the level of biomarker-rich early-phase trials. A practical scenario would be to assign a provisional baseline state, deliver one barrier-directed cycle, and then reassess at 2–4 weeks using ctDNA-MRD, imaging-based barrier measures, and repeat tissue or blood immune readouts. Improvement in MRD and barrier metrics would justify continuation or checkpoint addition, whereas persistent MRD, worsening exclusion, or emergence of APM loss would trigger switching or intensification along an alternative barrier track. Third, local delivery and perioperative strategies deserve special attention because the PMN is established before surgery; smart hydrogels, CCR2-axis intervention, and other local or perioperative approaches may offer a way to remodel both residual local disease and the metastatic window with lower systemic toxicity. By contrast, TLS induction should currently be regarded as an exploratory future direction rather than an established OS strategy, because direct evidence for spontaneous or therapeutically induced TLS formation in primary OS remains limited. Window-of-Opportunity studies are especially valuable in this setting because they can generate paired tissue and blood samples needed to test whether the hypothesized state transition truly occurs (Fig. 8).

Fig. 8.

Perioperative “Window of Opportunity” (WoO): biomarker-rich trial blueprint with pharmacodynamic reassessment points

In conclusion, decoding the biological roots of the OS “cold tumor” remains a formidable challenge, and turning it “hot” is better conceived as a staged, feedback-informed systems problem than as a single therapeutic maneuver. From T-cell exclusion to myeloid dominance, from dense fibrotic stroma to antigen loss, the OS TME erects multiple, coexisting barriers. The most credible path forward is therefore not empiric combination escalation, but biomarker-guided, human-in-the-loop precision stratification that links dynamic state assessment to rational therapeutic sequencing. Within that framework, the OS-TME states, RI, and WoO design proposed here should be viewed as conceptual tools to organize future studies, refine trial enrichment, and support mechanism-based translational progress rather than as validated clinical algorithms (Fig. 9).

Fig. 9.

Human-in-the-Loop, AI-Enabled “Diagnosis-Treatment-Monitoring” closed loop for state assessment and adaptive decision support

Authors’ contributions

Bai Yang, Shu Liu, Bingcheng Liu contributed equally to this work.- Conceptualization: Bai Yang, Xiao Ma, Tengfei Song- Methodology / Literature Search: Bai Yang, Shu Liu, Bingcheng Liu- Data Curation & Evidence Synthesis: Bai Yang, Shu Liu, Bingcheng Liu, Tianwen Ye- Visualization (Figures/Tables): Bai Yang, Tianwen Ye- Writing – Original Draft: Bai Yang, Shu Liu, Bingcheng Liu- Writing – Review & Editing: Xiao Ma, Tengfei Song- Supervision: Tengfei Song, Xiao Ma- Project Administration / Final Approval: Tengfei Song (Lead Contact), Xiao Ma.

Funding

This research received no external funding.

Data availability

This study did not generate or analyze any new data. All data utilized were sourced from previously published studies, which have been appropriately cited. Therefore, data sharing is not applicable to this article.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.