TGFβ attenuates tumour response to PD-L1 blockade by contributing to exclusion of T cells

Abstract

Therapeutic antibodies that block the programmed death-1 (PD-1)-programmed death-ligand 1 (PD-L1) pathway can induce robust and durable responses in patients with various cancers, including metastatic urothelial cancer. However, these responses only occur in a subset of patients. Elucidating the determinants of response and resistance is key to improving outcomes and developing new treatment strategies. Here we examined tumours from a large cohort of patients with metastatic urothelial cancer who were treated with an anti-PD-L1 agent (atezolizumab) and identified major determinants of clinical outcome. Response to treatment was associated with CD8⁺ T-effector cell phenotype and, to an even greater extent, high neoantigen or tumour mutation burden. Lack of response was associated with a signature of transforming growth factor β (TGFβ) signalling in fibroblasts. This occurred particularly in patients with tumours, which showed exclusion of CD8⁺ T cells from the tumour parenchyma that were instead found in the fibroblast- and collagen-rich peritumoural stroma; a common phenotype among patients with metastatic urothelial cancer. Using a mouse model that recapitulates this immune-excluded phenotype, we found that therapeutic co-administration of TGFβ-blocking and anti-PD-L1 antibodies reduced TGFβ signalling in stromal cells, facilitated T-cell penetration into the centre of tumours, and provoked vigorous anti-tumour immunity and tumour regression. Integration of these three independent biological features provides the best basis for understanding patient outcome in this setting and suggests that TGFβ shapes the tumour microenvironment to restrain anti-tumour immunity by restricting T-cell infiltration.

Extended Data Figure 1. Molecular correlates of outcome and tumour mutation burden (TMB)

a, Overlap of the efficacy-evaluable patient populations with assays used in this study (n = 326 for one or more of these assays). For gene expression analyses with respect to response, the complete RNAseq data set was used (n = 298). For gene expression analyses in the context of TMB or immune phenotype, the intersect between RNAseq and FMOne (n = 237) or immune phenotype (n = 244) was used, respectively. For mutation analysis around immune phenotypes, the intersect between FMOne and immune phenotype was used (n = 220). For associations between response or genes mutation status with TMB, the complete FMOne data set was used (n = 251). b, PD-L1 protein expression on tumour cells (TC), in contrast to expression on immune cells (IC; Fig. 1a), was not associated with response to atezolizumab (two-sided Fisher exact test; p = 0.72). c–d, Transcriptional correlates of PD-L1 protein expression on IC. c, Genes associated with PD-L1 immunohistochemistry (IHC) positivity on IC. Normalized log₂-transformed gene expression was compared with PD-L1 IC protein expression. Interferon-stimulated genes and previously reported CD8 T_eff and immune checkpoint molecule gene sets^, were among the most up-regulated (complete list of associated genes given in Supplementary Table S10). d, Association between CD8 T_eff signature score and PD-L1 IC. There is a significant positive relationship between the signature score and PD-L1 IC staining (likelihood ratio test p = 4.2 × 10⁻³⁵). e–f, Tumour neoantigen burden (TNB) is associated with outcome. e, Box plots showing the relationship between response status and TNB. Shown are number of mutations based on whole-exome sequencing, filtering for those mutations that are predicted to be expressed neoantigens. TNB is positively associated with response to atezolizumab (two-tailed t test p = 2.7 × 10⁻⁹). f, TNB, split into quartiles, is also associated with overall survival (OS; likelihood ratio test p = 0.00015). g, KEGG pathways enriched in genes whose expression is correlated with TMB. Shown are adjusted −log₁₀ p values for enrichment of KEGG gene sets significantly (FDR < 0.05) enriched in genes that are correlated with TMB (272 samples analysed). Sets inferred to reflect key underlying biological processes are coloured. Only the top seven genes per set (ranked by single-gene p value) are shown. h, Relationship between different gene expression signatures as well as the single-gene expression values for MKI67, a marker for proliferation. In the left corner, correlation between signature scores/gene expression is visualized (348 samples analysed). In the right corner, Pearson correlation coefficients are given. Gene set membership is given in Supplementary Table S8. i–j, APOBEC3A and APOBEC3B gene expression and its association with response and tumour mutation burden (TMB). Both APOBEC3A (two-tailed t test p = 0.015; i) and APOBEC3B (two-tailed t test p = 0.0025; j) exhibit higher mean expression in responders. For TMB, Pearson correlation coefficients and p values are given. For j, the two extreme expression outliers were excluded when calculating correlation between gene expression and TMB. k, Mutations in cell cycle regulator genes are associated with TMB. Genes are plotted in rows and patients (n = 293) in columns, marking patients with a mutation with a black rectangle. The upper bar plot depicts TMB in each patient. The final rows represent the mutation status of the pathway with or without TP53. Percentages to the left of the plot indicate prevalence. Genes with significant single gene associations with TMB are marked by an asterisk. Mutations in cell cycle regulator genes are associated with TMB with inclusion of TP53 (two-tailed t test p = 4.01 × 10⁻⁸), but not without inclusion of TP53 (two-tailed t test p = 0.0652156; Supplementary Table S4). l, Mutation status in cell cycle regulation (CCR) pathway by response. For each patient, it is determined whether they harbour any mutation in a gene belonging to the CCR pathway, except for TP53. Excluding TP53, there is no association between mutation status for the CCR pathway and response (two-sided Fisher test p = 0.31104; Supplementary Table S4). Sample sizes given in parentheses. Pan-F-TBRS: pan-fibroblast TGF-β response signature. CR, complete response; PD, progressive disease; PR, partial response; SD, stable disease; TMB, tumour mutation burden.

Extended Data Figure 2. Pathways associated with response and cancer-immune phenotypes

a, KEGG pathways significantly associated with response to atezolizumab (adj. p < 0.10; comparing 68 responders to 230 non-responders). The top seven genes per set are shown; complete lists are given in Supplementary Table S6. b–c, IFNG (b) and IFNGR1 (c) gene expression (y-axis) is significantly associated with response (two-tailed t test p = 9.1 × 10⁻⁵) and non-response (two-tailed t test p = 0.00012), respectively. d–e, TGFBR2 gene expression (y-axis) is significantly associated with non-response (two-tailed t test p = 0.00019, d) and, when split by quartiles, with reduced overall survival (likelihood ratio test p = 0.022, e). b–d: The numbers above the graphs specify sample numbers in each bin. CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease. e: 87 samples per quartile. Q1: lowest quartile, Q4: highest quartile. f, Histology of tumour-immune phenotypes: desert, excluded, and inflamed. g, Explained variance in patient response. Generalized linear models were fit to all efficacy-evaluable, immune phenotyped samples (n = 204) using binary response (CR/PR vs SD/PD) as the dependent variable and scores from single input or input combinations (x-axis) as independent variables. Percent explained variance of response is plotted on the y-axis. Comparisons between different models were made via likelihood ratio test. Horizontal bars describe likelihood ratio test results for two-biology vs. three-biology model comparisons. Stacked significance symbols for two-biology models show results of likelihood ratio test comparison to the first single-biology model and separately to the second single-biology model, in the same order as given in bar label on x-axis. (E.g., The T_eff, TMB model achieves three stars when compared to the T_eff singleton, but an “n.s.” when compared to the TMB singleton—due to dilution of its inflamed-specific signal in this all-sample analysis.) A model that includes both DNA (TMB) and RNA markers (CD8+ T_eff signature and Pan-F-TBRS) as well as interactions between the Pan-F-TBRS and both TMB and cancer-immune phenotype explains 50% of the variance observed in response, and it significantly improves on all singleton and two-biology models. This final bar is also given on the far right in Fig. 2f. n.s. p > 0.1; ^. p < 0.1; * p < 0.05; ** p < 0.01 ; *** p < 0.001. Exact likelihood ratio test p values: “Teff,TBRS” versus “Teff”: 0.0026, “Teff,TBRS” versus “TBRS”: 0.0032, “Teff,TMB” versus “Teff”: 4.9 × 10⁻⁸, “Teff,TMB” versus “TMB”: 0.2, “TBRS,TMB” versus “TBRS”: 6.6 × 10⁻⁸, “TBRS,TMB” versus “TMB”: 0.014, “Teff,TBRS,TMB” versus “Teff,TBRS”: 1.9 × 10⁻⁷, “Teff,TBRS,TMB” versus “Teff,TMB”: 0.0028, “Teff,TBRS,TMB” versus “TBRS,TMB”: 0.016. Teff: CD8 T effector gene signature, TMB: tumour mutation burden, (Pan-F-)TBRS: pan-tissue fibroblast TGF-β response genes.

Extended Data Figure 3. Comparison between Lund and TCGA subtyping schemes

a, Heatmap representing all patients evaluated, except for patients without defined response, arranged in columns and sorted first by molecular subtype, then by response to atezolizumab. For the left-hand panel, patients were sorted based on a TCGA-based subtyping scheme, for the right-hand panel, patients were sorted by a Lund-based subtyping scheme (like Fig. 3a). Immune cell (IC) and tumour cell (TC) PD-L1 status are given. In addition, tumour mutation burden (TMB) and mutation status (either mutated, black, or not-mutated; grey: patients without mutation data) for a few genes of interest are shown. The rows of the heatmap show expression (z-scores) of genes of interest, grouped into the following biologies/pathways: TCGA: TCGA subtyping genes, A: FGFR3 gene signature, B: CD8 T-effector signature, C: antigen processing machinery, D: Immune Checkpoint signature, E: MKI67 and cell cycle genes, F: DNA replication-dependent histones, G: DNA damage repair genes, H: TGF-β receptor and ligand, I: Pan-F-TBRS genes, J: angiogenesis signature, K: epithelial-mesenchymal transition (EMT) markers (for details on these signatures see Methods). CR, complete response; GU, genomically unstable; Inf, infiltrated; PD, progressive disease; PR, partial response; SCCL, basal/SCC-like; SD, stable disease; UroA, urothelial-like A; UroB, urothelial-like B, Pan-F-TBRS: pan-tissue fibroblast TGF-β response genes. b, FGFR3-related and WNT target genes as well as PPARG are significantly differentially expressed by Lund subtype (Wald p FGFR3-related: 2.7 × 10⁻⁴³, WNT target: 1.3 × 10⁻¹⁵, PPARG: 1.2 × 10⁻⁵³). Gene set membership is given in Supplementary Table S8. c–d, Distribution of Lund subtypes by cancer-immune phenotypes and response status. c, The fraction of patients within the different Lund subtypes (y-axis) is plotted by tumour-immune phenotype. There is a significant difference in Lund subtype composition between cancer-immune phenotypes (Chi-squared test p = 1.2 × 10⁻⁷). d, For excluded tumours, the fraction of patients within the different Lund subtypes (y-axis) is plotted by response status [responders: complete and partial response (CR/PR), non-responders: stable and progressive disease (SD/PD)]. There is a significant difference in Lund subtype composition between response groups (Chi-squared test p = 0.00061). The numbers above the graphs specify sample numbers in each bin. e, Assessment of MKI67 expression and signatures of interest as well as TMB relative to molecular subtypes. Biologies of interest were scaled before medians were calculated across the Lund (left) and TCGA (right) molecular subtypes (columns). Red means high, blue means low. CR, complete response; DNA rep., DNA replication; GU, genomically unstable; Inf, infiltrated; PD, progressive disease; PR, partial response; SCCL, basal/SCC-like; SD, stable disease; TMB, tumour mutation burden; UroA, urothelial-like A; UroB, urothelial-like B, Pan-F-TBRS: pan tissue fibroblast TGF-β response signature.

Extended Data Figure 4. Contrasting Lund and TCGA molecular subtyping

a, Tumour mutation burden (TMB; y-axis) is plotted against Lund and TCGA subtypes (x-axis). The Lund genomically unstable (two-tailed t test; p = 0.00018) and TCGA luminal II subtypes (p = 0.00024) have a higher median TMB. b, Patients are split into TMB low (grey) and high (black), based on median TMB, and the fraction of patients in these two groups is shown for the Lund and TCGA molecular subtypes. c–e, TGF-β as likely driver of differential response in Lund GU. c, Three patient subgroups: Lund GU but not TCGA luminal II, both GU and luminal II, or luminal II but not GU. d, CD8+ T_eff, Pan-F-TBRS and TMB by subgroup. e, Response differs significantly by subgroups (two-tailed Fisher exact p = 0.00062). The numbers above the graphs or in parentheses specify sample numbers in each bin. GU, genomically unstable; Inf, infiltrated; Pan-F-TBRS: pan-tissue fibroblast TGF-β response genes; SCCL, basal/SCC-like; T_eff: CD8 T effector signature; TMB: tumour mutation burden; UroA, urothelial-like A; UroB, urothelial-like B.

Extended Data Figure 5. Efficacy data of anti–TGF-β + anti–PD-L1 treatment in EMT6 and MC38 immune excluded tumour models

a, Fibroblast (PDGFRa, left panel) and T cell (CD3, right panel) parametric maps. Left image shows PDGFRa density (% positive pixels) and right shows T-cell density (cells/mm²). Scale bar: 1 millimetre. Representative images of eight biological replicates. b, Collagen (green) and T cells (CD3, red) stained by immunofluorescence. Representative images of five biological replicates. c, Collagen (green), T cells (CD3, white) and PDGFRa (red) in EMT6 tumours stained by immunofluorescence. Scale bar: 200 microns. d, PDGFRa (red) in EMT6 tumours stained by immunofluorescence. Scale bar: 200 microns. Representative images of four biological replicates. e, Quantification of TGF-β and PD-L1 RNA in whole EMT6 tumours by RNAseq. The tumours were inoculated orthotopically and collected when volume reached 300 mm³ (n = 5 mice; data from one experiment). f, Quantification of TGF-β protein within whole EMT6 tumours by ELISA. Tumours were collected 14 days after inoculation, flash frozen and lysed for protein quantification (n = 4 mice; data from one experiment). g, Balbc mice were inoculated with EMT6 tumour cells orthotopically. When tumour volumes reached ≈ 160 mm³ approximately nine days after inoculation, mice were treated with isotype control, anti–PD-L1, anti–TGF-β, or a combination of anti–PD-L1 with anti–TGF-β. Tumours were measured two times per week for approximately eight weeks by calliper. When tumour volumes fell below 32 mm³ (lowest limit of detection), they were considered complete response. Percentage of complete regressions across 2–6 independent studies (10 mice/group per study). h, Tumour weights at day 7 after initiation of treatment (n = 28 mice per treatment group; data from three independent experiments). i, CD8 depletion experiment. CD8 T cells were depleted before initiation of treatment. j, Quantification of TGF-β and PD-L1 RNA in whole MC38 tumours by RNAseq. The tumours were inoculated subcutaneously and collected when volume reached 300 mm³ (n = 5 mice; data from one experiment). k, Quantification of CD8 T cells in the centre and in the periphery of EMT6 and MC38 tumours from IHC stains. Data expressed as number of cells per tissue area (periphery is defined as 400–600 micron from the tumour edge, centre is the remaining distance to centre point). EMT6: n = 5 mice, MC38: n = 4 mice. l, Collagen (green) and T cells (CD3, red) in MC38 tumours stained by immunofluorescence. Scale bar, 1mm (left); 0.1mm (right). m, C57BL6 mice were inoculated with MC38 tumour cells subcutaneously. When tumour volumes reached ≈ 180 mm³ approximately eight days after inoculation, mice were treated with isotype control, anti–PD-L1, anti–TGF-β, or a combination of anti–PD-L1 with anti–TGF-β. Tumours were measured two times per week for approximately eight weeks by calliper. When tumour volumes fell below 32 mm³ (lowest limit of detection), they were considered complete response. Percentage of CR across 2 independent studies (1 for anti–TGF-β alone) shown with 10 mice per treatment group for each independent study. n, Tumour growth curves for each individual mouse are shown. The data are from one representative of two independent experiments with 10 mice per treatment group. All statistics are two-sided Mann-Whitney test compared to isotype group. * p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001.

Extended Data Figure 6. Changes in TME following anti–TGF-β + anti–PD-L1 treatment in EMT6 tumours

a, c, d, Cytofluorimetric analysis of T cells seven days after initiation of the treatment. The abundance of total T cells (a), total CD4+ cells (c) and the percentage of T regulatory cells (CD25+FOXP3+) in the CD4+ population (d) are shown. n = 15 mice for all treatment groups except for anti-TGF-β alone in which n = 10, data combined from three independent experiments expressed as fold change relative to the isotype cell/mg average. b, e, RNAseq analysis on whole tumours collected seven days after the initiation of treatment. Single-gene expression for IFNg, GzmB and Zap70 (b) Helios and Foxp3 (e) are shown (n = 8 mice per treatment group; data from one experiment). f, Distribution of tumour-infiltrating lymphocytes in tumours (T) as assessed by immunohistochemistry and digital imaging seven days after the initiation of treatment as above. Representative CD3 staining (brown). Dashed line indicates tumour boundaries. (n = 19 for all groups except anti–PD-L1/anti–TGF-β, in which n = 20; three independent experiments). Scale bar, 500 micron. g, Phospho-SMAD2 quantification by IHC at day 7 after initiation of treatment. (n = 9 or 10 mice per treatment groups; data from one experiment). h, Phosphoflow analysis of SMAD2/3 in tumours seven days after the initiation of treatment as above. Mean fluorescence intensity (MFI) of phospho-SMAD2/3 among total cells, CD45- or CD45+ cells are shown. Data are expressed as fold change (FC) relative to the isotype MFI average. 10 mice per treatment group from two independent experiments. i–k, RNAseq analysis on whole tumours collected seven days after the initiation of treatment. Three EMT signatures (i) as well as TGF-β response signatures for T cells (j) and macrophages (k) are also shown. (n = 8 mice per treatment group; data from one experiment). All statistics in the figure are two-sided Mann-Whitney test. * p < 0.05; ** p < 0.01; *** p < 0.001; ns, not significant compared to Isotype group.

Extended Data Figure 7. Explained variance in patient response

Generalized linear models were fit using binary response (complete or partial response vs. stable or progressive disease) as the dependent variable and scores from single input or input combinations (x-axis) as independent variables (number of samples: 236). Percent explained variance of response is plotted on the y-axis. Comparisons between different models were made via likelihood ratio test; a significant p value means that the additional variable contributed some independent information to the model. TMB’s association with response is significantly stronger than that of its proxy measurements (APOBEC3B and MKI67 expression or mutation in members of the DDR set). APOBEC3B and DDR gene set mutation provided no additional explanatory information independent of direct measurement of TMB. Combining TMB with MKI67 expression did marginally improve on TMB alone, possibly through MKI67’s negative association with TFG-β (Extended Data Fig. 1e,f). To test this hypothesis, we added MKI67 to a two-pathway model based on TMB and the Pan-F-TBRS, and confirmed that MKI67 does not add independent information to this two-pathway model. Further, there was no benefit from adding MKI67 to our full three-pathway model, shown in Fig. 2f and Extended Data Fig. 2g. DDR: DNA damage repair; TBRS: pan-tissue fibroblast TGF-β response genes; TMB, tumour mutation burden. n.s. p > 0.1; ^. p < 0.1; * p < 0.05. Exact likelihood ratio test p values: “TMB,DDR” versus “TMB”: 0.38, “TMB,APOBEC3B” versus “TMB”: 0.26, “TMB,MKI67” versus “TMB”: 0.029, “TMB,TBRS,MKI67” versus “TMB,TBRS”: 0.064.

Extended Data Figure 8. Relationship between different TGF-β related gene expression signatures and response

a, Correlation between different TGF-β related gene expression signatures. In the left corner, correlation between signature scores/gene expression is visualized, calculated based on the complete RNA sequencing data set of 348 samples. In the right corner, Pearson correlation coefficients are given. Gene set membership is given in Supplementary Table S8. See Methods for computation of signature scores. b, EMT signature expression is associated with response to atezolizumab in excluded tumours. Scores of three different EMT signatures—EMT1, EMT2 and EMT3—are significantly higher in non-responders (stable and progressive disease; SD/PD) than responders (complete and partial response; CR/PR) in excluded tumours (EMT1: p = 0.0102; EMT2: p = 0.0027: EMT3: p = 0.0063): there is no significant difference in signature scores in desert and inflamed tumours (all p = 1; two-tailed t test p values for each signature are Bonferroni corrected for 3 tests). The numbers above the graphs specify sample numbers in each bin. c, Explained variance in patient response. Generalized linear models were fit using binary response (complete or partial response vs. stable or progressive disease) as the dependent variable and scores from single input or input combinations (x-axis) as independent variables (number of samples: 233). Percent explained variance of response is plotted on the y-axis. Comparisons between different models were made via likelihood ratio test; a significant p value means that the additional variable contributed some independent information to the model. The Pan-F-TBRS’s association with response is the strongest among its correlates, i.e., three different epithelial-to-mesenchymal transition (EMT) signatures. None of these signatures provided additional explanatory information independent of Pan-F-TBRS. (Pan-F-)TBRS: pan tissue fibroblast TGF-β response signature, EMT: epithelial-to-mesenchymal transition.

Figure 1. Three core biological pathways are associated with response to atezolizumab

a, PD-L1 IC was associated with response (two-sided Fisher exact test p = 0.0038). IC2+ tumours had a significantly higher CR rate (p = 0.0006). b, CD8+ T_eff signature score is positively associated with response (two-tailed t test p = 0.0087), with association driven by the CR group (CR vs PR p = 0.0388, CR vs SD p = 0.0668, CR vs PD p = 0.0003). c, CD8+ T_eff signature quartiles (Q1: lowe are significantly associated with overall survival (likelihood ratio test p = 0.0092). d–e, TMB is positively associated with response to atezolizumab (two-tailed t test p = 6.9 × 10⁻⁷) and overall survival (likelihood ratio test p = 2.0 × 10⁻⁵). A similar plot for tumour neoantigen burden is given in Extended Data Fig. 1e,f. f–g, There is a significant association between DDR mutation status and (f) response (two-sided Fisher exact test p = 0.0117 excluding TP53) and (g) TMB, both with (two-sided Fisher exact test p = 6.01 × 10⁻⁸) and without inclusion of TP53 (p = 1.95 × 10⁻⁵) h–i, TGFB1 gene expression is significantly associated with non-response (two-tailed t test p = 0.00011) and reduced overall survival (likelihood ratio test p = 0.0096). j, The relationship between response and three core biological pathways. ^. p < 0.10; * p < 0.05; ** p < 0.01; *** p < 0.001. Sample sizes given in parentheses. Q1: lowest quartile, Q4: highest quartile.

Figure 2. TGF-β is associated with lack of response in the excluded tumour-immune phenotype

a, Combined CD8 IHC–trichrome stain. CD8+ T cells (DAB, brown) are primarily within collagenous stroma (blue). Rare CD8+ T cells are interspersed between tumour cells (green arrows). b, Distribution of tumour-immune phenotypes in IMvigor210. c, CD8+ T_eff score is significantly associated with response only in the inflamed phenotype (two-tailed t test p = 0.042). d, Pan-F-TBRS is significantly associated with response only in the excluded phenotype (two-tailed t test p = 0.0066). e, TMB is significantly associated with response in the excluded and inflamed phenotypes (two-tailed t test p = 0.00009 and 0.00258). f, Explained variance in patient response for generalized linear models fit using single core pathways or pathway combinations. In immune desert tumours (n = 57), core pathways showed negligible explanatory power. For excluded tumours (n = 91), Pan-F-TBRS and TMB were associated with response as singletons; combining the two provided statistically significant improvement over single terms (likelihood ratio p = 6.23 × 10⁻⁵ and 0.02125, respectively). For inflamed tumours (n = 56), CD8+ T_eff and TMB were associated with response as singletons; combining the two provided statistically significant improvement over single terms (likelihood ratio p = 0.00099 and 0.0557, respectively). Standard forward selection applied to all patients (grey) identified a three-pathway model as significantly better than all single- or two-pathway models (Extended Data Fig. 2g). n.s., nonsignificant; * p < 0.05; ** p < 0.01; *** p < 0.001. Sample sizes given in parentheses.

Figure 3. Relationship between molecular subtypes (Lund scheme) and core biological pathways

Rows of the heat map show gene expression (z-scores) grouped by pathway. Inset: response vs. Lund molecular subtype, showing that the genomically unstable (GU) subtype has a significantly higher response rate (two-sided Fisher exact test p = 1.6 × 10⁻⁵). Sample sizes given in parentheses.

Figure 4. Tumour regression and changes in TME following therapeutic anti–TGF-β + anti–PD-L1 treatment in EMT6 tumours

a, Tumour growth curves. b, Change in tumour volume compared to baseline. (6 independent experiments, 10 mice per group.) c–d, Fold change (FC) in total CD8+ T cell abundance (c) and CD8+ T cell GRANZYME B MFI by flow cytometry (d). (3 independent experiments, n = 15 for all groups except anti–TGF-β, where n = 10.) e, CD8+ T_eff signature (1 experiment, n = 8 per group). f, TIL localisation quantification by immunohistochemistry (3 independent experiments, n = 19 for all groups except anti–PD-L1/anti–TGF-β, where n = 20; Tukey HSD multiple comparison adjustment). g–h, Representative CD3 staining of tumour periphery (g) and centre (h). Scale bar is 200 microns. i–l, Pan-F-TBRS signature (i) and expression of fibroblast genes (j–l). (1 experiment, n = 8 per group.) All data shown in c–l from tumours collected at day 7 after treatment initiation. All p values based on two-sided Mann-Whitney tests unless otherwise indicated. * p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001.