An immunogenomic classification of solid tumours reveals subtype-specific therapeutic vulnerabilities for immunotherapy Authors: Yiming Zhao, Pei Wang, Zhiren Han, Zixuan Qiu, Xin Du, Qingliang Wen, Ziwei Zhou, Xiaorong Lin, Jiaxin Zhong, Beinan Han, Wenkui Fu, Keyi Sun, Herui Yao, Zhenkun Na, Canming Wang, Taobo Luo, Dan Su, Jian Zeng, Hai Hu, Man-Li Luo Published: 2026 ABSTRACT SummaryBackgroundThe efficacy of immune checkpoint blockade (ICB) is heterogeneous across patients. Tumour immune phenotype classification (immune-inflamed, -excluded, and -desert) represents a foundational but inadequate framework for predicting ICB efficacy. Here we aimed to develop an integrated immunogenomic classification to improve ICB response prediction and identify subtype-specific therapeutic vulnerabilities.MethodsWe analysed 13 public ICB cohorts and an in-house cohort. Using RNA-seq data, we developed ImmPred, a seven-gene classifier trained on IHC-defined immune phenotypes, and integrated it with TMB to define immunogenomic subtypes. Subtype-specific resistance mechanisms were investigated via pathway analysis and validated in syngeneic mouse models.FindingsPatients with cancers can be stratified into five immunogenomic subtypes with divergent responses to ICB, which are TMB-High (H) inflamed, TMB-Low (L) inflamed, TMB-H excluded, TMB-L excluded, and desert phenotypes. In immune-excluded tumours, MTAP deficiency contributes to ICB resistance in TMB-H excluded subtype and PRMT5 inhibitors enhances ICB efficacy in MTAP-KO B16-F10 and CT26 syngeneic mouse model, whereas TGF-β hyperactivation drives intrinsic resistance of TMB-L excluded subtype and TGF-β blockade potentiates anti-tumour immunity in MB49 and EMT6 mouse model. In TMB-H inflamed tumours, IFN-γ is a critical determinant of ICB efficacy, and TLR7 agonist, via enhancing IFN-γ signalling, improves anti-PD-L1 efficacy in MC38 mouse model. In TMB-L inflamed tumours, targeting COX-2–PGE2 axis with celecoxib sensitises ICB in LLC1 mouse model.InterpretationLeveraging clinical feasible RNA-seq and TMB analysis, our model exhibits robust predictive efficacy of ICB response in multiple cancers, enabling subtype-tailored therapeutic combinations to improve immunotherapy response.FundingThis work was supported by grants from National Key Research and Development Program of China (2021YFA1300602), National Natural Science Foundation of China (82025026, 82230091, 82472775), Guang Dong Basic and Applied Basic Research Foundation (2023A1515012412 and 2023A1515011214), Guangdong Science and Technology Department (2023B1212060013, 2023B1111030006), Key R&D Program of Zhejiang (2024C03160), Leading Innovative and Entrepreneur Team Introduction Program of Zhejiang Province (2024R01005 and 2025R01009).Keywords: Immune checkpoint blockade, Immunogenomic classification, Tumour immune phenotypes, Tumour mutational burden, Targeted therapy Research in context Evidence before this study Tumour immune phenotypes (inflamed, excluded, desert) based on CD8+ T cell spatial distribution has provided a foundational framework for predicting response to immune checkpoint blockade (ICB). However, this classification alone is insufficient to accurately predict ICB response, as the presence of T cells does not guarantee effective anti-tumour immunity. Tumour mutational burden (TMB), an established biomarker for ICB response, reflects cancer cell intrinsic characteristics, but its predictive value is inconsistent across cancer types. The framework that leverages both cancer cell-intrinsic features with the spatial architecture of the immune microenvironment is currently lacking to guide patient stratification and identify subtype-specific resistance mechanisms. Added value of this study We developed a clinically feasible immunogenomic classification integrating ImmPred, an RNA-seq-based classifier that defines tumour immune phenotypes (inflamed, excluded, desert) based on CD8+ T cell spatial distribution, with TMB. This approach stratified patients into five distinct subtypes, namely TMB-H inflamed, TMB-L inflamed, TMB-H excluded, TMB-L excluded, and desert subtype, with divergent responses to ICB across multiple solid tumours. Importantly, we uncovered subtype-specific resistance mechanisms and targetable vulnerabilities for combination strategies. In TMB-H excluded tumours, MTAP deficiency impaired T cell function, a vulnerability targetable by PRMT5 inhibition. In TMB-L excluded tumours, TGF-β hyperactivation drove ICB resistance and TGF-β blockade potentiates anti-tumour immunity. In TMB-H inflamed tumours, IFN-γ signalling was a critical determinant of efficacy, which could be enhanced by TLR7 agonism. In TMB-L inflamed tumours, the COX-2-PGE2 pathway promoted Tregs infiltration, and its inhibition sensitised tumours to ICBs. Implications of all the available evidence Our immunogenomic classification improves ICB response prediction using routinely accessible RNA-seq and TMB, while identifying biological vulnerabilities for four of the five subtypes. These findings provide a rationale for developing subtype-specific combination strategies in prospective clinical trials, moving beyond a one-size-fits-all approach to address distinct immunogenomic barriers. This framework offers a direct translational path to patient stratification and personalised immunotherapy, with the potential to improve clinical outcomes across multiple solid tumour types. Introduction Immune checkpoint blockade (ICB) therapies, utilising antibodies targeting cytotoxic T lymphocyte antigen 4 (CTLA4), programmed cell death 1 (PD-1) or PD-1 ligand 1 (PD-L1), have revolutionised the treatment of solid tumours including melanoma, non-small cell lung cancer, urothelial cancer, etc.1, 2, 3, 4 Nevertheless, durable response to ICBs occur in only 10%–40% of patients across cancer types.5, 6, 7 Therefore, elucidating biomarkers and mechanisms underpinning both therapeutic response and resistance to ICB remains imperative. The infiltration of immune cells in tumour microenvironments (TME) is one of the key determinants of ICB therapeutic responsiveness. Hot tumours, featuring immune-activation signatures and robust T cell infiltration, demonstrate favourable responses to ICBs, whereas cold tumours, comprising immune-excluded (T cell exclusion) and immune-desert (T cell absence) subtypes, exhibits ICB resistance.8,9 Current classification of tumours as hot or cold relies on the quantification of CD3+/CD8+ T cells both within the tumour center and at the invasive margins.10,11 However, T cell presence alone is insufficient to elicit an effective anti-tumour response, as bystander T cells or pro-tumour T cells can also infiltrate tumours.12,13 Therefore, the mere presence of tumour-infiltrating lymphocytes poorly predicts immunotherapy response, evidenced by inconsistent correlations between T cell infiltration and clinical outcomes in various cancer types including renal cancer and bladder cancer.14, 15, 16 Hot tumours require additional functional features, including PD-L1+ immune cells and pre-existing anti-tumour immunity, to collectively define ICB responsive phenotype.13 Growing evidence has demonstrated that cancer cell intrinsic characteristics also contribute to immunotherapy resistance, including low mutational burden, diminished PD-L1 expression, and insufficient antigenicity, etc. Tumours with a higher mutational burden are often immunogenic, characterised by increased expression of neo-antigens that are more potent to attract and activate cancer specific T cells, rather than bystander T cells incapable of recognising tumour epitopes,17 thereby conferring a higher probability of response to ICBs. However, tumour mutational burden (TMB) alone is an insufficient predictor of ICB response, as evidenced by the fact that TMB does not significantly correlate with ICB responsiveness in many cancer subtypes.18 Incorporation of TMB into current tumour immune phenotypes, guided by the theory of tumour immunity continuum, offers a promising framework that includes both the intrinsic cancer cell characteristics and features of the immune microenvironment for improved ICB response prediction. Importantly, TMB is relatively stable and less susceptible to short-term fluctuations influenced by biological regulatory pathways. Thus, combining TMB with TME assessment represents a clinically viable strategy to evaluate the response and resistance to ICB. In this study, we aimed to develop an integrated immunogenomic classification that combines tumour immune phenotypes (assessed by our RNA-seq-based classifier ImmPred) with TMB to improve the prediction of ICB response across multiple solid tumours. Furthermore, we sought to dissect the subtype-specific mechanisms underlying ICB resistance and identify therapeutic vulnerabilities to provide rational combination strategies to enhance the efficacy of immunotherapies. Methods Ethics All patients in the study were collected with informed consent and the study was approved by the Ethical Committee of Clinical Investigation from Zhejiang Cancer Hospital (Approval No. IRB-2023-843) and the Institute Research Ethics Committee of Sun Yat-Sen Memorial Hospital (Approval No. SYSKY-2025-743-01). All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) and were performed in accordance with institutional and national guidelines (Approval No. 2022000460). Patients To investigate the relationship between immune and genomic features and clinical benefits from immune checkpoint blockade, genomic and clinical data were retrieved from publicly available immunotherapy cohorts encompassing multiple cancer types treated with ICB monotherapy or combination regimens (Supplementary Table S1), which included melanoma,19, 20, 21, 22, 23 non-small cell lung cancer,23,24 urothelial cancer,15 glioblastoma,25 renal cancer,14,26 head and neck cancer,23 mesothelioma,27 gastric cancer,28 colorectal cancer,29 and pancreatic cancer.30 Genomic and clinical data from The Cancer Genome Atlas (TCGA) were obtained via cBioPortal (www.cbioportal.org). Tumour mutational burden (TMB) was classified as high if ≥ 10 mutations/Mb by panel sequencing or ≥175 mutations/exome by whole-exome sequencing, consistent with previous studies.31 The validation cohort of ImmPred includes 51 formalin-fixed paraffin-embedded (FFPE) tumour specimens (ZJCC cohort) collected within one month after surgical resection, including 32 lung adenocarcinoma specimens, 10 squamous cell lung cancer specimens, 1 small cell lung cancer specimens, 2 neuroendocrine lung cancer specimens, and 6 melanoma specimens (Supplementary Table S2). They were subjected to IHC staining for conventional immune phenotype classification and RNA sequencing for ImmPred-based immune phenotype classification. The in-house immunotherapy cohort were retrospectively collected, including 81 patients with diverse cancer types who received ICB therapy at Zhejiang Cancer Hospital and Sun Yat-Sen Memorial Hospital. These cases, referred to as the ZJCC&SYSMU cohort, were analysed by IHC staining to determine tumour immunophenotypes. Patients were included in ZJCC&SYSMU cohort if they met the following criteria: (1) histologically confirmed diagnosis of advanced or metastatic solid tumours (including but not limited to non-small cell lung cancer, melanoma, urothelial carcinoma, gastric cancer, and colorectal cancer); (2) received at least one dose of immune checkpoint inhibitor (anti-PD-1/PD-L1 or anti-CTLA-4 antibody) as monotherapy or in combination with other therapies between January 2018 and December 2023; (3) had available FFPE tumour specimens collected prior to ICB treatment; (4) had complete clinicopathological and follow-up data, including radiographic assessment of treatment response (assessed by RECIST v1.1); (5) aged 18 years or older. Exclusion criteria were: (1) patients who received concurrent radiotherapy or targeted therapy that could confound immunotherapy response assessment; (2) inadequate tumour tissue for IHC staining or RNA sequencing (tissue area <5 mm2 or tumour cellularity <20%); (3) presence of active autoimmune disease requiring systemic immunosuppressive therapy; (4) history of organ transplantation or HIV infection; (5) concomitant haematological malignancies or other primary tumours within 5 years prior to enrolment; (6) pregnancy or lactation. Detailed clinicopathological information of these patients is provided in Supplementary Tables S3–4. Sex was considered in the study design. For the publicly available cohorts (e.g., TCGA and other published immunotherapy datasets), sex data were retrieved from the original study databases. For the ZJCC&SYSMU cohort, sex was determined based on medical records and self-reporting at the time of admission. Both male and female patients were included to ensure representative immunogenomic classification. All participants in the ZJCC&SYSMU cohort were of Asian descent. Construction of T cell immunophenotype prediction model (ImmPred) A transcriptomic prediction model was developed using least absolute shrinkage and selection operator (LASSO) logistic regression. Firstly, immune-related genes were obtained from the ImmPort database (https://www.immport.org/),32 and stromal-related genes were curated from a previously published study.33 After excluding overlapping genes between the two lists, a combined panel of 1178 genes representing tumour immune and stromal characteristics was obtained. To ensure reliability, genes with low expression (TPM <0.1 in >95% of samples) were excluded, leaving 763 genes for analysis. Secondly, to identify genes most relevant to immunophenotypes, LASSO regression was applied to the 763-gene panel. This approach yielded 14 candidate genes that were preliminarily associated with T cell immunophenotypes. Patients were then randomly divided into training and internal validation cohorts at a 7:3 ratio. In the training set, the 14 candidate genes were further refined using bidirectional stepwise logistic regression, which identified the most parsimonious gene combination for phenotype prediction. The final classifier comprised seven genes (CD8A, IRF1, PDGFD, RABEP1, KLRC1, CD79A and SECTM1). Thirdly, the predictive accuracy of this 7-gene classifier was assessed in both the internal validation set (n = 79) and an independent external cohort (n = 51) based on the expression levels of the seven genes derived from bulk RNA-seq data. Then the prediction accuracy was compared with CD8 IHC-defined immune phenotypes (immune-inflamed, immune-excluded and immune-desert). For each immunophenotype, the following metrics were calculated: Sensitivity = TP/(TP + FN); Specificity = TN/(TN + FP); Positive predictive value (PPV) = TP/(TP + FP); F1 score = 2 × (PPV × Sensitivity)/(PPV + Sensitivity). TP (true positive) was defined as number of samples correctly classified as the target immunophenotype; TN (true negative): number of samples correctly classified as not belonging to the target immunophenotype; FP (false positive): number of samples incorrectly classified as the target immunophenotype; FN (false negative): number of samples incorrectly classified as not belonging to the target immunophenotype. To provide robust estimates of performance variability, 95% confidence intervals (CIs) for all metrics were calculated using bootstrap resampling with 2000 replicates. Whole transcriptome sequencing of FFPE tissues Primary tumour specimens were obtained from FFPE tissue blocks. Sections were dewaxed with xylene prior to RNA extraction. Total RNA was extracted using the RecoverAll™ Total Nucleic Acid Isolation Kit for FFPE (Invitrogen, CA, USA) following the manufacturer's protocol. RNA purity and quantification were evaluated using the NanoDrop 2000 spectrophotometer (Thermo Scientific, USA), while RNA integrity and fragment size distribution were assessed with an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). RNA-seq libraries were prepared using the VAHTS Universal V6 RNA-seq Library Prep Kit (Vazyme, Nanjing, China) according to the manufacturer's instructions, including depletion of ribosomal RNA. Paired-end sequencing was performed by OE Biotech Co., Ltd. (Shanghai, China), followed by downstream quality control and bioinformatic analysis. Mice Female C57BL/6 mice (6–8 weeks old) were purchased from GemPharmatech (Nanjing, China) and housed under SPF conditions in the Laboratory Animal Resource Center at Sun Yat-sen University. All cell lines in this study have been validated by the STR authentication and recently tested for mycoplasma contamination, with negative results confirmed for all lines. For tumour implantation, 5 × 105 B16-F10 and MB49 cells, or 1 × 106 MC38 and LLC1 were injected subcutaneously into the right flank of C57BL/6 mice. 1 × 106 EMT6 cells or 1.5 × 106 CT26 shMtap cells were injected subcutaneously into the right flank of BALB/c mice. Following tumour inoculation, animals were randomised into experimental groups when tumours reached a volume of approximately 100–200 mm3 (5–7 days post-inoculation). Tumour growth was monitored 2–3 times per week using digital callipers, and tumour volume was calculated using the formula: volume = (length × width2)/2. Mice were euthanised if tumours exceeded 2000 mm3 or if animals exhibited signs of distress or >20% body weight loss. For therapeutic studies, anti-PD-L1 monoclonal antibody (BioXcell, clone 10F.9G2, Cat. BE0101, RRID: AB_10949073) or IgG1 isotype control mAb (BioXcell, Cat. BE0083, RRID: AB_1107784) was administered intraperitoneally (i.p.) at 100 μg per mouse every 2 days. In the B16F10-MTAP KO model and CT26 shMtap mice were treated with the PRMT5 inhibitor GSK3326595 (Selleck, Cat. S8664) at 100 mg/kg by intragastric gavage once daily, either as monotherapy or in combination with anti-PD-L1 or isotype IgG. Anti-TGF-β mAb (mouse IgG1, clone 1D11; BioXcell, Cat. BE0057, RRID: AB_1107757) was administered intravenously (i.v.) at 10 mg/kg every 2 days in the MB49 and EMT6 tumour model. In the MC38 tumour model, mice were treated with the investigational TLR7 agonist CAN1012 (Clin-nov Medical), a compound currently in clinical development. Based on preclinical pharmacology demonstrating potent local immune activation with minimal systemic toxicity, CAN1012 was administered intratumorally at a dose of 5 μg per mouse every 2 days, beginning on day 7 post-tumour implantation. After two doses of CAN1012, animals received i.p. injections of either anti-PD-L1 mAb or isotype IgG control. Celecoxib (Selleck, Cat. S1261) was administered by intragastric gavage at 5 mg/kg every 2 days in the LLC1 tumour model. At the end of treatment, tumour tissues were harvested for IHC and FACS analysis to assess T-cell infiltration. Metabolomics Cells were washed twice with ice-cold PBS. 80% cold methanol was then added to the samples placed on dry ice. Cells were moved to a −80 °C freezer for 30 min and transferred to a 2 mL EP tube. The collected cell suspensions were vortexed at 4 °C for 30 min and then centrifuged at 12,000 g for 15 min at 4 °C to pellet cellular debris. The supernatants were transferred to fresh tubes and normalised to total protein content. Intracellular metabolite profiling was performed using liquid chromatography-mass spectrometry (LC-MS) according to standard protocols, enabling quantitative detection of targeted metabolites. Statistics Statistical analyses were performed using GraphPad Prism (GraphPad Software) and RStudio. For comparisons between two groups, two-tailed Student's t-tests were applied. For comparisons among more than two groups, one-way analysis of variance (ANOVA) was used, followed by either least significant difference (LSD) or Games-Howell post hoc tests, depending on variance homogeneity. Categorical variables between groups were assessed using Pearson's chi-square test and Fisher's exact test. All data were presented as mean ± standard deviation (SD). Survival analyses were conducted using the Kaplan–Meier method. For progression-free survival (PFS), the time-to-event was calculated from the start of ICB treatment to the date of disease progression (per RECIST v1.1), death from any cause, or last follow-up, whichever occurred first. For overall survival (OS), the time-to-event was calculated from the start of ICB treatment to the date of death from any cause or last follow-up. For the in-house cohort, strict inclusion criteria ensured complete data for all primary variables. For public cohorts, missing TMB data were handled using complete-case analysis for primary TMB-stratified analyses, with sample sizes clearly reported. For survival endpoints, PFS and OS were analysed separately using only cohorts reporting each endpoint; cohorts without survival data were included only in response rate analyses. No imputation was performed for missing survival endpoints. Patients who were lost to follow-up or were event-free at the data cut-off were censored at the date of last contact. Differences between survival curves were evaluated using the two-sided log-rank test. Hazard ratios (HR) and 95% confidence intervals (CI) were estimated using Cox proportional hazards models. For hypothesis-driven comparisons with a limited number of pre-specified endpoints (e.g., cytokine measurements, tumour growth curves), nominal p-values are reported without adjustment for multiple testing, consistent with standard practice in preclinical mechanistic studies. For exploratory high-dimensional analyses (pathway enrichment, GSEA, multi-gene or immune cells correlations), multiple testing corrections (FDR or Bonferroni) were applied as indicated. Nominal or adjusted p < 0.05 was considered statistically significant. Role of funders The funding source had no role in study design, data collection and interpretation, analysis, writing of this report, or in the decision to submit the paper for publication. Results An immunogenomic classification discriminating ICB outcomes of patients with cancers We aim to integrate the immune microenvironment features with cancer cell intrinsic characteristics for improved ICB response prediction. First, we develop a computational tool to use bulk RNA-seq data to classify tumour immune phenotypes (immune-inflamed, immune-excluded, and immune-desert), as there are very limited CD8+ T cell immunohistochemistry (IHC) data in public immunotherapy cohorts. We developed ImmPred, a Lasso-logistic regression model (Fig. 1A, Supplementary Fig. S1A and B and Supplementary Tables S5–S7), which was trained using IHC and RNA-seq data of IMvigor210 cohort. This seven-immune gene classifier (CD8A, IRF1, PDGFD, RABEP1, KLRC1, CD79A and SECTM1) was validated in both IMvigor210 (n = 79) and our in-house ZJCC cohorts (n = 51). In IMvigor210 cohort, ImmPred achieved an overall accuracy of 73.4% (58/79), with F1 scores of 0.65 (95% CI: 0.45–0.82) for desert, 0.75 (95% CI: 0.62–0.86) for excluded, and 0.71 (95% CI: 0.63–0.90) for inflamed subtype. In the ZJCC cohort, ImmPred achieved an overall accuracy of 92.2% (47/51), with F1 scores of 1.00 (95% CI: 1.00–1.00) for desert, 0.60 (95% CI: 0.35–0.85) for excluded, and 0.95 (95% CI: 0.88–0.99) for inflamed subtype. In both datasets, ImmPred consistently exhibited excellent performance for the inflamed subtype (F1: 0.71–0.95) and acceptable performance for the desert and excluded subtypes (Fig. 1A and Supplementary Fig. S1C and D). Classification of tumour microenvironment based on tumour immune phenotypes and tumour mutational burden. A. Diagram to construct the RNA sequencing-based tumour immune phenotype classifier ImmPred. In training cohort, the immune related genes identified in the LASSO regression model were subjected to bidirectional stepwise logistic regression to construct the final classifier which consisted of 7 genes. The predictive performance was validated in the internal (IMvigor210 cohort) and external cohort (ZJCC cohort). B. Response rate to ICB classified by tumour immune phenotypes in IMvigor210 cohort, ImmPred predicted pooled public datasets (Allen 2015, Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023, Braun 2020, Forde 2021, Bortolomeazzi 2021 and Prat 2017; Supplementary Table S1), in-house immunotherapy cohort (ZJCC&SYSMH), and all combined datasets. The p-values were calculated using Chi-square test. Sample sizes are indicated above bars. ∗∗p < 0.01, ∗∗∗∗p < 0.0001, ns: not significant. C–F. Kaplan–Meier curves of progression-free survival (PFS) in patients treated with PD-(L)1 inhibitors, stratified by tumour mutational burden (TMB-high vs. TMB-low) and tumour immune phenotypes, (C) desert, (D) excluded, and (E) inflamed (Liu 2019, Ravi 2023, Braun 2020, Zhao 2019; Supplementary Table S1), two-sided log-rank test was used for survival analysis. G. Response rate to ICB across three tumour immune phenotypes and TMB status in pooled cohorts (Supplementary Tables S1, S3 and S4). The p-values were calculated using Chi-square test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. H. Schematic representation of a tumour microenvironment classification integrating tumour immune phenotypes and TMB status. We then integrated ImmPred with TMB, the intrinsic feature of cancer cells, to predict ICB response. In IMvigor210 trial and in-house immunotherapy cohorts (ZJCC&SYSMU), IHC-based classification of tumour immune phenotype showed only a trend of correlation with ICB response, while RNA-seq-based classification in ImmPred predicted pooled public datasets (Allen 2015, Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023 longer progression-free survival (PFS) than TMB-L inflamed tumours, Braun 2020, Forde 2021, Bortolomeazzi 2021 and Prat 2017; Supplementary Table S1) and all samples combined datasets showed significant correlation with ICB response (p < 0.0001, Chi-square test) (Fig. 1B). When combining TMB with tumour immune phenotypes, TMB-H inflamed tumours showed markedly longer progression-free survival (PFS) than TMB-L inflamed tumours after PD-(L)1 inhibitors treatment (p = 0.024, two-sided log-rank test) (Fig. 1C). Similarly, TMB-H immune-excluded tumours demonstrated a trend toward longer PFS compared to TMB-L immune-excluded tumours (p = 0.32, two-sided log-rank test) (Fig. 1D), whereas no PFS difference was detected between TMB-H and TMB-L immune-desert tumours (p = 0.42, two-sided log-rank test) (Fig. 1E). The longest PFS was observed in TMB-H inflamed tumours (Fig. 1F), underscoring the synergistic effect of antigenicity and pre-existing immunity. Notably, the TMB-H group showed higher PD-(L)1 inhibitor response rates than TMB-L group in both inflamed and excluded tumours, but not in immune-desert tumours (Fig. 1G). Thus, combination of tumour immune phenotypes with TMB revealed five distinct TME subtypes with different clinical outcomes (Fig. 1H), which also unveils five distinct biological mechanisms underlying resistance to immunotherapy: (1) TMB-H inflamed tumours achieve durable responses due to synergistic antigen presentation and pre-existing immunity; (2) TMB-L inflamed tumours exhibit partial responsiveness limited by low neoantigen burden; (3) TMB-H excluded tumours fail to recruit immune cells despite high antigenicity; (4) TMB-L excluded tumours are resistant to ICB due to dual defects in antigenicity and immune infiltration; and (5) immune-desert tumours lack both T cell infiltration and antigenic triggers regardless of TMB status. Low expression of MTAP is a predictor of poor ICB responses in TMB-H excluded tumours As these five subtypes of tumours exhibit distinct responsiveness to ICB, we hypothesise that the heterogeneity in therapeutic outcomes may be driven by subtype-specific mechanisms underlying tumour-TME interactions. First, we dissected the mechanism by analysing signalling pathways enriched in TMB-H excluded tumours in IMvigor210 dataset. KEGG analysis demonstrated that cysteine and methionine metabolism was the most enriched pathway in responders, and PPAR signalling was the most enriched pathway in non-responders (Fig. 2A). Gene Set Enrichment Analysis (GSEA) showed that MTAP, CTH, TRDMT1, LDHC, LDHAL6A were top five ranked genes in cysteine and methionine metabolism pathway (Fig. 2B), and APOC3, AQP7, ACAA1, ACOX3, UBC were top five ranked genes in PPAR signalling pathway in IMvigor210 dataset (Fig. 2C). Impact of MTAP on the immunotherapy efficacy in TMB-H excluded subtype. A. KEGG pathways enriched in genes whose expression correlates with responses to atezolizumab in patients with TMB-H excluded cancers from IMvigor210 trial. Bars represent −log10p values for pathways with p < 0.05. B–C. Enrichment of the cysteine and methionine metabolism pathway (B) and the PPAR signalling pathway (C) in responders versus non-responders from urothelial cancer specimens with TMB-H excluded TME. Top five scoring genes for each pathway are indicated. D. KEGG pathways enriched in genes whose expression correlates with responses to PD-1 inhibitors in TMB-H excluded NSCLC. Bars represent −log10 p values for pathways with p < 0.05. E. Cysteine and methionine metabolism pathway enriched in responders compared to non-responders from NSCLC with TMB-H excluded TME. MTAP is among the top 5-scoring genes. F. Response rates stratified by MTAP expression (data sources: Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023 and IMvigor210; Supplementary Table S1). The p-values were calculated using Chi-square test. Sample sizes are indicated above bars. ∗p < 0.05, ∗∗p < 0.01. G–H. Kaplan–Meier analyses of progression-free survival (PFS) (G; Liu 2019, Ravi 2023) and overall survival (OS) (H; Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023, IMvigor210) in patients diagnosed of TMB-H excluded cancers with high versus low MTAP expression treated with PD-(L)1 inhibitors. Two-sided log-rank p values are shown. I. Response rates stratified by MTAP deletion status (data source: Liu 2019; Supplementary Table S1). The p-values were calculated using Chi-square test. Sample sizes are indicated above bars. ∗p < 0.05, ∗∗p < 0.01. J–K. Kaplan–Meier curves comparing PFS (J) and OS (K) between patients diagnosed of TMB-H excluded cancers with MTAP deletions and non-deletions treated with PD-(L)1 inhibitors (data source: Liu 2019; Supplementary Table S1). Two-sided log-rank p values are shown. Notably, MTAP (methylthioadenosine phosphorylase) was also among the top enriched genes in responders comparing to non-responders in a NSCLC cohort (Fig. 2D and E) treated with PD-(L)1 inhibitors. More importantly, MTAP was the best predictor of the overall survival (OS) after receiving ICBs among the top ten ranked genes in the cysteine and methionine metabolism and PPAR signalling pathway (Supplementary Fig. S2A–J). The pooled analysis of all cancer cohorts including melanoma, NSCLC and urothelial cancer showed that overall ICBs response was significantly higher in patients diagnosed of TMB-L excluded cancers with high MTAP expression (MTAP-H) compared with low MTAP expression (MTAP-L) (p < 0.01, Chi-square test) (Fig. 2F), with reduced PFS and OS also noted in low MTAP expression (Fig. 2G and H). As MTAP deletion was frequently observed in a wide spectrum of malignancies, and previous studies showed that MTAP deletions correlated with poor response to immunotherapies,34,35 we analysed MTAP gene status and found that MTAP deletion was associated with lower ICB response in patients with melanoma within the TMB-H excluded subtype (Fig. 2I–K). Notably, in other subtypes, neither low expression nor deletions of MTAP correlated with clinical outcomes of ICBs (Supplementary Fig. S2K–P), indicating MTAP deletion as a critical mediator of ICBs resistance specifically within TMB-H excluded tumours, wherein its loss abrogates the anticipated immunogenic advantage conferred by high tumour mutational burden. The subtype-specific mechanism also highlights methionine metabolism as a promising therapeutic vulnerability in this subtype. MTAP in cancer cells modulates T cell functions through methionine metabolism Previous studies have reported that MTAP loss impairs T cell function in melanoma and urothelial tumours.34 However, the immunomodulatory effects of MTAP on cancer cells remains to be elucidated. The bladder cancers are mainly immune-excluded subtypes with relatively high TMB.36 We silenced MTAP with siRNAs in human bladder cancer cell line MGH-U3 which highly expressed MTAP, and stably overexpressed MTAP in MTAP-loss UM-UC-3 human bladder cell line (Supplementary Fig. S3A–D). The transwell chemotaxis assays showed that knockdown of MTAP in cancer cells attenuated the migration of peripheral blood mononuclear cells (PBMCs)-derived T cells (Fig. 3A and B). Compared with the control cells, MTAP overexpression dramatically increased T cell migration (Fig. 3C and D). When MTAP knockdown MGH-U3 cells co-cultured with stimulated PBMCs, there was a more than 50% reduction of TNF-α and IFN-γ secretion, while MTAP overexpression in UM-UC-3 promoted the secretion of these two cytokines (Fig. 3E–L). Moreover, tumour cell produced chemokines associated with T cell migration (CCL5, CXCL9, CXCL10 and CXCL11) were significantly decreased upon MTAP knockdown (Fig. 3M), and dramatically increased in MTAP-overexpressing cells (Fig. 3N). MTAP in cancer cells modulates T cell functions through methionine metabolism. A–D. MTAP regulates T cell migration. Transwell migration assays were performed in triplicate, and migrated CD8+ T cells were quantified by flow cytometry. MTAP knockdown in MGH-U3 cells significantly reduced T cell migration (A–B), whereas MTAP overexpression in UM-UC-3 cells markedly enhanced T cell migration (C–D). One-way ANOVA with LSD's post-hoc test for B and Student's t-test for D; ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. E–L. Manipulating MTAP expression in cancer cells affects T cell functions. PBMCs were co-cultured with MTAP knockdown MGH-U3 or MTAP overexpressing UM-UC-3. Flow cytometry analysis revealed changes in T cell subsets and effector cytokine production, including TNF-α (E–H) and IFN-γ (I–L). One-way ANOVA with Games-Howell's post-hoc test for F, J and Student's t-test for H, L; ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. M–N. Effect of MTAP on chemotactic cytokine expression. RT-qPCR analysis showed that MTAP knockdown in MGH-U3 cells significantly decreased T cell-attracting chemokines (M), whereas MTAP overexpression increased their expression (N). Data are representative of three independent experiments. Statistical analysis was performed using one-way ANOVA (M) and Student's t-test (N). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. O. Schematic diagram of methionine and SAM cycles. P–S. Levels of intracellular metabolites, as detected by mass spectrometry 24 h and 48 h after transduction with si-NC, si-MTAP (P–Q) or MTAP expression vectors (R–S). Intracellular MTA and SAM levels were elevated following MTAP knockdown but decreased upon MTAP overexpression. The p-values were calculated using one-way ANOVA with Games-Howell's post-hoc test. T–V. Effects of SAM and MTA on CXCL9, CXCL10 and CXCL11 mRNA expression in MTAP overexpressing MGH-U3 cells under IFN-γ treatment. The p-values were calculated using one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. It has been shown that MTAP is a key metabolic enzyme involved in the methionine salvage pathway (Fig. 3O). Previous studies showed that MTA, secreted by MTAP-deleted cancer cells, accumulated in the TME and profoundly impaired T cell functions.34 However, in some cancer types, such as glioblastoma, MTAP deletion in tumour cells failed to accumulate MTA in the TME,37 because MTAP-expressing stromal cells extensively infiltrated in the TME and metabolised the MTA secreted by tumour cells. We performed LC-MS/MS analysis of metabolites in the methionine salvage pathway. Unlike the extracellular MTA, we found that MTAP knockdown caused an increase in the levels of intracellular MTA and SAM of MGH-U3 cells (Fig. 3P and Q), while MTAP overexpressing UM-UC-3 cells exhibited much lower intracellular MTA and SAM compared with the control cells (Fig. 3R and S). Moreover, supplementation of MTA in MTAP overexpressing cells prominently reduced CXCL9, CXCL10 and CXCL11 expression, while adding SAM only led to a slight reduction of these cytokines (Fig. 3T–V). To further establish the causal role of MTA in T cell dysfunction, we performed direct MTA supplementation experiments using MTAP-overexpressing UM-UC-3 cells co-cultured with PBMC-derived T cells. Exogenous MTA treatment significantly reversed the enhanced T cell chemotaxis induced by MTAP overexpression and reduced the secretion of TNF-α and IFN-γ by co-cultured T cells (Supplementary Fig. S3E–J). In contrast, supplementation with SAM resulted in minimal changes in cytokine production. These results indicate that intracellular accumulation of MTA, rather than SAM, in MTAP-deleted cancer cells impairs the anti-tumour effect of T cells. These findings not only resolve the paradox of poor immunotherapy response in TMB-H excluded tumours but also highlight the intracellular MTA as a potential therapeutic target to restore T cell-mediated anti-tumour immunity in this ICB resistant subtype. PRMT5 inhibitor enhances ICB efficacy in TMB-H excluded cancers with MTAP deletions Previous studies have identified PRMT5 as a synthetic lethal target of MTAP-deficient cancers.38 PRMT5 inhibitor GSK3326595 is potent in killing MTAP-deleted cancer cells via binding to both PRMT5 and SAM, an upstream metabolite of MTA, thereby reducing the concentration of free SAM and subsequently lowering the MTA level.39 We found that blocking PRMT5 with GSK3326595 in MTAP knockdown MGH-U3 cells could restore the mRNA level of CCL5, CXCL9, CXCL10 and CXCL11, compared with MTAP knockdown only (Fig. 4A). Furthermore, in the co-culture model of PBMC-derived T cell with MGH-U3 cells, the cytotoxicity effect was markedly reduced after MTAP knockdown, which could be restored by GSK3326595 treatment (Fig. 4B and C). PRMT5 inhibition enhances immunotherapy efficacy in the TMB-H excluded subtype. A. RT-qPCR analysis of chemotactic cytokines in MGH-U3 cells transfected with si-NC or si-MTAP treated with or without PRMT5 inhibitor. Statistical analysis was performed using one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. B–C. MTAP knockdown significantly reduced cytotoxicity, which was restored by PRMT5 inhibition. Tumour cells labelled with calcium green dye were co-cultured with stimulated PBMCs for 12 h, and dead cell count were calculated by flow cytometry. One-way ANOVA with LSD's post-hoc test was performed. ∗∗p < 0.01, ∗∗∗∗p < 0.0001. D–G. In vivo anti-tumour effect of PRMT5 inhibition combined with anti-PD-L1 therapy. MTAP-knockout B16F10 cells were implanted into C57BL/6 mice (n = 5/group) to establish a TMB-H excluded model. Treatment regimens are shown (D). Representative tumour images (E), tumour growth curves (F), and final tumour weights (G) are presented. Statistical analysis was performed by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. H–I. Representative IHC staining for Cd8+ T cells in allograft tumours from each treatment group: IgG + Vehicle, IgG + PRMT5i, anti-PD-L1 + Vehicle, and anti-PD-L1 + PRMT5i. Quantification of Cd8+ T cell infiltration is shown in (I). Scale bar, 20 μm. n = 5 per group. One-way ANOVA with LSD's post-hoc test was applied. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ns: not significant. J–O. Representative flow cytometry plots of Cd8+ T cells isolated from allografts showing expression of Perforin (J–K), IFN-γ (L–M), and Granzyme B (N–O). Combination therapy significantly increased effector molecule release. Statistical analysis was performed using one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. As loss of MTAP is a predictor of poor ICB responses in TMB-H excluded cancers, we test whether PRMT5 inhibitors increase the efficacy of ICBs in TMB-H immune excluded tumours. B16-F10 melanoma cells carried high somatic mutation,40 and MTAP-KO B16-F10 cells formed immune-excluded tumours in C57BL/6 mice (Supplementary Fig. 4C and D). Thus, we utilised this TMB-H excluded mouse model to explore whether PRMT5 inhibitor could improve the immunotherapy efficacy. The combined inhibition of PD-L1 and PRMT5 significantly reduced tumour growth and tumour weight compared to anti-PD-L1 or PRMT5 inhibitor only (Fig. 4D–G). Combinations of anti-PD-L1 and PRMT5 inhibitor also increased the abundance of tumour infiltrating CD8+ T cells (Fig. 4H and I), as well as the secretion of perforin, TNF-α and IFN-γ of CD8+ T cells (Fig. 4J–O). To further validate the generalisability of MTAP deficiency-driven immunotherapy resistance across different cancer types within the TMB-H excluded subtype, we established an additional in vivo model using CT26 murine colorectal cancer cells, which harbour a high TMB.41 We generated MTAP-knockdown CT26 cells via shRNA-mediated silencing and confirmed that these tumours exhibited an immune-excluded phenotype with reduced intratumoral CD8+ T cell infiltration (Supplementary Fig. S4E–G). Treatment of PRMT5 inhibitor GSK3326595 in combination with anti-PD-L1 also significantly suppressed tumour growth and reduced tumour weight compared to either monotherapy alone (Supplementary Fig. S4H–K). Together, these findings demonstrate that PRMT5 inhibition counteracts the immunosuppressive effects of MTAP deficiency in TMB-H excluded tumours by restoring T cell functionality, providing a strong rationale for combining PRMT5 inhibitors with immune checkpoint blockade in TMB-H excluded cancers. TGF-β blockade potentiates anti-tumour immunity and ICB efficacy in TMB-L excluded tumours TMB-L excluded tumours represent a clinically challenging subtype, characterised by low antigenicity and an immune-excluded microenvironment, therefore exhibiting dual barriers for effective immunotherapy. KEGG pathway analysis in IMvigor210 trial dataset demonstrated that complement and coagulation cascades, extracellular matrix receptor interaction and dilated cardiomyopathy are enriched in non-responders, and RNA degradation was enriched in responders to ICB in TMB-L excluded urothelial cancers (Fig. 5A, Supplementary Fig. S5A–B). GSEA identified top ranked genes in these pathways (Fig. 5B and C), which were mostly reported to regulate the integrin/TGF-β axis.42, 43, 44 Compared with other significantly enriched pathways such as complement and coagulation pathway and RNA degradation, TGF-β pathway genes, especially TGF-β1, were more abundantly expressed in immune cells (Fig. 5D). A pooled analysis of multiple cancer types showed that tumours with high TGF-β1 expression responded poorly to ICBs compared with those with low TGF-β1 expression in TMB-L excluded subtype (Fig. 5E). Additionally, TGF-β pathway genes (TGFB1, TGFB2, TGFB3, TGFBR1 and TGFBR2) negatively correlated with TMB in immune-excluded tumours of most cancer types in TCGA database (Fig. 5F), suggesting that TGF-β signalling was more activated in TMB-L excluded tumours. TGF-β attenuates immunotherapy efficacy in the TMB-L excluded subtype. A. KEGG pathways enriched in genes whose expression correlates with responses to atezolizumab in patients with TMB-L excluded from IMvigor210 trial. Bars indicate −log10p values for significantly enriched pathways (p < 0.05). B–C. GSEA enrichment plot analysis showed TGF-β related pathways in tumour specimens in non-responders compared to responders, including ECM-receptor interaction (B) and dilated cardiomyopathy (C). D. Dot plot showing distributions of top-scoring genes in TGF-β-related pathways, complement and coagulation cascades, and RNA degradation pathway. E. Response rates stratified by TGF-β1 expression in TMB-L excluded cancers. Patients with lower TGF-β1 expression exhibited significantly higher response rates than those with higher expression (data sources: Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023, and IMvigor210; Supplementary Table S1). Statistical significance was assessed by Chi-square test. ∗p < 0.05. F. Correlation between TGF-β pathway-related gene expression and TMB across TCGA cohorts. Each circle represents one gene–cohort pair; circle size reflects correlation coefficient magnitude, with blue indicating stronger negative correlation and red indicating stronger positive correlation. Spearman's correlations were used to calculate p values. G–J. In vivo efficacy of TGF-β blockade in combination with PD-L1 inhibition. MB49 cells were implanted into C57BL/6 mice (n = 5/group) to establish a TMB-L excluded model. Treatment schema is shown (G). Representative tumour images (H), tumour growth curves (I), and final tumour weights (J) are presented. Growth curves were analysed by one-way ANOVA with LSD's post-hoc test; tumour weights by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. K–L. Representative IHC staining of allograft tumours for Cd8+ T cells across treatment groups: IgG + PBS, IgG + anti-TGF-β, anti-PD-L1 + PBS, and anti-PD-L1 + anti-TGF-β. Quantification of CD8+ T cell infiltration is shown in (L). Scale bar, 20 μm. n = 5 per group. Statistical analysis was performed by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. We hypothesised that TGF-β signalling contributes to the ICB resistance in TMB-L excluded subtype. Mouse bladder cancer cell MB49 carries a KRAS G12D mutation, which was associated with a relatively low TMB, and formed an immunosuppressive TME according to previous studies.45, 46, 47 Therefore, we generated the TMB-L excluded animal model by grafting MB49 cells into syngeneic C57BL/6 mice to test whether TGF-β blockade could improve the responses to ICBs (Fig. 5G and Supplementary Fig. S5E). Compared to anti-PD-L1 treatment, dual inhibition of PD-L1 and TGF-β further reduced tumour growth and tumour weight dramatically (Fig. 5H–J), and led to increased infiltration of CD8+ T cells in tumours (Fig. 5K and L). We also established the EMT6 murine breast cancer model, which harbour a low mutational burden characteristic of excluded phenotype (Supplementary Fig. S5F).48 Consistently, treatment with the TGF-β antibody in combination with anti-PD-L1 significantly suppressed tumour growth and reduced tumour weight compared to either monotherapy alone (Supplementary Fig. S5G–J). Together, these findings identify TGF-β signalling as a dominant immunosuppressive mechanism in TMB-L excluded tumours, which indicates that stratifying patients with TMB-L excluded cancers based on TGF-β activation status may improve therapeutic efficacy with combined TGF-β and PD-L1 inhibition. IFN-γ further enhances ICB responses in TMB-H inflamed subtype Among the five subtypes, TMB-H inflamed tumours exhibit the best responsiveness to ICB therapy, due to their dual advantages of high tumour antigenicity and pre-existing T cell infiltration. Nevertheless, considerable heterogeneity in treatment response persists within this subtype, indicating that additional biological factors may underlie the therapeutic efficacy. GSEA in IMvigor210 dataset identified Type I diabetes mellitus (T1DM) and Ascorbate and aldarate metabolism associated with responses to ICB in TMB-H inflamed urothelial cancers (Fig. 6A and B). As T1DM pathway was also enriched in responders compared to non-responders in the NSCLC cohort (Fig. 6C and D), we further analysed the top ranked genes in this pathway. Interferon-γ (IFN-γ) was found to be among the top-scoring genes within T1DM pathway in both IMvigor210 and NSCLC cohorts (Fig. 6B and D). High level of IFNG correlated with better responses to ICBs in the pooled analysis of immunotherapy cohorts (Fig. 6E), and exhibited a trend with favourable PFS within TMB-H inflamed tumours (Fig. 6F). IFN-γ enhances ICB response in the TMB-H inflamed subtype. A and C. KEGG pathways enriched in genes whose expression correlated with responses to ICB in TMB-H inflamed patients from IMvigor210 trial (A) and NSCLC cohort (C). Bars indicate −log10p values for significantly enriched pathways (p < 0.05). B and D. GSEA enrichment plots showing the Type I diabetes mellitus pathway enriched in responders compared to non-responders of TMB-H inflamed subtype from IMvigor210 trial (B) and NSCLC cohort (D). IFNG ranked among the top five scoring genes. E. TMB-H inflamed cancers with higher IFNG expression showed significantly higher response rate compared to those with lower IFNG expression (data sources: Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023, Braun 2020 and IMvigor210; Supplementary Table S4). Statistical significance was determined by Chi-square test (∗p < 0.05). F. Kaplan–Meier curve of PFS in patients treated with PD-(L)1 inhibitors, stratified by IFNG expression (data sources: Liu 2019, Ravi 2023, and Braun 2020; Supplementary Table S1). Two-sided log-rank p values are shown. G. Response rates stratified by Teff score in TMB-H inflamed cancers. Higher Teff scores correlated with increased response rates; however, this association was not significant in pooled analyses of all patients (data sources: IMvigor210, Hugo 2016, Riaz 2017, Liu 2019, Ravi 2023, and Braun 2020; Supplementary Table S1). Chi-square test was used for significance testing. ∗p < 0.05, ns: not significant. H–K. In vivo efficacy of TLR7 agonist in combination with PD-L1 blockade. MC38 cells were implanted into C57BL/6 mice (n = 5/group) to establish a TMB-H inflamed model. Treatment schema is shown (H). Representative tumour images (I), growth curves (J), and final tumour weights (K) are presented. Growth curves were analysed by one-way ANOVA with LSD's post-hoc test; tumour weights by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ns: not significant. L–M. Representative IHC staining of allograft tumours for Cd8+ T cells in treatment groups: IgG + Vehicle, TLR7 agonist + IgG, anti-PD-L1 + Vehicle, and anti-PD-L1 + TLR7 agonist. Quantification of CD8+ T cell infiltration is shown in (M). Each group, n = 5. Scale bar, 20 μm. Statistical analysis was performed using one-way ANOVA with LSD's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. IFN-γ in the TME is mainly secreted by tumour-infiltrating effector T cells (Teff) to regulate anti-tumour immunity. We found that Teff signature correlated with responses to PD-(L)1 inhibitors specifically in TMB-H inflamed tumours, but not in the pooled analysis of all patients (Fig. 6G). Thus, we tested whether increasing the production of IFN-γ could further enhance ICB response in this subtype. We generated a TMB-H inflamed mouse model by using MC38 mouse colon cancer cells (Fig. 6H), as MC38 cells were reported to carry high mutations36 and formed an inflamed TME (Supplementary Fig. S6A). TLR7 activation has been reported to increase the production of IFN-γ, and TLR7 agonist shows promise in triggering robust anti-tumour response for cancer therapy.49 As anti-PD-L1 treatment reduced the tumour growth and weight of MC38 tumours in C57BL/6 mice, TLR7 agonist CAN1012 further sensitised the tumours to PD-L1 blockade (Fig. 6I–K). Combination of TLR7 agonist with anti-PD-L1 robustly activated the interferon-γ pathway, as detected by the elevated level of p-Stat1 (Supplementary Fig. S6B), which consequently enhanced the infiltration of CD8+ T cells into tumours (Fig. 6L and M). Our findings demonstrate that IFN-γ is a critical determinant of ICB efficacy in TMB-H inflamed tumours, and TLR7 agonist, via enhancing IFN-γ signalling, can potentiate ICB responses in this subtype. Inhibition of COX-2-PGE2 pathway improves anti-tumour immunity in TMB-L inflamed subtype Although CD8+ T cells are abundantly infiltrated in both TMB-H and TMB-L inflamed tumours, their responses to ICBs differs dramatically (Fig. 1C and G). To elucidate the difference between TMB-H and TMB-L inflamed tumours, we used xCell to compare 64 immune and stromal cell types in the TME. Analysis of the IMvigor210 dataset revealed a decreased infiltration of Th1, pro-B cells and basophils, but increased level of Treg in TMB-L inflamed TME (Fig. 7A and B), which was consistent with previous reports that Th1 cells positively correlated, while Tregs negatively correlated with immunotherapy efficacy.50,51 Given that the number of responders in TMB-L inflamed tumours is too low to be analysed in responders vs non-responders, we kept on comparing TMB-H and TMB-L inflamed tumours enriched pathways. Arachidonic acid metabolism emerged as the most significant pathway enriched in TMB-L inflamed tumours in IMvigor210 trial (Fig. 7C and Supplementary Fig. S6C). Furthermore, we identified COX2-PGE2 pathway as one of the downstream pathways of arachidonic acid metabolism. Notably, PTGS2, PTGES, PTGER1, PTGER2, PTGER3 and PTGER4 genes in COX2-PGE2 pathway exhibited a positive association with the abundance of Tregs, but a negative correlation with Th1 cells (Fig. 7D and E). The COX-2-PGE2 pathway dampens tumour responses to ICBs in the TMB-L inflamed subtype. A–B. Immune cell enrichment in TMB-L and TMB-H inflamed urothelial cancers, as estimated from transcriptomic profiles using the xCell algorithm. TMB-L inflamed tumours exhibited significantly higher frequencies of Tregs and reduced frequencies of Th1 cells compared to TMB-H inflamed tumours. Mann–Whitney U tests were applied. ∗p < 0.05, ns: not significant. C. KEGG pathway enrichment analysis of genes correlated with TMB in patients with inflamed cancers from IMvigor210 trial. Shown are −log10p values for significantly enriched pathways (p < 0.05). D–E. Correlation between COX2-PGE2 signalling related genes and Tregs (D) or Th1 cells (E) across cancer cohorts in TCGA. Each circle represents the correlation coefficient (r) between gene expression and immune cell abundance (estimated by xCell). Circle size indicates correlation strength, with blue denoting negative correlations and red denoting positive correlations. Spearman's correlations were used. F–I. In vivo efficacy of COX2 inhibition combined with PD-L1 blockade. LLC1 cells were implanted into C57BL/6 mice (n = 5/group) to establish a TMB-L inflamed model. Treatment schema is shown (F). Representative tumour images (G), growth curves (H), and tumour weights (I) are presented. Growth curves were analysed by one-way ANOVA with LSD's post-hoc test; tumour weights by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns: not significant. J–K. Representative IHC staining for FOXP3+ cells in allograft tumours across treatment groups: IgG + Vehicle, celecoxib + IgG, anti-PD-L1 + Vehicle, and anti-PD-L1 + celecoxib. Quantification of FOXP3+ Tregs is shown in (K). Each group, n = 5. Scale bar, 20 μm. Statistical significance was assessed by one-way ANOVA with Games-Howell's post-hoc test. ∗p < 0.05, ns: not significant. L–M. Representative IHC staining for CD4+ T cells in xenograft tumours across treatment groups: IgG + Vehicle, celecoxib + IgG, anti-PD-L1 + Vehicle, and anti-PD-L1 + celecoxib. Quantification of CD4+ T cells is shown in (M). Each group, n = 5. Scale bar, 20 μm. One-way ANOVA with Games-Howell's post-hoc test was used for statistical analysis. ∗p < 0.05, ∗∗∗p < 0.001, ns: not significant. Next, we explored whether inhibition of the COX2-PGE2 pathway could sensitise TMB-L inflamed tumours to ICB. LLC1 cancer cell has a relatively low TMB,36,52,53 and the LLC1 tumours formed in the mouse model exhibits the immune-inflamed feature (Supplementary Fig. S6D). Inhibition of COX2 with celecoxib could slightly reduce LLC1 tumour growth and weights in syngeneic C57BL/6 mice (Fig. 7G–I). Dual inhibition of COX2 and PD-L1 remarkably impeded tumour growth, and led to decreased infiltration of Tregs (Fig. 7J and K) and increased infiltration of Th1 cells (Fig. 7L and M). Our findings reveal that despite CD8+ T cell infiltration, TMB-L inflamed tumours exhibit distinct immunosuppressive features characterised by elevated Tregs and COX2-PGE2 pathway activation. The therapeutic efficacy of COX2 inhibition in the TMB-L inflamed model not only validates this pathway as a key mediator of ICB resistance but also highlights a clinically feasible strategy to reprogramme these low-response inflamed tumours into immunotherapy-sensitive microenvironments through targeted metabolic intervention. Discussion Accurate prediction of immunotherapeutic responsiveness is pivotal to patient stratification and inform more effective treatment strategies. Herein, we proposed a classification model and delineated five distinct subtypes by integrating RNA sequencing-based tumour immune phenotype classifier ImmPred with cancer cell intrinsic feature TMB to enhance the predictive accuracy of response to ICBs. Our analysis revealed that further stratification of the immune-inflamed and immune-excluded phenotypes into TMB-high and TMB-low subtypes could more effectively predict responses to ICBs. However, immune-desert tumours, irrespective of TMB levels, exhibit the poorest clinical outcome of immunotherapies, as the absence of initiation of adaptive immune response within this phenotype impedes anti-tumour immunity.16 Given the routine clinical availability of RNA-seq and TMB analysis, together with the potent predictive power, our classification model holds immediate translational value as a predictor of ICB response, with direct implications for treatment selection. There have been algorithms using bulk transcriptomic data for tumour immune phenotype classification, many of which stratify tumours by CD8+ T cell abundance. We have developed the RNA-seq-based classifier ImmPred, which is trained and validated using IHC classification of tumour-infiltrating CD8+ T cells in the TME, therefore comprehensively combining the density and spatial information of tumour-infiltrating lymphocytes and outperforming single-dimensional analysis. Meanwhile, our classification model highlights the significance of genomic feature TMB in patient stratification, jointly with the three tumour immune phenotypes, for improved prediction of ICB responses. Although our approach may not provide high-resolution view of cellular heterogeneity as single-cell RNA-sequencing, it is more widely practicable in clinical settings. Together, our immunogenomic classification model provides an integral assessment of cancer cell intrinsic characteristics and tumour-infiltrating lymphocytes in TME which enables accurate evaluation of clinical benefits to ICBs across cancer types. The ImmPred classifier, comprising seven genes (CD8A, IRF1, PDGFD, RABEP1, KLRC1, CD79A and SECTM1), was designed to capture both the presence and spatial distribution of T cells in the TME. CD8A directly reflects the abundance of CD8+ T cells. IRF1, a master transcription factor downstream of IFN-γ, regulates the expression of chemokines (CXCL9, CXCL10, CXCL11) that recruit effector T cells.54,55 PDGFD, produced by cancer-associated fibroblasts and tumour cells, contributes to desmoplasia and forming physical barriers that prevent T cell penetration into tumour islets, a hallmark of immune-excluded tumours.56,57 RABEP1 regulates endocytic trafficking and vesicle transport, with emerging evidence implicating its role in various immune functions.58,59 KLRC1 (NKG2A), expressed on NK and CD8+ T cells, delivers inhibitory signals upon binding to HLA-E.60 CD79A serves as a marker for B cells and tertiary lymphoid structures, which are increasingly recognised as key features of inflamed microenvironments and correlates of favourable immunotherapy responses.61 SECTM1 is a cytokine-like molecule involved in T cell trafficking and adhesion, promoting T cell infiltration by interacting with CD7 on T cells.62,63 Thus, this seven-gene panel captures not only the presence (CD8A, CD79A) and functional status (IRF1, KLRC1) of immune cells, but also the structural and stromal factors (PDGFD, RABEP1, SECTM1) that influence the spatial distribution of T cells, whether they successfully infiltrate tumour nests (inflamed), remain trapped in the stroma (excluded), or are completely absent (desert). Several TCGA-based immunophenotype classifications have been developed to quantify immune cell infiltration (e.g., CIBERSORT, xCell) or define distinct immune subtypes (e.g., C1–C6 immune subtypes) across cancers,64, 65, 66, 67 which provide valuable tools for profiling immune infiltration. However, they may overlook cancer cell intrinsic features or the spatial distribution of T cells, both of which are critical determinants of ICB response. Our model uniquely integrates TMB (antigenicity) with spatial T cell accessibility (inflamed/excluded/desert), not only improves the prediction of ICB response, but also offers subtype-specific vulnerabilities which have not been identified by immune infiltration-based methods. This positions our model not just as a descriptive taxonomy, but as a translational framework that provides practical combination immunotherapy strategies. Deciphering the mechanisms underlying each subtype of our classification model as well as identifying specific therapeutic targets can improve the tailoring of immunotherapies. Notably, in immune-excluded tumours, the ICB resistance mechanisms differ markedly between TMB-H and TMB-L excluded subtypes. Specifically, low expression or deletions of MTAP contribute to ICB resistance in TMB-H excluded tumours, whereas hyperactivation of TGF-β signalling pathway drives intrinsic resistance in TMB-L excluded tumours. Although previous studies have shown that deficiency of MTAP occurs frequently in tumours, we found that low expression or deletion of MTAP correlated significantly with lower ICB response only in TMB-H excluded tumours, but not in other subtypes. Moreover, our finding uncovered that intracellular MTA accumulation in MTAP deficient cells induced the inactivation of the interferon-γ pathway and impaired CD8+ T cell functions, therefore restoration of IFN-γ activity through pharmacological targeting of PRMT5 in MTAP deficient cancer cells significantly enhances the therapeutic efficacy of ICB in TMB-H excluded tumour models. In TMB-L excluded tumours, we demonstrated that inhibition of TGF-β pathway dramatically enhanced the efficacy of anti-PD-L1 treatment. Previous research has linked TGF-β signalling in cancer-associated fibroblasts to poor response to atezolizumab.15 Our study revealed that TGF-β signalling was specifically activated in TMB-L excluded tumours rather than in all subtypes. Thus, targeting the subtype-specific molecular drivers is essential for improving treatment efficacy of TMB-H and TMB-L excluded tumours. Immune-inflamed tumours are often referred to as “hot tumours” which show favourable ICB efficacy. However, TMB-H and TMB-L inflamed subtypes display starkly differential responses to immunotherapy, despite that they exhibit comparable levels of CD8+ T cell infiltration. We discover that this disparity is dictated by distinct underlying mechanisms of anti-tumour response. Our findings demonstrate that TMB-H inflamed tumours depend on IFN-γ to drive CD8+ T cells for anti-tumour responses, whereas in TMB-L inflamed tumours the abundance of Th1 and Tregs, modulated by COX-2-PGE2 pathway, is the key factor that affects ICB efficacy. Consistently, prior studies have highlighted the immunosuppressive role of COX-2-PGE2 signalling across various cancer types.68,69 In our TMB-L inflamed tumour models, inhibition of COX-2 by celecoxib led to increased accumulation of Th1 cells and decreased intratumoral Tregs, thereby enhancing responses to ICB treatments. Among the five immunogenomic subtypes, immune-desert tumours exhibit the poorest clinical outcomes. Unfortunately, we failed to identify therapeutic vulnerabilities to overcome ICB resistance in this subtype. High degree of biological and microenvironmental heterogeneity may exist in these tumours, making it challenging to pinpoint a single, dominant resistance mechanism using our current approach. Several potential mechanisms may underlie this phenotype. First, defective antigen presentation due to loss of MHC-I expression, B2M mutations, or impaired antigen processing machinery can render tumour cells invisible to T cells, preventing the initiation of adaptive immune responses.70 Second, immunosuppressive metabolic or stromal barriers may actively exclude dendritic cells and T cells from the tumour microenvironment. Third, primary resistance to type I interferon signalling may impair the cross-priming of T cells by dendritic cells, a critical step for initiating anti-tumour immunity.71 Addressing these mechanistic gaps represents a critical future direction. Promising therapeutic strategies include innate immune agonists (STING, TLR7/8, RIG-I) to activate dendritic cells and initiate de novo T cell responses,72, 73, 74 oncolytic viruses or cancer vaccines to provide in situ antigen presentation and T cell priming, myeloid cell reprogramming to alleviate immunosuppressive barriers,75 and epigenetic modulators (e.g., HDAC inhibitors, EZH2 inhibitors) to restore antigen presentation machinery.76,77 This study has several limitations. Firstly, the syngeneic mouse models in this study may not fully recapitulate the complexity of human cancer. Utilising humanised mouse models or direct clinical assessment will be essential to validate the therapeutic efficacy of our strategies in the future. Secondly, in our study TMB-H is defined by a uniform threshold (≧10 mutations/Mb) derived from previous pivotal pan-cancer studies.1 While this standardised cutoff facilitates cross-tumour comparison and has demonstrated clinical utility, we recognise that the intrinsic distribution of TMB varies substantially across different cancer types. Moreover, TMB measurements may vary across different sequencing platforms and bioinformatic pipelines. This technical variability could introduce noise into TMB classification. Future studies in larger, histology-specific cohorts will be essential to refine subtype-specific TMB cutoffs and optimise predictive accuracy for guiding personalised immunotherapy. Thirdly, some subgroup analyses, particularly those involving rare genetic alterations (e.g., MTAP deletion) or less common immunogenomic subtypes (e.g., immune-desert), were constrained by small sample sizes, yielding wide confidence intervals for hazard ratio estimates and indicating limited statistical power. Fourth, the survival analyses presented in this study are based on univariate Cox models and do not account for potential confounders. Although this approach was necessitated by the heterogeneity of pooled cohorts and the hypothesis-generating nature of subgroup analyses, we note that residual confounding may influence the observed associations. These findings should therefore be interpreted as hypothesis-generating and require validation in larger, dedicated cohorts. Lastly, the retrospective nature of the ZJCC&SYSMU validation cohort inherently introduces potential selection bias, which may affect the generalisability of our findings. Although we implemented stringent inclusion and exclusion criteria and consecutively enrolled eligible patients to minimise ascertainment bias, we acknowledge that unmeasured confounding factors, such as heterogeneous ICB regimens, and physician-led patient selection, could not be fully controlled. Prospective multicenter studies with predefined treatment protocols and larger, histologically stratified cohorts are warranted to validate the predictive accuracy of our immunogenomic classification and to confirm the clinical utility of the subtype-tailored therapeutic strategies identified in this study. In summary, leveraging ImmPred and TMB integrative assessment, we provide a clinical feasible classification with high predictive accuracy of ICB response in multiple cancers with unique patterns of intrinsic immune resistance. The subtype-tailored therapeutic combinations we propose hold promise for enhancing immune activation and overcoming immunosuppression, which will improve the efficacy and broaden clinical utility of immunotherapies. Contributors Conceptualisation, supervision, funding acquisition, writing, review and editing: M.L. and H.H.; methodology, conducting in vivo and in vitro experiments, clinical experiments, investigation, formal analysis: Y.Z., P.W., Z.H., and Z.Q.; writing original draft: Y.Z., P.W., and Z.H.; investigation, data curation, conducting in vitro experiments, clinical experiments: X.D., Q.W., Z.Z., K.S., C.W., and T.L.; validation, visualisation: X.L., J.Z., B.H., and Z.N.; resources, supervision: H.Y., Z.N., D.S., J.Z., and M.L.; software, project administration: Q.W. and W.F.; All authors have read and approved the final version of the manuscript. Data sharing statement RNA sequencing data of ZJCC cohort are deposited in NCBI Gene Expression Omnibus (GEO) dataset with the accession number GSE292421. Other data generated or analysed in this research can be made available upon request to the corresponding author. Declaration of interests The authors declare that they have no competing interests. Acknowledgements We thank all members of the Hu lab for helpful discussions. Thank Canwell Biotech Limited for providing CAN1012 (TLR7 agonist). This work was supported by grants from National Key Research and Development Program of China (2021YFA1300602), National Natural Science Foundation of China (82025026, 82230091, 82472775), Guangdong Basic and Applied Basic Research Foundation (2023A1515012412 and 2023A1515011214), Guangdong Science and Technology Department (2023B1212060013, 2023B1111030006), Key R&D Program of Zhejiang (2024C03160), Leading Innovative and Entrepreneur Team Introduction Program of Zhejiang Province (2024R01005 and 2025R01009). Footnotes Supplementary data related to this article can be found at https://doi.org/10.1016/j.ebiom.2026.106323. Supplementary data Supplementary Materials Western Blots STR-Merged References Associated Data Supplementary Materials Supplementary Materials Western Blots STR-Merged