Potentially curative treatments for HCC, such as liver transplantation, tumour resection, or ablation, are limited to early-stage tumours1,2. Multikinase inhibitors and anti-VEGF-R2 antibodies have been approved for use in advanced HCC1,2. Immunotherapy, which is thought to activate T cells or reinvigorate immune surveillance against cancer, showed response rates of 15–30% in patients with HCC5,8,9,10,11. Nivolumab and pembrolizumab (PD1-directed antibodies) have been approved for treatment of HCC3,4, although phase III trials failed to reach their primary endpoints to increase survival1,10,11. A combination of atezolizumab (anti-PDL1) and bevacizumab (anti-VEGF) demonstrated increased overall and progression-free survival in a phase III trial, making it a first-line treatment for advanced HCC5. The efficacy of immunotherapy might be affected by different underlying HCC aetiologies, with diverse hepatic environments distinctly regulating HCC induction and immune responses6. Hence, we lack biomarkers that correlate with treatment response to allow patient stratification12,13. Non-alcoholic fatty liver disease (NAFLD) is an HCC-causing condition that affects more than 200 million people worldwide14. Approximately 10–20% of individuals with NAFLD progress over time from steatosis to NASH14. Innate and adaptive immune-cell activation15,16,17, in combination with increased metabolites and endoplasmic reticulum stress16,18, are believed to lead to a cycle of hepatic necro-inflammation and regeneration that potentially leads to HCC19,20,21. NASH has become an emerging risk factor for HCC1,14,19, which led us to investigate the effects of immunotherapy in NASH–HCC22,23,24.

Hepatic CD8+PD1+ T cells increase in NASH

We fed mice with diets that cause progressive liver damage and NASH over 3–12 months (Extended Data Fig. 1a–c), accompanied by an increase in the frequency of activated CD8+ T cells expressing CD69, CD44 and PD1 (Extended Data Fig. 1d–g). Single-cell mapping of leukocytes showed altered immune-cell compositions in mice with NASH (Extended Data Fig. 1h, i) with strongly increased numbers of CD8+PD1+ cells (Fig. 1a, b, Extended Data Fig. 1j–m, o). Similarly, elevated CD8+ and PD1+ cells were found in a genetic mouse model of NASH17 (Extended Data Fig. 1n). Messenger RNA in situ hybridization and immunohistochemistry showed that increasing PDL1 expression in hepatocytes and non-parenchymal cells correlated with the severity of NASH (Extended Data Fig. 1p). Mass spectrometric characterization of CD8+PD1+ T cells from NASH-affected livers indicated enrichment in pathways involved in ongoing T cell activation and differentiation, TNF signalling, and natural killer (NK) cell-like cytotoxicity (Fig. 1c). Single-cell RNA sequencing (scRNA-seq) of cells expressing T cell receptor β-chains (TCRβ) from the livers of mice with NASH showed that CD8+ T cells had gene expression profiles related to cytotoxicity and effector-function (for example, Gzmk and Gzmm) and inflammation markers (for example, Ccl3) with elevated exhaustion traits (for example, Pdcd1 and Tox) (Fig. 1d, e). RNA-velocity analyses demonstrated enhanced transcriptional activity and differentiation from Sell-expressing CD8+ to CD8+PD1+ T cells (Extended Data Fig. 1q), indicating local differentiation. Thus, mice with NASH have increased hepatic abundance of CD8+PD1+ T cells with features of exhaustion and effector functions.

Fig. 1: NASH progression is associated with increased, activated CD8+PD1+ T cells.
figure 1

a, CD8 and PD1 staining (right) and quantification (left) of T cells from mice fed normal diet (ND), CD-HFD or western-style diet with trans fat (WD-HTF) by immunohistochemistry. b, Immunofluorescence-based detection of PD1 (green), CD8 (red) and CD4 (yellow) cells. Scale bar, 100 μm. c, GSEA of hepatic CD8+PD1+ T cells sorted from TCRβ+ cells by mass spectrometry. FDR, false discovery rate; NES, normalized enrichment score. df, t-distributed stochastic neighbour embedding (tSNE) of TCRβ+ cells (d), differential gene expression by scRNA-seq (e) and scheme for experiment (f): mice were fed CD-HFD for 13 months and then treated with anti-PD1 for 8 weeks before measurement of tumour incidence. Mem, memory CD4 T cells. g, Livers from treated and untreated mice after CD-HFD. Arrowheads, tumours or lesions. Scale bar, 10 mm. h, Quantification of CD8+ cell in liver by immunohistochemistry. Details of sample sizes, biological replicates and statistical tests are given in Methods and Source Data. a, h, P values shown above brackets.

Source data

The high numbers of T cells in NASH suggest that anti-PD1-targeted immunotherapy may serve as an efficient therapy for NASH–HCC. Thirty per cent of C57BL/6 mice fed a choline-deficient high-fat diet (CD-HFD) for 13 months developed liver tumours with a similar load of genetic alterations to human NAFLD–HCC or NASH–HCC (Extended Data Fig. 2a, b). NASH mice bearing HCC (identified using MRI) were allocated to anti-PD1 immunotherapy or control arms (Fig. 1f). None of the pre-existing liver tumours regressed in response to anti-PD1 therapy (Fig. 1g, h, Extended Data Fig. 2c). Rather, we observed increased fibrosis, unchanged liver damage, slightly increased incidence of liver cancer and unaltered tumour loads and sizes after anti-PD1 treatment (Extended Data Fig. 2 d–h). In anti-PD1-treated mice, liver tumour tissue contained increased numbers of CD8+/PD1+ T cells and high levels of cells expressing Cxcr6 or Tnf mRNA (Extended Data Fig. 2i–n). We found no regression of NASH-induced liver tumours upon anti-PDL1 immunotherapy (Extended Data Fig. 3a–f). By contrast, other (non-NASH) mouse models of liver cancer (with or without concomitant damage) reacted to PD1 immunotherapy with tumour regression25, suggesting that lack of response to immunotherapy was associated specifically with NASH–HCC (Extended Data Fig. 3g–i). Thus, NASH precluded efficient anti-tumour surveillance in the context of HCC immunotherapy. Similarly, impaired immunotherapy has been described in mouse models with NASH and secondary liver cancer25,26.

CD8+ T cells promote HCC in NASH

As CD8+PD1+ T cells failed to execute effective immune surveillance, but rather showed tissue-damaging potential, we reasoned that CD8+ T cells might be involved in promoting NASH–HCC. We depleted CD8+ T cells in a preventive setting in mice with NASH but without liver cancer (CD-HFD fed for 10 months). CD8+ T cell depletion significantly decreased liver damage and the incidence of HCC in these mice (Fig. 2i, Extended Data Fig. 4a–j, n). Similar results were obtained after co-depletion of CD8+ and NK1.1+ cells (Fig. 2i, Extended Data Fig. 4a–f, n). This suggests that as well as lacking immune surveillance functions, liver CD8+ T cells also promote HCC in mice with NASH. Next, we investigated the effect of anti-PD1 therapy on HCC development in mice with NASH. Anti-PD1 immunotherapy aggravated liver damage (Fig. 2g, Extended Data Fig. 7c) and increased hepatic CD8+PD1+ T cells, with only minor changes in liver CD4+PD1+ T cells or other immune-cell populations (Extended Data Fig. 4a–o). Anti-PD1 immunotherapy also caused a marked increase in liver-cancer incidence, independent of changes in liver fibrosis (Fig. 2i). Mice lacking PD1 (Pdcd1−/−) showed an increase in incidence of, and earlier onset of, liver cancer, along with increased liver damage and elevated numbers of activated hepatic CD8+ T cells with increased cytokine expression (IFNγ, TNF) (Extended Data Fig. 5a–g). In summary, CD8+PD1+ T cells triggered the transition to HCC in mice with NASH, probably owing to impaired tumour surveillance and enhanced T cell-mediated tissue damage27. Despite a strong increase in CD8+PD1+ T cells within tumours, therapeutic PD1- or PDL1-related immunotherapy failed to cause tumour regression in NASH–HCC.

Fig. 2: Resident-like CD8+PD1+ T cells drive hepatocarcinogenesis in a TNF-dependent manner upon anti-PD1 treatment in NASH.
figure 2

a, b, RNA-velocity analyses of scRNA-seq data showing expression (a) and correlation of expression (b) along the latent time of selected genes in CD8+ T cells from mice with NASH. Latent time (pseudo-time by RNA velocity): dark colour, start of RNA velocity; yellow, end point of latent time. Kendall’s τ, gene expression along latent time. c, Principal component analysis (PCA) plot of hepatic CD8+ and CD8+PD1+ T cells sorted by mass spectrometry from TCRβ+ cells from mice fed for 12 months with ND, CD-HFD or CD-HFD and treated for 8 weeks with anti-PD1 antibodies. d, e, UMAP representations showing FlowSOM-guided clustering (d, left), heat map showing median marker expression (d, right), and quantification of hepatic CD8+ T cells (e) from mice fed for 12 months with ND or CD-HFD and treated for 8 weeks with IgG or anti-PD1 antibodies. f, Quantification of CellCNN-analysed flow cytometry data for hepatic CD8+ T cells from mice fed for 12 months with CD-HFD and treated for 8 weeks with IgG or anti-PD1 antibodies. g, h, NAS evaluation (g) and quantification of hepatic CD8+PD1+CXCR6+ T cells (h) from mice fed with ND for 12 months or fed with CD-HFD for 12 months and treated for 8 weeks with anti-PD1, anti-PD1 + anti-CD8, anti-TNF, anti-PD1 + anti-TNF, anti-CD4, or anti-PD1 + anti-CD4 antibodies. Kendall’s τ, gene expression along latent time. i, Quantification of tumour incidence in mice as in g, h. Details of sample sizes, biological replicates and statistical tests are given in Methods and Source Data. ei, P values shown above brackets.

Source data

We used an immune-mediated cancer field (ICF) gene-expression signature associated with the development of human HCC28 to understand the tumour-driving mechanisms of anti-PD1 immunotherapy. Preventive anti-PD1 treatment was strongly associated with the pro-tumorigenic immunosuppressive ICF signature (for example, Ifng, Tnf, Stat3, Tgfb1), capturing the traits of T cell exhaustion, pro-carcinogenic signalling, and mediators of immune tolerance and inhibition. Depletion of CD8+ T cells led to significant downregulation of the high-infiltrate ICF signature and diminished TNF in non-parenchymal cells (Extended Data Fig. 5h, i). Gene set enrichment analysis (GSEA), mRNA in situ hybridization, and histology of tumours developed in NASH mice that were treated prophylactically with anti-PD1 corroborated these data, showing increased CD8+ T cell abundance and enrichment for genes involved in inflammation-related signalling, apoptosis, and TGFβ signalling (Extended Data Fig. 5j–l). Anti-PD1 treatment triggered the expression of p62 (Extended Data Fig. 5m), which has been shown to drive hepatocarcinogenesis29. Array comparative genomic hybridization identified no significant differences in chromosomal deletions or amplifications between tumours from anti-PD1-treated mice or control mice (Extended Data Fig. 5n). In summary, hepatic CD8+PD1+ T cells did not cause tumour regression during NASH, but rather were linked to HCC development, which was enhanced by anti-PD1 immunotherapy.

We next analysed the hepatic T cell compartment for correlations with inflammation and hepatocarcinogenesis. Comparison of CD8+PD1+ T cells with CD8+ T cells by scRNA-seq showed that the former showed higher expression of genes associated with effector function (for example, increased Gzma, Gzmb, Gzmk, Prf1; reduced Sell, Klf2), exhaustion (for example, increased Pdcd1, Tox; reduced Il7r, Tcf7) and tissue residency (for example, increased Cxcr6, low levels of Ki-67) (Extended Data Fig. 6a–c). Notably, there was no difference in the transcriptome profiles of CD8+PD1+ T cells in NASH mice after anti-PD1 immunotherapy (Extended Data Fig. 6c), indicating that the number of T cells rather than their functional properties were changed. RNA-velocity blot analyses corroborated these data (Fig. 2a, Extended Data Fig. 6d–f). Similar patterns of markers (for example, IL7r, Sell, Tcf7, Ccl5, Pdcd1, Cxcr6, and Rgs1) correlated with latent time and overall transcriptional activity in NASH mice that received either treatment (Fig. 2a, b, Extended Data Fig. 6e, f). Mass spectrometry-based analyses of CD8+ or CD8+PD1+ T cells isolated from NASH mouse livers confirmed these findings (Fig. 2c, Extended Data Fig. 6g).

We characterized the transcriptome profiles of PD1+CD8+ T cells by uniform manifold approximation and projection (UMAP) analysis of high-parametric flow-cytometry data, dissecting the CD8+PD1+ and CD8+PD1 subsets (Fig. 2d). This revealed that CD8+PD1+ cells expressed high levels of effector (for example, Gzmb, Ifng, Tnf) and exhaustion markers (for example, Eomes, Pdcd1, Ki-67low). In particular, CD8+PD1+TNF+ cells were more abundant upon anti-PD1 treatment (Fig. 2e). Convolutional neural network analysis and manual gating validated this result (Fig. 2f, Extended Data Fig. 6j, k). CD8+PD1+ T cells were non-proliferative in anti-PD1-treated NASH mice; this result was supported by in vitro experiments, in which anti-PD1 treatment led to increased T cell numbers in the absence of proliferation (Extended Data Fig. 6l, m). Notably, CD8+PD1+ T cells from NASH mice showed reduced levels of FOXO1, which indicates an enhanced tissue-residency phenotype30, potentially combined with boosted effector function, as indicated by higher calcium levels in CD8+PD1+ T cells (Extended Data Fig. 6n, o). Single-cell RNA-seq analysis also showed that CD8+PD1+ T cells from NASH mice had a tissue residency signature (Extended Data Fig. 6b). Thus, upon anti-PD1 immunotherapy in NASH mice, CD8+PD1+ T cells accumulated to high numbers in the liver, revealing a resident-like T cell character with increased expression of CD44, CXCR6, EOMES and TOX and low levels of CD244 expression, but lacking expression of TCF1/TCF7, CD62L, TBET, and CD127 (Extended Data Fig. 6p–u). In summary, anti-PD1 immunotherapy increased the abundance of CD8+PD1+ T cells with a residency signature in the liver.

To investigate the mechanisms that drive the increased NASH–HCC transition in the preventive anti-PD1 treatment-setting, we treated NASH-affected mice with combinations of treatments. Both anti-CD8–anti-PD1 and anti-TNF–anti-PD1 antibody treatments ameliorated liver damage, liver pathology and liver inflammation (Fig. 2g, Extended Data Fig. 7), and decreased the incidence of liver cancer compared to anti-PD1 treatment alone (Fig. 2i). By contrast, anti-CD4–anti-PD1 treatment did not reduce the incidence of liver cancer, the NAFLD activity score (NAS), or the number of TNF-expressing hepatic CD8+ or CD8+PD1+CXCR6+ T cells (Fig. 2g–i, Extended Data Fig. 7). However, both the number of tumours per liver and tumour size were reduced, suggesting that depletion of CD4+ T cells or regulatory T cells might contribute to tumour control (Extended Data Fig. 8a, b). The incidence of tumours was directly correlated with anti-PD1 treatment, alanine aminotransferase (ALT), NAS, number of hepatic CD8+PD1+ T cells, and TNF expression (Extended Data Fig. 8c–e). These data suggested that CD8+PD1+ T cells lacked immune-surveillance and had tissue-damaging functions27, which were increased by anti-PD1 treatment, possibly contributing to the unfavourable effects of anti-PD1 treatment on HCC development in NASH.

Augmented CD8+PD1+ T cells in human-NASH

We next investigated CD8+ T cells from healthy or NAFLD/NASH-affected livers. In two independent cohorts of patients with NASH, we found enrichment of hepatic CD8+PD1+ T cells with a residency phenotype (by flow cytometry and mass cytometry) (Fig. 3a, b, Extended Data Fig. 9a–j, Supplementary Tables 1,2). The number of hepatic CD8+PD1+ T cells directly correlated with body-mass index and liver damage (Extended Data Fig. 9b). To investigate similarities between mouse and human T cells from livers with NASH, we analysed liver CD8+PD1+ T cells from patients with NAFLD or NASH by scRNA-seq. This identified a gene expression signature that was also found in liver T cells from NASH mice (for example, PDCD1, GZMB, TOX, CXCR6, RGS1, SELL) (Fig. 3c, d, Extended Data Fig. 9k, l). Differentially expressed genes were directly correlated between patient- and mouse-derived hepatic CD8+PD1+ T cells (Fig. 3d). Velocity-blot analyses identified CD8+ T cells expressing TCF7, SELL and IL7R as root cells, and CD8+PD1+ T cells as their endpoints (Fig. 3e, f), indicating a local developmental trajectory of CD8+ T cells into CD8+PD1+ T cells. The amount of gene expression and velocity magnitude, which indicate transcriptional activity, were increased in CD8+PD1+ T cells from mice and humans with NASH (Fig. 3e). The expression of specific marker genes (for example, IL7R, SELL, TCF7, CCL5, CCL3, PDCD1, CXCR6, RGS1 and KLF2) along the latent time in patients with NAFLD or NASH differed from that seen in control participants (Fig. 3g), and correlated with the expression patterns seen in CD8+ T cells from NASH mice (Fig. 3h). Thus, scRNA-seq analysis demonstrated a resident-like liver CD8+PD1+ T cell population in patients with NAFLD or NASH that shared gene expression patterns with hepatic CD8+PD1+ T cells from NASH mice.

Fig. 3: Hepatic resident-like CD8+PD1+ T cells are increased in livers of patients with NAFLD patients.
figure 3

a, b, UMAP representation showing the FlowSOM-guided clustering of CD45+ cells (a) and flow cytometry plots (b, left) and quantification (b, right) of CD8+PD1+CD103+ cells derived from hepatic biopsies of healthy individuals or patients with NAFLD or NASH (Supplementary Table 2). Populations in b: violet, CD8+; red, CD8+PD1+CD103+. Treg cells, regulatory T cells. c, UMAP representations and analyses of differential gene expression by scRNA-seq of CD3+ cells from control individuals or patients with NAFLD or NASH. MAITs, mucosal-associated invariant T cells. d, Correlation of significant differentially expressed genes in liver-derived CD8+PD1+ T cells compared to CD8+PD1 T cells from mice fed with CD-HFD for 12 months and patients with NAFLD/NASH. Shading shows 95% CI. eh, Expression (e) and transcriptional activity (f) of velocity analyses of scRNA-seq data, and gene expression (g) and correlation (h) of expression along the latent-time of selected genes along the latent-time of liver-derived CD8+ T cells from patients with NAFLD or NASH in comparison to control or NASH mouse liver-derived CD8+ T cells. Root cells: yellow, root cells; blue, cells furthest from the root by RNA velocity. End points: yellow, end-point cells; blue, cells furthest from defined end-point cells by RNA velocity. Latent time (pseudo-time by RNA velocity): dark colour, start of RNA velocity; yellow, end point of latent time. RNA velocity flow (top): blue cluster, start point; orange cluster, intermediate; green cluster, end point. Arrow indicates cell trajectory. Details of sample sizes, biological replicates and statistical tests are given in Methods and Source Data. b, e, P values shown above brackets.

Source data

Different stages of NASH severity are considered to herald the development of liver cancer31. Indeed, different fibrosis stages (F0–F4) in patients with NASH correlated directly with the expression of PDCD1, CCL2, IP10 and TNF, and the degree of fibrosis correlated with the numbers of CD4+, PD1+, and CD8+ T cells (Extended Data Fig. 10a–d, Supplementary Table 3). Moreover, PD1+ cells were absent from healthy livers but present in the livers of patients with NASH or NASH–HCC, but the number of these cells did not differ with the underlying fibrosis level (Extended Data Fig. 10e, Supplementary Tables 46). Species-specific effects, such as the absence in mice of cirrhosis or burnt-out NASH (a condition found in some patients with NASH–HCC32), and their possible influence on immunotherapy may make it difficult to translate findings from preclinical models of NASH to human NASH. However, in tumour tissue from patients with NASH-induced HCC—treated with anti-PD1 therapy—we found increased numbers of intra-tumoral PD1+ cells compared to patients with HCC and viral hepatitis (Extended Data Fig. 10f). Thus, we found a shared gene-expression profile and increased abundance of unconventionally activated hepatic CD8+PD1+ T cells in human NASH tissue.

Lack of immunotherapy response in human NASH–HCC

To explore the concept of disrupted immune surveillance in NASH after anti-PD1 or anti-PDL1 treatment, we conducted a meta-analysis of three large randomized controlled phase III trials of immunotherapies in patients with advanced HCC (CheckMate-45911, IMbrave1505 and KEYNOTE-24010). Although immunotherapy improved survival in the overall population (hazard ratio (HR) 0.77; 95% confidence interval (CI) 0.63–0.94), survival was superior to the control arm in patients with HBV-related HCC (n = 574; P = 0.0008) and HCV-related HCC (n = 345; P = 0.04), but not in patients with non-viral HCC (n = 737; P = 0.39) (Fig. 4a, Extended Data Fig. 10g, Supplementary Table 7). Patients with viral aetiology (HBV or HCV infection) of liver damage and HCC showed a benefit from checkpoint inhibition (HR 0.64; 95% CI 0.48–0.94), whereas patients with HCC of a non-viral aetiology did not (HR 0.92; 95% CI 0.77–1.11; P of interaction = 0.03 (Fig. 4a)). Subgroup analysis of first-line treatment compared to a control arm treated with sorafenib (n = 1,243) confirmed that immunotherapy was superior in patients with HBV-related (n = 473; P = 0.03) or HCV-related HCC (n = 281; P = 0.03), but not in patients with non-viral HCC (n = 489; P = 0.62; Extended Data Fig. 10h–j). We acknowledge that these results were derived from a meta-analysis of trials that included different lines of treatment and patients with heterogeneous liver damage, and did not differentiate between alcoholic liver disease and NAFLD or NASH. Nevertheless, the results of this meta-analysis supported the notion that stratification of patients according to the aetiology of their liver damage and ensuing HCC identified patients who responded well to therapy.

Fig. 4: PD1 and PDL1 targeted immunotherapy in advanced HCC has a distinct effect depending on disease aetiology.
figure 4

a, Meta-analysis of 1,656 patients (Supplementary Table 7). Immunotherapy was initially assessed and then analysed according to disease aetiology: non-viral (NASH and alcohol intake) vs viral (HBV and HCV) (top). Heterogeneity: τ2 = 0.00; χ2 = 0.14, degrees of freedom (d.f.) = 2 (P = 0.93); I2 = 0%. Test for overall effect: Z = 0.87 (P = 0.39). Separate meta-analyses were subsequently performed for each of the three aetiologies: non-viral (NASH and alcohol intake), HCV and HBV (bottom). Heterogeneity: τ2 = 0.03; χ2 = 3.67, d.f. = 2 (P = 0.16); I2 = 46%. Test for overall effect: Z = 3.13 (P = 0.002). Diamonds represent estimated overall effect based on the meta-analysis random effect of all trials. Inverse variance and random effects methods were used to calculate HRs, 95% CIs, P values, and the test for overall effect; calculations were two-sided. b, NAFLD is associated with a worse outcome in patients with HCC treated with PD(L)1-targeted immunotherapy. A total of 130 patients with advanced HCC received PD(L)1-targeted immunotherapy (Supplementary Table 8). c, Validation cohort of patients with HCC treated with PD(L)1-targeted immunotherapy. A total of 118 patients with advanced HCC received PD(L)1-targeted immunotherapy (Supplementary Table 10). b, c, Log-rank test. Details of sample sizes, biological replicates and statistical tests are given in Methods and Source Data.

Source data

To specifically characterize the effect of anti-PD(L)1 immunotherapy with respect to underlying liver disease, we investigated a cohort of 130 patients with HCC (patients with NAFLD n = 13; patients with other aetiologies n = 117) (Supplementary Table 8). NAFLD was associated with shortened median overall survival after immunotherapy (5.4 months (95% CI 1.8–9.0 months) versus 11.0 months (95% CI 7.5–14.5 months); P = 0.023), even though patients with NAFLD had less frequent macrovascular tumour invasion (23% versus 49%), and immunotherapy was more often used as a first-line therapy in these patients (46% versus 23%; Fig. 4b). After correction for potentially confounding factors that are relevant for prognosis, including severity of liver damage, macrovascular tumour invasion, extrahepatic metastases, performance status, and alpha-fetoprotein (AFP), NAFLD remained independently associated with shortened survival of patients with HCC after anti-PD1-treatment (HR 2.6; 95% CI 1.2–5.6; P = 0.017, Supplementary Table 9). This finding was validated in a further cohort of 118 patients with HCC who were treated with PD(L)1-targeted immunotherapy (patients with NAFLD n = 11; patients with other aetiologies n = 107) (Supplementary Table 10). NAFLD was again associated with reduced survival of patients with HCC (median overall survival 8.8 months, 95% CI 3.6–12.4 months) compared to other aetiologies of liver damage (median overall survival 17.7 months, 95% CI 8.8–26.5 months; P = 0.034) (Fig. 4c). Given the relatively small number of patients with NAFLD in both cohorts, these data need prospective validation. However, collectively these results indicate that patients with underlying NASH did not benefit from checkpoint-inhibition therapy.

Liver cancer develops primarily on the basis of chronic inflammation. The latter can be activated by immunotherapy to induce tumour regression in a subset of patients with liver cancer. However, the identification of patients who will respond to immunotherapy for HCC remains difficult. Our data identify a non-viral aetiology of liver damage and cancer (that is, NASH) as a predictor of unfavourable outcome in patients treated with immune-checkpoint inhibitors. The better response to immunotherapy in patients with virus-induced HCC than in patients with non-viral HCC might be due to the amount or quality of viral antigens or to a different liver micro-environment, possibly one that does not impair immune surveillance. These results might also have implications for patients with obesity and NALFD or NASH who have cancer at other organ sites (for example, melanoma, colon carcinoma, or breast cancer) and are at risk for liver damage and the development of liver cancer in response to systemically applied immunotherapy. Overall, our results provide comprehensive mechanistic insight and a rational basis for the stratification of patients with HCC according to their aetiology of liver damage and cancer for the design of future trials of personalized cancer therapy.


Mice, diets, and treatments

Standard mouse diet feeding (ad libitum water and food access) and treatment regimens were as described previously17. Male mice were housed at the German Cancer Research Center (DKFZ) (constant temperature of 20–24 °C and 45–65% humidity with a 12-h light–dark cycle). Mice were maintained under specific pathogen-free conditions and experiments were performed in accordance with German law and the governmental bodies, and with approval from the Regierungspräsidium Karlsruhe (G11/16, G129/16, G7/17). Tissues from inducible knock-in mice expressing the human unconventional prefoldin RPB5 interactor were received from N. Djouder17,33. The plasmids for hydrodynamic tail-vein delivery have been described previously34,35,36,37. For interventional studies, male mice fed a CD-HFD were treated with bi-weekly for 8 weeks by intravenous injection of 25 μg CD8-depleting antibody (Bioxcell, 2.43), 50 μg NK1.1-depleting antibody (Bioxcell, PK136), 300 μg anti-PDL1 (Bioxcell, 10F.9G2), 200 μg anti-TNF (Bioxcell, XT3.11), 100 μg anti-CD4 (Bioxcell, GK1.5), or 150 μg anti-PD1 (Bioxcell, RMP1-14). PD1−/− mice were kindly provided by G. Tiegs and K. Neumann. Mice for Extended Data Fig. 3g were treated with anti-PD1 antibody (Bioxcell, RMP1-14) or isotype control (Bioxcell, 2A3) at an initial dose of 500 μg intraperitoneally (i.p.) followed by doses of 200 μg i.p. bi-weekly for 8 weeks. Mice for Extended Data Fig. 3h were treated i.p. with anti-PD1 (200 μg, Bioxcell, RMP1-14) or IgG (200 μg, Bioxcell, LTF-2). The treatment regimen for Extended Data Fig. 3i was as described elserwhere38.

Intraperitoneal glucose tolerance test and measurement of serum parameters were as described previously17.

Magnetic resonance Imaging

MRI was done in the small animal imaging core facility in DKFZ using a Bruker BioSpec 9.4 Tesla (Ettlingen). Mice were anaesthetized with 3.5% sevoflurane, and imaged with T2-weighted imaging using a T2_TurboRARE sequence: TE = 22 ms, TR = 2,200 ms, field of view (FOV) 35 × 35 mm, slice thickness 1 mm, averages = 6, scan time 3 min 18 s, echo spacing 11 ms, rare factor 8, slices 20, image size 192 × 192 pixels, resolution 0.182 × 0.182 mm.

Multiplex ELISA

Liver homogenates were prepared as for western blotting17 and cytokines or chemokines were analysed on a customized ELISA according to the manufacturer’s manual (Meso Scale Discovery, U-PLEX Biomarker group 1, K15069L-1).

Flow cytometry and FACS

Isolation and staining of lymphocytes

After perfusion and mechanical dissection, livers were incubated for up to 35 min at 37 °C with collagen IV (60 U final concentration (f.c.)) and DNase I (25 μg/ml f.c.), filtered at 100 μm, and washed with RPMI1640 (11875093, Thermo Fisher). Next, samples underwent a two-step Percoll gradient (25%/50% Percoll/HBSS) and centrifugation for 15 min at 1,800g and 4 °C. Enriched leukocytes were then collected, washed, and counted. For re-stimulation, cells were incubated for 2 h at 37 °C under 5% CO2 with 1:500 Biolegend´s Cell Activation Cocktail (with brefeldin A) (423304) and 1:1,000 Monensin Solution (420701). Live/dead discrimination was done using DAPI or ZombieDyeNIR according to the manufacturer’s instructions with subsequent staining of titrated antibodies (Supplementary Tables 1214). Samples for flow cytometric-activated cell sorting (FACS) were sorted and samples for flow cytometry were fixed using eBioscience IC fixation (00-8222-49) or FOXP3 Fix/Perm kit (00-5523-00) according to the manufacturer’s instructions. Intracellular staining was performed in eBioscience Perm buffer (00-8333-56). Cells were analysed using BD FACSFortessa or BD FACSSymphony and data were analysed using FlowJo (v10.6.2). For sorting, FACS Aria II and FACSAria FUSION were used in collaboration with the DKFZ FACS core facility.

For UMAP and FlowSOM plots, BD FACSymphony data (mouse and human) were exported from FlowJo (v10). Analyses were performed as described elsewhere39.

Single-cell RNA-seq and metacell analysis (mouse)

Single-cell capturing for scRNA-seq and library preparation were done as described previously40. Libraries (pooled at equimolar concentration) were sequenced on an Illumina NextSeq 500 at a median sequencing depth of ~40,000 reads per cell. Sequences were mapped to the mouse genome (mm10), using HISAT (version 0.1.6); reads with multiple mapping positions were excluded. Reads were associated with genes if they were mapped to an exon, using the Ensembl gene annotation database (Ensembl release 90). Exons of different genes that shared a genomic position on the same strand were considered to represent a single gene with a concatenated gene symbol. The level of spurious unique molecular identifiers (UMIs) in the data was estimated by using statistics on empty MARS-seq wells and excluded rare cases with estimated noise >5% (median estimated noise overall for experiments was 2%). Specific mitochondrial genes, immunoglobulin genes, genes linked with poorly supported transcriptional models (annotated with the prefix “Rp-”), and cells with fewer than 400 UMIs were removed. Gene features were selected using Tvm = 0.3 and a minimum total UMI count >50. We carried out hierarchical clustering of the correlation matrix between those genes (filtering genes with low coverage and computing correlation using a down-sampled UMI matrix) and selected the gene clusters that contained anchor genes. We used K = 50, 750 bootstrap iterations, and otherwise standard parameters. Subsets of T cells were obtained by hierarchical clustering of the confusion matrix and supervised analysis of enriched genes in homogeneous groups of metacells41.

Velocity and correlation analyses of scRNA-seq data

Velocyto (0.6) was used to estimate the spliced and unspliced counts from the pre-aligned bam files42. RNA velocity, latent time, root, and terminal states were calculated using the dynamical velocity model from scvelo (0.2.2)43. Kendall’s rank correlation coefficient (τ) was used to correlate the expression patterns of biologically significant genes with latent time.

Preparation for mass spectrometry, data acquisition, and data analysis

After FACS purification, cells were resuspended in 50% (vol/vol) 2,2,2-trifluoroethanol in PBS pH 7.4 buffer and lysed by repeated sonication and freeze–thaw cycles. Proteins were denatured at 60 °C for 2 h, reduced using dithiothreitol at a final concentration of 5 mM (30 min at 60 °C), cooled to room temperature, alkylated using iodoacetamide at 25 mM (30 min at room temperature in the dark), and diluted 1:5 using 100 mM ammonium bicarbonate, pH 8.0. Proteins were digested overnight by trypsin (1:100 ratio, 37 °C), desalted using C18-based stage-tips, dried under vacuum, resuspended in 20 μl HPLC-grade water with 0.1% formic acid, and measured using A380.

We used 0.5 μg of peptides for proteomic analysis on a C18 column using a nano liquid chromatography system (EASY-nLC 1200, Thermo Fisher Scientific). Peptides were eluted using a gradient of 5–30% buffer B (80% acetonitrile and 0.1% formic acid) at a flow rate of 300 nl/min at a column temperature of 55 °C. Data were acquired by data-dependent Top15 acquisition using a high-resolution orbitrap tandem mass spectrometer (QExactive HFX, Thermo Scientific). All MS1 scans were acquired at 60,000 resolution with AGC target of 3 × 106, and MS2 scans were acquired at 15,000 resolution with AGC target of 1 × 105 and maximum injection time of 28 ms. Analyses were performed using MaxQuant (, mouse UniProt Isoform fasta (Version: 2019-02-21, number of sequences 25,233) as a source for protein sequences. One per cent FDR was used for controlling at the peptide and protein levels, with a minimum of two peptides needed for consideration of analysis. GSEA was performed using ClusterProfiler (3.18)44 and gene sets obtained from WikiPathway ( and MSigDB (,46,47.

Histology, immunohistochemistry, scanning, and automated analysis

Histology, immunohistochemistry, scanning, and automated analysis have been described previously17. Antibodies used in this manuscript are described in Supplementary Table 12. For immunofluorescence staining, established antibodies were used, coupled with the AKOYA Biosciences Opal fluorophore kit (Opal 520 FP1487001KT, Opal 540 FP1494001KT, Opal 620 FP1495001KT). For mRNA in situ hybridization, freshly non-baked 5 μm formalin-fixed paraffin-embedded sections were cut and stained according to the manufacturer’s (ACD biotech) protocol for manual assay RNAscope, using probes PDL1 (420501), TNF (311081) and CXCR6 (871991).

Isolation of RNA and library preparation for bulk RNA sequencing

RNA isolation17 and library preparation for bulk 3′-sequencing of poly(A)-RNA was as described previously48. Gencode gene annotations version M18 and the mouse reference genome major release GRCm38 were derived from Dropseq tools v1.1249 were used for mapping the raw sequencing data to the reference genome. The resulting UMI-filtered count matrix was imported into R v3.4.4. Before differential expression analysis with Limma v3.40.650 sample-specific weights were estimated and used as coefficients alongside the experimental groups as a covariate during model fitting with Voom. t-test was used for determining differentially (P < 0.05) regulated genes between all possible experimental groups. GSEA was conducted with the pre-ranked GSEA method46 within the MSigDB Reactome, KEGG, and Hallmark databases ( Raw sequencing data are available at European Nucleotide Archive ( under the accession number PRJEB36747.

Stimulation of CD8 T cells

Stimulation of CD8 T cells was as described elsewhere27.

Flow cytometry of human biopsies

Analysis of patient material (Supplementary Table 1) was performed on liver tissue (needle biopsies or resected tissue, BIOFACS Study KEK 2019-00114), which were obtained from the patient collection nAC-2019-3627 (CRB03) from the biological resource centre of CHU Grenoble-Alpes (nBRIF BB-0033-00069). Tissue samples were minced using scalpels, incubated (with 1 mg/ml collagenase IV (Sigma Aldrich), 0.25 μg/ml DNase (Sigma Aldrich), 10% FCS (Thermo Fisher Scientific), RPMI 1640 (Seraglob)) for 30 min at 37 °C, stopping enzymatic reactions with 2 mM EDTA (StemCell Technologies, Inc.) in PBS. After filtering through a 100-μm cell strainer, cells were resuspended in FACS buffer (PBS, EDTA 2 mM, FCS 0.5%) with Human TruStain FcX (Fc Receptor Blocking Solution) (Biolegend), incubated for 15 min at 4 °C and stained with antibodies (Supplementary Table 13).

Flow cytometry of human samples (Extended Data Fig. 9f) was approved by the local ethical committee (AC-2014-2094 n 03).

High-throughput RNA-seq of human samples

As previously reported, RNA-seq analysis was performed using the data from 206 snap-frozen biopsy samples from 206 patients diagnosed with NAFLD in France, Germany, Italy, and the UK and enrolled in the European NAFLD Registry (GEO accession GSE135251)51,52. Samples were scored for NAS by two pathologists53. Alternate diagnoses were excluded, including excessive alcohol intake (30 g per day for males, 20 g for females), viral hepatitis, autoimmune liver diseases, and steatogenic medication use. Patient samples were grouped: NAFL (n = 51) and NASH with fibrosis stages of F0/1 (n = 34), F2 (n = 53), F3 (n = 54) and F4 (n = 14). Collection and use of data of the European NAFLD Registry were approved by the relevant local and/or national Ethical Review Committee51. A correction for sex, batch, and centre effects was implemented. Pathway enrichment and visualization were as described elsewhere52,54,55.

Immunohistochemistry of NAFLD/NASH cohort

Sixty-five human FFPE biopsies from patients with NAFLD were included (Supplementary Table 3). Sequential slides were immunostained with antibodies against human CD8 (Roche, SP57, ready-to-use), PD1 (Roche; NAT105, ready-to-use), and CD4 (Abcam, ab133616, 1:500). All staining was performed on the VENTANA BenchMark autostainer at 37 °C. Immunopositive cells were quantified at 400× magnification in the portal tract and the adherent parenchyma.

Isolation of cells for scRNA-seq data analysis (human)

Analyses used liver samples from patients undergoing bariatric surgery at the Department of Surgery at Heidelberg University Hospital (S-629/2013). Samples were preserved by FFPE for pathological evaluation and single cells were generated by mincing, using the Miltenyi tumour dissociation kit (130-095-929) per the manufacturer’s instructions, filtering through a 70-μm cell strainer and washing. ACK lysis using the respective buffer (Thermo Fischer Scientific A1049201) was performed, and samples were stored in FBS with 20% DMSO until further processing (scRNA-seq analysis and mass cytometry).

Cells were thawed in a 37 °C water bath, washed with PBS + 0.05 mM EDTA (10 min, 300g at 4 °C), Fc receptor-block (10 min at 4 °C), stained with CD45-PE (3 μl, Hl30, 12-0459-42) and Live/Dead discrimination (1:1,000, Thermofischer, L34973), washed and sorted on a FACSAria FUSION in collaboration with the DKFZ FACS. Library generation was performed according to the manufacturer’s protocol (Chromium Next EM Single Cell 3′GEM, 10000128), and sequencing was performed on an Illumina NovaSeq 6000. De-multiplexing and barcode processing were performed using the Cell Ranger Software Suite (Version 4.0.0) and reads were aligned to human GRCh3856. A gene–barcode matrix containing cell barcodes and gene expression counts was generated by counting the single-cell 3′ UMIs, which were imported into R (v4.0.2), where quality control and normalization were executed using Seurat v357. Cells with more than 10% mitochondrial genes, fewer than 200 genes per cell, or more than 6,000 genes per cell were excluded. Matrices from 10 samples were integrated with Seurat v3 to remove batch effects across samples. PCA analysis of filtered gene–barcode matrices of all CD3+ cells, visualized by UMAP (top 50 principal components), and identification of major cell types using the highly variable features and indicative markers were performed. Pairwise comparisons of CD4+ T cells versus CD4+PD1+ T cells and CD8+ T cells versus CD8+PD1+ T cells were performed using the results of differential expression analysis by DESeq2 (v1.28.1)58, setting CD4+/CD8+ T cells as controls. Volcano plots were then generated using EnhancedVolcano (v1.6.0)59 to visualize the results of differential expression analysis.

Mass cytometry data analysis (human)

Antibody conjugates for mass cytometry were purchased from Fluidigm, generated in-house using antibody labelling kits (Fluidigm X8, MCP9), or as described before60,61. Antibody cocktails for mass cytometry were cryopreserved as described before62. Isolation of cells is described in ‘Isolation of cells for scRNA-seq data analysis (human)’. Cells were thawed, transferred into RPMI + benzonase (14 ml RPMI + 0.5 μl benzonase), and centrifuged for 5 min at 500g. The cell pellet was resuspended in 1 ml CSM-B (CSM (PBS 0.5% BSA 0.02% sodium azide) +1 μl benzonase), filtered through a 30-μm cell strainer, adjusted to 3 ml, counted, resuspended in 35 μl CSM-B and incubated for 45 min at 4 °C, and 100 μl CSM-B was added. Cells were pooled and stained with a surface antibody cocktail (Supplementary Table 15) for 30 min at 4 °C. Dead cell discrimination was performed with mDOTA-103Rh (5 min, room temperature). For intracellular staining, the FOXP3 intracellular staining kit from Miltenyi Biotec was used per the manufacturer’s instructions, followed by staining for intracellular targets for 30 min at room temperature. Cells were washed, resuspended in 1 ml of iridium intercalator solution, and incubated for 25 min at room temperature. Cells were washed with CSM, PBS, and MilliQ water, adjusted to a final concentration of 7.5 × 105 cells/ml and supplemented with 4-element EQ beads. The sample was acquired on a Helios mass cytometer and raw data were EQ-Bead-normalized using Helios mass cytometer and Helios instrument software (version 6.7). Compensation was performed in CATALYST (v1.86)63 and FlowCore (1.50.0). De-barcoding and gating of single, live CD45+ cells were performed using FlowJo (v10.6.2). Then, data from CD45+ cells were imported into Cytosplore 2.3.1 and transformed using the arcsinh(5) function. Major immune cell lineages were identified at the first level of a two-level hierarchical stochastic neighbour embedding (HSNE) analysis with default perplexity and iteration settings. HSNE with the same parameters was run on CD3+ cells to identify T cell phenotypes. Gaussian mean shift clustering was performed in Cytosplore and a heat map of arcsinh(5)-transformed expression values of all antibody targets was generated. Cell type identification was based on the transformed expression values and clusters showing high similarity were merged manually.

Histological and immunohistochemical analysis of NASH–HCC cohort

Four healthy samples, 16 samples from patients with NASH cases, and non-tumoral tissue adjacent to HCC tumours from patients of the following aetiologies were selected: NASH (n = 26), viral hepatitis (n = 19 HCV, n = 3 HBV), alcohol (n = 5), and other (n = 2). All samples were obtained from International Genomic HCC Consortium with IRB approval. After heat-induced antigen retrieval (10 mM sodium citrate buffer (pH 6.0) or Universal HIER antigen retrieval reagent (ab208572) for 15 min (3 × 5 min), the reaction was quenched using 3% hydrogen peroxide. Samples were washed with PBS and incubated with anti-CD8 (Cell Signaling, Danvers, MA) or anti-PD1 (NAT105, ab52587). DAB (3,3′-diaminobenzidine) was used as a detection system (EnVision+ System-HRP, Dako). PD1-positive cases were defined by considering median positivity by immunohistochemistry64 and using a cutoff of ≥1% of PD1-positive lymphocytes among all lymphocytes present on each slide. Analysis of human samples from the Department of Pathology and Molecular Pathology, University Hospital Zurich (Extended Data Fig. 10), was approved by the local ethics committee (Kantonale Ethikkommission Zürich, KEK-ZH-Nr. 2013-0382 and BASEC-Nr. PB_2018-00252).

Search strategy, selection criteria, and meta-analysis of phase III clinical trials

The literature search was done through MEDLINE on PubMed, Cochrane Library, Web of Science, and, using the following searches: ‘checkpoint inhibitors’, ‘HCC’, ‘phase III’, between January 2010 and January 2020, and complemented by manual searches of conference abstracts and presentations. Single-centre, non-controlled trials, studies with insufficient data to extract HRs or 95% confidence intervals, and trials including disease entities other than HCC were excluded. As conference abstracts were not excluded, quality assessment of the included studies was not performed. Three studies5,10,11 fulfilled the criteria and were included in the quantitative synthesis (Extended Data Fig. 10). The primary outcome of the meta-analysis was overall survival, defined as the time from randomization to death. HRs and CIs related to overall survival were extracted from the papers or conference presentations5,10,11. Pooled HRs were calculated using the random-effects model and we used the DerSimonian–Laird method to estimate τ2, and the generic inverse variance was used for calculating weights 65. To evaluate heterogeneity among studies, Cochran’s Q test and I2 index were used. P < 0.10 in the Q-test was considered to indicate substantial heterogeneity. I2 was interpreted as suggested in the literature: 0% to 40% might not represent significant heterogeneity; 30% to 60% may represent moderate heterogeneity; 50% to 90% may represent substantial heterogeneity; 75% to 100% represents considerable heterogeneity. All statistical pooled analyses were performed using RevMan 5.3 software.

A cohort of patients with HCC treated with PD(L)1-targeted immunotherapy

The retrospective analysis was approved by local Ethics Committees. Data from this cohort were published previously66. Patients with liver cirrhosis and advanced-stage HCC treated with PD(L)1-targeted immune checkpoint blockers from 12 centres in Austria, Germany, Italy, and Switzerland were included. The χ2 test or Fisher’s exact test were used to compare nominal data. Overall survival was defined as the time from the start of checkpoint inhibitor treatment until death. Patients who were still alive were censored at the date of the last contact. Survival curves were calculated by the Kaplan–Meier method and compared by using the log-rank test. Multivariable analysis was performed by a Cox regression model. Statistical analyses were performed using IBM SPSS Statistics version 25 (SPSS Inc., Chicago, IL).

A validation cohort of patients with HCC treated with PD1-targeted immune checkpoint blockers

A multi-institutional dataset that included 427 patients with HCC treated with immune checkpoint inhibitors between 2017 and 2019 in 11 tertiary-care referral centres specialized in the treatment of HCC was analysed. Clinical outcomes of this patient cohort have been reported elsewhere67,68. Inclusion criteria were: 1) diagnosis of HCC made by histopathology or imaging criteria according to American Association for the Study of Liver Disease and European Association for the Study of the Liver guidelines; 2) systemic therapy with immune checkpoint inhibitors for HCC that was not amenable to curative or loco-regional therapy following local multidisciplinary tumour board review; 3) measurable disease according to RECIST v1.1 criteria at commencement of treatment with immune checkpoint inhibitors. One hundred and eighteen patients with advanced-stage HCC were recruited with Child–Pugh A liver functional reserve, and documented radiologic or clinical diagnosis of cirrhosis. Ethical approval to conduct this study was granted by the Imperial College Tissue Bank (reference number R16008).

Statistical analyses

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment. Data were collected in Microsoft Excel. Mouse data are presented as the mean ± s.e.m. Pilot experiments and previously published results were used to estimate the sample size, such that appropriate statistical tests could yield significant results. Statistical analysis was performed using GraphPad Prism software version 7.03 (GraphPad Software). Exact P values lower than P < 0.1 are reported and specific tests are indicated in the legends.

Sample sizes, biological replicates and statistical tests

Fig. 1a: PD1, n = 5 mice/group; CD8, ND n = 6 mice; CD-HFD n = 6 mice; WD-HTF n = 5 mice. Scale bar, 100 μm. Fig. 1b: n = 3 mice/group. Scale bar, 100 μm. Fig. 1c: ND n = 4 mice, CD-HFD n = 6 mice. Fig. 1d, e: n = 3 mice/group. Fig. 1f: tumour incidence: CD-HFD, n = 19 tumours/lesions in 25 mice; CD-HFD + anti-PD1, n = 10 tumours/lesions in 10 mice. Fig. 1h: ND, n = 3 mice; CD-HFD, n = 13 mice; CD-HFD + anti-PD1, n = 8 mice; intra-tumoral staining: CD-HFD, n = 11 mice; CD-HFD + anti-PD1, n = 8 mice. Data in Fig. 1a, h were analysed by two-tailed Student’s t-test. Data in Fig. 1f were analysed by two-sided Fisher’s exact test.

Fig. 2a, b: n = 3 mice/group. Fig. 2c: CD8+: ND, n = 6 mice; CD-HFD + IgG, n = 5 mice; CD-HFD + anti-PD1, n = 6 mice; CD8+PD1+: ND, n = 4 mice, CD-HFD + IgG, n = 6 mice; CD-HFD + anti-PD1, n = 6 mice. Fig. 2d, e: ND, n = 4 mice; CD-HFD + IgG, n = 8 mice; CD-HFD + anti-PD1, n = 6 mice. Fig. 2f: CD-HFD + IgG, n = 6 mice; CD-HFD + anti-PD1, n = 4 mice. Fig. 2g: ND, n = 30 mice; CD-HFD, n = 47 mice; CD-HFD + anti-PD1, n = 35 mice; CD-HFD + anti-PD1/anti-CD8, n = 9 mice; CD-HFD + anti-TNF, n = 10 mice; CD-HFD + anti-PD1/anti-TNF, n = 11 mice; CD-HFD + anti-CD4, n = 8 mice; CD-HFD + anti-PD1/anti-CD4, n = 8 mice. Fig. 2h: CD8+PD1+CXCR6+: ND, n = 30 mice; CD-HFD, n = 47 mice; CD-HFD + anti-PD1, n = 35 mice; CD-HFD + anti-PD1/anti-CD8, n = 9 mice; CD-HFD + anti-TNF, n = 10 mice; CD-HFD + anti-PD1/anti-TNF, n = 11 mice; CD-HFD + anti-CD4, n = 8 mice; CD-HFD + anti-PD1/anti-CD4, n = 8 mice. Fig. 2j: tumour incidence: CD-HFD, n = 32 tumours/lesions in 87 mice; CD-HFD + anti-CD8, n = 2 tumours/lesions in 31 mice; CD-HFD + anti-CD8/NK1.1, n = 0 tumours/lesions in 6 mice; CD-HFD + anti-PD1, n = 33 tumours/lesions in 44 mice; CD-HFD + anti-PD1/anti-CD8, n = 2 tumours/lesions in 9 mice; CD-HFD + anti-TNF, n = 3 tumours/lesions in 10 mice; CD-HFD + anti-PD1/anti-TNF, n = 3 tumours/lesions in 11 mice; CD-HFD + anti-CD4, n = 3 tumours/lesions in 9 mice; CD-HFD + anti-PD1/anti-CD4, n = 8 tumours/lesions in 9 mice. All data are shown as mean ± s.e.m. Data in Fig. 2e, g, h were analysed by one-way ANOVA and Fisher’s LSD test. Data in Fig. 2f were analysed by two-tailed Mann–Whitney test. Data in Fig. 2j were analysed by two-sided Fisher’s exact test.

Fig. 3a, b: control, n = 6 patients; NAFLD/NASH, n = 11 patients. Fig. 3c: control, n = 4 patients; NAFLD/NASH, n = 7 patients. Fig. 3d–h: mouse, n = 3; human, n = 3. All data are shown as mean ± s.e.m. Data in Fig. 3b, f were analysed by two-tailed Mann–Whitney test. Data in Fig. 3d were analysed by two-tailed Spearman’s correlation.

Fig. 4a: Hazard ratios are represented by squares, the size of the square represents the weight of the trial in the meta-analysis. Cochran’s Q-test and I2 were used to calculate heterogeneity. Fig. 4b: Kaplan–Meier curve displays overall survival of patients with NAFLD versus those with any other aetiology; all 130 patients were included in these survival analyses (NAFLD n = 13; any other aetiology n = 117). Fig. 4c: Kaplan–Meier curve displays overall survival of patients with NAFLD versus those with any other aetiology (NAFLD n = 11; any other aetiology n = 107). Data in Fig. 4b, c were analysed by Kaplan–Meier method and compared using log rank test.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.