STING signalling compensates for low tumour mutation burden to drive anti-tumour immunity

Summary

Background

While mutation-derived neoantigens are well recognized in generating anti-tumour T cell response, increasing evidences highlight the complex association between tumour mutation burden (TMB) and tumour infiltrating lymphocytes (TILs). The exploration of non-TMB determinants of active immune response could improve the prognosis prediction and provide guidance for current immunotherapy.

Methods

The transcriptomic and whole exome sequence data in The Cancer Genome Atlas were used to examine the relationship between TMB and exhausted CD8+ T cells (Tex), as an indicator of tumour antigen-specific T cells across nine major cancer types. Computational clustering analysis was performed on 4510 tumours to identify different immune profiles. NanoString gene expression analysis and single cell RNA-seq analysis using fresh human breast cancer were performed for finding validation.

Findings

TMB was found to be poorly correlated with active immune response in various cancer types. Patient clustering analysis revealed a group of tumours with abundant Tex but low TMB. In those tumours, we observed significantly higher expression of the stimulator of interferon genes (STING) signalling. Dendritic cells, particularly those of BATF3+ lineage, were also found to be essential for accumulation of Tex within tumours. Mechanistically, loss of genomic and cellular integrity, marked by decreased DNA damage repair, defective replication stress response, and increased apoptosis were shown to drive STING activation.

Interpretation

These results highlight that TMB alone does not fully predict tumour immune profiles, with STING signalling compensating for low TMB in non-hypermutated tumours to enhance anti-tumour immunity. Translating these results, STING agonists may benefit patients with non-hypermutated tumours. STING activation may serve as an additional biomarker to predict response to immune checkpoint blockades alongside TMB. Our research also unravelled the interplay between genomic instability and STING activation, informing potential combined chemotherapy targeting the axis of genomic integrity and immunotherapy.

Funding

City of Hope Christopher Family Endowed Innovation Fund for Alzheimer’s Disease and Breast Cancer Research in honor of Vineta Christopher; Breast Cancer Alliance Early Career Investigator Award; National Cancer Institute of the National Institutes of Health under award number R01CA256989 and R01CA240392.

Research in context.

Evidence before this study

What drives immune infiltration into tumours remains a key open question in oncology. It is widely accepted that somatic mutation-derived neoantigens are major targets of tumour-reactive CD8+ T cells, which have phenotypically been characterized as exhausted T cells (Tex). High tumour mutation burden (TMB) and abundance of Tex have both been widely connected with response to immune checkpoint blockades (ICBs). Low TMB tumours are considered to have more immunosuppressive or ‘cold’ tumour microenvironments, rendering them less responsive to ICBs. However, clinical data have also shown mixed responses of high TMB tumours to ICBs. Moreover, several studies found that mutation burden is comparable between T cell-inflamed and non-T cell-inflamed tumours, and some low TMB tumours can still possess active T cell responses.

Added value of this study

Our study showed a lack of significant correlation between TMB and abundance of Tex across nine different cancer types. In low TMB tumours, we observed that upregulation of stimulator of interferon genes (STING) signalling could enhance tumour-specific T cell responses and drive an inflamed tumour microenvironment. Furthermore, loss of genomic and cellular integrity was associated with STING activation in non-hypermutated tumours, serving as a compensatory mechanism to drive immune response in low TMB tumours.

Implications of all the available evidence

The complex relationship between TMB and immune response suggests TMB alone is not sufficient to predict anti-tumour immunity. Contrary to the belief that low TMB exclusively results in non-T cell inflamed tumours, our study reveals that STING signalling is an additional important marker that can compensate for low tumour antigen levels to enhance the tumour-specific T cell response, offering valuable insights for immunotherapy.

Introduction

Tumour mutation burden (TMB) is generally accepted as the primary driver of T cell responses in various cancer types.1, 2, 3, 4 More specifically, somatic mutation-derived tumour ‘neoantigens’ processed by antigen presentation machinery and loaded onto major histocompatibility complex (MHC) molecules for peptide presentation are accepted to elicit tumour-specific T cell responses.⁵ Tumours harbouring higher numbers of immunogenic antigens have increased potential to facilitate anti-tumour immune responses and respond to immune checkpoint blockades (ICBs), as prominently demonstrated by FDA approval of the anti-programmed cell death protein 1 (PD-1) antibody pembrolizumab for any cancer type with high TMB (≥10 mut/Mb).6, 7, 8 However, several studies have shown that TMB does not correlate directly with inflammatory status that is primarily manifested by T cell infiltration into the tumour microenvironment (TME), as both inflamed low TMB tumours and non-inflamed high TMB tumours may develop.^9,10 Furthermore, neoantigen-specific T cells have been identified across a wide range of tumour types with low or high TMB.11, 12, 13 Thus, factors beyond TMB that drive anti-tumour immunity remain incompletely understood.

While tumour-infiltrating T cell subsets are highly heterogeneous, exhausted CD8+ T cells (Tex) are now widely accepted as the tumour-specific fraction.14, 15, 16 Tex arise after repeated exposure to chronic antigen.^17,18 In multiple human tumour types, Tex have been demonstrated to be both tumour antigen and neoantigen specific, highlighting a clear connection between presence of somatic mutations and accumulation of Tex within the TME.^15,19 Recently, several studies have identified the presence of Tex within human tumours to be tightly associated with robust immune response featured by increased cytotoxic molecules, enhanced antigen presentation, and infiltrations of other immune components.20, 21, 22 Tex have also been shown to be an important marker for predicting therapeutic responses to ICBs.^23,24 Differentiation of memory precursors into Tex involves increased expression of checkpoint molecules such as PD-1, impairment of cytokine production, and loss of memory potential.^25,26 Central to this process is epigenetic and transcriptional rewiring of T cells,²⁷ which enables the identification of Tex through distinct gene expression patterns.28, 29, 30

In this study, we sought to identify additional features of the TME beyond TMB that could govern the magnitude of tumour-specific T cells within human tumours. We use Tex as the indicator for anti-tumour immunity, rather than bulk CD8 T cells which are known to also include bystander T cells.^15,31,32 We developed a distinct Tex transcriptional signature for an unbiased computational assessment of known variables of anti-tumour immunity across 4510 human tumour samples of nine major cancer types. Our analyses showed that TMB correlates poorly with Tex across all cancer types, and that high numbers of Tex can be found within both low TMB and high TMB tumours. Importantly, upregulation of stimulator of interferon genes (STING) signalling in non-hypermutated TMB tumours is strongly linked with increased levels of Tex. These findings suggest that STING activation may serve as a compensatory mechanism for inducing tumour-specific T cells within low TMB tumours. We further validated these bioinformatics findings experimentally using fresh breast tumours via fluorescence-activated cell sorting (FACS) analysis, NanoString, and single cell RNA-seq. Mechanistically, we found that diminished DNA damage repair, defective replication stress response and increased apoptotic rate led to loss of genomic and cellular integrity, and subsequently STING activation, thereby linking high cancer cell turnover and death rates with increased immune activation.

Methods

Tissue sample collection and processing

Tissues were collected from consenting patients with breast cancer who were undergoing stand-of-care at City of Hope. Patient characteristics can be found in Supplemental Table S1. The evaluation of tumour expression for oestrogen receptor-positive (ER), progesterone receptor-positive (PR), and human epidermal growth factor receptor-positive (HER2) was carried out by clinical pathologists. Fresh tumour tissue specimens were obtained through surgical resection and immediately placed in tube containing cold HBSS (Life Technologies, Thermo Fisher Scientific). These specimens were then transported on ice to the laboratory within 1 h of the surgery for further processing. To prepare single-cell suspension, tissues were first minced into smaller pieces and a gentleMACS Dissociator (Miltenyi Biotec) was used for mechanical dissociation. Enzymatic treatment was then carried out using 0.2 Wunsch U/mL Liberase TM (Roche) and 10 U/mL DNase (MilliporeSigma) in RPMI for up to 1 h, if necessary. In cases where red blood cells (RBCs) were present and needed to be removed, RBC lysis was performed using RBC Lysis Buffer (BioLegend).

Flow cytometry

PD-1+ CD39+ CD8+ Tex were quantified by flow cytometry as described before.³⁰ Briefly, single-cell suspensions were stained at room temperature in PBS with 2% FBS. Optimized antibody cocktails contained antibodies to identify Tex within tumour infiltrating T cells: CD3 (UCHT1, RRID: AB_893299), CD8 (SK1 RRID: AB_10551438), CD45RA (HI100 RRID: AB_647424), CCR7 (G043H7 RRID: AB_2561753), PD-1 (EH1.2 RRID: AB_2033989), CD39 (A1 RRID: AB_940423). All antibodies were titrated for optimal signal:noise ratios using peripheral blood mononuclear cells. The acquisition of samples was performed on a BD Biosciences Fortessa instrument operated through FACSDiva 6.1.3. Photomultiplier tube voltages were established using BD Biosciences CS&T Beads. Compensation was calculated with single-stained OneComp compensation beads (eBioscience, Thermo Fisher Scientific). Tex abundance was quantified based on the proportion of PD1+CD39+CD8+ T cells among the total CD8+ T cells. Using a cutoff of 1.5%, the samples were grouped into Tex-high and Tex-low groups (Supplemental Table S2).

NanoString gene expression analysis

Gene expressions were quantified using NanoString technology for 31 ER + HER2-human breast cancer samples. RNA was extracted from 10 μm thick slices of unbaked formalin-fixed paraffin-embedded (FFPE) tissue, utilizing Qiagen miRNeasy FFPE kits. To detect RNA transcripts, the NanoString PanCancer Immune Panel was employed, and nCounter technology from NanoString Technologies was utilized. The concentration of RNA was assessed using the Nanodrop spectrophotometer ND-1000 and Qubit 3.0 Fluorometer from Thermo Fischer Scientific. Furthermore, RNA fragmentation and quality control were assessed using the 2100 Bioanalyzer from Agilent. The total RNA was then subjected to overnight hybridization at 65 °C for 14–18 h following the manufacturer’s recommendations. Post-hybridization, the probe-target mixture was purified using the nCounter Prep Station, and quantification was performed with the nCounter Digital Analyzer. For data analysis, quality control, and normalization, the nSolver Analysis Software version 4.0 and nCounter Advanced Analysis Software version 2.0.115 were utilized. The measured gene expression values were normalized to the geometric mean of 40 housekeeping genes. NanoString genomic data have been deposited under the GEO accession number GSE190169.

Single cell RNA sequencing and analysis

Single cell sequencing of n = 14 ER+ and triple negative breast tumour samples was conducted using the 10× Genomics Chromium platform, following recommended procedures. The CellRanger software was employed for aligning sequence reads to the human genome and counting aligned transcripts for each cell. Raw counts for each tumour sample were then filtered, normalized, and scaled using the Seurat R package (v4.3.0).³³ Specifically, nonviable cells were defined as cells with more than 20% mitochondrial gene expression; empty well or duplets were defined as cells expressing fewer than 200 genes or more than 2500 genes. After discarding those cells, a total of 50818 cells were collected for the downstream analysis. The top 2000 most variable genes were selected for sample integration and principal component analysis (PCA). For integration, all tumour samples were combined using Seurat’s standard integration workflow, which involved using the “FindIntegrationAnchors” and “IntegrateData” functions with 50 dimensionality. PCA was applied to the integrated data object. Two heuristic methods available in Seurat, the modified Jack Straw procedure and ranking variance method, were used to select the top 30 principal components for subsequent nonlinear dimensional reduction using uniform manifold approximation and projection (UMAP) and clustering analysis. Expression heatmap was plotted with the selection of top ten gene markers of each cluster. The normalized gene expression of genes of interest in cells was then compared between the Tex-high and Tex-low tumours determined by flow cytometry in ER + samples only (n = 9).

Public dataset of ER + breast tumour samples (n = 75) from Pal et al. (GSE161529)³⁴ was analysed by the same workflow and parameters as the above quality control and dataset integration. A total of 363062 cells were collected, and a resolution of 0.1 was selected for “FindClusters” function in the cell clustering analysis. An expression heatmap using “DoHeatmap” function with the top ten gene markers was followed by down sampling with the fewest cell count to capture each cluster due to the bulky cell number.

Public databases for gene expression and somatic mutations

mRNA expression (FPKM) and somatic mutation data (MAF) of ten major cancer types (n = 4686) assessed by the Cancer Genome Atlas (TCGA)³⁵ were downloaded and organized from genomic data commons (GDC) of the National Cancer Institute. mRNA expression data (FPKM) was further log2 transformed. METABRIC (Molecular Taxonomy of Breast Cancer International Consortium) mRNA and clinical dataset (Illumina HT 12 platform) with total 1992 samples were downloaded from European Genome-Phenome Archive (data set IDs EGAD00010000210 and EGAD00010000211).^36,37 Breast cancer tumours were sub-grouped by the oestrogen receptor positive (ER+) and triple negative (TNBC) by hormone receptor status when shown. Colorectal cancer tumours were sub-grouped into microsatellite unstable (CRC-MSI) and microsatellite stable (CRC-MSS) when shown.

Gene signature analysis

The Tex signature and Treg signature were generated from our previous study.³⁰ We further developed a Tirosh Tex signature,²⁸ Schumacher Tex signature,²⁹ IFNγ signature,³⁸ T cell signature,⁹ STING signature,³⁹ apoptosis signature,⁴⁰ replication stress response defect (RSRD) signature developed from cell models with key RSR signalling factors ATR, ATM, CHEK1, and CHEK2 depletion (filtered by log fold change ≥ 2),⁴¹ mismatch repair signature,⁴² nucleotide excision repair signature,⁴² necrosis signature (Gene Ontology gene list), autophagy signature,⁴² pyroptosis signature⁴³ by collecting available gene expression patterns from previous publications (Supplemental Table S3). Signature scores were calculated using sig.score function from genefu R package (v.2.26.0). Patients of all cancer types except pancreatic cancer were grouped into Tex-high, Tex-medium, and Tex-low by top 1/3 and bottom 1/3 cutoffs. TMB from each tumour sample was used to perform correlation analysis using ggscatter function in ggpubr R package (v.0.4.0). Previously predicted neoantigens for TCGA deposited samples were downloaded from TSNAdb platform (http://biopharm.zju.edu.cn/tsnadb/) using NetMHCpan 4.0.⁴⁴ MAF files were input and analysed using maftools (v2.10.05) R package.⁴⁵ To detect and compare the affected oncogenic pathways in each cluster, OncogenicPathway function⁴⁵ was used to show the fraction of sample affected of eight major oncogenic pathways in each cluster within major cancer type.

Unsupervised differentiation of tumour microenvironment subtypes using known immune-related variables

A selection of ten variables was made to offer a comprehensive depiction of the composition of inflamed TME. These variables comprise of canonical genes or signatures that signify different immune subsets: TMB, Tex signature, BATF3, CD274, CD3G, CD8A, IFNγ signature, CD79A, STING signature, and CD68 of 4510 TCGA tumour samples were scaled for the unsupervised clustering analysis. Principal component analysis (PCA) using sklearn. decomposition was performed, and top principal components (n = 8) were selected based on the elbow method (Supplemental Fig. S6a). The cluster number was determined by the k-means algorithm (scikit-learn.cluster.KMeans, v1.0.2) python package based on elbow method using Distortion (matplotlib.pyplot, v3.5.3) (Supplemental Fig. S6b). A UMAP projection was used to visualize the generated clusters (umap-learn, v0.5.3) python package as outlined in the preprint of the UMAP projection.⁴⁶ Markers for each cluster were generated by Wilcoxon rank-sum statistics (scipy.stats.ranksums v1.7.3). Mean values of the cluster markers were used to display the features of each cluster as shown.

To further characterize each cluster, gene set enrichment analysis (GSEA) analysis was performed by GSEA software (v4.2.2) between clusters using Hallmark and KEGG (Kyoto Encyclopedia of Genes and Genomes) gene sets with default settings.^40,42,43 CIBERSORT data were obtained from TCGA and major types of immune cells were compared among clusters to depict immune composition.⁴⁷

LASSO regression analysis

To compare the value of immune-related variables (as predictor variables) in predicting Tex abundance (as response variable), a logistic regression model LASSO (least absolute shrinkage and selection operator) using glmnet (v4.1-6) R package was performed. To first remove the redundant variables, a variance inflation factor (VIF) was calculated and filtered using five as cutoff. The optimal coefficient of each variable was then generated from the model with lowest mean squared error (MSE). The coefficient of each variable represented its weight in the prediction model. The most important predictor variable with highest coefficient was further validated using a cross-validation approach, in which one predictor variable was removed at each time to calculate the R² of the model including all the rest variables. The most important predictor was then identified when the regression model generated the smallest R² value without the corresponding variable.

Statistics

Statistical analyses were performed using “stat_compare_means” from ggpubr (v0.4.0) R package. Comparison of correlation coefficient was performed using cocor (v1.1-4) R package.⁴⁸ Unpaired Student’s t tests for two groups and Holm-Sidak multiple comparison was used for one-way ANOVA for multiple groups were used in parametric data. Wilcoxon rank-sum test for two comparison and Kruskal–Wallis with Dunn’s for multiple comparison were used in data for which the normality assumption is violated. P < 0.05 were considered as significant difference. Spearman test was used in correlation analysis. Calculated P values are displayed as ∗, P < 0.05; ∗∗, P < 0.01; ∗∗∗, P < 0.001; ∗∗∗∗, P < 0.0001. The upper and lower quartiles, and the median were shown in the box plots. Experiment-specific detailed statistical methods are described in corresponding figure legends and Methods sections.

Ethics

Fresh tumour and peripheral blood were obtained from patients who gave institutional review board-approved (IRB-approved) written informed consent prior to inclusion in the study (City of Hope IRB 05091, IRB 07047, and IRB 14346).

Role of funders

Funders had no role in study design, data collection, data analysis, interpretation, or writing of the manuscript.

Results

Tumours with increased exhausted CD8+ T cells have abundant T cell infiltration and IFNγ signalling

We selected ten of the most common cancer types to examine their tumour mutation and immune profiles with data publicly available from TCGA as described in Fig. 1a. To start, we selected ten variables including TMB and key immune components to portray comprehensive tumour immune profiles. Next, we performed unsupervised clustering analysis to capture different features of tumour immune microenvironments. By comparing these clusters and performing deconvolution analysis, we identified crucial determinants that could drive active anti-tumour immune response. While previous studies have assessed the relationship between general T cell infiltration, TMB, and other immune variables,^9,10 we aimed to do so more specifically using a Tex signature as a specific marker of tumour-specific T cells.^15,19,23,49 To validate the value of this approach, we examined associations of the abundance of various previously identified tumour-infiltrating CD8+ T cell subsets with IFNγ signalling, which is widely accepted to reflect anti-tumour T cell effector activity.^50,51 Correlations between IFNγ signature and those of four major CD8+ T cell subsets (exhausted, resident, effector, and central memory) generated from our single cell RNA-seq dataset³⁰ were performed across all cancer types combined (Supplemental Fig. S1a). Indeed, we found that Tex showed a significantly higher correlation coefficient with IFNγ signature compared with other CD8+ T cell subsets (Supplemental Fig. S1b). The same findings were also identified individually within each cancer type (Supplemental Fig. S1c). Our data suggests that Tex could reflect the active anti-tumour immunity than other CD8+ T cells that could be considered as bystanders. In addition, we investigated the correlation of Tex level with CYT index and CD8/Treg ratio, which has been shown to reflect the immune cytolytic activity and homeostatic tumour microenvironment.^52,53 The significantly positive correlations of Tex abundance with CYT index and CD8/Treg ratio further indicate Tex could reflect an inflammatory TME (Supplemental Fig. S1d–g). We extended the validation of these findings beyond our own Tex signature, developed from breast cancer to Tex signatures identified from other cancer types. Our Tex signature showed strong positive correlation (R > 0.65, Spearman’s rank test, P < 0.0001) with Tex subset generated by Schumacher et al. in a NSCLC dataset and by Tirosh et al. in a melanoma dataset^28,29 within each respective cancer type, as well as all cancer types combined (Supplemental Fig. S2a–h).

Fig. 1.

Tumours with increased exhausted CD8+ T cells have abundant T cell infiltration and IFN signalling. (a) Experimental design. Most common cancer types with somatic mutation and transcriptomic data from TCGA were subjected to unsupervised clustering analysis to reveal diverse tumour immune profiles. Through the comparison of these different tumour immune profiles, critical factors to drive active tumour immunity were identified. (b) Tex level within each cancer type grouped into Tex-high, Tex-medium, and Tex-low based on the upper and lower 1/3 cutoff across all cancer types (n = 4510) and ordered by mean value (indicated by black triangle). (c) Tex group distributions in various cancer types. (d) The correlation between Tex and IFNγ signature and CD3G expression across all cancer types combined (n = 4510) (Spearman’s rank test, P < 0.001). (e) The correlation indicated by the R value between Tex level and IFNγ signature and CD3G within each cancer type (Spearman’s rank test, P < 0.001).

To examine Tex abundance from a pan-cancer perspective, we next classified all tumours into Tex “high”, “medium”, or “low” subgroups based on top 1/3, middle 1/3, and bottom 1/3 across all cancer types (Fig. 1b). Tex levels as well as CD3+ and CD8+ T cells in pancreatic cancer were dramatically lower than all other cancer types (Supplemental Fig. S2i–k), consistent with other findings that have shown dramatic immune exclusion and immunosuppression in the pancreatic tumour microenvironment.⁵⁴ Given the scarcity of T cells found within the pancreatic tumour TME, we removed these samples from the remainder of our study. Among other types of cancer, lung adenocarcinoma, TNBC, and CRC-MSI exhibited the highest fractions of tumours classified as Tex-high, while prostate cancer, CRC-MSS, and liver cancer generally were composed of Tex-low tumours (Fig. 1c). While tumour-infiltrating Tex levels are highly variable across these cancer types, abundant Tex can be found in tumours of any cancer type. We further validated the use of our Tex signature by examining the association between Tex groups and IFNγ signalling, CD3+ T cells across all cancer types (Fig. 1d) or within each cancer type (Fig. 1e and Supplemental Fig. S3). Tex demonstrated strong and significant positive correlations across these measurements (Fig. 1d and e). Notably, we did not observe tumours with high levels of CD3+ T cells but low Tex levels, demonstrating that Tex generation is a key feature of heavily T cell-infiltrated tumours. In further support of this, we also observed a significant positive correlation between Tex and a previously generated T cell infiltration signature⁹ across several cancer types (Supplemental Fig. S3). These data collectively confirmed the central role of Tex in driving T cell abundance and IFNγ related inflammation within the TME.

Tumour mutation burden does not predict T cell-inflamed microenvironment status

We then examined the relationship between TMB, Tex levels, and tumour immune profiles. As expected, total mutation load varies across cancer types (Supplemental Fig. S4a).⁵⁵ Within all cancer types except CRC-MSI, TMB was not higher in Tex-high tumours as compared to Tex-low tumours (Supplemental Fig. S4b). In contrast, Lung-LUAD Tex-low group had higher TMB than Tex-high tumours, which further indicated the non-linear relationship between TMB and infiltration of tumour specific T cells. We then showed that TMB weakly correlated with Tex levels across all tumour types, with low absolute Spearman’s correlation coefficients (|R| < 0.25) (Fig. 2a–c and Supplemental Fig. S4c). Similarly, CD8+ T cells and IFNγ signalling correlated poorly with TMB within each cancer type, as well as all cancer types combined (Fig. 2b and c). Since only a small fraction of somatic mutations are actually presented on HLA molecules as peptide:MHC complexes,⁵⁶ we then examined immune profiles in the context of predicted neoantigen by binding affinities based on somatic mutation data and HLA allele information generated from a public immunogenomic database.⁴⁴ Much like TMB, neoantigen levels showed poor correlation with CD8+ T cell, IFNγ signalling, or Tex abundance either within each cancer type or across all cancer types combined (|R| < 0.2, Spearman’s rank test) (Supplemental Fig. S5).

Fig. 2.

Total mutation burden shows no correlation with T cell-inflamed tumour microenvironment. (a) Correlation analysis between Tex sig.score and TMB in relatively inflamed tumour types (Spearman’s rank test). (b) R value of the correlation of CD8A, IFNγ, and Tex with TMB within each tumour type (Spearman’s rank test). (c) Correlation analysis between Tex sig.score and CD8A expression and TMB including all types of cancer (n = 4510) (Spearman’s rank test). (d) Summary of the correlation of BATF3, CD8A, and TMB with Tex in each cancer type, coloured by Tex group and ordered by the mean of Tex signature score within each cancer type.

BATF3+ conventional type 1 DCs have been shown to be required for effector CD8+ T cell accumulation in tumours and cytotoxic activity.^57,58 Furthermore, BATF3+ DCs have been shown to be highly positively correlated with degree of T cell infiltration.^9,58 We therefore examined the relationship between Tex, CD8A, and BATF3 in context of TMB. While TMB was comparable between Tex-high and Tex-low groups, Tex-high tumours demonstrated both higher abundance of CD8+ T cells and BATF3-lineage DCs than Tex-low tumours (Fig. 2d). Together these results demonstrate a poor relationship between Tex and TMB, but a strong relationship between Tex and tumour localized BATF3+ DCs.

Discordant relationship between TMB and TME immune features

To comprehensively investigate the complex relationship between Tex, TMB, and immune features of the TME, we next performed an unsupervised computational analysis including all cancer types. We incorporated input variables thought to be important in defining tumour immune profiles: TMB, Tex, BATF3, CD274 (PD-L1),⁵⁹ CD3G (T cells), CD8A (CD8+ T cells), IFNγ signature,³⁸ CD79A (B cells),⁶⁰ STING signature,³⁹ and CD68 (macrophages).⁶¹ We included these major immune populations and signalling molecules involved in tumour immunity to explore associations between these factors as well as how they interact to form different immune patterns. Such an approach enables a better understanding of the pivotal drivers of effective anti-tumour immunity. Through unsupervised clustering analysis using k-means algorithm, six clusters of TME immune profiles were identified, with heterogenous distribution across cancer types (Fig. 3a and b and Supplemental Fig. S6c and d). Of the six TME immune profiles, clusters 0,1, and 2 were relatively ‘cold’, while clusters 3, 4, and 5 were relatively ‘hot’ (Fig. 3c–f).

Fig. 3.

Unsupervised clustering analysis shows distinct TME immune profiles. (a) Six major clusters generated by selected variables (shown in 3C) including nine tumour types (n = 4510). Coloured by Tex levels to show the TME inflammatory status from deep blue (‘cold’) to deep red (‘hot’). Each dot indicates one patient. (b) Cluster composition within each tumour type. (c) Comparison of the average expression of each variable among clusters with Z-score normalisation. (d) Comparison of the expression of the variables with Z-score normalisation visualized by individual patient samples. (e) Quantitative comparison of major variables across each cluster. TMB level is shown with log2 transformation of mutation per megabase. (f) Overlay of the expression of major variables onto UMAP visualized in each cluster.

In our observation of these six clusters, there existed three distinct levels of TMB. Cluster 0, 1, and 4 had low TMB, cluster 2 and 5 had intermediate TMB, and cluster 3 had high TMB. No consistent positive or negative correlation between TMB and immune infiltrate was observed (Fig. 3c). Irrespective of TMB status, our TME profiles revealed several notable features of tumour immune composition. Across all cancer types, we observed strong positive correlation in abundances of Tex, CD8+ T cells, general T cells (CD3G), BATF3+ DCs, B cells (CD79A), IFNγ signalling, and PD-L1 (CD274). Notably, macrophage abundance (CD68) was an immune variable uniquely upregulated in TME profile 2, an immune cold group enriched in esophageal, kidney, and liver cancer. This finding likely points to a prominent role of immunoinhibitory macrophages in these tumour types.⁶² In contrast, BATF3+ DCs were most abundant in immune-rich groups 3,4, and 5 but absent in immune-poor groups 0, 1, and 2. These findings support the notion that BATF3+ DCs are required for accumulation of T cells within tumours.

STING activation compensates for mutation burden-low tumours to drive active immune response

Notably, STING signalling was less elevated in the intermediate TMB group 2 compared to the more inflamed low TMB group 4 (Fig. 3c–e). Across the three low TMB groups 0, 1, and 4, increased STING signalling progressively associated with increased abundance of Tex (Fig. 4a). A variety of inflammatory cytokines and chemokines have been determined to be driven by STING signaling.63, 64, 65 In line with STING pathway expression, we observed a gradual rise in expression of type 1 interferon-related molecules IFNAR1, IFNAR2, and IFNB1 in cluster 0, 1, and 4 (Fig. 4b). Similarly, we found increased expression of chemokines CCL2, CCL20, CCL26, CXCL10, cytokines IL12A, IL12B, IL6, and TNFA, and the inflammatory feedback molecule IDO1 among these three groups. These findings were recapitulated individually in low TMB tumours extracted from each cluster for cancer type specific analysis: CRC-MSS, ER + BC, kidney cancer, and prostate cancer (Supplemental Fig. S7a). In light of the critical role of chemokines in lymphocyte infiltration as highlighted by the enriched pathway identified through GSEA analysis (Fig. 5a), we proceeded to examine additional genes involved in chemokine signalling within these clusters. Markedly augmented expression of genes involved in chemokine signalling was noted among these low TMB groups, including both chemokine and downstream signalling transduced by chemokine receptors (Fig. 4c). As CXCL10 has been shown to recruit CD8+ T cells and other immune cells to the TME,⁵⁸ we next examined the relative abundance of tumour-infiltrating immune populations via CIBERSORT. We found progressively increased numbers of DCs, pro-inflammatory macrophage populations (M1), CD4+ T cells, CD8+ T cells, and B cells among clusters 0, 1, and 4 (Fig. 4d and Supplemental Fig. S7b).

Fig. 4.

STING activation compensates for non-hypermutated tumours to drive active immune response. (a) TMB, STING, and Tex levels compared among low TMB clusters (One-way ANOVA. ns, not significant; ∗∗∗∗, P < 0.0001). (b) The expression of STING-dependent genes compared among low TMB clusters (One-way ANOVA. ∗∗, P < 0.01; ∗∗∗∗, P < 0.0001). (c) Chemokine pathway genes comparison among low TMB clusters. (d) CIBERSORT analysis of relative abundance of major immune populations compared among low TMB clusters (Kruskal–Wallis rank-sum test. ∗∗∗∗, P < 0.0001). (e) TMB, STING, and Tex levels compared between intermediate TMB clusters (Student’s t test. ns, not significant; ∗∗∗∗, P < 0.0001). (f) The expression of STING-dependent genes compared between intermediate TMB clusters (Student’s t test. ∗∗, P < 0.01; ∗∗∗∗, P < 0.0001). (g) Chemokine-related genes comparison between intermediate TMB clusters. (h) CIBERSORT analysis of relative abundance of immune populations compared between intermediate TMB clusters (Wilcoxon’s rank-sum test. ∗∗∗∗, P < 0.0001).

Fig. 5.

Loss of genomic and cellular integrity regulates STING activation. (a) Mismatch repair signalling and (b) STING signalling compared between CRC-MSS and CRC-MSI. (c) Mismatch repair signalling and (d) STING signalling compared among clusters in low TMB group in CRC-MSS tumour samples. (e) The correlation between TMB and STING signalling across all types of cancer combined. RSRD signature compared among (f) low TMB clusters, (g) intermediate TMB clusters. Apoptosis signature compared among (h) low TMB clusters, (i) intermediated TMB clusters. (j) The correlation between apoptosis signalling and STING signalling across all types of cancer combined. Statistics generated by unpaired Student’s t test (a, b, g, i), 1-way ANOVA with Holm-Šídák multiple comparisons test (c, d, f, h), and nonparametric Spearman rank correlation (e, j). P values are indicated in the figure.

In the same fashion, we next compared genes downstream of STING between two intermediate TMB groups with distinct inflammatory status: cluster 2 and cluster 5. Despite similar TMB levels, both STING and Tex were significantly higher in cluster 5 (Fig. 4e). Again, we found robustly higher levels of pro-inflammatory cytokines and chemokines (Fig. 4f and g). Consistent with previous findings, cluster 5 also demonstrated an increase in major anti-tumour immune populations and decrease in M2 macrophages (Fig. 4h). All findings were found to be consistent within individual cancer types from each cluster group (Supplemental Fig. S8). To validate these findings in an independent dataset, we utilized the METABRIC breast tumour public dataset. We found a significant positive correlation between STING signalling and Tex levels (R = 0.79, Spearman’s rank test, P < 0.0001) in non-hypermutated invasive ductal carcinoma (IDC) (TMB < 10), with no correlation between TMB and Tex levels (R = −0.063, Spearman’s rank test, P = 0.016) (Supplemental Fig. S9).

We next assessed if STING amplification of immune activation was also observable in metastatic tumours.⁶⁶ To examine this, we specifically evaluated melanoma tumours metastatic to distant solid organs within the TCGA dataset. Here we also found Tex and STING signalling were significantly correlated (R = 0.7, Spearman’s rank test, P < 0.0001), with no correlation between Tex and TMB (R = 0.054, Spearman’s rank test, P = 0.75) (Supplemental Fig. S10).

Loss of genomic and cellular integrity associates with STING activation

To explore potential pathways that may promote STING signalling, we further compared different TME immune profiles in both low and intermediate TMB groups by GSEA (Supplemental Fig. S11). Upregulated gene sets in clusters with high STING signalling (cluster 4 and 5) were mainly immune-related pathways, such as T cell receptor signalling, chemokine signalling, and antigen-processing and presentation. As expected, we observed upregulation of cytosolic DNA sensing pathways, which is a primary feature of STING signalling via cyclic GMP-AMP synthase (cGAS) sensing of cytosolic nucleic acids.⁶⁷ We also observed upregulation of other pattern recognition receptor (PRR) pathways, such as nod-like receptor signalling and toll-like receptor signalling, which have been shown to synergize with cGAS activation of the STING pathway.^64,68

Beyond immune-related genes, we observed increased levels of genes associated with DNA damage repair (DDR) pathways in the low STING signalling cluster (cluster 0) (Supplemental Fig. S11a). This suggested that the opposite may also occur in other tumours: diminished/dysregulated DNA damage repair could initiate STING activation. TREX1 (three-prime repair exonuclease 1), a critical component in DNA damage repair machinery, is known to remove accumulated cytosolic DNA as a result of DNA damage or replication errors.⁶⁹ Mutation of TREX1 has been associated with high level of type I interferon and several autoimmune diseases.^70,71 We found decreased TREX1 expression in both low and intermediate TMB clusters with enhanced STING signalling, indicating the independent role of TREX1 from TMB in STING activation (Supplemental Fig. S12a and b). We further examined DNA repair pathways implicated in two notable contexts: mismatch repair (MMR) in colorectal cancer⁷² and nucleotide excision repair (NER) in melanoma.⁷³ As predicted, CRC-MSI tumours displayed lower gene expression involved in MMR pathway (e.g., MLH1) and higher STING signalling as compared to CRC-MSS tumours (Fig. 5a and b, Supplemental Fig. S12c). Even within CRC-MSS, MMR progressively decreased across low TMB clusters as STING signalling increased (Fig. 5c and d). A similar inverse relationship between NER and STING signalling was observed in melanoma tumours. In both low and intermediate TMB groups, genes involved NER pathway were decreased in tumours with increased STING signalling (Supplemental Fig. S13). Together, these findings suggest that deficiencies in DNA damage repair could enhance STING activation independently of total mutation load (Fig. 5e).

We next investigated the relationship between STING signalling and DNA replication stress, which is now increasingly viewed as a feature of genomic instability in cancer cells.⁷⁴ Replication stress response (RSR) machinery has been shown to overcome the stalling during DNA synthesis to maintain genomic integrity.⁷⁵ Previously, a gene signature associated with RSR defects (RSRD) was developed using cell models from oncogene-induced depletion of key RSR factors.⁴¹ Mirroring our findings of an increased lack of DDR genes associated with elevated STING signalling, we found a progressive increase in RSRD signature expression in clusters with increased STING signalling in both low and intermediate TMB clusters (Fig. 5f and g). These results were also found to be consistent within individual cancer types (Supplemental Fig. S14). These findings indicate that deficient RSR and loss of genomic integrity may also contribute to STING activation, which is in agreement with previous findings demonstrating defective RSR leads to accumulation of immunostimulatory cytosolic DNA.⁷⁶ In an attempt to identify key mutational events that might drive reduced genomic integrity and STING signalling, we next assessed oncogenic pathways across each TME immune profile within each cancer type (Supplemental Fig. S15). However, no clear pattern of mutational profiles could be identified, reflecting the heterogeneous and redundant relationships between specific gene mutations, genomic integrity, and STING signalling.

Beyond genomic instability, loss of cellular integrity induced by rapid cell turnover, cell death, and subsequent release of tumour cytosolic DNA are also known to drive STING signaling.^64,77 Consistent with this, we found upregulation of genes related to apoptosis signalling in clusters with high STING signalling (cluster 4 and 5) in GSEA analysis (Supplemental Fig. S11). We also observed a gradual rise in apoptosis signalling in clusters with increased STING activation in low TMB group and within individual cancer types (Fig. 5h and Supplemental Fig. S16a). Similar findings were also evident in intermediate TMB group and within individual cancer types (Fig. 5i and Supplemental Fig. S16b). Apoptosis and STING signalling were strongly correlated across all tumour types (Fig. 5j, R = 0.48, Spearman’s rank test, P < 0.0001). Similar findings were also observed in other cellular death pathways, such as necrosis, pyroptosis, and autophagy (Supplemental Fig. S17). Since apoptosis and other cellular death events may act both upstream and downstream of STING activation,^78,79 further studies are necessary to identify how critical features of cell death are linked to STING-dependent inflammation. Collectively, our data suggest that loss of genomic and cellular integrity could enhance STING signalling triggered by increased tumour cytosolic DNA.

Experimental validation from fresh breast tumours via FACS analysis, NanoString, and single cell RNA-seq

Since ER + breast tumours are almost exclusively non-hypermutated tumours (99% based on TCGA database, Supplemental Fig. S4a), we sought to validate the above findings using ER+ human breast tumour samples (n = 31). Expression of STING pathway genes was quantified by NanoString technology. Tex levels of each sample were determined by flow cytometry and stratified into Tex-high and Tex-low groups. Tex-high tumours were associated with higher expression of STING pathway genes (Fig. 6a). STING signature was significantly correlated with Tex signature with available genes quantified by NanoString (Fig. 6b). Though not the full Tex gene list were captured by NanoString, we proved the signature of available genes were highly correlated with the full signature in both TCGA and METABRIC databases (R ≥ 0.94, Spearman’s rank test, P < 0.0001) (Supplemental Fig. S18b–d).

Fig. 6.

Validation by fresh breast tumours analysed via NanoString and single cell RNA-seq. (a) STING-related genes quantified by NanoString compared between Tex-high (n = 18) and Tex-low (n = 13) tumour samples (Student’s t test. ∗, P < 0.05; ∗∗, P < 0.01). (b) Correlation between STING signature and Tex signature from NanoString dataset (n = 31) (Spearman’s rank test). Single cell RNA-seq quantified the expression of STING signature compared between Tex-high (n = 5) and Tex-low (n = 4) tumour samples in (c) dendritic cells and (d) cancer cells (Student’s t test). Single cell RNA-seq quantified the expression of STING-related individual genes compared between Tex-high (n = 5) and Tex-low (n = 4) tumour samples in (e) dendritic cells (n = 840) and (f) cancer cells (n = 8599) (Wilcoxon’s rank-sum test. ∗, P < 0.05; ∗∗, P < 0.01; ∗∗∗, P < 0.001; ∗∗∗∗, P < 0.0001).

Another batch of ER+ breast tumour samples (n = 9) were further analysed with single cell RNA-seq data to identify cell subsets that exhibited active STING signalling in Tex-high vs. Tex-low samples (as quantified by flow cytometry). We captured and annotated the most common immune and tumour populations (Supplemental Fig. S19a and c). STING signalling was differentially activated within two cell types: dendritic cells and cancer cells (Supplemental Fig. S19b and d). We further validated these findings using ER + breast tumour from Pal et al. single cell RNA-seq datasets (n = 75) (Supplemental Fig. S20).³⁴ Since previous studies also demonstrated DCs and tumour cells as the primary cell types for STING activation,^80,81 we then compared STING signalling in DCs and cancer cells between Tex-high and Tex-low tumours. We observed both STING signature and STING-related genes were upregulated in Tex-high tumours compared to Tex-low tumours (Fig. 6c–f).

STING signalling is a central node of robust anti-tumour immunity

Our findings indicated that STING signalling plays an important role in driving T cell infiltration into non-hypermutated tumours but is not essential in high TMB tumours such as cluster 3. This is supported by a significantly positive correlation between STING signalling and Tex level in non-hypermutated tumours (TMB < 10) (Fig. 7a). Further, we classified tumours into six groups based on TMB (5 and 10 mut/megabase as cutoffs to define TMBlo, TMBint, or TMBhi) and STING signalling (median of STING sig.score as cutoff to define STINGlo or STINGhi). We also identified BATF3+ DCs within the TME as another critical requirement for T cell activation. Based on these results, we next examined Tex levels in context of STING activation status (STINGlo or STINGhi) and BATF3 expression. We found that within each degree of TMB stratification, increased STING was associated with increased Tex, and this was additionally associated with increased tumour BATF3 expression (Fig. 7b). Since we also observed concurrent expression of immune variables in more inflamed tumour groups, we next asked which immune variables were most critical for Tex accumulation. To do so, we performed LASSO regression of selected immune variables to predict Tex. The variables IFNγ signalling, CD8A, and CD3G were determined to be redundant based on their VIF value exceeding the cutoff of 5, so they were removed from subsequent regression model (Supplemental Fig. S21). The coefficient of each remaining variable was then calculated from the fitted LASSO regression model. Results showed that STING signature exhibited the highest coefficient, indicating its importance in predicting Tex levels, while the coefficient of TMB was lowest among all other variables (Fig. 7c). To validate the most significant predictor variable with the highest coefficient, a cross-validation approach was employed. This method involved removing one predictor variable at a time and calculating R² of the model with all remaining variables. This identified the most important predictor by determining which variable, when removed, resulted in the lowest R² value in the regression model. Results indicated that the prediction accuracy without STING was lowest than other single exclusive scenarios, while the prediction accuracy without TMB was the highest among all the scenarios (Fig. 7d). Together, these results identify STING signalling as a key feature of Tex accumulation and immune activation within the TME (Fig. 7e).

Fig. 7.

STING signalling is a more important factor to predict active anti-tumour immunity than TMB. (a) Correlation between STING signalling and Tex level in non-hypermutated tumours (TMB<10) (Spearman’s rank test). (b) Tex levels compared among groups stratified by TMB (TMB 5 and 10 as cutoffs) and STING (median as cutoff) signalling and coloured based on BATF3 expression (Student’s t test. ∗∗∗∗, P < 0.0001). (c) Coefficient of each variable in the LASSO regression model for Tex prediction. (t statistics for each variable as ordered: 27.21, 25.41, 21.65, 22.16, 25.86, −1.64. ∗∗∗, P < 0.001) (d) R² value of LASSO regression model to predict Tex without the corresponding genes or signatures. (e) Graphical summary of the key identified variables that drive accumulation of tumour-specific Tex T cells and related immune activation in the TME.

Discussion

Factors that drive anti-tumour immune responses remain key open questions in oncology. While TMB is widely accepted as a key driver of TIL into tumours, our findings demonstrate that TMB alone does not dictate T cell activation or immune composition within the TME. Beyond TMB, we showed that in tumours with low or intermediate TMB, activation of STING signalling indicated by increased production of type I interferon and pro-inflammatory cytokines can augment the anti-tumour immune response. Additionally, increased chemokine secretion downstream of STING activation, especially CXCL10, can facilitate the recruitment of immune cells into tumours. We additionally validated these findings by analysing fresh luminal breast tumour samples that were sequenced using NanoString and single cell RNA-seq techniques. The results showed the upregulation of STING signalling in Tex-high tumours compared with that in Tex-low tumours. Further, single cell RNA-seq data specifically revealed DCs and cancer cells exhibited the STING activation in Tex-high tumours, which provides deeper insights into the cells responsible for driving STING signalling and subsequent downstream immune activation. Mechanistically, we provide evidence that STING activation may result from loss of genomic integrity, reflected by decreased DNA damage repair and increased RSRD. A parallel increase in apoptosis rate further helps fuel cGAS/STING signalling via increased availability of cytosolic DNA. In contrast, tumours with robust DNA repair machinery and intact RSR may fail to stimulate STING signalling, allowing for immune escape.

This perspective suggests a previously unidentified link between loss of genomic integrity and generation of tumour-specific T cell responses, as reflected by intratumoural Tex. We further showed that these events are common across many cancer types, indicating potential combination strategies that target DDR pathways⁸² and T cell activation may be effective across non-hypermutated tumours (such as CRC-MSS) that generally do not respond to immunotherapy. In contrast, we found that enhanced MMR pathway to maintain genomic stability in colorectal cancer may lead to the failure of STING activation. Our findings align with previous research that suggest mutant MLH1-induced defective MMR could drive the accumulation of tumour cytosolic DNA and subsequent STING activation.⁸³ In addition, we demonstrated that defective RSR also associates with increased STING signalling, similar to a prior study that observed increased RSRD led to accumulation of immunostimulatory tumour DNA. Our findings demonstrate an important relationship between tumour genomic integrity and host immunity, which also provides the rationale for combination therapy of STING agonists and ICBs. In tumours with cold TME and low TMB, STING agonists would be needed to first prime the T cell response, initiating the cancer–immunity cycle, before ICBs could be effective. For tumours with high TMB, our results indicate immune activation driven by an abundance of neoantigens may be less dependent on STING activation. This suggests that clinical response to STING agonists may be more substantial in patients with low-TMB tumours than those with high-TMB tumours.

While we show the relationship between STING signalling and loss of cellular integrity, including apoptosis signalling and other cell death pathways, it was unclear if these events associated with increased STING signalling is upstream or downstream of STING activation. Evidence for both have been reported. Mitochondrial DNA in apoptotic cells has been shown to elicit cGAS/STING signaling.⁸⁴ STING activation has been linked to promoting apoptosis⁸⁵ and cell death.⁸⁶ In addition, endoplasmic reticulum (ER) stress is known to directly stimulate STING and is also associated with apoptosis.^87,88 An improved understanding of how the interconnected events of PRR signalling, ER stress, and cell death govern inflammation and anti-tumour immunity is needed.

In summary, this study investigated the complex, non-linear relationship between TMB and accumulation of Tex within tumours. Currently, high TMB is approved by FDA as an indication for pembrolizumab in the treatment of unresectable or metastatic tumours.89, 90, 91 However, increasing evidence indicate a complex association between TMB and overall response rates to ICB immunotherapy.^92,93 Our results suggest that STING pathway related biomarkers may identify additional tumours with low/intermediate TMB that may respond to ICB, as the majority of do not carry high TMB tumours. Furthermore, we provide new evidence for an important interplay between genomic integrity and STING activation in leading to more inflamed tumours and amplifying tumour-specific T cells, which provides more mechanistic insights into the combined chemotherapy or radiation therapy and immunotherapy.

Contributors

JT: Methodology, Investigation, Visualization, Writing – original draft. CAE: Conceptualization, Methodology, Investigation, Data access and verification, Writing – review & editing. WG: Conceptualization, Methodology, Visualization, Data access and verification, Writing – review & editing. JMS: Conceptualization, Supervision, Writing – review & editing. PPL: Conceptualization, Supervision, Writing – review & editing. All authors read and approved the final version of the manuscript.

Data sharing statement

NanoString genomic data are deposited in GEO under accession number GSE190169. Single-cell sequencing data are deposited in GEO under the accession number GSE248288 (3’ sequence, n = 4) and GSE255107 (5’ sequence, n = 11). All scripts used in this publication are available in https://github.com/joycetan817/TMB_STING_pancancer.

Declaration of interests

All authors declare no conflict of interest.