NNMT-associated metabolic–thromboinflammatory–immune co-activation in heterogeneous CTC clusters: a hypothesis-generating computational framework” Authors: Jiayang Gong, Zhe Wang, Yun Shi, Minlan Ren, Tingting Li, Gang Li, Rui Peng Published: 2026 ABSTRACT BackgroundHeterogeneous circulating tumor cell (CTC) clusters interact with platelets, neutrophils, stromal cells, and immune cells, forming a protective state that may facilitate survival in the circulation and metastatic dissemination. Nicotinamide N-methyltransferase (NNMT) is associated with metabolic reprogramming, stromal activation, epithelial–mesenchymal plasticity, and immune suppression. However, its coordinated relationship with platelet/coagulation, neutrophil/ neutrophil extracellular trap (NET), and immune exhaustion programs in CTC-associated states has not been systematically evaluated. Therefore, we constructed a hypothesis-generating framework integrating mechanistic evidence, pan-cancer computational analyses, and single-cell transcriptomic analyses.MethodsWe performed a mechanistic evidence synthesis combined with TCGA PanCancer bulk RNA-seq analysis and cross-dataset evaluation of eight publicly available multi-cancer single-cell RNA-seq cohorts. Module scores were constructed for metabolic/EMT, platelet/coagulation, neutrophil/NET, immune checkpoint/exhaustion, and cytotoxic/NK cell programs. Correlation analyses across cancer types, feature-level pseudotime analysis, ligand– receptor mapping, and in silico perturbation simulations were applied to evaluate associations between NNMT and the composite tri-axial state. In addition, exploratory supervised machine learning models were used to determine whether axis-related features and ligand–receptor features could discriminate single CTCs from clustered or leukocyte-associated CTC states. Leave-one-cancer-out cross-validation was further used to assess the reproducibility of axis- associated survival risk across cancer types. No new experimental, animal, or clinical intervention data were generated.ResultsAcross 483,590 cells from eight publicly available single-cell datasets, NNMT expression was positively correlated with the composite tri-axial score across multiple cancer contexts. Feature-level pseudotime analysis indicated convergence of NNMT-associated metabolic, platelet/coagulation, neutrophil/NET, and immune exhaustion programs toward a common trajectory endpoint, although this analysis does not establish temporal causality. In silico NNMT perturbation preferentially implicated stromal and fibroblast-like populations as candidate responsive compartments. Ligand–receptor analyses further suggested an interconnected communication network involving CTC/epithelial and CAF/endothelial compartments, platelet/coagulation bridging, myeloid/NET recruitment, and downstream T/NK cell exhaustion.ConclusionThese integrated findings support NNMT-associated metabolic–thromboinflammatory–immune co-activation as a candidate feature of heterogeneous CTC-associated states and provide mechanistic rationale for a “disaggregation–exposure–clearance” strategy combining αIIbβ3 inhibition, NNMT inhibition, and PD-1/PD-L1 blockade.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16692-x. Introduction The majority of cancer-related deaths are caused by metastasis, with approximately 90% of cancer mortality attributed to metastatic disease [1, 2]. Circulating tumor cells (CTCs) detach from primary lesions and enter the bloodstream. To establish secondary colonization, they must withstand shear stress, anoikis, and immune surveillance, with only a small fraction successfully surviving these conditions [3–5]. Recent studies have shown that CTCs migrating as clusters exhibit markedly enhanced metastatic efficiency. These clusters exist as homotypic aggregates (CTC–CTC interactions) or heterotypic aggregates (CTC interactions with neutrophils, platelets, myeloid-derived suppressor cells, and other blood components). The metastatic potential of CTC clusters is 23–50 times higher than that of single CTCs, and approximately 97% of metastatic events are mediated by clustered CTCs [6–8]. Single-cell transcriptomic and proteomic analyses demonstrate distinct signaling programs between clustered and single CTCs, with prominent activation centered on the Src/FN1 axis. This pathway promotes intracluster adhesion and protects tumor cells from immune clearance [9]. Recent single-cell sequencing studies further revealed epigenetic and transcriptional heterogeneity underlying cluster cohesion [10]. The classical “seed and soil” hypothesis, together with molecular mechanistic studies, has profoundly shaped our understanding of organ-specific metastatic colonization [11, 12]. Advances in microfluidic technologies have enabled isolation and characterization of rare cell clusters [13, 14]. Heterotypic CTC clusters exhibit marked heterogeneity (reported heterogeneity index approximately 69%). Platelet coating forms a protective barrier that shields CTCs from shear stress and suppresses NK cell cytotoxicity via TGF-β signaling [15–17]. Neutrophils actively participate in these processes. They release neutrophil extracellular traps (NETs), which capture and protect CTCs, while secreting immunosuppressive mediators such as ARG1 and IL-10, forming an immune-protective niche that restricts cytotoxic T cell and NK cell access [18–21]. Platelets contribute additional immune evasion mechanisms via the CD155–TIGIT axis, suppressing NK cell-mediated innate immune surveillance [22]. Within CTC clusters, adhesion molecules such as VCAM-1 and integrins activate survival signaling pathways and promote resistance to anoikis [23, 24]. NETs further facilitate tumor progression by establishing a pro-metastatic inflammatory microenvironment [25]. Masked CTCs embedded in platelet or neutrophil aggregates can modulate immune checkpoint blockade responsiveness and reduce T cell sensitivity [26, 27]. Aspirin is associated with bleeding risk [28, 29], highlighting the need for stratified multi-dimensional strategies to disrupt CTC clusters [30, 31]. Metastasis is a multistep process including dissemination, survival in circulation, and colonization, underscoring the necessity of multi-target interventions [32–34]. NNMT catalyzes the conversion of nicotinamide (NAM) to 1-methylnicotinamide (MNAM), consuming the methyl donor S-adenosylmethionine (SAM), thereby affecting epigenetic and metabolic regulation [35, 36]. In multiple cancers, high NNMT expression is associated with CTC stemness, chemoresistance, and metastatic potential [37–39]. In cancer-associated fibroblasts (CAFs), NNMT promotes H3K27me3 hypomethylation, enhancing complement secretion and recruitment of immunosuppressive myeloid-derived suppressor cells (MDSCs), thereby suppressing anti-tumor immunity systemically [40]. Pericyte-derived NNMT promotes CTC–neutrophil cluster formation via the CXCL5/CXCR2 axis, and NNMT inhibition has been shown to disrupt these clusters and suppress metastasis [37]. NNMT therefore represents a central node linking metabolic reprogramming and immune evasion in the pre-metastatic niche [41]. A proposed triple-therapy rationale includes: (1) αIIbβ3 inhibitors disrupting platelet–CTC adhesion and removing the platelet “cloak” [29, 42]; (2) NNMT inhibitors modulating CTC cluster survival capacity [35, 36, 39, 43]; (3) PD-1/PD-L1 antibodies activating T and NK cell-mediated clearance of CTCs [27, 44]. Methods and mechanisms Methods TCGA pan-cancer analysis We analyzed UCSC Xena TCGA PanCancer RNA-seq expression matrices and TCGA-CDR clinical annotation datasets. Only primary tumor samples were included, and cancer types with sample sizes ≥ 50 were retained, resulting in 9,596 primary tumor samples across 30 cancer types. NNMT was defined as the anchor gene. We constructed module scores for EMT/metabolism, platelet/coagulation, neutrophil/NET, immune checkpoint/exhaustion, and cytotoxic/NK programs. To avoid self-correlation, NNMT was excluded from the EMT/metabolism module in TCGA analyses. Within each cancer type, gene features were z-score normalized and averaged to compute module scores. The composite tri-axis score was defined as the mean of normalized EMT/metabolism, platelet/coagulation, neutrophil/NET, and immune checkpoint/exhaustion scores. Spearman correlation analysis was performed, and false discovery rate was adjusted using the Benjamini–Hochberg method. Samples were stratified into NNMT-high and NNMT-low groups based on median expression within each cancer type, and differences were assessed using Wilcoxon rank-sum tests. Overall survival analysis was exploratory and conducted using Cox proportional hazards models. Multi-cancer single-cell validation and in silico perturbation analysis We downloaded single-cell RNA sequencing datasets from GEO, including GSE109761, GSE131907, GSE132465, GSE149614, GSE154778, GSE176078, GSE234129, and GSE72056, representing eight cancer types. Feature genes were extracted from expression matrices, and z-score normalization was performed within each dataset. Module scores were then computed. PCA/SVD analysis was used to construct NNMT/tri-axis pseudotime trajectories at the feature level, and samples were divided into five equal-frequency bins. In silico perturbation of NNMT, the platelet/αIIbβ3 axis, and the immune checkpoint axis was performed by reducing target variables to the 10th percentile of cohort-specific distributions. Downstream module responses and composite protective scores were estimated based on expression–score slope relationships. Combined perturbation effects were computed by summing contributions across all three axes. Machine learning analysis We selected datasets based on relevance to CTC biology, GEO/TCGA/Xena resources, single-cell atlases, immunotherapy response datasets, and clinical survival cohorts. In GSE109761_human, we trained an L2-regularized logistic regression classifier to distinguish single CTCs from clustered or leukocyte-associated CTCs. Predictors included NNMT expression, EMT/metabolism, platelet/coagulation, neutrophil/NET, immune checkpoint/exhaustion, cytotoxic/NK scores, and cluster adhesion scores. Extended models incorporated CXCL5–CXCR2, VCAM1–ITGA4, platelet–integrin, immune checkpoint, and TGF-β ligand–receptor axes. Model performance was evaluated using repeated stratified five-fold cross-validation. For cross-cancer generalization, leave-one-cancer-out cross-validation was performed in TCGA PanCancer datasets, using 3-year overall survival mortality as the binary endpoint. Cancer types with insufficient sample size or events were excluded. Axis-based features were integrated into a cross-cancer single-cell atlas, combined with perturbation simulations to identify candidate responsive cell populations. We interpreted and distinguished literature-supported observations from model-derived inferences and therefore organized them into mechanistic-framework categories (Table 1).Table 1Evidence hierarchy applied throughout this studyEvidence levelScope in the present studyRequired wording and interpretationEstablished evidenceCTC clusters display greater metastatic competence than single CTCs; platelet cloaking and neutrophil/NET-associated protection contribute to CTC survival, immune evasion, and disseminationUse “has been demonstrated,” “has been reported,” or “is supported by direct experimental/clinical evidence,” with citations. Do not generalize beyond the tumor type, model, or context examinedInferred mechanismNNMT is linked to EMT and metabolic reprogramming, CAF/MDSC remodeling, and CXCL5/CXCR2-mediated myeloid recruitment; the present computational analyses identify coactivation and communication patternsUse “supports,” “is consistent with,” “suggests,” or “may mediate.” State that the mechanism is inferred from integrated literature and computational evidence rather than directly proven in the triple-combination settingSpeculative interactionPotential pharmacologic or biological synergy among NNMT inhibition, αIIbβ3 inhibition, and PD-1/PD-L1 blockadeUse “we speculate,” “may interact,” “could cooperate,” or “may be synergistic.” Do not state that synergy has been demonstrated; formal perturbation, combination-index, and in vivo studies are requiredTherapeutic hypothesisThe proposed Disintegration–Exposure–Elimination cascade as a staged intervention against heterogeneous CTC clustersUse “we propose” or “we hypothesize,” and explicitly state that prospective experimental and clinical validation is required. Do not present the cascade as an established regimen or efficacy claim Table 1Evidence hierarchy applied throughout this studyEvidence levelScope in the present studyRequired wording and interpretationEstablished evidenceCTC clusters display greater metastatic competence than single CTCs; platelet cloaking and neutrophil/NET-associated protection contribute to CTC survival, immune evasion, and disseminationUse “has been demonstrated,” “has been reported,” or “is supported by direct experimental/clinical evidence,” with citations. Do not generalize beyond the tumor type, model, or context examinedInferred mechanismNNMT is linked to EMT and metabolic reprogramming, CAF/MDSC remodeling, and CXCL5/CXCR2-mediated myeloid recruitment; the present computational analyses identify coactivation and communication patternsUse “supports,” “is consistent with,” “suggests,” or “may mediate.” State that the mechanism is inferred from integrated literature and computational evidence rather than directly proven in the triple-combination settingSpeculative interactionPotential pharmacologic or biological synergy among NNMT inhibition, αIIbβ3 inhibition, and PD-1/PD-L1 blockadeUse “we speculate,” “may interact,” “could cooperate,” or “may be synergistic.” Do not state that synergy has been demonstrated; formal perturbation, combination-index, and in vivo studies are requiredTherapeutic hypothesisThe proposed Disintegration–Exposure–Elimination cascade as a staged intervention against heterogeneous CTC clustersUse “we propose” or “we hypothesize,” and explicitly state that prospective experimental and clinical validation is required. Do not present the cascade as an established regimen or efficacy claim Evidence hierarchy applied throughout this study Mechanism No prior studies have evaluated combined NNMT inhibition, αIIbβ3 inhibition, and PD-1/PD-L1 blockade. Therefore, we adopted a mechanism-inference and evidence-integration framework (Fig. 1).Fig. 1*Metabolic-immunologic-thrombotic triple intervention against heterogeneous CTC clusters. (1) Platelets and neutrophils participate in the formation of heterogeneous circulating tumor cell (CTC) clusters, establish an immune-evasion niche containing NETs, ARG1, IL-10, TGF-β, and PD-L1, and increase metastatic efficiency by 23- to 50-fold. (2) The three model components target three pathways: the metabolic axis (NNMT inhibitor), the immune axis (PD-1/PD-L1 antibody), and the thrombotic axis (αIIbβ3 inhibitor). Together, they are hypothesized to disrupt platelet-mediated CTC camouflage and reactivate T cells and NK cells. (3) Three complementary steps: 1) αIIbβ3 inhibition may weaken CTC clusters; 2) NNMT inhibition may facilitate the disaggregated tumor cells; and 3) PD-1/PD-L1 blockade and NK-cell surveillance achieve immune-mediated clearance, have been reported in selected preclinical models* *Metabolic-immunologic-thrombotic triple intervention against heterogeneous CTC clusters. (1) Platelets and neutrophils participate in the formation of heterogeneous circulating tumor cell (CTC) clusters, establish an immune-evasion niche containing NETs, ARG1, IL-10, TGF-β, and PD-L1, and increase metastatic efficiency by 23- to 50-fold. (2) The three model components target three pathways: the metabolic axis (NNMT inhibitor), the immune axis (PD-1/PD-L1 antibody), and the thrombotic axis (αIIbβ3 inhibitor). Together, they are hypothesized to disrupt platelet-mediated CTC camouflage and reactivate T cells and NK cells. (3) Three complementary steps: 1) αIIbβ3 inhibition may weaken CTC clusters; 2) NNMT inhibition may facilitate the disaggregated tumor cells; and 3) PD-1/PD-L1 blockade and NK-cell surveillance achieve immune-mediated clearance, have been reported in selected preclinical models* Mechanistic rationale and computational validation Immune protective mechanisms of heterogeneous CTC clusters The rationale for therapeutic targeting. CTC clusters (Fig. 2) have evolved multi-factorial protective barriers that function as a key “protective shield” during tumor metastasis [6, 17, 45, 46]. Mechanistically, these clusters depend on collective protection. To clarify the mechanistic description, the following sections focus on two common and clinically relevant heterotypic forms: CTC–platelet clusters and CTC–neutrophil clusters, while acknowledging that CTC clusters also include broader homotypic and heterotypic configurations. The former functions as a “cloaking layer,” whereas the latter acts as an immune “barrier”; their interaction generates synergistic defensive effects [18, 47].Fig. 2CTC Clusters’ Cellular composition and immunosuppressive features. *Cellular composition and immunosuppressive features of heterogeneous CTC clusters. Circulating tumor cells (CTCs), platelets (PLTs), neutrophils (NEUs), proteoglycan (PT), adhesion molecules such as VLA-4, VCAM-1, and CD44, suppression of natural killer (NK)-cell function (↓NKG2D, ↓IFNγ, CD155–TIGIT), T-cell exhaustion (↑PD-1, ↑TIM-3), and key immune checkpoints (PD-1/PD-L1, CD155–TIGIT, TIM-3/Galectin-9). Double-positive T cells (DPT: CD44⁺CD8⁺, PD-1⁺TIM-3⁺) CTC Clusters’ Cellular composition and immunosuppressive features. *Cellular composition and immunosuppressive features of heterogeneous CTC clusters. Circulating tumor cells (CTCs), platelets (PLTs), neutrophils (NEUs), proteoglycan (PT), adhesion molecules such as VLA-4, VCAM-1, and CD44, suppression of natural killer (NK)-cell function (↓NKG2D, ↓IFNγ, CD155–TIGIT), T-cell exhaustion (↑PD-1, ↑TIM-3), and key immune checkpoints (PD-1/PD-L1, CD155–TIGIT, TIM-3/Galectin-9). Double-positive T cells (DPT: CD44⁺CD8⁺, PD-1⁺TIM-3⁺) CTC-platelet clusters: physical shielding and immune signal suppression In the circulatory system, platelets are among the most important “partners” of CTCs [24, 48, 49]. Experimental studies indicate that in CTC-positive samples from patients with lung cancer, breast cancer, or melanoma, approximately 60%–83% of CTCs are associated with platelet clusters [15, 17]. The adhesion mechanisms include two main processes: First, platelet surface P-selectin recognizes corresponding ligands on CTCs. Second, activated integrin αIIbβ3 mediates binding via fibrinogen as a molecular “glue” between the two cell types [29, 50]. In mouse models, oral administration of the GPIIb/IIIa antagonist XV454 completely inhibited platelet aggregation, significantly reduced tumor cell-induced thrombocytopenia (P < 0.01), and reduced lung metastatic nodules by 80%–85% (P < 0.001), demonstrating the feasibility of targeting αIIbβ3 to inhibit metastasis [51]. Platelets provide multiple advantages to CTCs:(1) reduction of mechanical injury caused by shear stress [17]; (2) secretion of cytokines that suppress NK cell function and reduce surface activating receptor NKG2D expression [15, 16, 52, 53]; (3) induction of EMT through platelet-derived growth factors and TGF-β, thereby enhancing invasiveness and metastatic potential [17, 24, 54]; (4) “education” of CTCs via paracrine signaling to reinforce stem-like phenotypes [55]. CTC–platelet clusters also escape T-cell immune surveillance. Platelet surface PD-L1 can engage PD-1 on T cells and suppress T-cell function [27, 33]. According to our predictions, PD-1/PD-L1 antibodies may weaken the immune protective effect of CTC–platelet clusters and represent a potential intervention target [56]. CTC-neutrophil clusters: NETs entrapment and inflammatory microenvironment remodeling In tumor contexts, neutrophils exhibit immune dysfunction [14, 57, 58]. CTC surface VCAM-1 binds VLA-4 on neutrophils [18, 46, 54]. Neutrophils release neutrophil extracellular traps (NETs), which physically capture CTCs and promote cluster formation [18, 59, 60]. Previous studies have shown that nanomedicine delivery systems can inhibit CTC–neutrophil cluster formation by blocking adhesion, neutralizing NETs, or delivering chemotherapeutic agents [31]. NET formation can be triggered by tumor-derived factors, thereby amplifying CTC capture [19]. CTC–neutrophil clusters enhance metastatic dissemination. Direct cell–cell contact increases cyclin D1 levels in CTCs, where cyclin D1 is a key regulator of cell cycle progression. This promotes CTC entry into proliferative states [52]. Neutrophils secrete ARG1 (arginase 1), IL-10, and reactive oxygen species (ROS) [61]. These factors suppress immune responses by depleting arginine required for T cell function (ARG1), and directly inhibiting T cell and NK cell activation and cytotoxicity (IL-10 and ROS) [62]. Neutrophils can release two enzymes, cathepsin G and elastase [63], which degrade components of the vascular basement membrane and facilitate intravascular migration of CTCs [64]. A novel “partner” of CTCs is CD4⁺CD8⁺ double-positive T cells (DPTs) DPTs are considered a T-cell lineage that exhibits immune desensitization upon interaction with CTCs. CTC surface VCAM-1 interacts with VLA-4 on DPTs, promoting adhesion and heterotypic cluster formation. Blocking VLA-4 using neutralizing antibodies significantly disrupts CTC–DPT clusters, suppresses metastasis, and prolongs survival in mouse models [65]. To connect the cluster-level evidence described above with single-cell evidence, we performed additional analyses and mapped curated ligand–receptor axes among CTC/epithelial, stromal/endothelial, platelet/coagulation, myeloid/NET, and T/NK-exhaustion subpopulations (Fig. 3).Fig. 3Ligand–receptor communication network at the subpopulation level. *Edge colors indicate the sender compartment, and node colors indicate the major cell category*. Results: A continuous protective communication chain was observed among CTC/epithelial, CAF/endothelial, platelet/coagulation, myeloid/NET, and T/NK-exhaustion compartments. CTC/CAF/endothelial cells recruit myeloid/NET cells through chemotactic and adhesion signals, coagulation–platelet bridging forms a platelet coating, and platelet/coagulation and myeloid/NET compartments further transmit immunosuppressive and exhaustion signals to T/NK cells. These results are highly consistent with an”immune-protective niche for CTC clusters” Ligand–receptor communication network at the subpopulation level. *Edge colors indicate the sender compartment, and node colors indicate the major cell category*. Results: A continuous protective communication chain was observed among CTC/epithelial, CAF/endothelial, platelet/coagulation, myeloid/NET, and T/NK-exhaustion compartments. CTC/CAF/endothelial cells recruit myeloid/NET cells through chemotactic and adhesion signals, coagulation–platelet bridging forms a platelet coating, and platelet/coagulation and myeloid/NET compartments further transmit immunosuppressive and exhaustion signals to T/NK cells. These results are highly consistent with an”immune-protective niche for CTC clusters” At the upstream recruitment and adhesion stage, VCAM1–ITGA4 signaling from CAF_endothelial—> myeloid_NET was relatively strong, with a mean LR score of 0.678 and support from n = 8 datasets; CXCL5–CXCR2 signaling from CAF_endothelial—> myeloid_NET was likewise supported, with a mean LR score of 0.404 and n = 9. Within CTC_epithelial—> myeloid_NET, the CXCL5–CXCR2 score was 0.444 (n = 8), and the VCAM1–ITGA4 score was 0.347 (n = 7). These results suggest that CTCs and surrounding stromal/endothelial components may promote CTC–neutrophil/NET cluster formation through chemotaxis of CXCR2⁺ myeloid cells and ITGA4-mediated adhesion. At the platelet–coagulation bridging stage, FGA/FGB/FGG–ITGA2B/ITGB3 signaling from CTC_epithelial—> platelet_coagulation was relatively strong, with a mean LR score of 0.670 and n = 7; SELP–PSGL1 signaling from platelet_coagulation—> myeloid_NET was among the strongest signals in the entire network, with a mean LR score of 0.786 and n = 8. These findings suggest that fibrinogen–αIIbβ3 and P-selectin–PSGL1 may jointly participate in platelet coating of CTCs and in linking platelets with myeloid/NET cells, directly supporting αIIbβ3 as one of the targets of the triple intervention proposed in this study. At the downstream immune-exhaustion stage, myeloid_NET—> T_NK_exhaustion constituted the most concentrated communication module, including LGALS9–HAVCR2 (score = 0.742, n = 11), CD274–PDCD1 (score = 0.699, n = 11), PVR–TIGIT (score = 0.632, n = 11), and TGFB1–TGFBR (score = 0.575, n = 11). Meanwhile, platelet_coagulation—> T_NK_exhaustion also included SELP–PSGL1 (score = 0.700, n = 10) and TGFB1–TGFBR (score = 0.576, n = 11). This indicates that T/NK exhaustion is not driven solely by tumor-cell PD-L1, but is a multiaxis immunosuppressive outcome jointly shaped by platelet/coagulation and myeloid/NET compartments. Cell–cell communication analysis provided mechanistic support: (1) the NNMT-associated CTC metabolic/EMT state may lie upstream and promote protective communication between CTCs and stromal/endothelial, platelet, and myeloid/NET cells; (2) the αIIbβ3 axis performs the platelet–coagulation bridging function; and (3) the PD-1/PD-L1 and TIGIT/TIM3/TGF-β axes constitute the final immune-exhaustion output. These findings further support our hypothesis. Phenotypic heterogeneity of CTC clusters and dependence on EMT state EMT should be conceptualized as a continuum rather than a binary process. Epithelial and hybrid epithelial/mesenchymal CTCs may retain sufficient cell–cell adhesion for collective cohesion while acquiring enhanced motility, stress tolerance, and metastatic plasticity. Established studies show that circulating tumor cell (CTC) clusters encompass multiple cellular configurations, including homotypic tumor-cell clusters, CTC–platelet clusters, CTC–neutrophil clusters, and CTC-associated aggregates containing CD4⁺CD8⁺ double-positive T cells (DPTs) or other immune-cell subsets. Homotypic clusters are maintained primarily through tumor cell–tumor cell adhesion, whereas heterotypic clusters acquire additional structural, inflammatory, and immunoregulatory features from associated blood or immune cells. Platelets can stabilize and shield CTCs through αIIbβ3–fibrinogen interactions and P-selectin-dependent adhesion; neutrophils can promote cluster assembly through VCAM1–ITGA4 interactions and neutrophil extracellular trap (NET)-mediated capture; and CTC–DPT aggregates are associated with VCAM1–VLA-4-dependent adhesion and a PD-1⁺TIM-3⁺ exhausted T-cell phenotype. Thus, the cellular composition of an individual CTC cluster may determine not only its metastatic potential but also its sensitivity to specific therapeutic interventions [6–10, 15–18, 29, 46, 49, 53, 58, 59, 64]. This compositional heterogeneity means that not all CTC clusters should be assumed to respond uniformly to the proposed three-axis therapeutic strategy. Platelet-rich clusters may be particularly sensitive to αIIbβ3 blockade because their structural integrity and immunoprotective coating depend at least in part on platelet-integrin signaling. By contrast, homotypic tumor-cell clusters maintained primarily by direct tumor cell–tumor cell adhesion may not be adequately disaggregated by antiplatelet therapy alone. Neutrophil-rich clusters may depend more strongly on CXCL5–CXCR2, VCAM1–ITGA4, and NET-related pathways, whereas DPT-containing clusters may be more responsive to interventions targeting VLA-4-mediated adhesion, immune-checkpoint signaling, or both. Complete mesenchymal transition may weaken stable epithelial junctions and promote single-cell motility, invasion, and interactions with platelets, neutrophils, stromal cells, or vascular endothelium. Fully mesenchymal CTCs may have strong migratory capacity and heterotypic-cell interaction capacity without necessarily being the most capable of forming stable homotypic aggregates [9, 10, 17, 23, 24, 54]. This distinction is particularly important when interpreting platelet-induced EMT or NNMT-associated EMT programs, because induction of mesenchymal features should not automatically be equated with increased homotypic clustering capacity. Instead, EMT state may separately regulate collective cohesion, heterotypic adhesion, intravascular transport, immune evasion, and metastatic seeding. Likewise, the role of nicotinamide N-methyltransferase (NNMT) should be interpreted in a cell-state and compartment-specific context. NNMT may support the protective and pro-metastatic features of CTC clusters through several biologically distinct cellular sources, including tumor cells, specific CTC-associated tumor-cell states, cancer-associated fibroblast (CAF)-like stromal cells, endothelial cells, and pericyte-like populations. These compartments are unlikely to be functionally equivalent. In tumor cells, NNMT-mediated S-adenosylmethionine depletion and changes in NAD⁺-related metabolism, epigenetic plasticity, stemness, and anoikis resistance may promote persistence of specific metastasis-initiating states. In CAF-like stromal populations, NNMT may contribute to the formation of an immunosuppressive microenvironment and enhance myeloid-cell recruitment or functional polarization. Pericyte-derived NNMT has also been proposed to promote CTC–neutrophil cluster formation through the CXCL5–CXCR2 axis [35–41, 44]. Endothelial and other perivascular compartments may further participate in tumor-cell adhesion, intravasation, intravascular retention, and assembly of heterotypic circulating aggregates. Therefore, the biological or therapeutic significance of elevated NNMT activity in one cellular compartment should not be assumed to be identical to that in another. We found that the largest predicted reductions in the composite protection score occurred in colorectal-cancer stromal cells and lung fibroblasts, followed by breast-cancer CAF and endothelial-cell populations. These findings suggest that stromal/fibroblast-like and vascular-associated compartments are candidate NNMT-dependent populations. They also support the view that the role of NNMT in tumor cells or CTC-associated populations is more selective and constrained by cell state. Virtual-knockout results should be regarded as a tool for screening candidate cellular dependencies. The results support advancing the proposed three-axis strategy within a biomarker-guided, CTC-cluster-subtype-specific framework. Variables for stratification include cluster composition, epithelial/mesenchymal state, NNMT expression and its cellular source, degree of platelet coverage, neutrophil association, NET burden, adhesion-molecule profile, and immune-checkpoint activity. Only by integrating these features can we identify which CTC-cluster subtypes are most likely to respond to NNMT inhibition, αIIbβ3 blockade, PD-1/PD-L1 inhibition, or rational combinations of these interventions. Such stratification can help distinguish clusters for which physical disaggregation is the principal therapeutic objective from those requiring metabolic reprogramming, disruption of heterotypic adhesion, restoration of immune surveillance, or simultaneous targeting of multiple protective mechanisms. NNMT: metabolic reprogramming as a novel dimension for CTC cluster intervention Biological function of NNMT and regulation of tumor metabolism Studies have shown that NNMT (Fig. 4) uses the methyl donor SAM to catalyze the methylation of nicotinamide [36, 43]. NNMT expression is elevated in multiple solid tumors, including colorectal cancer, gastric cancer, pancreatic cancer, and non-small-cell lung cancer. High expression is associated with poor prognosis and resistance to chemotherapy or targeted therapy [37, 39, 58]. In hepatocellular carcinoma, NNMT occupies a key metabolic node at the intersection of NAD⁺ biology and the methionine cycle [66].Fig. 4Metabolic pathways promoting CTC–neutrophil clusters and NNMT-driven maintenance of stemness. Results: We found that NNMT catalyzes the reaction NAM + SAM → MNAM + SAH. SAM depletion causes histone hypomethylation (for example, reduced H3K27me3) and activates EMT. NNMT activity in pericytes induces CXCL5 signaling through CXCR2 on neutrophils, promoting CTC–neutrophil cluster formation; NNMT in tumor cells or cancer-associated fibroblasts (CAFs) consumes SAM and promotes EMT, stemness, and resistance to anoikis. NNMT inhibitors (NNMTi) can block both pathways, disrupt CTC–neutrophil clusters, and reverse the pro-metastatic phenotype, requiring direct validation in the proposed combination Metabolic pathways promoting CTC–neutrophil clusters and NNMT-driven maintenance of stemness. Results: We found that NNMT catalyzes the reaction NAM + SAM → MNAM + SAH. SAM depletion causes histone hypomethylation (for example, reduced H3K27me3) and activates EMT. NNMT activity in pericytes induces CXCL5 signaling through CXCR2 on neutrophils, promoting CTC–neutrophil cluster formation; NNMT in tumor cells or cancer-associated fibroblasts (CAFs) consumes SAM and promotes EMT, stemness, and resistance to anoikis. NNMT inhibitors (NNMTi) can block both pathways, disrupt CTC–neutrophil clusters, and reverse the pro-metastatic phenotype, requiring direct validation in the proposed combination The tumor-promoting mechanisms of NNMT include consumption of the methyl donor SAM, leading to reduced nicotinamide levels and accumulation of the byproduct MNAM; activation of pro-EMT transcription factors such as Snail and Twist; impairment of NAD⁺-dependent deacetylases such as SIRT1, reducing cellular metabolism; and activation of NF-κB signaling to promote inflammation, [37, 67, 68] High NNMT expression in CTCs can enhance stem-cell-like features and increase resistance to anoikis [43, 47]. NNMT knockout can trigger a sharp increase in reactive oxygen species and induce ferroptosis by depleting NAD(P)⁺ and glutathione reserves, thereby blocking breast-cancer tumor growth and metastasis [69]. NNMT is a key factor linking metabolic stress with the epigenetic plasticity of metastasis-initiating cells [49, 70]. Research progress and clinical potential of NNMT inhibitors NNMT inhibitors remain at the preclinical validation stage [37, 38, 71]. The small-molecule NNMT inhibitors AG-670 and AO-022 may act by affecting mitochondrial oxygen consumption [72]. NNMT inhibition can reduce cell viability, increase ROS levels, and activate apoptotic pathways [73]. A representative NMT inhibitor is Zelenirstat, an orally administered dual NMT1/2 inhibitor [74, 75]. These drugs do not cause severe myelosuppression or alopecia [74, 76]. Several novel NNMT inhibitors with improved selectivity and oral bioavailability have entered late-stage preclinical development [77]. Theoretical association between NNMT inhibitors and CTC clusters Tumor-metabolism inference suggests that upregulated NNMT expression maintains a ‘stem-cell-like’ state in CTCs, accompanies enhanced intercellular adhesion, and promotes CTC aggregation into clusters [36, 37, 73]. NNMT occupies a central position in cellular methylation homeostasis [38]. NNMT inhibition has been associated with changes in epithelial–mesenchymal plasticity in selected models; whether this improves immune recognition of heterogeneous CTC clusters remains unproven [37]. NNMT inhibition may modulate the tumor immune microenvironment, enhance PD-1/PD-L1 blockade, and reduce MDSC migration to the premetastatic niche, supporting its potential role in the triple combination [40, 78, 79]. Potential complementary or antagonistic interactions framework of the triple therapy The present analyses do not demonstrate additivity or synergy. These interactions must be tested using factorial perturbation, combination-index analysis, and in vivo dose–response experiments. The steps are disaggregation, exposure, and clearance. The first step is ‘disaggregation’: αIIbβ3 blockers such as Tirofiban or Abciximab occupy platelet GPIIb/IIIa receptors and prevent fibrinogen binding [29, 50]. Such inhibitors may disrupt CTC–neutrophil clusters, although this requires experimental validation [27, 29, 80]. In mouse models, the oral GPIIb/IIIa antagonist XV454 suppressed tumor-cell-induced thrombocytopenia and pulmonary metastasis, indicating that αIIbβ3-inhibitor monotherapy can interfere with CTC–platelet interactions [51]. Multiple integrins participate in the platelet–tumor-cell adhesion axis, and αIIbβ3 is the most clinically actionable target [81]. The second key step is ‘exposure,’ mediated by an NNMT inhibitor. NNMT inhibitors may reduce S-adenosylmethionine (SAM) levels or disrupt NAD⁺ metabolism, inducing metabolic stress in CTCs; this stress may lower intracellular anti-apoptotic protein levels and weaken intercellular adhesion within clusters [36, 37]. NNMT inhibition may produce a phenotypic ‘reversal’ in CTCs that have undergone epithelial–mesenchymal transition (EMT), restoring some epithelial features; cells with epithelial features may be more readily recognized by the immune system [37, 43, 82]. Pericytes with high NNMT expression can promote aggregation of colorectal-cancer cells, neutrophils, and endothelial cells through the CXCL5/CXCR2 axis, driving CTC–neutrophil cluster formation. Inhibiting NNMT in pericytes may therefore also disrupt such clusters. In mouse models, NNMT inhibitors can reduce tumor burden and metastasis and can restore the efficacy of immune-checkpoint blockade by reducing CAF-mediated MDSC recruitment and reactivating CD8⁺ T cells [40, 78]. The final ‘clearance’ step is performed primarily by PD-1/PD-L1 antibodies. First, these antibodies can restore the immune-surveillance capacity and activity of T cells and NK cells [27, 44, 79]. Mechanistically, antiplatelet drugs first ‘dismantle’ the platelet ‘shield,’ NNMT inhibitors further expose CTC clusters and modulate their immunosuppressive microenvironment, and T cells activated by PD-1/PD-L1 blockade may then recognize and clear CTCs [6, 83, 84]. Second, NK-cell activity may also be restored, reinforcing the ‘clearance’ process [37]. Clinical studies indicate that PD-L1 expression on CTCs may serve as a biomarker for predicting response to PD-1/PD-L1 immunotherapy. Among patients with advanced lung cancer, those with PD-L1-positive CTCs had significantly higher objective response rates (84.2% vs. 36.8%) and longer progression-free survival (16 months vs. 4 months) after chemoimmunotherapy than patients with PD-L1-negative CTCs, suggesting that CTCs are important effector targets of immune-checkpoint inhibitors [85]. Known associations between CTC PD-L1 status and ICB benefit in other solid tumors further support this biomarker strategy [6, 30, 86, 87]. System-level interactions, potential antagonism, and tumor-type dependence The proposed triple regimen should not be understood as three independent modules that necessarily and uniformly act synergistically. Platelets, neutrophils, stromal cells, endothelial cells, and lymphocytes constitute a coupled and adaptive system; intervention in one compartment may redistribute cell trafficking, inflammatory signaling, and selective pressure in others. Therefore, the clinically meaningful outcome will depend on the system-level net balance among CTC-cluster disruption, maintenance of vascular homeostasis, preservation of effector-cell fitness, and removal of residual immune barriers, rather than solely on inhibition of any single target [17, 40, 86]. At the thromboinflammatory level, αIIbβ3 inhibition may reduce fibrinogen-dependent platelet–CTC adhesion and thereby weaken platelet coating. Platelets maintain endothelial integrity, seal inflammatory microinjury, and coordinate leukocyte recruitment and extravasation. Insufficient inhibition may leave a residual platelet shield, whereas excessive inhibition or inappropriate timing may increase vascular leakage or bleeding and alter neutrophil and monocyte recruitment, endothelial activation, and subsequent inflammatory dynamics. Platelet function, endothelial-injury markers, and NET-related indices should therefore be evaluated under short-acting, exposure-controllable dosing regimens rather than relying on platelet counts alone [15–17, 77, 87]. At the metabolic level, NNMT inhibition may weaken the adaptive capacity of tumor cells, CAFs, and pericytes by altering nicotinamide/SAM flux, epigenetic plasticity, and CXCL5–CXCR2-dependent myeloid-cell recruitment. However, this intervention is not intrinsically tumor-cell specific. NAD⁺ salvage synthesis and methionine/SAM-dependent methylation are also important for T- and NK-cell proliferation, mitochondrial fitness, chromatin state, cytokine production, and cytotoxic activity. Systemic NNMT inhibition may therefore relieve stromal immunosuppression while simultaneously redistributing nicotinamide and methyl-donor reserves in immune cells, with direction and magnitude varying according to inhibitor selectivity, dose, tissue exposure, and baseline metabolic state. Before assuming a uniform immune-sensitizing effect, cell-type-resolved pharmacodynamic studies should separately quantify NAD⁺/NADH, SAM/SAH, MNAM, mitochondrial function, and effector programs in CTCs, CAFs/pericytes, CD8⁺ T cells, and NK cells [35–39, 62–64]. An additional uncertainty is whether metabolic and microenvironmental compensation may attenuate the effects of NNMT inhibition. Suppression of NNMT may redirect nicotinamide toward the NAMPT-mediated NAD⁺ salvage pathway, thereby partially maintaining intracellular NAD⁺ availability and metabolic fitness. Tumor cells and stromal populations may also adapt through remodeling of methionine/SAM and one-carbon metabolism, activation of alternative methylation programs, or increased reliance on fatty-acid oxidation and other mitochondrial energy pathways. Such adaptations could preserve stress tolerance even when the NNMT-associated metabolic state is weakened. Compensation may also occur at the multicellular level. Reduced platelet-dependent protection may be accompanied by increased reliance on neutrophil extracellular traps or alternative myeloid-recruitment signals, including CXCL1-, CXCL2-, or CXCL8-associated pathways. Consequently, a platelet-dominant CTC-cluster state could shift toward a neutrophil/NET-dominant protective state rather than undergoing complete disaggregation. These possibilities remain hypothetical and should be evaluated using paired metabolomic profiling, NAD⁺/NADH and SAM/SAH measurements, isotope-tracing experiments, transcriptional and protein-level analysis of compensatory pathways, and longitudinal assessment of platelet and neutrophil composition within CTC clusters. At the immune level, PD-1/PD-L1 blockade is a permissive intervention rather than a treatment that is sufficient by itself. Reactivation of exhausted lymphocytes can translate into CTC killing only when tumor antigens are adequately presented, functionally competent T-cell or NK-cell populations are present, and effector cells can spatially approach and contact CTCs. Even after PD-1 signaling is blocked, persistent platelet/fibrin coating, NETs, ARG1-mediated nutrient depletion, TGF-β signaling, or compensatory TIGIT and TIM-3 pathways may still prevent effective recognition and immune-synapse formation. Thus, immune activation without sufficient removal of the coating or disruption of NETs may fail to achieve clearance and may instead aggravate inflammation without eliminating protected CTC populations [18–22, 27, 44, 73, 75, 85]. Some tumors may predominantly form platelet-dominant clusters, whereas others may be driven primarily by neutrophil/NET programs or NNMT-high CAF/pericyte compartments; mixed states are likely. The proposed regimen should therefore be developed as a biomarker-guided strategy rather than a universally applicable pan-cancer therapy. Candidate stratification variables include markers of platelet coating (ITGA2B/ITGB3, SELP, and fibrinogen-related signaling), neutrophil/NET markers (CXCL5/CXCR2, MPO, ELANE, and citrullinated histone H3), stromal and perivascular programs (NNMT, FAP, and PDGFRB), and markers of immune competence (HLA class I/antigen-presentation machinery, IFN-γ signaling, PD-L1, and T/NK cytotoxicity scores). Analysis of CTC-cluster composition before treatment and early during treatment should help determine whether all three components are required, whether sequential rather than concurrent administration should be used, and whether a residual dominant axis requires additional targeted intervention. Within this systems framework, the objective is not to suppress every pathway, but to remove the protective niche sufficiently while preserving vascular integrity and antitumor immune fitness. Prospective toxicity and risk framework for the triple strategy (Table 2) Table 2Prospective toxicity and prevalidation risk framework for the proposed NNMT inhibitor–αIIbβ3 inhibitor–PD-1/PD-L1 blockade strategyRiskSourcePre-validation indicatorsBleeding/thrombocytopeniaαIIbβ3 inhibition superimposed on cancer-associated thrombocytopenia and possible ICI-related immune thrombocytopeniaSerial platelet count and change from baseline; PT and APTT; standardized bleeding scoreImmune-related adverse eventsPD-1/PD-L1 blockade, with possible amplification or phenotypic alteration by metabolic modulationCytokine panel; liver enzymes and bilirubin; colitis markers; organ-specific immune-toxicity surveillanceSystemic metabolic effectsNNMT inhibition in tumor, stromal, endothelial, and immune compartments, with potential perturbation of NAD + and methyl-donor metabolismNAD +/NAM/MNAM and SAM/SAH ratios; redox and viability assays; T-cell activation, proliferation, and cytotoxic functionTumor-type-specific riskDependence on platelet cloaking, neutrophil/NET support, baseline cytopenias, and organ-specific metastatic biology differs among cancer typesCTC-cluster phenotype; platelet-cloaking/adhesion score; platelet/NET score; baseline marrow, hepatic, and coagulation statusAbbreviations: APTT Activated partial thromboplastin time, CTC Circulating tumor cell, ICI Immune-checkpoint inhibitor, NAM Nicotinamide, MNAM 1-methylnicotinamide, NAD⁺ Nicotinamide adenine dinucleotide, NET Neutrophil extracellular trap, PT Prothrombin time, SAM S-adenosylmethionine, SAH S-adenosylhomocysteine Prospective toxicity and prevalidation risk framework for the proposed NNMT inhibitor–αIIbβ3 inhibitor–PD-1/PD-L1 blockade strategy Abbreviations: APTT Activated partial thromboplastin time, CTC Circulating tumor cell, ICI Immune-checkpoint inhibitor, NAM Nicotinamide, MNAM 1-methylnicotinamide, NAD⁺ Nicotinamide adenine dinucleotide, NET Neutrophil extracellular trap, PT Prothrombin time, SAM S-adenosylmethionine, SAH S-adenosylhomocysteine Because no direct safety data currently exist for the proposed triple combination, toxicity should be evaluated within a prospective interaction framework rather than described as a collection of readily manageable adverse events. The principal risks arise from overlapping pharmacodynamic effects, the uncertain systemic consequences of NNMT inhibition, and differences among cancer types in their dependence on platelet- and neutrophil-mediated protection. Risk assessment should begin with characterization of each single agent, proceed to pairwise combinations, and only then evaluate the triple regimen in ex vivo whole-blood systems, immunocompetent metastatic models, and patient-derived CTC-cluster platforms. Dose escalation should occur only when platelet reduction, coagulation abnormalities, clinically meaningful bleeding, organ-specific immune toxicity, and loss of T-cell function or metabolic homeostasis all remain below prespecified thresholds. Sequential dosing may be tested as a design variable, but should not be described as an established risk-mitigation strategy before two-drug and three-drug interaction data are available. Patient selection should likewise be biomarker guided rather than based on a broad assumption of ‘low-risk cancer types.’ Ideal preclinical subjects should have a demonstrable CTC-cluster/platelet/NET phenotype, adequate baseline platelet reserve, and detectable activity of the PD-1/PD-L1 axis. Conversely, baseline thrombocytopenia, active bleeding, uncontrolled autoimmune disease, marked hepatic dysfunction, or absence of the target cluster phenotype should be regarded as candidate exclusion criteria or high-risk features in early translational validation. The purpose of this framework is not to claim that these risks are easy to manage, but to specify the measurements required before the strategy can advance. Preclinical testability and translational considerations and safety considerations Preclinical accessibility of individual agents (Table 3) Table 3Preclinical and clinical evidence for agents targeting the triple-axis strategy Table 3Preclinical and clinical evidence for agents targeting the triple-axis strategy Preclinical and clinical evidence for agents targeting the triple-axis strategy Drugs such as PD-1 or PD-L1 antibodies have been approved for multiple solid tumors, have well-characterized safety profiles, and are supported by extensive experience in combination therapy [30, 80, 6]. αIIbβ3 inhibitors such as Tirofiban and Eptifibatide are approved for acute coronary syndromes, are administered intravenously, and have short half-lives; Abciximab is a longer-acting agent. Their principal risk is bleeding, and particular caution is required in patients with cancer who already have low platelet counts [29, 50, 88]. Platelet inhibition can disrupt CTC–platelet clusters and may improve immune surveillance by enhancing T-cell activity, providing an additional theoretical rationale for combining αIIbβ3 inhibitors with immunotherapy [89]. No NNMT inhibitor has yet received formal marketing approval. Several NNMT inhibitors are summarized in Supplementary Table 1 [36, 37, 72, 90]. Clinical data for some NMT inhibitors indicate that targeting metabolic pathways is feasible [86, 91]. NNMT inhibitors have attracted broad attention because they can effectively eliminate circulating tumor cell clusters (CTC clusters) in preclinical models [37]. In high-grade serous ovarian cancer, NNMT inhibitors reduced tumor burden and metastasis and restored the efficacy of immune-checkpoint blockade by remodeling the immunosuppressive function of CAFs, suggesting broad translational potential for combining NNMT-targeted strategies with immunotherapy. Real-world pharmacovigilance data on the safety of short-term GPIIb/IIIa-inhibitor use in the oncology setting have begun to emerge [92]. Potential overlapping toxicity risks and mitigation strategies (Fig. 5) Proposed experimental validation workflow for circulating tumor cell (CTC) clusters and a hypothetical sequential treatment timeline. *A Experimental validation workflow, including phenotypic analysis of CTC clusters, multicellular co-culture, microfluidic shear-flow analysis, immune killing assays, and animal metastasis models; (B) Hypothetical sequential treatment timeline. * Result:Using a multi-feature analytical framework, we anticipate that high-risk CTC clusters with enhanced adhesive capacity, immune evasion, and metabolic activity can be identified. Our combination treatment may effectively reduce cluster size, decrease platelet coverage and neutrophil adhesion, enhance CD8⁺ T cell– and NK cell–mediated cytotoxicity, and suppress immune checkpoint expression. In vivo, we expect these treatments to reduce CTC burden as well as lung and liver metastases, while sequential therapy may improve safety The triple-combination regimen may generate toxicity through several partially overlapping mechanisms. First, inhibition of αIIbβ3 suppresses platelet aggregation, while PD-1/PD-L1 blockade can, in rare cases, induce immune-mediated thrombocytopenia. Their combination may therefore increase platelet depletion and bleeding risk [27, 29, 93]. Second, PD-1/PD-L1 blockade may induce excessive or dysregulated immune activation. NNMT inhibition alters metabolic signaling across tumor cells, stromal compartments, and immune cells, and this may further modulate the magnitude or duration of immune responses [5, 44, 94]. Third, NNMT inhibitors may produce systemic metabolic toxicity as well as gastrointestinal toxicity. Zelenirstat has reported dose-limiting diarrhea. When combined with immunotherapy or antiplatelet therapy, adverse effects may become more pronounced [74, 91]. As illustrated in Fig. 5, these risks should be evaluated within an integrated, staged preclinical framework rather than being treated as isolated safety signals. The first step is phenotypic characterization of patient-derived CTC clusters, including epithelial, mesenchymal, adhesion-related, immune checkpoint, chemokine receptor, and NNMT-associated markers to identify biologically distinct and potentially high-risk subtypes. Subsequently, tumor cells, platelets, neutrophils, and T/NK cells can be used to reconstruct heterogeneous CTC clusters in vitro. Under physiologically relevant shear flow conditions, comparative experiments should evaluate vehicle, monotherapy, dual therapy, and triple therapy, as well as sequential regimens. Outcomes should include cluster size, platelet coverage, and neutrophil attachment. In downstream cytotoxic assays, investigators should assess tumor cell killing, IFN-γ and granzyme B production, and expression of PD-1, TIM-3, and TIGIT. In sequential strategies, short-term αIIbβ3 inhibition may be used first to reduce platelet shielding of CTC clusters. If platelet counts remain stable and no clinically significant bleeding is observed, PD-1/PD-L1 blockade and NNMT inhibition can be introduced [47, 95]. Patient selection should consider baseline platelet status, bleeding risk, immune-related comorbidities, and whether CTC clusters display platelet-dependent or neutrophil-dependent phenotypes [23, 96]. Pan-cancer bulk RNA-seq and single-cell validation of NNMT-related triple-axis coactivation Using two types of TCGA-related data—bulk data and single-cell data packages—we constructed two complementary levels of validation evidence. Bulk analysis was used to test NNMT/three-axis co-activation across cancer types, whereas single-cell analysis assessed whether this state was retained at the resolution of cell states and cellular compartments (Fig. 6).Fig. 6A: Cell-level Spearman correlations between NNMT expression and feature scores across cancer types;B: Endpoint enrichment of the NNMT-axis pseudotrajectory program. Single-cell resolution, N = 483,590;C shows patient-level paired comparisons between CTC clusters and pools of single CTCs in GSE51827; D compares single CTCs, CTC-cluster tumor cells, and leukocyte-associated CTCs in GSE109761; E summarizes median differences in intracluster cohesion, platelet/coagulation, checkpoint/exhaustion, and composite three-axis scores.*. Results: Eight public cohorts showed a consistent positive correlation between NNMT expression and the composite three-axis score across cancer types. The strongest coupling was observed in breast cancer (Spearman rho = 0.315), colorectal cancer (rho = 0.268), melanoma (rho = 0.253), hepatocellular carcinoma (rho = 0.245), and lung cancer (rho = 0.229). Along the feature-level NNMT-axis pseudotrajectory, the composite three-axis score increased from T0_low to T4_high in every cancer type, with the largest endpoint increases in lung cancer (T4–T0 = 1.463), hepatocellular carcinoma (1.461), and the CTC/breast-cancer model (1.451) A: Cell-level Spearman correlations between NNMT expression and feature scores across cancer types;B: Endpoint enrichment of the NNMT-axis pseudotrajectory program. Single-cell resolution, N = 483,590;C shows patient-level paired comparisons between CTC clusters and pools of single CTCs in GSE51827; D compares single CTCs, CTC-cluster tumor cells, and leukocyte-associated CTCs in GSE109761; E summarizes median differences in intracluster cohesion, platelet/coagulation, checkpoint/exhaustion, and composite three-axis scores.*. Results: Eight public cohorts showed a consistent positive correlation between NNMT expression and the composite three-axis score across cancer types. The strongest coupling was observed in breast cancer (Spearman rho = 0.315), colorectal cancer (rho = 0.268), melanoma (rho = 0.253), hepatocellular carcinoma (rho = 0.245), and lung cancer (rho = 0.229). Along the feature-level NNMT-axis pseudotrajectory, the composite three-axis score increased from T0_low to T4_high in every cancer type, with the largest endpoint increases in lung cancer (T4–T0 = 1.463), hepatocellular carcinoma (1.461), and the CTC/breast-cancer model (1.451) CTC-specific validation of single CTCs, CTC clusters, and leukocyte-associated CTC states To directly test whether the proposed protective-axis program was enriched in CTC-specific states, we performed focused reanalysis of GSE51827 and GSE109761. In GSE51827, the intracluster-cohesion score was higher in CTC clusters than in patient-matched pools of single CTCs (median difference = + 0.91, Wilcoxon p = 0.017). In GSE109761, compared with single CTCs, tumor cells from CTC clusters had higher intracluster-cohesion, platelet/coagulation, checkpoint/exhaustion, and composite three-axis scores; leukocyte-associated CTCs showed the most consistent increases in checkpoint/exhaustion and composite three-axis scores relative to single CTCs (Fig. 7). Thus, CTC-focused datasets were directionally consistent with the multi-cancer analysis: aggregation and association with immune cells represent a stronger protective-axis state rather than merely reflecting background tumor heterogeneity (Fig. 8).Fig. 7A Cell populations ranked by the predicted reduction in the composite three-axis score after virtual knockout; (B): cancer-level NNMT-positive proportions and NNMT–three-axis coupling; (C): estimated changes in the composite protection score in each dataset after NNMT knockdown, platelet/αIIbβ3-axis perturbation, checkpoint-axis perturbation, or combined perturbation; (D): cell types ranked by the strongest combined-perturbation response, with stacked bars showing the contribution of each single-axis component; (E): responses of representative ligand–receptor gene axes.*. Results: Virtual-knockout analysis screened candidate NNMT-dependent compartments. The composite three-axis score predicted the largest reductions in colorectal-cancer stromal cells and lung fibroblasts (median difference = − 0.407), followed by breast-cancer CAFs and endothelial cells (median difference = − 0.323). These findings provide an order of priority for experimental validation of candidate sensitive cell populationsFig. 8A Heatmap of cross-cancer correlations for the NNMT axis in TCGA pan-cancer data; (B). TCGA pan-cancer three-axis differential dot plot; C. TCGA pan-cancer NNMT-axis correlation bubble plot. Results: Pan-cancer analysis showed that NNMT expression was positively correlated with the composite three-axis score in 29 of 30 TCGA cancer types, and all 29 reached FDR < 0.05. At the module level, NNMT repeatedly showed positive correlations with the EMT/metabolism, platelet/coagulation, neutrophil/NET, checkpoint/exhaustion, and cytotoxicity/NK modules. The median cross-cancer Spearman rho values for these modules and the composite three-axis score were 0.376, 0.358, 0.427, 0.381, 0.389, and 0.500, respectively A Cell populations ranked by the predicted reduction in the composite three-axis score after virtual knockout; (B): cancer-level NNMT-positive proportions and NNMT–three-axis coupling; (C): estimated changes in the composite protection score in each dataset after NNMT knockdown, platelet/αIIbβ3-axis perturbation, checkpoint-axis perturbation, or combined perturbation; (D): cell types ranked by the strongest combined-perturbation response, with stacked bars showing the contribution of each single-axis component; (E): responses of representative ligand–receptor gene axes.*. Results: Virtual-knockout analysis screened candidate NNMT-dependent compartments. The composite three-axis score predicted the largest reductions in colorectal-cancer stromal cells and lung fibroblasts (median difference = − 0.407), followed by breast-cancer CAFs and endothelial cells (median difference = − 0.323). These findings provide an order of priority for experimental validation of candidate sensitive cell populations A Heatmap of cross-cancer correlations for the NNMT axis in TCGA pan-cancer data; (B). TCGA pan-cancer three-axis differential dot plot; C. TCGA pan-cancer NNMT-axis correlation bubble plot. Results: Pan-cancer analysis showed that NNMT expression was positively correlated with the composite three-axis score in 29 of 30 TCGA cancer types, and all 29 reached FDR < 0.05. At the module level, NNMT repeatedly showed positive correlations with the EMT/metabolism, platelet/coagulation, neutrophil/NET, checkpoint/exhaustion, and cytotoxicity/NK modules. The median cross-cancer Spearman rho values for these modules and the composite three-axis score were 0.376, 0.358, 0.427, 0.381, 0.389, and 0.500, respectively The triple-intervention strategy produced a stronger predicted effect than any individual treatment axis. Across eight single-cell cohorts comprising 483,590 cells, the combined perturbation outperformed the best single-axis perturbation in every analyzed dataset. The strongest combined responses were enriched primarily in endothelial, stromal/fibroblast-like, and exhausted/proliferative T-cell compartments. Metabolic adaptation, platelet/coagulation shielding, and checkpoint-mediated immunosuppression constituted complementary, nonredundant modules. Predictor variables were reduced to the dataset-specific 10th percentile, and downstream changes were estimated from positive expression–score slopes. Bulk and single-cell analyses reached consistent conclusions: NNMT was associated with a recurrent metabolic–thromboinflammatory–immune co-activation state; at single-cell resolution, this state was associated with stromal/fibroblast-like compartments. Feature-gene coverage was high in all datasets. After stratification within each cancer type according to the median NNMT expression level, NNMT-high tumors showed FDR-significant differences in 29 tumor types. Cancer types with relatively large differences included BLCA, LUSC, READ, COAD, ESCA, and THCA. These findings suggest that the NNMT-associated metabolic, thromboinflammatory, and immune-checkpoint axes constitute a recurrent tumor-ecological state across cancer types. Machine-learning-based prioritization of high-risk CTC states and identification of candidate triple-axis responders Review of local and public data packages yielded 18,093 candidate files; after screening, 4,558 biologically relevant data files belonging to 135 dataset groups were retained. Data directly usable as machine-learning inputs included GSE109761, GSE51827, GSE111065, GSE202650, the existing eight-cohort single-cell atlas, and TCGA/Xena bulk-validation outputs. In GSE109761_ human, 281 CTC-labeled cells were used for supervised learning, including 150 clustered/leukocyte-associated CTCs and 131 single CTCs. The model combining axis features with ligand–receptor features achieved a cross-validated AUROC of 0.711 ± 0.065, AUPRC of 0.741 ± 0.063, and balanced accuracy of 0.668 ± 0.058. The composite three-axis score alone still carried significant information (AUROC = 0.638; bootstrap 95% CI, 0.575–0.702; permutation-test p = 0.0002). Because CTC-cluster labels were not uniformly available across cancer types, cross-cancer machine learning was performed at the TCGA PanCancer bulk-transcriptome level using leave-one-cancer-out cross-validation for the risk of death within 3 years of overall survival. After excluding samples with uncertain 3-year survival status and cancer types with insufficient numbers of events or non-events, 4,949 evaluable samples across 26 cancer types were retained, of which 1,937 died within 3 years. The axis-core model, integrating NNMT expression, intracluster-cohesion score, metabolism/EMT score, platelet/coagulation score, neutrophil/NET score, checkpoint/exhaustion score, cytotoxicity/NK score, and composite three-axis score, performed best: pooled AUROC = 0.575, AUPRC = 0.427, balanced accuracy = 0.565, mean held-out-cancer AUROC = 0.560, and median = 0.565. The NNMT single-feature model showed a similar but slightly weaker signal, with a pooled AUROC of 0.571 and AUPRC of 0.424, whereas the model using only the composite three-axis score had limited discrimination, with a pooled AUROC of 0.517 and AUPRC of 0.404. Thus, the pan-cancer signal was modest and unsuitable for diagnosis, but it suggested that the NNMT-associated three-axis program carries reproducible risk information beyond individual cancer cohorts. In the axis-core logistic-regression model, the highest standardized positive coefficient was for the composite three-axis score (β = 0.564), followed by NNMT expression (β = 0.303) and the intracluster-cohesion score (β = 0.072). These coefficients support the use of the integrated three-axis score and NNMT expression as exploratory screening features for cross-cancer risk states (Fig. 9 Machine-learning-based prioritization of high-risk CTC states and identification of candidate triple-axis responders).Fig. 9A TCGA PanCancer leave-one-cancer-out machine-learning validation; (B). ranking heatmap of CTC-label machine-learning models; (C). cross-validated ROC curve for the best-performing CTC-label machine-learning model; (D). feature heatmap of GSE109761 CTC-labeled cells ordered by cross-validated machine-learning probability; (E). TCGA PanCancer leave-one-cancer-out validation of the axis-core model; (F). pooled held-out-sample ROC and precision–recall analyses across TCGA cancer types; G. standardized coefficients of the final TCGA PanCancer axis-core logistic-regression model. Results: GSE109761_human included 281 CTC-labeled cells, of which 150 were clustered/leukocyte-associated CTCs and 131 were single CTCs. The combined model based on axis and ligand–receptor features performed best in repeated stratified five-fold cross-validation, with an AUROC of 0.711 ± 0.065, AUPRC of 0.741 ± 0.063, and balanced accuracy of 0.668 ± 0.058. The composite three-axis score alone retained significant discriminatory ability, with an AUROC of 0.638, bootstrap 95% CI of 0.575–0.702, and permutation-test p value of 0.0002. We used leave-one-cancer-out cross-validation to evaluate cross-cancer generalizability in the TCGA PanCancer cohort, with death within 3 years of overall survival as the binary endpoint. The final analysis included 4,949 evaluable samples across 26 cancer types, including 1,937 deaths within 3 years. Results: The model showed a certain degree of discriminatory ability and carried reproducible risk information at the cross-cancer level. Although performance was moderate and exploratory, the model may be suitable as a screening tool for candidate high-risk cells and potential treatment responsiveness A TCGA PanCancer leave-one-cancer-out machine-learning validation; (B). ranking heatmap of CTC-label machine-learning models; (C). cross-validated ROC curve for the best-performing CTC-label machine-learning model; (D). feature heatmap of GSE109761 CTC-labeled cells ordered by cross-validated machine-learning probability; (E). TCGA PanCancer leave-one-cancer-out validation of the axis-core model; (F). pooled held-out-sample ROC and precision–recall analyses across TCGA cancer types; G. standardized coefficients of the final TCGA PanCancer axis-core logistic-regression model. Results: GSE109761_human included 281 CTC-labeled cells, of which 150 were clustered/leukocyte-associated CTCs and 131 were single CTCs. The combined model based on axis and ligand–receptor features performed best in repeated stratified five-fold cross-validation, with an AUROC of 0.711 ± 0.065, AUPRC of 0.741 ± 0.063, and balanced accuracy of 0.668 ± 0.058. The composite three-axis score alone retained significant discriminatory ability, with an AUROC of 0.638, bootstrap 95% CI of 0.575–0.702, and permutation-test p value of 0.0002. We used leave-one-cancer-out cross-validation to evaluate cross-cancer generalizability in the TCGA PanCancer cohort, with death within 3 years of overall survival as the binary endpoint. The final analysis included 4,949 evaluable samples across 26 cancer types, including 1,937 deaths within 3 years. Results: The model showed a certain degree of discriminatory ability and carried reproducible risk information at the cross-cancer level. Although performance was moderate and exploratory, the model may be suitable as a screening tool for candidate high-risk cells and potential treatment responsiveness Network pharmacology-based validation of the triple-axis strategy In order to explore the pathway among Triple-Axis, we design Network Pharmacology-Based Validation (Fig. 10).Fig. 10Network-pharmacology target-network analysis of the triple-intervention strategy. (A) drug–target–disease network; (B), GO Biological Process enrichment bar chart; (C), GO Biological Process enrichment bubble plot; (D), core targets selected according to degree values; (E), KEGG pathway-enrichment bubble plot; (F), mechanism-curated protein–protein interaction network of three-axis targets.*Results: We found that the direct intersecting targets of the triple-intervention strategy mainly included NNMT, ITGA2B, ITGB3, PDCD1, and CD274, corresponding respectively to NNMT-mediated metabolic regulation, platelet adhesion, and immune-checkpoint pathways. Our PPI network suggested that EMT/adhesion, platelet coagulation, TGF-β signaling, and immune exhaustion may jointly participate in maintaining CTC clusters. GO enrichment results showed that the enriched biological processes were concentrated primarily in regulation of cell adhesion, leukocyte-mediated immunity, leukocyte migration, and cytotoxicity. KEGG enrichment results showed pathways related to neutrophil extracellular trap formation, integrin signaling, cell-adhesion molecules, TGF-β signaling, and Th17-cell differentiation. These results support the possibility that the triple-intervention strategy may act on the protective niche of CTC clusters at multiple levels—metabolism, thrombosis, and immunity Network-pharmacology target-network analysis of the triple-intervention strategy. (A) drug–target–disease network; (B), GO Biological Process enrichment bar chart; (C), GO Biological Process enrichment bubble plot; (D), core targets selected according to degree values; (E), KEGG pathway-enrichment bubble plot; (F), mechanism-curated protein–protein interaction network of three-axis targets.*Results: We found that the direct intersecting targets of the triple-intervention strategy mainly included NNMT, ITGA2B, ITGB3, PDCD1, and CD274, corresponding respectively to NNMT-mediated metabolic regulation, platelet adhesion, and immune-checkpoint pathways. Our PPI network suggested that EMT/adhesion, platelet coagulation, TGF-β signaling, and immune exhaustion may jointly participate in maintaining CTC clusters. GO enrichment results showed that the enriched biological processes were concentrated primarily in regulation of cell adhesion, leukocyte-mediated immunity, leukocyte migration, and cytotoxicity. KEGG enrichment results showed pathways related to neutrophil extracellular trap formation, integrin signaling, cell-adhesion molecules, TGF-β signaling, and Th17-cell differentiation. These results support the possibility that the triple-intervention strategy may act on the protective niche of CTC clusters at multiple levels—metabolism, thrombosis, and immunity Discussion We propose a triple-treatment regimen for heterogeneous circulating tumor cell (CTC) clusters [31, 37, 97]. The innovation of this strategy lies in three aspects. First, the three targets act on distinct biological processes, including metabolism, platelet activity, and immune evasion, and are mechanistically complementary. Second, synergy: removal of the platelet barrier may expose tumor antigens and create conditions for immune-mediated clearance. Third, each individual agent is already supported by preclinical or clinical evidence [30, 44, 98]. A major gap in this approach is that no experimental studies have yet demonstrated the efficacy of combining NNMT inhibitors with the other two drug classes. Future studies are required using mouse CTC cluster xenograft models or microfluidic co-culture systems to validate in vivo feasibility [31, 99]. A previous ovarian cancer study using spatial transcriptomics and single-cell RNA sequencing mapped the role of NNMT in the tumor microenvironment. This study developed highly selective NNMT inhibitors that reduced tumor burden, suppressed metastasis, and restored immune checkpoint blockade efficacy in multiple mouse models, providing strong proof-of-concept for NNMT targeting in clinical translation. Most NNMT inhibitors remain in early development stages, with limitations such as poor selectivity, low membrane permeability, or suboptimal pharmacokinetics (PK) [40, 74, 75]. Evidence suggests that the composition of CTC–leukocyte clusters is associated with subtype and population background, and biomarker-guided stratification may optimize therapeutic strategies. During treatment, platelet-dependent states may shift toward neutrophil-dependent states, promoting resistance [37, 61]. Innovative liquid biopsy technologies detecting PD-L1-positive CTCs enable dynamic monitoring and patient stratification, thereby improving therapeutic precision [100]. Advances in single-cell multi-omics of CTC clusters provide rich resources for understanding resistance mechanisms [31]. Long-term use of αIIbβ3 inhibitors increases bleeding risk; thus, they may be more suitable for short-term “debulking” or “shield removal” strategies.PD-1/PD-L1 antibodies require continuous administration to maintain T-cell effector function. The optimal timing and dosing of NNMT inhibitors remain unclear. To translate this concept into clinical practice, rational sequencing and dosing strategies must be designed [29–31, 50]. New pharmacokinetic modeling approaches have been introduced to optimize sequential administration of antiplatelet therapies in cancer patients [101]. It must be recognized that CTC clusters represent only one stage of the metastatic cascade [30]. Even complete clearance of circulating CTCs cannot prevent progression of micrometastatic lesions already established in distant organs.VEGFR1⁺ bone marrow-derived progenitor cells, chemokines, and exosomes contribute to formation of pre-metastatic niches, preconditioning distant sites for colonization [102–107]. Therefore, the most appropriate clinical application of the triple therapy may be as adjuvant treatment after primary tumor resection in patients with high risk of recurrence [5, 44]. It may also be combined with anti-angiogenic agents or stromal modulators targeting micrometastatic niches [25, 31, 76]. Animal studies show that the antiplatelet agent ifetroban reduces lung metastasis by 67%, liver metastasis by 60%, and circulating tumor cells by more than 50%, indicating strong potential for antiplatelet strategies in adjuvant settings [55, 56]. Clinical trials evaluating perioperative antiplatelet interventions are beginning to define the risk–benefit balance in adjuvant therapy [108]. In preclinical models, modulation of CAF-targeted pre-metastatic niches may synergize with metabolic inhibitors [109]; LOX and exosomes have also been identified as key mediators of niche formation [110–113]. We performed complementary computational validation using TCGA and single-cell datasets. We found that NNMT-high tumors in TCGA are broadly enriched for composite tri-axis activity. Single-cell analyses further showed that NNMT-associated co-activation is localized to cell-state trajectories and stromal/fibroblast-like compartments. We also found that the αIIbβ3/platelet axis reduces platelet-mediated shielding, while PD-1/PD-L1 blockade restores immune clearance. In machine learning analyses, TCGA pan-cancer leave-one-cancer-out models showed only limited discriminative ability for 3-year overall survival mortality risk. Importantly, this endpoint is derived from bulk tumor transcriptomes rather than direct CTC cluster labels. These results support exploratory cross-cancer risk prioritization and hypothesis generation regarding NNMT-associated tri-axis states [6, 27, 31, 37, 43, 99, 114–116, 117]. Conclusions CTC clusters utilize multiple protective mechanisms, including platelet shielding, neutrophil/NET-associated inflammation, metabolic adaptation, and checkpoint-mediated immune evasion. We propose a triple intervention strategy combining NNMT inhibition, αIIbβ3/platelet axis blockade, and PD-1/PD-L1 immunotherapy to target these non-redundant vulnerabilities. We integrated TCGA PanCancer and multi-cancer single-cell analyses. The results support the presence of NNMT-associated tri-axis co-activation across cancer types. The findings identify stromal/fibroblast-like compartments as candidate NNMT-dependent populations. These observations provide both mechanistic and computational support for the proposed “disaggregation–exposure–clearance” strategy. Prior to clinical translation, future studies are required involving experimental perturbation, longitudinal CTC profiling, and biomarker-guided preclinical validation. Future directions It should be recognized that prospective clinical studies of perioperative antiplatelet and metabolic interventions are emerging, and innovative trial designs using liquid biopsy endpoints are being proposed [108, 118]. These developments provide a theoretical foundation for future clinical trials. Existing studies also suggest that AI-guided CTC cluster morphology and molecular feature analysis may help predict response to triple combination therapy and enable individualized dosing strategies [119, 120]. Successful implementation will require close multidisciplinary collaboration. Researchers should incorporate imaging flow cytometry (imFC)-based morphological modules into microfluidic and in vivo studies, including blinded image analysis and longitudinal sampling. The minimum detection system should quantify cluster composition, including adjacency relationships between CTCs and platelets, erythrocytes, and immune cells; cluster size and 3D geometric characteristics; CTC morphology; and nuclear abnormalities. These image-derived variables should be integrated with pathway module scores and therapeutic response outcomes. This integration may help determine whether the molecular state described here corresponds to a reproducible physical phenotype, and whether therapy can convert protected CTC clusters into exposed and more vulnerable states [123]. Based on existing findings, we designed a ‘hypothetical clinical-trial protocol draft’ to strengthen the conceptual framework of the triple strategy (Supplementary Material 1) [1, 6, 10, 19, 22, 23, 25, 30, 31, 38, 44, 50, 51, 65, 86, 92, 96, 100, 118, 119, 121–123]. To further discuss the use of single-cell sequencing to distinguish the two subtypes, we dynamically adjusted the intensity and sequence of the triple intervention according to the dependence of each subtype, transforming the clinical strategy from a ‘static combination’ into a ‘dynamic precision strike’ (Supplementary Material 2). Terminology definition "NNMT inhibitor" in this article broadly refers to interventional molecules targeting NNMT enzymatic activity or its upstream/downstream signaling pathways. As specific NNMT inhibitors are still in early-stage development, some discussions reference clinical data from NMT inhibitors as "proof-of-concept" validation, but it is crucial to note the distinct targets. Statement on theoretical deduction The synergistic effect between NNMT inhibitors and αIIbβ3 inhibitors within the proposed triple therapy currently constitutes a theoretical hypothesis and has not been experimentally validated. This article aims to propose a research hypothesis and construct a theoretical framework rather than report a verified therapeutic regimen. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 1. Supplementary Material 2.