Natural killer cell-mediated immunosurveillance modulates liver cancer evolution through cancer stemness enhancement and lipid metabolism reprogramming

Liang Shi; Boqiang Liu; Ruya Sun; Jing He; Hao Shen; Weiqi Li; Yi Wang; Weijun Zhao; Chenqi Jin; Binghan Jin; Yuanshi Tian; Lingfeng Ma; Lidan Hou; Mingyu Chen; Jiaying Shen; Bo Shen; Xiao Liang; Hong Yu; Yifan Wang; Di Wang; Xiujun Cai

doi:10.1038/s41467-026-74360-x

. 2026 Jun 11;17:7427. doi: 10.1038/s41467-026-74360-x

Natural killer cell-mediated immunosurveillance modulates liver cancer evolution through cancer stemness enhancement and lipid metabolism reprogramming

Liang Shi ^1,^2,^3,^✉,^#, Boqiang Liu ^1,^2,^3,^✉,^#, Ruya Sun ^4,^#, Jing He ^1,^2,³, Hao Shen ^1,^2,³, Weiqi Li ¹, Yi Wang ^1,^2,³, Weijun Zhao ^1,^2,³, Chenqi Jin ¹, Binghan Jin ⁵, Yuanshi Tian ⁶, Lingfeng Ma ⁷, Lidan Hou ^1,^2,³, Mingyu Chen ^1,^2,³, Jiaying Shen ⁷, Bo Shen ^1,^2,³, Xiao Liang ^1,^2,³, Hong Yu ^1,^2,^3,⁸, Yifan Wang ^1,^2,^3,^✉, Di Wang ^9,^✉, Xiujun Cai ^1,^2,^3,^✉

PMCID: PMC13408697 PMID: 42277059

Abstract

Tumor evolution enables liver cancer cells to acquire survival advantages and evade therapy-induced cell death. However, the role of natural killer (NK) cells in liver cancer evolution remains unclear. Here, we establish immune-humanized spatiotemporal liver cancer models and integrate single-cell, spatial transcriptomic, and CRISPR/Cas9 screening analyses to investigate this process. We demonstrate that early NK cell-mediated immunosurveillance promotes tumor cell state transition and impairs subsequent adaptive immune responses. Mechanistically, NK cells induce lipid metabolic reprogramming, particularly cholesterol accumulation, and enhance tumor stemness, both of which promote liver cancer evolution. Furthermore, combined anti-LAG-3 treatment and liver X receptor activation suppress tumor evolution and improve the efficacy and durability of immune checkpoint blockade in advanced liver cancer. Collectively, our findings identify that NK cell-mediated early immunosurveillance promotes liver cancer evolution and suggest immunometabolic therapy as a potential strategy for advanced liver cancer.

Subject terms: Cancer, Immunology, Immunosurveillance, NK cells, Translational immunology

NK cell function in liver cancer has been shown and may influence the progression of liver cancer. Here the authors use mouse liver cancer models and spatial transcriptomics in an immune-humanized liver cancer mouse model to show that NK cells can influence the early progression of liver cancer by altering lipid metabolism and tumour stemness, subsequently demonstrating that inhibition of these pathways can enhance immune cell-mediated tumour control.

Introduction

Tumor evolution, which defines and explains tumorigenesis and cancer progression through a Darwinian lens, demands that oncologists examine malignancies, including liver cancer, from a higher-dimensional perspective¹. Although the concept of tumor evolution has been discussed for decades, its clinical applications remain challenging. One of the reasons is that the factors triggering and reshaping tumor evolution are largely unknown². Certain conditions, particularly those involving therapeutics, have been shown to trigger tumor evolution. For instance, anti-tumor drugs can drive the evolutionary process within tumors, leading to treatment resistance and disease progression³. Recent studies have provided additional insights into the relationship between tumor immunology and tumor evolution⁴. Immunoediting has been recognized as an intrinsic force capable of shaping tumor evolution, though the details remain obscure and its clinical applications are still limited⁴. NK cells are known for their ability to rapidly induce innate immune responses⁵. In human peripheral tissues, liver possesses a large portion of NK cells, which represent up to 30−40% of all hepatic lymphocytes⁶. Meanwhile, there is a close relationship between NK cells and both liver precancerous disease and liver cancer⁷. It has been observed that innate immunosurveillance, primarily mediated by NK cells within tumors, can conduct immune editing and modify the clonal heterogeneity of tumors⁸.

However, the precise immune factors and mechanisms that initiate and shape liver cancer evolution during early immunosurveillance remain largely unclear. In this work, we develop several models of hepatocellular carcinoma (HCC) to identify potential crucial immune factors that determine tumor evolutionary trajectories at the onset of immunosurveillance. By simulating and analyzing the liver cancer evolution at cellular, animal, and patient levels, we confirm that NK cell-mediated immunosurveillance is a key force driving tumor evolution. And, we successfully identify the features and mechanism of NK cell-induced tumor evolution, which involve the enhancement of liver cancer stemness and lipid metabolism reprogramming, respectively. In addition, we highlight a potential strategy to combat evolved liver cancer using a combination of immune checkpoint inhibitors (ICIs) and metabolism agonists.

Results

Spatial distribution of NK cells positively correlates with early transcriptome evolution of liver cancer

To biologically simulate this early-stage immunosurveillance in liver cancer and identify potential key immune cells mediating early liver cancer evolution, we established a spatiotemporal murine model. Given that genetically induced tumor models often fail to capture nascent lesions that initially interact with immune cells, and direct injection of tumor cell suspensions poorly mimics early immune infiltration, we instead orthotopically xenografted immunodeficient-derived human liver tumor tissue into recipient mice with a humanized immune system (hHSC-NCG-hIL15, Fig.1A and Fig. S1A).

Fig. 1 — A Schematic representation of the immune-humanized mouse orthotopic liver cancer model. Wild-type HepG2 tumors derived from the same NCG mice (severely immunodeficient mice, completely lacking functional T, B, and NK cells) were orthotopically implanted into the livers of hHSC-NCG-hIL15 mice (immune-humanized mice). Each week following implantation, mice from designated time points were sacrificed, and spatial transcriptomic analysis was performed. B Visium captured spots colored by region annotation. Regions are labeled based on H&E images and gene expression profiles as hTumor (human-origin tumor), peri-hTumor (hTumor surrounding area), mHep (mouse-origin hepatocyte), mRBC (mouse-origin red blood cell), and mStromal (mouse-origin stromal cell) regions. Sample labels (e.g., 1W1, 1W2, 2W1, 2W2) indicate independent immune-humanized mice analyzed at the indicated time points (weeks) after tumor implantation. C, D PCA embedding of Visium captured spots within the hTumor region, colored by corresponding sample (C) or cluster (D). Each dot represents a ST spot and is colored by corresponding sample (C) or cluster (D). E Transcriptional similarity between the corresponding spots of each hTumor cluster and HepG2 WT cell, measured by Spearman’s correlation coefficient. Data compared using two-sided Student’s t-test (P = 1.57×10^-96). F, G. The spatial distribution of clusters identified in the hTumor area (hC1–hC7) in each tumor tissue (F). Proportions of hTumor clusters in ST samples showed that the tumor cells in early stage are mainly composed of hC1 and hC2 clusters, with hC2 becoming the dominant group as the tumor progresses (G). H, I The 1W1 sample was partitioned into three spatial regions based on transcriptional clustering and spatial localization patterns: hC1-enriched, hC2-enriched, and peri-hTumor (H). Tumor purity scores were estimated using SpatialDecon for spots within each region. The inferred tumor purity values, which are inversely related to immune cell infiltration, are shown for the indicated regions (I, P = 0.0003, two-sided unpaired Student’ s t-test). J Spatial transcriptome of 1W1 sample colored by transcriptional similarity to wild-type HepG2 cell. K SpatialDecon analysis was performed on the 1W1 sample to estimate the relative abundance of mixed cell types across spatially resolved gene expression profiles. The inferred abundance of humanized immune cell populations, including NK cells, naïve B cells, naïve CD4⁺ T cells, and macrophages, is shown. L Immunofluorescence staining of CD56, NCR1 (NK cell), CD68 (macrophage), CD4 (CD4⁺ T cell) and CD19 (B cell) in 1W1 sample. Scale bars, 100 μm. M Beeswarm (left) and violin plot (right) displaying the NK activity score in hTumor-hC1, hTumor-hC2 and peri-hTumor relevant spots in 1W1 samples. Data compared using two-sided Student’s t-test (P = 0.0269, 1.956 × 10⁻⁵, respectively). N, O. Based on spatial proximity, the human-derived tumor regions of the 1W1 sample were stratified, where the “Core” area overlapped with the hC2-enriched region, while Circles 1-4 represented the hC1-enriched region (N). Expression level of canonical immune signatures in defined regions, from “Circle” 1-4 to “Core” (O). P, Q An evolutionary trajectory axis was defined based on the spatial distribution of the hC1 and hC2 sub-populations (P, left). Along this axis, the predominant cell populations gradually transitioned from hC1 to hC2 (P, right). Concurrently, the expression dynamics of NK cell signature genes, such as GNLY and GZMB, along this trajectory were quantified (Q).

The immunodeficient-derived human liver tumor tissue (P1) was generated by intrahepatic injection of HepG2 cell suspension (P0) into NCG mice (See details in Method section). Two weeks after injection, liver tumor tissues (P1) were collected by surgical approach and surgically implanted into liver of humanized recipient mice with a size of 1 mm × 1 mm × 1 mm. We selected the HepG2 cell line for this experiment because, compared to other HCC cell lines established from adult HCC patient, HepG2 cells exhibit more epithelial similarity to hepatocytes (Fig. S1B) and can be recognized and targeted by the immune system as liver tumor cells¹. We believed that these characteristics make HepG2 cells particularly advantageous for early-phase studies in liver cancer.

Each week following implantation, mice from designated time points were sacrificed, and spatial transcriptomic (ST) analysis was performed (Fig. 1A and Fig. S1C–S1E). Based on the marker gene expression level and corresponding H&E images, ST spots were first classified into the human-origin tumor (hTumor, as shown in Eq. [2]), peri-tumor (peri-hTumor), mouse-origin hepatocyte (mHep), mouse-origin red blood cell (mRBC) and mouse-origin stromal cell (mStromal) regions (Figure S1F). We observed that, in cases where tumors persisted, a marked proliferation occurred between the 3^rd and 4^th weeks (Fig. 1B, Fig. S1C–S1E). In temporal transcriptome analysis, we identified substantial changes over time (Fig. 1C). Through a shared nearest neighbor (SNN) based clustering algorithm, seven humanized sub-populations (hC1–hC7) were identified in the hTumor regions, with hC1 showing the highest transcriptional similarity to the original liver tumor cells (Fig. 1D, E). Notably, hC1 predominated during the 1^st and 2^nd weeks, while hC2 emerged and gradually became dominant in the 3^rd and 4^th weeks (Fig. 1F, G, and Fig. S1G). These findings suggest that the period between the 1^st and 2^nd weeks is crucial for transcriptome transition or potential tumor evolution. Importantly, the same modeling procedure was performed in NCG mice, and tumors were collected at corresponding time points. Transcriptomic analysis showed that tumors from NCG mice displayed markedly different transcriptional profiles compared with those from immune-humanized mice, suggesting that the observed effects depend on the presence of the immune system (Fig. S1H, S1I).

To investigate the role of the immune system in the early transcriptomic evolution of tumors, we focused on the transition from hC1 to hC2 that occurred during the 1^st and 2^nd weeks. We selected the most representative sample, 1W1, and divided it into three main regions based on cluster distribution: hC1, hC2, and peri-hTumor (Fig. 1H). Tumor purity of each spot was quantified via SpatialDecon², which revealed a significantly higher infiltration of immune cells in the hC2 and peri-hTumor regions (Fig. 1H, I). This uneven distribution was negatively correlated with transcriptional similarity to original HepG2 cells, suggesting a potential role of immune cells in driving transcriptomic shifts (Fig.1J). A similar pattern was observed in the 1W2 sample (Fig. S1J, S1K). These findings indicate an association between early immune cell infiltration and tumor transcriptomic shifts.

To identify the specific immune cell types responsible for this phenotype, we examined the spatial distribution of various humanized immune cell signals in the 1W1 and 1W2 samples. Among all non-tumor components, NK cells exhibited the strongest signals of tumor infiltration, accompanied by moderate signals from naïve B cells, naïve CD4⁺ T cells, and macrophages (Fig. 1K and Fig. S1L). However, immunofluorescence staining of an adjacent section from the 1W1 tumor confirmed significant intratumoral localization only for NK cells (Fig. 1L). Next, we used a spatial deconvolution algorithm to further resolve immune cell infiltration³. The analysis revealed a markedly higher presence of NK cells in the hC2 and peri-hTumor regions, whereas other immune cell types did not exhibit similar spatial patterns (Fig. 1M and Fig. S1M–S1P)⁴.

Furthermore, to better investigate the gene expression tendency in the 1W1 sample, we stratified the spatial spots of hC1-enriched region into Circle 1−4 based on corresponding distance to the hC2-enriched “Core” area (Fig. 1N). With the evolutionary spatial bins radiating outward to the “Core” area, only the NK cell signals (CD56, GZMB, and GNLY) gradually enhanced (Fig. 1O). We then defined a spatial axis extending from the center of the hC1 region toward the hC2 region, which reflected the trajectory of the hC1-to-hC2 transition (Fig. 1P). Notably, this trajectory also aligned with the spatial distribution of NK cells, further supporting a strong spatial correlation between NK cell infiltration and tumor subpopulation transition (Fig. 1Q).

Overall, the above findings suggest that the spatial distribution of NK cells is positively associated with liver cancer subpopulation transitions. Specifically, early infiltration of NK cells within the hC2 region appears to have the potential to mediate the transition from hC1 to hC2. Given the critical role of NK cells in early tumor immunosurveillance^5,6, we speculate that NK cells may start to influence liver cancer evolution at a very early stage (1^st week time point in hHSC-NCG-hIL15 model).

NK cell mediated-immune pressure induces liver cancer evolution in murine models

To investigate the potential role of NK cells in the liver cancer evolution, we utilized two orthotopic xenograft murine models. The P1 generation of immunodeficient-derived mouse liver tumor tissue was established by intrahepatic injection of Hepa1-6 mouse HCC cells into NCG mice. Through sequential passaging of these tumors from P1 to P4 in either Rag1^−/− or NCG mice, we generated tumor lineages that were either exposed to NK cell surveillance (in Rag1^−/− mice) or not (in NCG mice) (Fig. 2A, see details in the Method section). A similar strategy was employed in immune-humanized mice (hHSC-NCG-hIL15) with human liver tumor cells HepG2, where another key cytotoxic surveillance cell type, T cells, which infiltrated significantly less than NK cells during early immunosurveillance in our model (Fig. 1K and Fig. S2A), were either suppressed using cyclosporin A or Teplizumab, or left untreated (Fig. 2B). Then, P4 tumors were reimplanted crosswise into severely immunodeficient (NCG mice) and relatively immunocompetent mice (Rag1^−/− or hHSC-NCG-hIL15).

The results showed that P4 tumors derived from Rag1^−/− or hHSC-NCG-hIL15 mice (with or without T cell suppression by cyclosporin A or Teplizumab) exhibited a higher degree of malignancy, as evidenced by accelerated growth (Fig. 2C, D and Fig. S2B, S2C), increased metastasis (Fig. 2E–G and Fig. S2D–S2F) and reduced overall survival (OS, Fig. 2H, I). Notably, these aggressive features persisted regardless of whether the NK-conditioned P4 tumors were implanted into severely immunodeficient (NCG) or relatively immunocompetent hosts (Rag1^−/− or hHSC-NCG-hIL15 mice). Tumor growth was assessed by measuring tumor volume in subcutaneous models, metastasis was evaluated using in vivo imaging system (IVIS) following orthotopic liver implantation, and overall survival was monitored in mice bearing orthotopic liver tumors.

Next, key histological markers confirmed higher malignancy of P4 tumors derived from Rag1^−/− or hHSC-NCG-hIL15 mice (Fig. 2J, K). Additionally, transcriptome analyses in P4 tumors revealed substantial differences in tumor immunology pathways, including a significant loss of MHC-I-related signals (Fig. 2L, M and Fig. S2G, S2H), suggesting enhanced immune escape in P4 tumors derived from Rag1^−/− or hHSC-NCG-hIL15 mice. As expected, P4 tumor cells developed in the presence of NK cells exhibited strong resistance to NK cell-mediated cytotoxicity, accompanied by consistent downregulation of NK cell-activating ligands (Fig. 2N and Fig. S2I–S2L).

More importantly, we specifically depleted NK cells in C57BL/6 mice using anti-Asialo GM1 antibodies (Fig. S2M). Subsequently, Hepa1-6 P1 tumors, originally derived from NCG mice (Fig. 2A), were orthotopically and sequentially passaged from P1 to P4 in either C57BL/6 mice treated with isotype control or anti-Asialo GM1 (NK cell-depleted). Transcriptomic sequencing of the P4 tumors revealed that those derived from NK cell-depleted mice exhibited a less evolved (reduced activation of immune evasion-related pathway and decreased activity of canonical oncogenic pathway) transcriptional profile compared to the isotype control group (Fig. 2O). These findings indicate that NK cells possess the capacity to induce liver cancer evolution independently of T cells. Using a similar depletion strategy, macrophage removal in this model did not result in a comparable alteration in tumor evolutionary features relative to isotype controls (Fig. S2N–S2O).

The loss of MHC-I molecules is a strong unfavorable indicator during tumor evolution and suggests an evasion of adaptive immune surveillance⁷. To further investigate this, we developed a vaccinated murine model by subcutaneously injecting Hepa1-6 mouse HCC cells. P4 tumors from Rag1^−/−/NCG mice were implanted side by side into pre-sensitized C57BL/6 mice (Fig. 2P). As anticipated, NK cell-conditioned P4 tumors showed a much higher tumor formation rate and larger tumor volume compared to the control group in pre-sensitized C57BL/6 mice, indicating that NK cell conditioning conferred increased resistance to adaptive immunity (Fig. 2Q and Fig. S2P)⁸.

These results suggest that, although other factors may also contribute to tumor evolution, NK cell-mediated immune pressure is associated with in vivo evolutionary changes in liver cancer. This evolution results not only in enhanced immune evasion, but also in intrinsic, tumor cell-autonomous properties including proliferation, invasion and metastasis⁹.

Alterations of liver cancer cell state mediated by NK cells lead to evolution

Recently, single-cell studies have provided additonal insights into comprehensive tumor evolution, particularly from the perspective of cell state transitions and tumor cell plasticity¹⁰. Results from spatiotemporal transcriptome analysis suggested that new sub-populations can emerge and remodel liver tumor composition (Fig. 1D–G), implying cancer cell states transition.

For further investigation, we established an in vitro model to induce the evolution of liver cancer cell lines (Fig. 3A). After more than 20 generations of exposure to NK cells under defined conditions (See details in Method section), we obtained NK cell-resistant liver cancer cells (NK-R cells), compared to wild-type cells (WT cells, non-NK-exposed). The NK-R liver cancer cells displayed robust resistance to NK cell cytotoxicity (Suppl. Movies 1–4, Fig. 3B, C and Fig. S3A, S3B), along with extensive down-regulation of NK-activating ligands and MHC-I molecules (Fig. S3C, S3D). Although IFN-γ is typically known to influence tumor immunity by upregulating MHC-I expression¹¹, our data showed that while transient NK cell cytotoxicity could temporarily elevate MHC-I levels in NK-R cells, the overall expression of MHC-I remained suppressed (Fig. S3E). As a possible consequence, NK-R cells also displayed increased resistance to T cell-mediated killing, as demonstrated by in vitro OVA-T cell cytotoxicity assay (Fig. S3F).

Furthermore, akin to observations in murine models, NK-R liver cancer cells also underwent comprehensive evolution in non-immunity-related tumorigenic characteristics, including growth, angiogenesis, metastasis, and tumor stemness (Fig. S3G–S3K). Additionally, morphological changes were observed, such as increased membrane blebbings¹² and an epithelial-mesenchymal transition appearance, alongside a significant decrease in immunological synapses between NK-R cells and NK cells (Fig. 3D, E). We also noted that NK cells were inhibited after contacting NK-R tumor cells, suggesting the presence of immunosuppressive signals from NK-R cells (Fig. S3L–S3N). These consistent phenotypic alterations collectively indicate that in vitro NK-R cells represent a suitable model for further investigation in liver cancer evolution.

Considering the high level of homogeneity of liver cancer cell lines, we conducted single-cell RNA sequencing (scRNA-seq) analysis of HepG2 NK-R and HepG2 WT cells to define the state of each tumor cell. The rationale for selecting HepG2 cells here has been previously outlined, as we consider them particularly advantageous for investigating early liver cancer evolution. The scRNA-seq results identified a sharp transcriptional discrepancy between NK-R and WT cells (Fig. 3F). Then, cells were clustered into cC1 to cC6 sub-populations (cell line clusters) via SNN-based clustering (Fig. 3G). We noticed that cC1-cC2 predominantly distributed in WT tumor cells, while cC5-cC6 concentrated in NK-R cells, with cC3-cC4 representing intermediate states (Fig. S3O). Analysis of the top differentially expressed genes demonstrated that NK-R cells had increased expression related to tumor stemness (such as CD24 and SOX4), lipid metabolism (such as FABP1 and ANGPTL8), and cellular redox regulation (such as GSTP1 and SELENOH), along with decreased expression of liver-specific metabolic molecules (such as AKR1C1, AKR1C2, HAMP and MT1E), indicating a loss of original epithelial traits (Fig. S3P). Consistently, typical oncological pathways were highly activated in NK-R cells (Fig. 3H).

These results all support that transcriptional evolution occurred in our in vitro NK-mediated model, and highlight a valuable in vitro model for investigating NK cell-mediated liver cancer evolution. This evolutionary process endows liver cancer cells with enhanced aggressiveness across multiple dimensions, fundamentally reshaping their cellular states.

Liver cancer patients undergo NK cell-mediated tumor evolution

To specifically observe the immune intrahepatic and early HCC evolution, the mouse models we employed above are all representing a healthy liver background. Given the complexity of various precancerous liver backgrounds in HCC patients, we need to carefully validate our findings regarding potential evolutionary phenotypes in real-world settings. We started with an analysis of The Cancer Genomic Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort. HCC patients from the TCGA cohort were classified based on transcriptomic similarity to either HepG2 NK-R or HepG2 WT cells. As expected, patients whose profiles more closely resembled the NK-R or less resembled the WT exhibited poorer prognoses. (Fig. 3I). Furthermore, by calculating the correlation between transcriptional similarity to cell line clusters (cC1 to cC6) and overall OS corresponding to TCGA samples from each cluster view, we observed a stepwise deterioration of OS from cC1 to cC6 (Fig. 3J). These results suggest that the cluster classification derived from the NK-R/WT cell model aligns with clinical outcomes in HCC patients. Specifically, cC1-like patients are associated with the most favorable prognosis, whereas progression toward cC6-like corresponds to the poorer prognosis.

Next, we established a new cohort of HCC patients from our center, categorizing each as either sensitive or resistant to cytotoxicity of NK cells (NK-S patients or NK-R patients, Fig. 3K). Briefly, primary HCC cells were collected from at least three distinct regions or sections of each tumor sample immediately after surgical resection, and subjected to cytotoxicity assays at an effector-to-target (E: T) ratio of 5:1. We defined NK-S patients as those whose primary HCC cells received NK92-MI cell-mediated cytotoxicity ≥1%, whereas NK-R patients received cytotoxicity < 1%. As anticipated, the majority of HCC patients had tumors resistant to NK cells (NKS: NKR = 5: 27), implying that many liver cancers may have already undergone immune training and evolutionary adaptation prior to clinical diagnosis. (Fig. 3K and Table 1).

Table 1.

The Clinical Characteristics of NK-S/R HCC Patients

Patients	Age	Sex	Tumor Size (cm)	HBV	CNLC Stage	Differentiation	Cytotoxicity%
NK-S1	69	M	2.9 × 2.7 × 2.2	+	IA	well-moderately	15.0
NK-S2	30	M	7.0 × 5.2 × 5.0	+	IB	moderately	2.9
NK-S3	70	M	8.0 × 6.0 × 5.0	+	IB	well-moderately	23.0
NK-S4	68	F	3.3 × 3.0 × 3.0	+	IB	moderately-low	32.5
NK-S5	73	M	6.1 × 5.7 × 5.4 2.2 × 2.2 × 1.8	+	IIA	well-moderately	5.6
NK-R1	79	M	4.8 × 3.5 × 2.3	+	IB	moderately	< 1.0
NK-R2	74	F	1.8 × 1.7 × 1.4	+	IA	well	< 1.0
NK-R3	67	M	7.0 × 5.8 × 4.5	+	IB	moderately-low	< 1.0
NK-R4	53	M	3.1 × 1.9 × 1.6	+	IB	moderately	< 1.0
NK-R5	32	M	8.9 × 8.5 × 7.2	+	IB	moderately	< 1.0
The above samples were submitted for single-cell RNA sequencing
NK-R6	40	M	2.5 × 1.9 × 2.0	+	IA	moderately	< 1.0
NK-R7	51	M	3.5 × 2.5 × 2.1	+	IB	well-moderately	< 1.0
NK-R8	53	M	17 × 13 × 10	+	IB	well	< 1.0
NK-R9	64	F	4.5 × 4.2 × 4.2	−	IB	moderately	< 1.0
NK-R10	39	M	3.7 × 3.7 × 3.4	+	IB	poorly	< 1.0
NK-R11	42	F	6.1 × 5.1 × 5.1	+	IB	moderately	< 1.0
NK-R12	73	M	5.2 × 4.5 × 4	+	IB	moderately	< 1.0
NK-R13	76	M	5.0 × 4.2 × 2.3	+	IB	well-moderately	< 1.0
NK-R14	74	F	3.2 × 2.8 × 2.6	−	IB	moderately	< 1.0
NK-R15	58	M	1.7 × 1.6 × 1.0	+	IA	moderately	< 1.0
NK-R16	76	M	3.5 × 2.5 × 1.9	+	IB	poorly	< 1.0
NK-R17	60	M	5.5 × 4.2 × 4.0	+	IB	well	< 1.0
NK-R18	44	M	1.8 × 1.8 × 2.0	+	IA	moderately	< 1.0
NK-R19	70	F	1.6 × 1.3 × 1.0	+	IA	poorly	< 1.0
NK-R20	64	M	3.6 × 2.5 × 2.3	+	IB	moderately-poorly	< 1.0
NK-R21	59	M	6.5 × 5.5 × 5.5	+	IB	well-moderately	< 1.0
NK-R22	71	M	4.0 × 3.5 × 2.5	+	IB	well	< 1.0
NK-R23	64	M	3.3 × 3.2 × 2.7	−	IB	moderately	< 1.0
NK-R24	71	M	5.0 × 4.5 × 4.0	+	IB	moderately	< 1.0
NK-R25	57	M	8.5 × 7.0 × 6.0	+	IB	well	< 1.0
NK-R26	49	M	3.1 × 2.5 × 2.2	+	IB	moderately	< 1.0
NK-R27	64	M	6.0 × 5.0 × 5.0	−	IB	moderately	< 1.0

Open in a new tab

To match HCC patient transcriptome with in vitro cell line we established, we selected five pairs of NK-S/NK-R samples with similar clinical features for scRNA-seq (Table 1 and Fig. S3Q). Principal component analysis (PCA) of patients’ epithelial cells revealed a clear demarcation between NK-S/NK-R tumors, except for patient S2, whose primary tumor cells showed only 2.9% NK-mediated cytotoxicity, compared to > 5% in all other NK-S patients (Table 1 and Fig. 3L). Next, we developed a machine learning classifier named the NK-based Tumor Evolution Prediction Model (NKTEPM), trained on the transcriptomic profiles of HepG2 WT and NK-R cells (Fig. 3M). Using NKTEPM, we classified each single human HCC cell as HepG2-WT-enriched, HepG2-NKR-enriched, mixed or unassigned state. The results showed that HepG2-NKR-enriched cells predominated in NK-R patients, while HepG2-WT-enriched cells were more prevalent in NK-S patients (Fig. 3N). Specifically, in NK-S patients, 59.01% of HCC cells exhibited high transcriptional similarity to HepG2-WT, whereas only 6.16% were classified as HepG2-NKR-enriched. Conversely, in NK-R patients, 59.69% of HCC cells were categorized as HepG2-NKR-enriched, with only 11.74% resembling HepG2-WT cells (Fig. 3N). Moreover, according to the most recent follow-up data, no NK-S patients have experienced recurrence, while 33.3% (9/27) of NK-R patients had tumor recurrence within 30 months after radical resection (Fig. S3R). These findings suggest that whether liver cancer cells have undergone NK cell-mediated evolution and their sensitivity to NK cell cytotoxicity may serve as critical indicators for HCC clinical outcomes. In parallel, our in vitro cell line evolution model (Fig. 3A) has been effectively validated by clinical cohorts, supporting its suitability for further mechanistic investigations.

Briefly, NK cell-induced transcriptional alterations consistently define this evolutionary process in both experimental models and clinical HCC patients across diverse liver pre-cancerous backgrounds. This transcriptomic reprogramming reshapes tumor cell states and facilitates the emergence and dominance of newly evolved subpopulations with fitness advantages.

Enhancement of cancer stemness is a prominent feature of NK cell-mediated tumor evolution

Using well-validated tumor cell line subpopulations (cC1–cC6), we conducted pseudotime analysis using Monocle¹³ and identified a clear evolutionary trajectory from cC1 to cC6 (Fig. 4A and Fig. S4A), which aligned with the increasing malignancy grades observed across clusters in HCC patients (Fig. 3I). Interestingly, among the various biological traits emerging along this trajectory, cancer stemness emerged as the most robust marker distinguishing HepG2-NKR from HepG2-WT cells, as well as cC1 from cC6, outperforming even the expression profiles of NK cell activation ligands (Fig. 4B, C and Fig. S4B, S4C). Cancer stemness was assessed using CytoTRACE analysis¹⁴ (a computational framework that can predict the differentiation state of single cell) and stemness-related gene signature scores. Along the Monocle-predicted trajectory from cC1 to cC6, we observed a gradual increase in cancer stemness accompanied by a progressive downregulation of NK cell activation ligands and MHC-I molecules (Fig. 4D). Consistently, NK-R patients from cohort of our center (Fig. 3J) also exhibited significantly elevated cancer stemness compared to NK-S patients (Fig. 4E), along with a marked loss of hepatocyte-related signatures (Fig. S4D).

Next, we revisited our early evolution model in humanized mice (Fig. 1A), which shows promise in identifying the initial signals or traits at the onset of intrahepatic immunosurveillance. We observed a gradual increase in CytoTRACE-predicted cancer stemness in implanted human liver tumors over time, beginning as early as the 1^st and 2^nd weeks—a period we defined as the onset of humanized immunosurveillance (Fig. 4F). Specifically, tumor cells collected at the 4^th week exhibited the highest degree of dedifferentiation, with this trend progressively developing from week 1 to week 4 (Fig. 4F). Furthermore, tumor subpopulations derived from the humanized mouse model also displayed a stepwise increase in cancer stemness from hC1 to hC7 (Fig. S4E).

Most importantly, when examining cancer stemness from a spatial trajectory perspective using the circle bin map (Fig. 4G), we found that the hC2 region exhibited higher levels of cancer stemness compared to the hC1 region (Fig. 4H). Alongside the elevated NK cell density and increased cancer stemness, we also noted a coordinated downregulation of hepatocyte markers and NK cell activation ligands from hC1 to hC2 (Fig. 4I, J). Additionally, a similar NK cell-induced enhancement of cancer stemness was observed across other solid tumor types (Fig. S4F, S4G).

In summary, we demonstrated that the enhancement of liver cancer stemness is a prominent feature of NK cell-induced tumor evolution, initiating as early as the initial stages in our model. These findings provide further insights into the intricate relationship between the immune microenvironment, cancer stemness, and the intrinsic plasticity of tumor cells¹⁵.

Cholesterol metabolism mediates liver cancer evolution induced by NK cells

To identify the mechanism of NK-mediated evolution, we developed a CRISPR/Cas9 whole-genome knockout library to pinpoint key regulators in tumor cells under NK-induced stress (Fig. S5A, see details in Method section). Briefly, wild-type HepG2 cells were transduced with a lentiviral vector containing a pooled library of guide RNAs targeting the entire genome (19,050 genes and 1864 miRNAs, with six gRNAs per gene). Following stable integration, the transduced HepG2 cells were either exposed to NK92-MI-mediated cytotoxicity or left untreated. Genomic DNA from both NK-stimulated and control cells was then extracted and sequenced for comparison.

Analysis of the top positively selected genes revealed that NK cells can induce significant changes in pathways related to lipid metabolism, epigenetic regulation, response to chemical stress, and WNT signaling (Fig. 5A and Fig. S5B). Furthermore, we also observed that, in the previously obtained PCA dimensional reduction result of spatial transcriptome, genes with positive loadings on PC_1 (PC_1 genes) are also closely related to lipid metabolism (Fig. S5C, S5D). Given the profound impact of metabolism in oncology, we hypothesized that there might be a rapid reprogramming of lipid metabolism when NK cells trigger liver cancer evolution.

Fig. 5 — A Using HepG2 cell-NK cell in vitro killing assay, a CRISPR/Cas9 whole-genome library screening was conducted. Functional enrichment analysis was performed on the top 100 candidate differential genes identified from the positive selection results (Benjamini-Hochberg test). These positively selected genes should have the potential function in enhancing tumor sensitivity to NK cell cytotoxicity. B Dot plot showing expression changes of all REACTOME metabolism pathways in hC2 versus hC1 (sub-population of humanized spatiotemporal liver cancer evolution model), colored by corresponding Cohens’ d value of two-sided t-test. hC2-hC1 comparison was performed by two-sided Student’s t-test. C Ridge plot displays distribution of lipid metabolism activity in spots from different hTumor clusters, with dash line indicating median value (P = 5.25 × 10⁻⁸⁸, two-sided unpaired Student’s t-test). D The positive candidate genes identified from CRISPR/Cas9 whole-genome library screening (HepG2 cell-NK cell in vitro killing assay) were intersected with genes associated with lipid metabolism. Subsequently, functional enrichment analysis was conducted on this intersected gene set (Benjamini-Hochberg test). E Network displays clusters of the top 200 enriched GO, KEGG terms, which was calculated via over-representation analysis and ranked by corresponding adjusted P value in tumor cells from NK-R patients versus NK-S patients (Benjamini-Hochberg test). Name of cluster is assigned based on the main biological theme of relevant terms. F. Intracellular total cholesterol and related metabolites of HepG2 WT and NK-R cells were quantified using targeted lipidomics mass spectrometry. The upper panel shows a heatmap of cholesterol peak intensities detected by mass spectrometry across different samples, while the lower panel displays the absolute quantification of cholesterol content in each sample (n = 4, P = 0.0088, two-sided unpaired Student’s t-test). Box plots show the median (centre line), the 25th and 75th percentiles (box bounds), and the minimum and maximum values (whiskers). G Bodipy probe (green) was used to detect intracellular neutral lipid of HepG2 cells, with Hoechst dye (blue) used to label cell nuclei. Scale bars, 5 μm. H Total cholesterol content of 2 × 10⁶ tumor cells was measured using the Amplex Red cholesterol assay kit and standardized (n = 3, mean ± SD, P = 0.0008, two-sided unpaired Student’s t-test). I Bodipy probe (green) was used to detect neutral lipid in NK-S and NK-R patients, with Hoechst dye (blue) used to label cell nuclei. Scale bars, 100 μm. J Total cholesterol content of 50 mg tumor tissues was measured using the Amplex Red cholesterol assay kit and standardized (n = 5, biological replicates, mean ± SD, P = 0.0026, two-sided unpaired Student’s t-test). K Single-cell transcriptomic analysis of cholesterol metabolism-related pathways in HepG2 WT and NK-R cells, as well as tumor cells from NK-S and NK-R HCC patients. Cholesterol metabolism pathways were subdivided into cholesterol uptake and cholesterol efflux processes. LXR-associated signaling, a central regulatory axis of hepatic cholesterol efflux, was prominently enriched. Data were compared using a two-sided Student’s t-test. L Heatmap (top) and smoothed line plots (bottom) display the expression trends of cholesterol metabolism-associated genes (SCAP, SREBF2, NR1H2, NR1H3, ABCA1, and MYLIP) as they vary with distance from the hC2-enriched region. The definition of circles and evolutionary trajectories in this panel is based on Fig. 1N, 1P. M, N Tumor cells were treated with LXR agonists (HepG2 NK-R with GW3965 [2 uM, P = 0.0025], T0901317 [100 nM, P = 0.0034], LXR-623 [2 uM, P = 0.0363], (M) or LXR inhibitors (HepG2 WT with GSK2003 [50 nM, P = 0.0342], SR9238 [200 nM, P = 0.0335], (N) for 48 h, followed by assessment of changes in tumor cell sensitivity to NK cell cytotoxicity (n = 3, biological replicates, mean ± SD, one-way ANOVA test). O Tumor evolution inhibiting assay: HepG2 WT sham treatment group served as the control group. The NK-induction group was stimulated with NK-92MI cells for five rounds (maintain about 30% survival rate of live HepG2 cells after each round of induction). The evolution inhibition group was based on the NK-induction group, with continuous addition of a low dose of GW3965. Compared with control group, NK-induction group exhibited resistance to NK cell-mediated cytotoxicity (P = 0.0009), which was effectively abolished by LXR agonist GW3965 (P = 0.3567, n = 3, biological replicates, mean ± SD, one-way ANOVA test). P Heatmap displaying Jensen-Shannon Divergence (JSD) measured similarity between samples from different groups, with transcriptome-based hierarchical clustering of all sample annotated (n = 3, biological replicates). Q, R Tumor stemness (Q, P = 0.0031, 0.0877, 0.0004, respectively) and cholesterol efflux capacity (R) of indicated group were evaluated by GSVA based on transcriptomic data (n = 3, biological replicates, One-way ANOVA test).

To investigate this, we calculated metabolic signals via scMetabolism¹⁶ and compared between hC1 and hC2, with results specifically highlighting lipid metabolism pathways (Fig. 5B). Lipid metabolic activity was also found to increase progressively along the evolutionary trajectory from hC1 to hC7 (Fig. 5C). To further dissect this, we examined all lipid-related genes that were positively selected in the CRISPR/Cas9 screen and observed a significant enrichment of cholesterol-associated pathways (Fig. 5D). Network analysis of the top 200 most significantly enriched Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathways between the NK-R and NK-S HCC patients also revealed the critical role of cholesterol metabolism in this situation (Fig. 5E).

Consistently, targeted lipidomics mass spectrometry analyzed intracellular total cholesterol and related metabolites of HepG2 WT and NK-R cells and revealed cholesterol accumulation in evolved HepG2-NKR cells compared to HepG2-WT cells (Fig. 5F). This phenotype was further validated by BODIPY staining and cholesterol quantification assays in both HepG2 cells and HCC patient samples (Fig. 5G–J). Membrane lipidomic profiling of the WT/NK-R cells also revealed significant alterations in membrane lipid composition, notably in cholesterol abundance (Fig. S5E).

Next, we conducted an in-depth investigation into cholesterol metabolism signals and discovered significant changes related to cholesterol transport and Liver X receptors (LXR)-associated genes in the context of NK cell-mediated models (Fig. 5K). More specifically, we found that genes involved in cholesterol efflux exhibited significant alterations in both the cell-based models and patient cohorts, and that LXR-associated signaling-a central regulatory axis of hepatic cholesterol efflux-was prominently enriched. This was consistent with our earlier analysis (Fig. 5D, E). LXRs are key regulators of cellular and systemic cholesterol homeostasis by controlling cholesterol efflux pathways.¹⁷. In evolved liver cancer cells (including NK-R cells generated by in vitro NK cell-stimulation, NK-R tumor isolated from HCC patient samples, and tumor derived from serial passaging in immunocompetent mice), we observed a marked downregulation of LXR expression (Fig. S5F), which occurred at very early in a spatiotemporal manner (1^st week time point in hHSC-NCG-hIL15 model), along with suppression of other LXR-associated mediators, such as ABCA1 and MYLIP (Fig. 5L). Furthermore, in vitro induction of liver cancer evolution by NK cell or perforin stimulation also resulted in increased intracellular cholesterol and reduced LXR expression (Fig. S5G–S5K). Similar trends were observed in other solid tumor types (Figs. S4F and S5L).

To assess the functional role of LXR signaling, we treated NK-R/WT liver cancer cells with LXR agonists/inhibitors. Notably, NK cell-mediated cytotoxicity against these cells varied in parallel with changes in cholesterol levels (Fig. 5M, N and Fig. S5M, S5N). More importantly, treatment with the LXR agonist GW3965 not only restored sensitivity of evolved liver cancer cells to NK cell cytotoxicity but also maintained a transcriptomic profile more similar to wild-type rather than evolved cells (Fig. 5O, P, Fig. S5O). Also, GW3965 treatment successfully modulated cancer stemness and promoted cholesterol efflux (Fig. 5Q-5R).

These findings suggest that liver cancer cells preferentially reprogram lipid metabolism, particularly by retaining more intracellular cholesterol, to evolve and survive the stress induced by NK cells. This accumulation of cholesterol primarily results from the suppression of the LXR-mediated efflux process. Activating cholesterol efflux process impeded liver cancer evolution in our models, offering a potential target for better controlling evolved tumor cells.

LXR activation enhances the efficacy of Anti-LAG-3 immunotherapy in HCC

To explore the potential of LXR activation therapy in slowing tumor evolution, we analyzed LXR expression using the TCGA pan-cancer transcriptome profilers. The results confirmed a general downregulation of LXR across nearly all cancer types, accompanied by a marked increase in cancer stemness signatures, including in HCC (Fig. 6A). Additionally, LXR reduction and heightened cancer stemness along with tumor progression were also observed in several TCGA cohorts (Fig. 6B and Fig. S6A).

Fig. 6 — A Heatmap displaying the transcriptome level of signature genes in tumor tissues versus normal tissues across TCGA cohorts (right). The pie chart (left) illustrates the proportion of different alteration patterns of the target gene across all cancer types (red indicates the proportion of cancers with gene upregulation, blue with downregulation). B Expression trends of the target gene (related to cholesterol efflux and tumor stemness) across different tumor stages in the TCGA-LIHC cohort. C Representative spatial transcriptomic maps from HCC patients treated with immune checkpoint inhibitors (GSE238264), analyzed using the NKTEPM algorithm (left). Tumor cell states were classified based on transcriptional similarity to HepG2 WT-like or HepG2-NKR-like signatures, and their spatial distribution is shown for responder and non-responder samples. A bar plot summarizes the proportions of HepG2-WT-like and HepG2-NKR-like tumor cells across all 7 patients (right, mean ± SD, P = 0.0003, two-sided Two-way ANOVA). D Ridge plot showing the expression level of previously reported cell state signatures in NK cells⁴ and CD8⁺ T cells¹⁹ from NKS and NKR patients, with median value indicated by dash line (For NK cell, P = 1.0762 × 10⁻³⁶, 1.0576 × 10⁻²⁸, 3.7638 × 10⁻²⁰, respectively; For CD8⁺ T cells, P = 3.6275 × 10⁻¹²², 8.6077 × 10⁻²⁰, 2.6678 × 10⁻³³, respectively; two-sided unpaired Student’s t-test). E CellPhoneDB inferred cell-cell interactions playing co-inhibitory function between tumor cell and NK cell (left), tumor cell and CD8⁺ T cell (right). Color and size of each dot represent mean expression value of relevant ligand-receptor genes in corresponding patient group. F scRNA-seq analysis of LAG3 expression in NK (P = 8.50 × 10⁻¹⁴) and CD8⁺ T (P = 1.30 × 10⁻⁹²) cells from HCC patients stratified into NK-S and NK-R groups (two-sided unpaired Student’s t-test). G Schematic of treatment model in hLAG3-C57BL/6 mice: Tumors (Hepa1-6 from NCG or *Rag1*^−/− mice) are implanted at equal volumes into the bilateral axillae of mice. One week post implantation, respective treatment regimens commence for each group. Tumor volumes are monitored every 3 days over a period of 3 weeks. H, I Tumor volume trends for each group were represented by distinct colored curves (H) and subjected to quantitative analysis (I, two-sided paired t-test was used for NCG vs. *Rag1*^−/− tumors, control-control, P = 0.0002; while two-sided unpaired t-test for within-group comparisons, NCG control-dual-drug, P = 0.0001; *Rag1*^−/−: control-GW3965, P = 0.0054, control-Relatlimab, P = 0.0002, control-dual-drug, P = 0.0001, n = 6, mean ± SD).

Current ICI therapies for liver cancer continue to suffer from unsatisfactory response rates, driven by a variety of contributing factors¹⁸. We then analyzed the spatial transcriptomes of seven HCC patients who underwent ICI treatment (GSE238264, nivolumab, PD-1 antibody) using our NKTEPM algorithm. Notably, responders exhibited a distinct population of HepG2 WT-like tumor cells, while non-responders’ tumors were exclusively composed of HepG2-NKR-like cells (Fig. 6C and Fig. S6B). This phenotype implied that evolved HCC mediated by NK cell could be more resistant to current ICI treatment. Next, we examined LXR expression in cancer patients undergoing systemic ICI treatment for some clues. We found that ICI-nonresponsive patients typically exhibited lower LXR expression levels (Fig. S6C). This led us to hypothesize that LXR activation could support immunotherapy by inhibiting tumor evolution.

We then revisited the immune microenvironment in NK-S/NK-R HCC patients and observed that both infiltrating NK cells and CD8⁺ T cells were functionally inhibited in NK-R patients (Fig. 6D)¹⁹. Meanwhile, the coinhibitory signaling between evolved tumor cells and NK or CD8⁺ T cells through the LAG-3/FGL-1 axis was significantly enhanced (Fig. 6E and Fig. S6D), with elevated LAG-3 expression in both NK and CD8⁺ T cells (Fig. 6F).

Recently, the combination of anti-LAG-3 and anti-PD-1 has garnered significant attention and is considered a promising approach to enhance the efficacy of immune checkpoint inhibitors in solid malignancies²⁰. However, anti-LAG-3 monotherapy has shown unsatisfactory results in clinical settings, with no clear explanation. To further evaluate the clinical impact of our findings on liver cancer evolution, we designed a combined regimen to explore the potential of Relatlimab (anti-hLAG3) plus GW3965 (LXR agonist) in controlling HCC (Fig. 6G). By treating Hepa1-6 P4 tumors from Rag1^−/− or NCG mice on hLAG3-C57BL/6 backgrounds, we observed that monotherapy with either Relatlimab or GW3965 failed to achieve satisfactory efficacy (Fig. 6H, I). However, the combination of this immunotherapy and the metabolic agonist appeared to have a synergistic effect in the mouse model, particularly against evolved HCC (Fig. 6H, I).

IHC analysis showed that tumors from Rag1^−/− mice exhibited higher stemness marker expression than those from NCG mice, with minimal NK cell infiltration observed in both models. CD8⁺ T cells were nearly absent in Rag1^−/− tumors but detectable at low levels in NCG tumors. Combination therapy reduced stemness marker expression and increased intratumoral NK and CD8⁺ T cell infiltration (Fig. S6E). Consistently, tumors from Rag1^−/− mice displayed higher total cholesterol levels compared to NCG-derived tumors. Treatment with Relatlimab plus GW3965 markedly decreased intratumoral cholesterol accumulation (Fig. S6F).

While the evolution described here should not be entirely blocked due to the essential role of NK cell-mediated immunosurveillance, the potential combined regimen of anti-LAG-3 antibody and LXR agonist offers a glimmer of hope in this challenging scenario. By using an LXR agonist to inhibit the ongoing evolution, anti-LAG-3 treatment demonstrated improved efficacy in immunotherapy.

Discussion

Liver cancer, particularly HCC, presents a unique challenge due to its distinct immune microenvironment and genomic background, which complicate the understanding of its evolutionary trajectories²¹. In our study, we utilized various cell lines and mouse models to simulate the potential early evolutionary pathways associated with the immune system. However, it is important to note that many of our results were generated from HepG2 cells and mice with a healthy liver background. As previously mentioned, HepG2 cells have been widely used to investigate various aspects of liver cancer biology. More importantly, the careful validation of our findings with additional models, including different HCC patient cohorts, ensures that our conclusions are broadly applicable. The same consideration applies to the issue of different precancerous liver conditions. While the primary aim of our study was to explore the potential link between early tumor immunity mediated by NK cells and liver cancer evolution, the specific influence of liver precancerous conditions on tumor progression warrants in-depth investigation in future studies, as these conditions involve a broad array of factors. Fortunately, the clinical patient analysis in our study encompassed these conditions and may offer valuable insights for subsequent research.

Although researchers have made efforts to dissect the interaction between early-stage tumor cells and the immune system using various methodologies, we believe that NK cells exert relatively stronger and more direct stress on tumor cells due to their cytotoxicity, which usually leads to cell death. This phenotype is likely to be more dominant in HCC, given the high abundance of NK cells in the liver²². In contrast, macrophages and other potential players may create a pro-tumorigenic niche for early tumors²³, thereby reshaping and driving tumor evolution through distinct mechanisms. Moreover, the cytotoxic activity of NK cells-which, based on our results, appears critical for reshaping HCC evolution-suggests that conventional circulating NK cells are more likely to mediate this pro-tumorigenic influence in our study, rather than the immunosuppressive liver-resident NK cells. However, the functions of these immune cells, often regarded as extrinsic factors that influence the evolutionary trajectory of HCC but remaining largely unexplored, should not be considered in isolation²¹. This highlights the need for more advanced models and methodologies to fully understand their role.

Cancer stemness, a key feature during tumor development, was previously described as “cancer stem cells (CSCs)”. Recent updates indicate that enhancement of tumor stemness is closely associated with tumor cell states and other characteristics that contribute to tumor evolution¹⁵. In the context of immune-mediated tumor evolution, metabolic reprogramming toward increased cholesterol accumulation may therefore create a permissive membrane and signaling environment for stemness acquisition. Consistently, our transcriptomic analyses show concurrent upregulation of cholesterol-associated metabolic programs, stemness signatures, and Wnt/β-catenin signaling in NK cell–selected tumor cells and evolutionarily advanced tumors^24,25. These findings suggest that lipid metabolic rewiring may actively contribute to the emergence of stem-like tumor states during early tumor evolution under immune pressure.

In our study, a primary challenge that liver cancer cells must overcome is NK cell-mediated intrahepatic immunosurveillance. Unfortunately, surviving liver cancer cells evolve, acquiring stemness, immune escape, and other traits necessary for their proliferation. One of the most concerning features of HCC evolution mediated by NK cells is its potential resistance to T cell-mediated immunosurveillance, especially considering that a significant portion of efforts made by oncologists are focused on T cells in liver cancer.

Fortunately, we have identified a significant upregulation of LAG-3 in both infiltrated NK cells and CD8⁺ T cells during such evolution. This finding provides an opportunity to test the efficacy of LAG-3-based ICI treatments in HCC. Relatlimab, the only approved hLAG-3 monoclonal antibody, has been used in combination with the PD-1 inhibitor nivolumab for treating unresectable or metastatic melanoma²⁶. Given its promising prospect in solid malignancies but limited efficacy as a monotherapy, our results suggest a potential new direction for developing combination regimens, particularly for evolved HCC.

In our study, cholesterol metabolism was found to be involved in NK-induced evolution. Our study suggests that LXR agonists could act as sensitizers for ICI therapy by inhibiting the ongoing evolution of HCC cells induced by the immune system or immunotherapy. Notably, other groups have initiated clinical trials to evaluate the combination of the LXRβ agonist abequolixron and PD-1 inhibitors in solid tumors (ClinicalTrials.gov ID NCT05911308 and NCT02922764), and our findings support such efforts from an evolutionary perspective. At the same time, given the potential metabolic adverse effects associated with metabolism-targeting agents, their suitability for long-term use requires careful evaluation. Additional evidence from rigorously designed clinical trials will be essential to define the therapeutic window, safety profile, and optimal treatment strategies for such combinations.

While we have illustrated early immunosurveillance-mediated liver cancer evolution in terms of phenotype, mechanism, and clinical relevance, several limitations require further study. First, we need to verify our theory in different malignancies, particularly comparing NK cell-enriched tissues with NK cell-deficient tissues²². Second, in-depth investigations into the epigenetic or genetic regulation of NK-mediated evolution are important, although separating genetics and epigenetics within the complex context of tumor evolution is challenging²⁷. Third, the strategy involving anti-LAG-3 and LXR agonists requires further investigation, particularly in relation to enhancing the anti-PD-1 regimen. Finally, if uncontrollable and unfavorable evolution begins too early for physicians to detect and block, the evolution triggered by therapeutic interventions—whether drug, surgery, radiation, or other treatments will become the main issue. In the future, considering the evolutionary side effects of specific treatments should be integral to assessing therapeutic efficacy.

Methods

Patients

A cohort of 32 HCC patients categorized as either vulnerable or resistant to NK cells was established, which were underwent surgical treatment at the Sir Run Run Shaw Hospital from December 2021 to February 2023. Tumor specimens were obtained from surgically resected primary HCC lesions, and viable tumor tissue was sampled from the active central region of the tumor immediately after resection. Extraction of tumor primary cells was completed within 2 h after specimen resection. Tumor cell viability was assessed using trypan blue staining, and tumor samples with viability greater than 50% underwent NK cell-primary tumor cell cytotoxicity assays. Patients were classified as NK-S or NK-R based on the cytotoxicity results. To minimize batch effects, simultaneous cytotoxicity experiments will be conducted using NK92MI cells against primary tumor cells from patients, alongside NK92MI-HepG2 assays. Standardization procedures will be implemented accordingly.

This study conformed to the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Sir Run Run Shaw Hospital. All patients signing Informed Consent before surgery for the use of their tissues for scientific research.

Mice

C57BL/6 mice were purchased from Qizhen Experimental Animal Technology (Hangzhou, CN). NOD/ShiLtJGpt-Prkdc^em26Cd52Il2rg^em26Cd22/Gpt (NCG) mice, C57BL/6Gpt-Rag1^em1Cd3259/Gpt (Rag1^−/−) mice, hHSC-NCG-hIL15 (immune-humanized) mice, and C57BL/6JGpt-H11^{em1Cin(Tcra&Tcrb)}/Gpt (OT-1) mice were purchased from GemPharmatech (Nanjing, CN). C57BL/6-hLAG3 mice were purchased from Biocytogen (Nantong, CN). Mice were housed under specific pathogen-free conditions with a 12 h light/dark cycle at an ambient temperature of 22 ± 2 °C and relative humidity of 50–60%. This study was approved by Zhejiang University Laboratory Animal Welfare and Ethics Review Committee.

Materials (antibodies, primers and etc.)

The relevant information is provided in Supplementary Data. 1.

Spatiotemporal orthotopic tumor evolution model in immune-humanized mice

In preliminary experiments, we found that direct intrahepatic injection of HepG2 cell suspension into immune-humanized hHSC-NCG-hIL15 mice failed to form tumors efficiently. To improve tumor formation rate, 2 × 10⁶ HepG2 cells were first subcutaneously injected into 8-week-old male NCG mice, which are severely immunodeficient, to allow tumor formation in an immune-free environment and minimize early interaction with immune cells.

After two weeks, subcutaneous tumors were surgically excised, and necrosis-free regions were dissected into 1 mm × 1 mm × 1 mm fragments. These tumor fragments were orthotopically implanted into the right hepatic lobe of 13-week-old immune-humanized hHSC-NCG-hIL15 mice. All implanted tumor tissues were derived from the same donor mouse to ensure consistency among recipients. No significant bleeding or other complications were observed during the surgical procedures.

Following implantation, mice were sacrificed at designated time points (1–4 weeks post-implantation). Liver tumors were harvested for spatial transcriptomic profiling and molecular analyses to investigate tumor progression and dynamic interactions with human immune components. This model allowed us to trace the temporal transcriptomic evolution of tumor and to identify distinct molecular subpopulations and critical transition windows during early-stage tumor development.

Two independent experimental batches were conducted. In each batch, four immune-humanized mice (hHSC-NCG-hIL15) were used, and each mouse was sacrificed at a distinct post-implantation time point (1, 2, 3, or 4 weeks). At the 3-week time point, tumor formation was observed in only one mouse across the two batches. Therefore, samples obtained for downstream analyses were designated as 1W1 and 1W2, 2W1 and 2W2, 3 W (single sample), and 4W1 and 4W2.

Sequential orthotopic tumor passage model

2 × 10⁶ liver cancer cells (Hepa1-6 or HepG2 cells, P0) were orthotopically injected into the liver of 8 weeks old NCG mice, with the injection volume kept below 50 µL. After 2 weeks, the tumors were surgically removed and defined as P1 tumor. Subsequently, tumor fragments (2 mm × 2 mm × 2 mm) were carefully separated from non-necrotic regions of P1 tumor and orthotopically implanted into the right hepatic lobe of recipient mice (NCG, Rag1^−/−, or hHSC-NCG-hIL15), and no significant active bleeding was observed throughout the entire surgical procedure. After another 2 weeks, tumors were harvested and serially passaged to generate P2 to P4 tumors.

To inhibit T cell function in immune-humanized mice, we pre-treated the mice with Cyclosporin A (15 mpk, i.p., every two days) and Teplizumab (0.2 mpk, i.p., every two weeks) starting 1 week prior to tumor implantation and continued throughout the experiment. Tumor tissues were subsequently collected for RNA extraction, protein analysis, or sequencing, depending on the experimental objectives.

In vitro tumor evolution model

WT liver cancer cells (including HepG2 and HCCLM3) were co-cultured with NK92MI cells (human NK cell line) at an effector-to-target (E:T) ratio empirically determined to result in approximately 30% survival of the tumor cells after 24 h of incubation. The E:T ratio was not fixed, but dynamically optimized to achieve a consistent selective pressure, with the goal of approximately 70% killing efficiency. At the initiation of the induction process, the E:T ratio was typically in the range of 1–2:1, and was gradually increased up to 10:1 over successive rounds of selection as tumor cells acquired increased resistance. Following co-culture, the surviving tumor cells were gently washed, collected, and allowed to recover and proliferate in fresh complete medium for 48–72 h until reaching about 80% confluence. The surviving cells were subsequently passaged and subjected to the next cycle of NK cell-mediated induction. In parallel, control WT cells were passaged under standard conditions, with approximately 70% of the population randomly removed during each passage to simulate population bottlenecking without immune selection pressure.

The stimulation and recovery cycle was repeated more than 20 times. After this serial stimulation, we obtained tumor cell that exhibited stable resistance to NK92MI-mediated cytotoxicity, hereafter referred to as NK-R cells. These NK-R cells maintained their resistance phenotype over multiple passages, even in the absence of continuous NK cell exposure.

Cell culture

HepG2, HCCLM3, Hepa1-6, and HEK293T were cultured in Dulbecco’s Modified Eagle Medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin-streptomycin under standard conditions (37 °C in a humidified incubator with 5% CO₂). Cells were routinely passaged at 70-80% confluence using 0.25% trypsin-EDTA (Gibco) and were regularly tested to confirm the absence of mycoplasma contamination. NK92MI cells were maintained as a suspension culture in NK cell-specific medium (Alpha MEM supplemented with 12.5% FBS, 12.5% horse serum, and 1% penicillin-streptomycin) under standard conditions, and all experiments were performed using cells in logarithmic growth phase.

Tumor vaccination murine model

To establish an adaptive immune response against Hepa1-6 cells, 8-week-old male C57BL/6 mice were subcutaneously inoculated in the upper axillary region with 5 × 10⁵ Hepa1-6 cells. After 3 weeks, the primary tumors were surgically excised to allow immune memory development.

For tumor rechallenge, tumor fragments (2 mm × 2 mm × 2 mm) derived from P4-generation tumors previously passaged in immunodeficient NCG or Rag1^−/− mice were subcutaneously implanted in parallel into the lower axillary regions of the vaccinated mice. Tumor engraftment and growth were monitored and recorded over 4 weeks to assess tumor formation rates and volumetric changes.

In vitro cytotoxicity assay

Tumor cells were labeled with carboxyfluorescein succinimidyl ester (CFSE) according to the manufacturer’s instructions. Labeled tumor cells were seeded at a density of 1 × 10⁵ cells per well. Effective cells (NK92MI [human NK cell line] or primary murine NK cell [isolated from the spleens of C57BL/6 mice] or purified OT-1 T cells [isolated from the spleens of OT-1 mice]) were then added to the wells at varying E:T ratios. The co-culture was incubated at 37 °C in a humidified incubator with 5% CO₂ for 4–24 h, depending on the experimental design. After incubation, cells were stained with propidium iodide (PI) to identify dead cells. The samples were analyzed using flow cytometry. CFSE-positive tumor cells and PI-positive dead cells were quantified to determine the percentage of tumor cell lysis (1).

cytotoxicity % = \frac{Experimental PI / CFSE positive cells - Spontaneous PI / CFSE positive cells}{Total CFSE positive cells} \times 100 %

For T cell cytotoxicity assay, OVA-specific TCR transgenic OT-1 cells were isolated from OT-1 mice, and the OVA-derived peptide SIINFEKL was used to activate OT-1 cells. OVA-Hepa1-6 cells were generated by the lentiviral vector OVA.

NK cell depletion

NK cell depletion was performed using anti-Asialo GM1 antibody. Mice received intravenous injections starting one week prior to tumor implantation, followed by continued administration twice per week to maintain sustained NK cell depletion throughout tumor growth and serial passaging experiments. Each injection consisted of 20 uL of anti-Asialo GM1 antibody diluted in sterile PBS. The control group was treated with IgG isotype antibody using the same dosing schedule. The efficiency of NK cell depletion was validated by flow cytometric analysis of NK cell populations (FSC/SSC gating was used to exclude debris and identify lymphocytes based on size and granularity after red blood cell lysis, DX5⁺ cells were then quantified within the lymphocyte gate) in peripheral blood, confirming effective and sustained NK cell depletion during the experimental period.

Scanning Electron Microscopy (SEM) Imaging

Tumor cells were seeded on glass coverslips at a density of 5 × 10⁴ cells per well. NK92MI cells were added to the wells at an E:T ratio of 2:1. The co-culture was incubated for 1 h at 37 °C. Then, sample were washed with phosphate buffer (0.1 M, pH 7.4), and fixed in 2.5% glutaraldehyde solution overnight at 4 °C. Next, samples were fixed with 1% osmium solution for 1.5 h and washed 3 times with PBS, followed by a graded series of ethanol (50%, 70%, 90% and 100%) and dried by Critical Point Dryer CPD 300 (Leica). Finally, dried samples were sputter coated with Pt-palladium (8 mA, 60 s) by high vacuum ion sputtering apparatus (Quorum Q150T ES plus) and observed with a scanning electron microscope Nova Nano 450 (Thermo FEI) at 5 kV.

CRISPR/Cas9 whole-genome knockout library screening

HepG2 WT cells are transduced with a lentiviral vector containing a pooled library of guide RNAs (gRNAs) targeting the entire genome (19,050 genes and 1864 miRNAs, 6 × gRNAs/gene). After selection with 2 µg/mL puromycin to ensure stable integration, the cells are subjected to the corresponding conditions (NK92MI-cytotoxicity group: E:T ration = 5:1, the NK92MI cells and dead cells were washed away when the killing rate reached about 50%, then live cells were collected and counted. Control group: Routine culture, the excess cells were randomly discarded when the cells were collected, ensuring the remaining cell count matched with NK92MI-cytotoxicity group). Genomic DNA of surviving cells is extracted. The integrated gRNA sequences are amplified by PCR and sequenced. The sequencing data is analyzed to identify gRNAs that are enriched (cytotoxicity enhancement) or depleted (cytotoxicity resistance).

Cholesterol measurement by Amplex Red assay

Total cholesterol levels in cells or tissue samples were measured using the Amplex Red Cholesterol Assay Kit (Thermo Fisher Scientific) according to the manufacturer’s instructions. Briefly, cultured cells were harvested and lysed, or tumor tissues were homogenized in assay buffer to obtain clarified lysates. Protein concentrations were determined for normalization where indicated.

Sample lysates were incubated with the Amplex Red reagent mixture containing cholesterol oxidase, cholesterol esterase, and horseradish peroxidase, allowing cholesterol-dependent generation of the fluorescent product resorufin. Fluorescence intensity was measured using a microplate reader (excitation/emission: 560/590 nm). Cholesterol content was calculated based on a standard curve and normalized to cell number or protein content as specified in the figure legends.

Quantitative real-time PCR (RT-qPCR)

Total RNA was extracted from cells or tissue samples using standard RNA isolation methods. RNA concentration and purity were assessed prior to reverse transcription. Complementary DNA (cDNA) was synthesized from total RNA (2 ug) using a reverse transcription kit according to the manufacturer’s instructions. RT-qPCR was performed using SYBR Green–based detection on a real-time PCR system. Gene-specific primers were used as indicated. Relative gene expression levels were calculated using the 2^−ΔΔCt method, with housekeeping genes serving as internal controls for normalization. All reactions were performed in technical triplicates, and at least three independent biological replicates were analyzed for each condition.

Bulk RNA sequencing

For bulk RNA sequencing, tumor cell or tissue samples were collected as indicated, and total RNA was extracted according to standard procedures. RNA quality and quantity were assessed prior to sequencing. Library preparation and high-throughput RNA sequencing were performed by a commercial service provider (Lianchuan Biotechnology Co., Ltd., Hangzhou, China) following standard Illumina protocols.

Briefly, poly(A) + RNA was enriched from total RNA, fragmented, and reverse-transcribed to generate cDNA libraries, which were then end-repaired, adaptor-ligated, and PCR-amplified. Sequencing was conducted on an Illumina platform to generate paired-end reads. Raw sequencing data were subjected to quality control, adaptor trimming, and alignment to the reference genome by the service provider, and gene-level expression matrices were generated.

Downstream bioinformatic analyses were performed by the authors. Differential gene expression was assessed by comparing normalized expression levels of target genes between experimental groups. For pathway-level analyses, Gene Set Variation Analysis (GSVA) and Gene Set Enrichment Analysis (GSEA) were conducted using curated gene sets to evaluate the relative activation of specific signaling pathways. All analyses were performed using R software with appropriate packages, and statistical significance was assessed as described in the figure legends.

scRNA sequencing and snRNA sequencing

Library construction

For scRNA-seq, single-cell suspensions were loaded to 10x Chromium to capture 10000 single cells according to the manufacturer’s instructions of 10X Genomics Chromium Single-Cell 3 kit (V3). For snRNA-seq, nuclei were isolated with Nuclei EZ Lysis buffer (NUC-101; Sigma-Aldrich) supplemented with protease inhibitor (5892791001; Roche) and RNase inhibitor (N2615; Promega and AM2696; Life Technologies). The following cDNA amplification and library construction steps were performed according to the standard protocol. Libraries were sequenced on an Illumina NovaSeq 6000 sequencing system (paired-end multiplexing run,150 bp) by LC-Bio Technology Co. Ltd. (HangZhou, China) at a minimum depth of 20,000 reads per cell.

Read alignment and quality control

Raw fastq data of scRNA-seq and snRNA-seq were preprocessed using CellRanger software (v7.1.0, 10x Genomics) against the pre-built GRCh38 human reference genome (GENCODE v32/Ensembl98). The cell-by-gene count metrices obtained from CellRanger’s output were subsequently imported into R (v4.2.0) with Seurat package (v4.2.0) for downstream analyses²⁸. Quality control was performed to filter out low-quality cells. For HepG2 samples (scRNA-seq), cells with high percentage of mitochondrial genes (percent.mt ≥ 20%), cells with Unique Molecular Identifier count (nUMI) less than 4000 or more than 50,000, cells with detected genes (nGene) less than 2000 or more than 80,000 were removed. For patient samples (snRNA-seq), considering the high cross-patient heterogeneity, threshold for nUMI and nGene of each sample was decided by runScStatistics of scCancer package (v2.2.1) respectively, also cells with nUMI ≤ 400, nGene ≤ 500 or percent.mt ≥ 4% were discarded²⁹.

scRNA-seq data processing

The filtered count matrices were integrated and processed mainly with Seurat package. Data normalization, highly variable gene (HVG) detection, data scaling, and principle component analysis (PCA) was conducted with Seurat’s NormalizeData (normalization.method = ‘LogNormalize’), FindVariableFeatures (nfeatures = 5000), ScaleData (regress out ‘nCount_RNA’), and runPCA function. Batch effect was then corrected by harmony algorithm (v0.1.1) based on precomputed PCA dimensions³⁰. Subsequently, the top 25 harmony dimensions were taken as input for Seurat’s RunUMAP to achieve 2D Uniform Manifold Approximation and Projection (UMAP) visualization. Following that, shared nearest neighbor (SNN) based cell clustering was run by Seurat’s FindNeighbors and FindClusters (resolution = 0.3) function based on the top 25 harmony embeddings.

snRNA-seq processing and annotation

Considering the high cross-patient variability, integration of snRNA-seq was conducted following Seurat’s ‘anchor-based’ integration workflow. In brief, for each sample, data normalization (normalization.method = ‘LogNormalize’), HVG detection (nfeatures = 2000), and data scaling (regress out ‘nCount_RNA’ and ‘percent.mt’) were performed respectively as described before. After that, anchors for data integration were identified by FindIntegrationAnchors with ‘cca’ reduction. All patient samples were integrated together with IntegrateData function based on the top 30 dimensions. PCA analysis was then performed with RunPCA. Following that, UMAP visualization, cell clustering was carried out by RunUMAP FindNeighbors and FindClusters (resolution = 0.1) based on the top 30 PCA dimensions. Cell types were annotated manually based on canonical marker gene expression. For specific cell types, including epithelial, NK cell, and T cell, above procedure was repeated for subtype annotation.

Spatial transcriptome sequencing

Library construction

The samples were sectioned coronally with a cryostat at 10 mm thickness (Leica, CM1950). Tissue sections were then fixed on a spatially barcoded array with 3.6–3.8% formaldehyde solution at room temperature for 10 min, then treated for one min with isopropanol and airdried. To acquire the appropriate areas for Visium Spatial slides, we quick H&E staining of series of sections. Tissue fixation, HE staining, and bright-field imaging were used to establish optimal conditions for permeabilization. Tissue sections were placed on Visium spatial tissue optimization slides (10x Genomics) with six capture areas covered, one as a positive control and one as a negative control. After tissue staining, imaging and pre-permeabilize, tissues were immediately permeabilized and followed by in situ reverse transcription. Finally, libraries were sequenced on the Illumina Nextseq platform using paired-end sequencing.

Read alignment

Raw fastq data and H&E images of spatial transcriptome (ST) were processed using SpaceRanger (v2.1.1, 10x Genomics). Considering the hybrid genetic background of immune-humanized mice, a customized reference genome combining human GRCh38 (GENCODE v32/Ensembl98) and mouse mm10 (GENCODE vM23/Ensembl98) reference genome was constructed using SpaceRanger’s mkref function for read alignment. Besides, considering that transcripts within tumor region should be mainly of human origin, a separate sequence alignment against the GRCh38 reference genome was also performed. At the same time, species classification (human/mouse) of sequencing reads was evaluated using Xenome. To screen out human transcripts dominated spots for subsequent analysis, we measure the proportion of human-origin transcripts in each spot by defining the human ratio with the following formula:

human ratio = \frac{number of Xenome classfied human derived reads}{number of SpaceRanger quantified human / mouse derived reads} \times 100 %

Region annotation and quality control

The SpaceRanger-generated cloupe file was then imported into Loupe Browser software (v7.0.1, 10x Genomics). For each sample, the graph-based clustering was performed based on the alignment result obtained via the hybrid reference genome. According to the H&E images as well as the gene expression of each cluster, spots outside the main region were first discarded and labeled as ‘drop’, then the left spots were preliminarily classified into human-origin tumor (hTumor), peri-tumor (peri-hTumor), mouse-origin hepatocyte (mHep), mouse-origin red blood cell (mRBC), and mouse-origin stromal cell (mStromal) regions. In brief, we classified spot clusters presenting high ‘human ratio’ as well as high expression of canonical HCC marker genes (GPC3, AFP, EPCAM, ALB) as ‘hTumor’, and manual adjustments were performed to this region according to corresponding H&E images. To facilitate the study of immune infiltration in the tumor periphery, we categorized the region extending three spots away from the ‘hTumor’ area, which exhibit moderate ‘human ratio’, as ‘peri-hTumor’. Regions further away from the ‘peri-hTumor’ area were classified as ‘mHep’, which representing specific expression of mouse hepatocyte signatures (Alb, Ttr). At the same time, in the 4 W samples, we observed significant intratumoral heterogeneity within the ‘hTumor’ area. Thus, clusters within the ‘hTumor’ regions but with high expression of red blood cell (RBC) marker genes (Hbb-a1, Hba-a2) or stromal signature genes (Col1a1, Vim) were further classified as ‘mRBC’ and ‘mStromal’ region respectively. In the downstream analysis, we mainly focused on the hTumor and surrounding peri-hTumor area, thus the spot-by-gene count matrix generated by SpaceRanger based on the GRCh38 reference genome was then imported into R and processed with Seurat (v5.0.1)³¹. Low-quality spots, including spots with high percentage of mitochondrial genes (percent.mt ≥ 10%) or hemoglobin genes (precent.hb ≥ 2%) and spots with nUMI ≤ 500 were all excluded from further analysis.

Data integration and clustering

Profile of all samples was integrated after data normalization performed with SCTransform function. Then we extracted spots within the hTumor regions and performed PCA and clustering via Seurat’s runPCA, FindNeighbors (dims = 1:11), and FindClusters (resolution = 0.1) function. hTumor regions spots was then classified into hC1-hC6 clusters.

ST analyses

SpatialDecon (v1.8.0) was run to estimate tumor purity and immune infiltration within ST spots (lower tumor purity indicates increased infiltration of immune and stromal cells)². Annotation of the hC2-enriched ROI area (‘Core’) was performed using SPATA2 toolkit (v2.0.4)³². Spots surrounding the selected ‘Core’ region was then screened and binned into Circle 1-4 by distance to the “Core” area via plotSurfaceIAS function. Spatial trajectory was constructed with SPATA2’s createSpatialTrajectories function.

DEG analyses and pathway enrichment

Differentially expressed genes (DEGs) analysis: for scRNA/snRNA-seq data, DEGs were identified by Seurat’s FindAllMarkers function, with threshold set to be FDR < 0.25 and |log₂FC | > 0.25; For bulk-seq data, DEGs were identified by the R package “limma” (version 3.58.1). The criteria for screening included |log₂(FC)| > 1 and an adjusted P-value < 0.05. P-value correction was performed using the Benjamini-Hochberg (BH) method. To further explore potential functions played by DEGs, over-representation analysis (ORA) was carried out by enrichGO, enrichKEGG functions of clusterProfiler package (v4.11.0.2)³³. Gene Set Enrichment Analysis (GSEA) analysis was performed by fgsea (v1.24.0) on pre-rank gene list against gene set obtained via msigdbr (v7.5.1)³⁴. Enrichment and network visualization of relevant pathways was achieved by aPEAR (v1.0)³⁵.

Calculation of signature score

Signature gene sets associated with NK activation ligand, HCC stemness, NK function state, hepatocyte characteristics, and T cell phenotype were obtained from previous research (Suppl. Data 2). For comparison of cholesterol metabolism between groups, signature gene sets were obtained via msigdbr (v7.5.1) msigdbr: MSigDB Gene Sets for Multiple Organisms in a Tidy Data Format. Signature score of specific gene set was calculated using Seurat’s AddModuleScore function.

Metabolism analyses

Signature score of metabolism pathways in single-cell resolution was quantified by scMetabolism (v0.2.1) against the REACTOME gene sets¹⁶.

Survival analyses

The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) dataset was obtained from GDC data portal (https://portal.gdc.cancer.gov/). Low-expression genes with the sum of FPKM across all samples less than 10 were discarded. Primary tumor samples with available overall survival (OS) information and a survival time less than 5 years were included (n = 365). To evaluate transcriptional similarity between TCGA-LIHC samples and HepG2 groups (WT or NK-R), overlapping genes between the two datasets were first identified. Only genes present in both datasets were retained for similarity analysis. Spearman’s correlation coefficient was then computed between each TCGA-LIHC sample and each HepG2 group. The resulting correlation coefficient was defined as the transcriptomic similarity score. TCGA-LIHC samples were then stratified into high-similarity (‘high’) and low-similarity (‘low’) groups based on the transcriptomic similarity to each HepG2 group. Kaplan–Meier survival curves were generated, and statistical significance was determined using the log-rank test. Survival analyses were performed using the R packages survival (v3.5.3) and survminer (v0.4.9). For each HepG2 cluster (cC1-cC6), Pearson’s correlation was then utilized to examine the relationship between overall survival and transcriptomic similarity to HepG2 cluster across TCGA-LIHC samples.

Trajectory analyses

For HepG2 cells and patients’ epithelial cells, Monocle3 (v1.0.0) algorithm were utilized to infer trajectory and corresponding pseudotime of single cells¹³. Gene expression trend along with tumor evolution was predicted based on Monocle3 calculated pseudotime. The HepG2 trajectory was verified by Slingshot (v2.6.0), which shows high consistency³⁶. CytoTRACE toolkit (v0.3.3), which mainly focus on cell differentiation state, was also applied to construct cellular trajectory and quantify stemness level of cells¹⁴.

Cell-cell communication analysis

For both patient groups (NK-S, NK-R), expression matrix and cell type information of all cells were extracted and imported into CellPhoneDB (statistical analysis mode) respectively to infer potential interactions between different cell types. After that, state of coinhibitory and costimulatory ligand-receptor (LR) pairs predefined by CellPhoneDB within each group were screened out via plot_cpdb function provided by ktplots (v1.2.3)³⁷. Cell state-specific LR pairs in both groups were subsequently visualized using ggplot2 (v3.5.1).

Cell state classifier

To verify that the NK-trained tumor cell evolution observed in HepG2 cells in vitro also exists in in-vivo patient level, the single cell classification software scPred (v1.9.2) was employed³⁸. Following the standard workflow, we first divided previously obtained HepG2 clusters into three cell states (WT-enriched, Mixed, NKR-enriched) based on their cell components. Next, feature space for classifier training was created via getFeatureSpace. Following that, classifiers for each cell state were trained by trainModel based on Support Vector Machines (SVM) model with a radial kernel. Model performance was evaluated by calculation (get_probabilities()) and visualization (plot_probabilities()) of training probabilities for HepG2 cells. Finally, all patients’ epithelial cells were classified and aligned onto a low-dimensional space using the pre-trained classifier by scPredict function.

Lentiviral transduction

Lentiviral particles were produced in HEK293T cells using a standard three-plasmid packaging system, as previously described in ref. ³⁹. Briefly, HEK293T cells were co-transfected with the lentiviral transfer plasmid (pLVX, or pCDH), the packaging plasmid psPAX2, and the envelope plasmid pMD2.G using Lipofectamine 3000, according to the manufacturer’s instructions.

Culture medium was replaced 6–8 h after transfection, and lentiviral supernatants were collected at 48 hours post-transfection, clarified by centrifugation at 300 × g for 5 minutes, and filtered through a 0.45 μm filter. Viral supernatants were either used immediately or aliquoted and stored at −80 °C until use. For lentiviral transduction, target cells were seeded at 50% confluence and incubated with lentiviral supernatant supplemented with 8 μg/mL polybrene. After 24 h, the medium was replaced with fresh complete culture medium. Stable polyclonal cell populations were selected using appropriate antibiotics (e.g., 2 μg/mL puromycin) for 5–7 days, depending on the resistance marker encoded by the lentiviral vector.

H&E staining

Paraffin-embedded tissue sections (4–5 μm thick) were deparaffinized in xylene and rehydrated through a graded ethanol series to distilled water. Sections were stained with hematoxylin for nuclear visualization, followed by differentiation in acid alcohol and bluing in running tap water. Subsequently, sections were counterstained with eosin to visualize cytoplasmic and extracellular components. After staining, sections were dehydrated through graded ethanol, cleared in xylene, and mounted with a permanent mounting medium. H&E-stained sections were examined and imaged using a bright-field microscope according to standard histopathological procedures.

Immunohistochemistry

Immunohistochemistry was performed on paraffin-embedded tissue sections (4–5 μm thick). Sections were deparaffinized in xylene and rehydrated through a graded ethanol series. Antigen retrieval was carried out by heating sections in citrate buffer (10 mM sodium citrate, pH 6.0) or EDTA buffer (pH 9.0), depending on the antibody, using a microwave.

Endogenous peroxidase activity was quenched by incubating sections with 3% hydrogen peroxide in methanol for 10 minutes. Sections were then blocked with 10% normal goat serum for 60 minutes at room temperature to reduce nonspecific binding.

Sections were incubated with the indicated primary antibodies overnight at 4 °C. After washing with PBS, sections were incubated with horseradish peroxidase (HRP)-conjugated secondary antibodies for 30–60 min at room temperature. Signal detection was performed using 3,3′-diaminobenzidine (DAB) as the chromogen, and nuclei were counterstained with hematoxylin.

Following dehydration and clearing, sections were mounted and imaged using a bright-field microscope. Image acquisition parameters were kept constant for all samples within the same experiment.

Immunofluorescence

Tissues were harvested, fixed in 4% paraformaldehyde, and embedded in paraffin according to standard procedures. Paraffin-embedded tissues were sectioned at a thickness of 4–5 μm. Tissue sections were deparaffinized in xylene and rehydrated through a graded ethanol series to distilled water.

Antigen retrieval was performed by heating sections in citrate buffer (10 mM sodium citrate, pH 6.0) using a microwave or water bath. After cooling to room temperature, sections were permeabilized with 0.3–0.5% Triton X-100 in PBS for 10 minutes and blocked with 10% goat serum in PBS containing 0.1% Tween-20 (PBST) for 1 hour at room temperature to reduce nonspecific binding.

Sections were incubated with the indicated primary antibodies diluted in blocking buffer overnight at 4 °C. After washing with PBST, sections were incubated with species-appropriate secondary antibodies (Goat anti-Mouse or Goat anti-Rabbit IgG) for 60 min at room temperature in the dark. Nuclei were counterstained with DAPI. After final washes, sections were mounted using antifade mounting medium and imaged using a fluorescence or confocal microscope. Image acquisition parameters were kept constant across samples within the same experiment.

Statistics and reproducibility

GraphPad Prism v.9.5.1 was used to calculate statistical significance. Where statistics were derived, experiments were independently performed at least 3 biological replicates. Results are reported as mean ± standard deviation (SD). Comparisons were performed using paired/no-paired Student’s t test, one-way ANOVA test, two-way ANOVA test, Benjamini-Hochberg test, or Log-rank test, as indicated in individual figures. All statistical tests were two-sided. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Reporting summary

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

Supplementary information

Supplementary Information^{(9.7MB, pdf)}

41467_2026_74360_MOESM2_ESM.pdf^{(6.1KB, pdf)}

Description of Additional Supplementary Files

Supplementary Data 1^{(27.8KB, docx)}

Supplementary Data 2^{(13.5KB, xlsx)}

Supplementary Movie 1^{(194.7KB, mp4)}

Supplementary Movie 2^{(50.1KB, mp4)}

Supplementary Movie 3^{(349KB, mp4)}

Supplementary Movie 4^{(80.1KB, mp4)}

Reporting Summary^{(99.4KB, pdf)}

Transparent Peer Review file^{(1.1MB, pdf)}

Source data

Source Data^{(30.7KB, xlsx)}

Acknowledgements

We thank Prof. Linrong Lu and Dr. Xuexiao Jin for their technical support in immunology for this study. We thank the Center of Cryo-Electron Microscopy (CCEM), Zhejiang University for technical assistance on scanning electron microscopy. We appreciate LC-Bio Technology Co., Ltd. (scRNA-seq, snRNA-seq and spatial transcriptome), LipidALL Technologies Co., Ltd. (Lipidome) for their technology service.

Author contributions

X.C., D.W., L.S. and B.L. designed the study. L.S. and B.L. performed experiments, collected data, and interpreted results. R.S., B.L. and J.H. performed bioinformatics analysis and interpreted the data. J.H., H.S., B.J., Y.T., W.Z., W.L., Y.W., L.M. and C.J. assisted B.L. to performed experiments and collecting data. L.H., M.C., J.S., B.S., X.L., H.Y. and YF.W. provided expertise and essential material. L.S. and B.L. wrote the manuscript; R.S. and D.W. revised the manuscript. All of the authors read and commented on the manuscript.

Peer review

Peer review information

Nature Communications thanks Domenico Mavilio and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

Dr. Liang Shi receives funding support from Fundamental Research Funds for the Zhejiang Provincial Universities (No. 226-2023-00011), National Natural Science Foundation of China (No. 82203426/82472742), Zhejiang Natural Science Foundation (No. LY23H160021/MS26H160105), and Clinical Top-notch Personnel Support of Zhejiang University. Dr. Boqiang Liu receives funding support from the National Natural Science Foundation of China (No. 82403159), China National Postdoctoral Program for Innovative Talents (No. BX20230316), Zhejiang Natural Science Foundation (No. LQN25H160014), and China Postdoctoral Science Foundation (No. 2024M752893). Prof. Hong Yu receives funding support from Science and Technology Program of Zhejiang Province (No. 2024C03201) and the National Natural Science Foundation of China (No. 82270670). Dr. Mingyu Chen receives funding support from the Medical and Health Science and Technology Program of Zhejiang Province (No. 2024KY1126/2024KY1139). Prof. Xiujun Cai receives funding support from the Key Research and Development Project of Zhejiang Province (No. 2021C03061) and Zhejiang Clinical Research Center of Minimally Invasive Diagnosis and Treatment of Abdominal Diseases (No. 2018E50003).

Data availability

All raw single-cell and spatial transcriptomic datasets generated in this study have been deposited in the Genome Sequence Archive (GSA) at the China National Center for Bioinformation (https://ngdc.cncb.ac.cn/). Specifically, spatial transcriptomic data from the immune-humanized mouse HCC tumor evolution model are available under accession HRA014878. Single-cell RNA-seq data from HepG2 WT and NK-R cell lines are available under accession HRA014870. Single-cell RNA-seq data from NK-sensitive and NK-resistant HCC patient tumors are available under accession HRA010660. All other data are available in the article and its Supplementary files or from the corresponding author upon request. Source data are provided with this paper.

Code availability

All codes used in this study are publicly available and have been appropriately cited in the Methods section. The corresponding scripts and software packages can be accessed and downloaded through the referenced publications or official repositories.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Liang Shi, Boqiang Liu, Ruya Sun.

Contributor Information

Liang Shi, Email: liang_shi@zju.edu.cn.

Boqiang Liu, Email: liuboqiang@zju.edu.cn.

Yifan Wang, Email: anwyf@zju.edu.cn.

Di Wang, Email: diwang@zju.edu.cn.

Xiujun Cai, Email: srrsh_cxj@zju.edu.cn.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-74360-x.

References

1.Dixon, E. D. et al. Inhibition of ATGL alleviates MASH via impaired PPARalpha signalling that favours hydrophilic bile acid composition in mice. J. Hepatol.10.1016/j.jhep.2024.09.037 (2024). [DOI] [PubMed] [Google Scholar]
2.Danaher, P. et al. Advances in mixed cell deconvolution enable quantification of cell types in spatial transcriptomic data. Nat. Commun.13, 385 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Martini, R. et al. African ancestry-associated gene expression profiles in triple-negative breast cancer underlie altered tumor biology and clinical outcome in women of African descent. Cancer Discov.12, 2530–2551 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Xue, R. et al. Liver tumour immune microenvironment subtypes and neutrophil heterogeneity. Nature612, 141–147 (2022). [DOI] [PubMed] [Google Scholar]
5.Ichise, H. et al. Functional visualization of NK cell-mediated killing of metastatic single tumor cells. eLife10.7554/eLife.76269 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Lopez-Soto, A., Gonzalez, S., Smyth, M. J. & Galluzzi, L. Control of metastasis by NK cells. Cancer cell32, 135–154 (2017). [DOI] [PubMed] [Google Scholar]
7.Dersh, D., Holly, J. & Yewdell, J. W. A few good peptides: MHC class I-based cancer immunosurveillance and immunoevasion. Nat. Rev. Immunol.21, 116–128 (2021). [DOI] [PubMed] [Google Scholar]
8.Tremain, A. C. et al. Synthetically glycosylated antigens for the antigen-specific suppression of established immune responses. Nat. Biomed. Eng.7, 1142–1155 (2023). [DOI] [PubMed] [Google Scholar]
9.Hanahan, D. Hallmarks of cancer: new dimensions. Cancer Discov.12, 31–46 (2022). [DOI] [PubMed] [Google Scholar]
10.Gargiulo, G., Serresi, M. & Marine, J. C. Cell states in cancer: drivers, passengers, and trailers. Cancer Discov.14, 610–614 (2024). [DOI] [PubMed] [Google Scholar]
11.Zhang, S. et al. Systemic Interferon-gamma Increases MHC Class I Expression and T-cell Infiltration in Cold Tumors: Results of a Phase 0 Clinical Trial. Cancer Immunol. Res.7, 1237–1243 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Weems, A. D. et al. Blebs promote cell survival by assembling oncogenic signalling hubs. Nature615, 517–525 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Cao, J. et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature566, 496–502 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Gulati, G. S. et al. Single-cell transcriptional diversity is a hallmark of developmental potential. Science367, 405–411 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Loh, J. J. & Ma, S. Hallmarks of cancer stemness. Cell Stem Cell31, 617–639 (2024). [DOI] [PubMed] [Google Scholar]
16.Wu, Y. et al. Spatiotemporal immune landscape of colorectal cancer liver metastasis at single-cell level. Cancer Discov.12, 134–153 (2022). [DOI] [PubMed] [Google Scholar]
17.Bovenga, F., Sabba, C. & Moschetta, A. Uncoupling nuclear receptor LXR and cholesterol metabolism in cancer. Cell Metab.21, 517–526 (2015). [DOI] [PubMed] [Google Scholar]
18.Cillo, A. R. et al. Blockade of LAG-3 and PD-1 leads to co-expression of cytotoxic and exhaustion gene modules in CD8(+) T cells to promote antitumor immunity. Cell187, 4373–4388 e4315 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Chu, Y. et al. Pan-cancer T cell atlas links a cellular stress response state to immunotherapy resistance. Nat. Med. 29, 1550–1562 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Andrews, L. P. et al. LAG-3 and PD-1 synergize on CD8(+) T cells to drive T cell exhaustion and hinder autocrine IFN-gamma-dependent anti-tumor immunity. Cell187, 4355–4372 e4322 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Craig, A. J., von Felden, J., Garcia-Lezana, T., Sarcognato, S. & Villanueva, A. Tumour evolution in hepatocellular carcinoma. Nat. Rev. Gastroenterol. Hepatol.17, 139–152 (2020). [DOI] [PubMed] [Google Scholar]
22.Bjorkstrom, N. K., Ljunggren, H. G. & Michaelsson, J. Emerging insights into natural killer cells in human peripheral tissues. Nat. Rev. Immunol.16, 310–320 (2016). [DOI] [PubMed] [Google Scholar]
23.Casanova-Acebes, M. et al. Tissue-resident macrophages provide a pro-tumorigenic niche to early NSCLC cells. Nature595, 578–584 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Niehrs, C. The complex world of WNT receptor signalling. Nat. Rev. Mol. Cell Biol.13, 767–779 (2012). [DOI] [PubMed] [Google Scholar]
25.Erazo-Oliveras, A., Munoz-Vega, M., Salinas, M. L., Wang, X. & Chapkin, R. S. Dysregulation of cellular membrane homeostasis as a crucial modulator of cancer risk. FEBS J.291, 1299–1352 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Zhao, Y. et al. Model-informed clinical pharmacology profile of a novel fixed-dose combination of nivolumab and relatlimab in adult and adolescent patients with solid tumors. Clin. Cancer Res.: Off. J. Am. Assoc. Cancer Res.10.1158/1078-0432.CCR-23-2396 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Curtis, C. et al. The genomic and transcriptomic architecture of 2000 breast tumours reveals novel subgroups. Nature486, 346–352 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell184, 3573–3587.e3529 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Guo, W. et al. scCancer: a package for automated processing of single-cell RNA-seq data in cancer. Brief. Bioinform10.1093/bib/bbaa127 (2021). [DOI] [PubMed] [Google Scholar]
30.Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods16, 1289–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Hao, Y. et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat. Biotechnol.42, 293–304 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Ravi, V. M. et al. Spatially resolved multi-omics deciphers bidirectional tumor-host interdependence in glioblastoma. Cancer Cell40, 639–655.e613 (2022). [DOI] [PubMed] [Google Scholar]
33.Wu, T. et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innov. (Camb.)2, 100141 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst.1, 417–425 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Kerseviciute, I. & Gordevicius, J. aPEAR: an R package for autonomous visualization of pathway enrichment networks. Bioinformatics10.1093/bioinformatics/btad672 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Street, K. et al. Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics19, 477 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Efremova, M., Vento-Tormo, M., Teichmann, S. A. & Vento-Tormo, R. CellPhoneDB: inferring cell-cell communication from combined expression of multi-subunit ligand-receptor complexes. Nat. Protoc.15, 1484–1506 (2020). [DOI] [PubMed] [Google Scholar]
38.Alquicira-Hernandez, J., Sathe, A., Ji, H. P., Nguyen, Q. & Powell, J. E. scPred: accurate supervised method for cell-type classification from single-cell RNA-seq data. Genome Biol.20, 264 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Shi, L. et al. A tumor-suppressive circular RNA mediates uncanonical integrin degradation by the proteasome in liver cancer. Sci. Adv.10.1126/sciadv.abe5043 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information^{(9.7MB, pdf)}

41467_2026_74360_MOESM2_ESM.pdf^{(6.1KB, pdf)}

Description of Additional Supplementary Files

Supplementary Data 1^{(27.8KB, docx)}

Supplementary Data 2^{(13.5KB, xlsx)}

Supplementary Movie 1^{(194.7KB, mp4)}

Supplementary Movie 2^{(50.1KB, mp4)}

Supplementary Movie 3^{(349KB, mp4)}

Supplementary Movie 4^{(80.1KB, mp4)}

Reporting Summary^{(99.4KB, pdf)}

Transparent Peer Review file^{(1.1MB, pdf)}

Source Data^{(30.7KB, xlsx)}

Data Availability Statement

[CR1] 1.Dixon, E. D. et al. Inhibition of ATGL alleviates MASH via impaired PPARalpha signalling that favours hydrophilic bile acid composition in mice. J. Hepatol.10.1016/j.jhep.2024.09.037 (2024). [DOI] [PubMed] [Google Scholar]

[CR2] 2.Danaher, P. et al. Advances in mixed cell deconvolution enable quantification of cell types in spatial transcriptomic data. Nat. Commun.13, 385 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR3] 3.Martini, R. et al. African ancestry-associated gene expression profiles in triple-negative breast cancer underlie altered tumor biology and clinical outcome in women of African descent. Cancer Discov.12, 2530–2551 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Xue, R. et al. Liver tumour immune microenvironment subtypes and neutrophil heterogeneity. Nature612, 141–147 (2022). [DOI] [PubMed] [Google Scholar]

[CR5] 5.Ichise, H. et al. Functional visualization of NK cell-mediated killing of metastatic single tumor cells. eLife10.7554/eLife.76269 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Lopez-Soto, A., Gonzalez, S., Smyth, M. J. & Galluzzi, L. Control of metastasis by NK cells. Cancer cell32, 135–154 (2017). [DOI] [PubMed] [Google Scholar]

[CR7] 7.Dersh, D., Holly, J. & Yewdell, J. W. A few good peptides: MHC class I-based cancer immunosurveillance and immunoevasion. Nat. Rev. Immunol.21, 116–128 (2021). [DOI] [PubMed] [Google Scholar]

[CR8] 8.Tremain, A. C. et al. Synthetically glycosylated antigens for the antigen-specific suppression of established immune responses. Nat. Biomed. Eng.7, 1142–1155 (2023). [DOI] [PubMed] [Google Scholar]

[CR9] 9.Hanahan, D. Hallmarks of cancer: new dimensions. Cancer Discov.12, 31–46 (2022). [DOI] [PubMed] [Google Scholar]

[CR10] 10.Gargiulo, G., Serresi, M. & Marine, J. C. Cell states in cancer: drivers, passengers, and trailers. Cancer Discov.14, 610–614 (2024). [DOI] [PubMed] [Google Scholar]

[CR11] 11.Zhang, S. et al. Systemic Interferon-gamma Increases MHC Class I Expression and T-cell Infiltration in Cold Tumors: Results of a Phase 0 Clinical Trial. Cancer Immunol. Res.7, 1237–1243 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Weems, A. D. et al. Blebs promote cell survival by assembling oncogenic signalling hubs. Nature615, 517–525 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Cao, J. et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature566, 496–502 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Gulati, G. S. et al. Single-cell transcriptional diversity is a hallmark of developmental potential. Science367, 405–411 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Loh, J. J. & Ma, S. Hallmarks of cancer stemness. Cell Stem Cell31, 617–639 (2024). [DOI] [PubMed] [Google Scholar]

[CR16] 16.Wu, Y. et al. Spatiotemporal immune landscape of colorectal cancer liver metastasis at single-cell level. Cancer Discov.12, 134–153 (2022). [DOI] [PubMed] [Google Scholar]

[CR17] 17.Bovenga, F., Sabba, C. & Moschetta, A. Uncoupling nuclear receptor LXR and cholesterol metabolism in cancer. Cell Metab.21, 517–526 (2015). [DOI] [PubMed] [Google Scholar]

[CR18] 18.Cillo, A. R. et al. Blockade of LAG-3 and PD-1 leads to co-expression of cytotoxic and exhaustion gene modules in CD8(+) T cells to promote antitumor immunity. Cell187, 4373–4388 e4315 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Chu, Y. et al. Pan-cancer T cell atlas links a cellular stress response state to immunotherapy resistance. Nat. Med. 29, 1550–1562 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR20] 20.Andrews, L. P. et al. LAG-3 and PD-1 synergize on CD8(+) T cells to drive T cell exhaustion and hinder autocrine IFN-gamma-dependent anti-tumor immunity. Cell187, 4355–4372 e4322 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Craig, A. J., von Felden, J., Garcia-Lezana, T., Sarcognato, S. & Villanueva, A. Tumour evolution in hepatocellular carcinoma. Nat. Rev. Gastroenterol. Hepatol.17, 139–152 (2020). [DOI] [PubMed] [Google Scholar]

[CR22] 22.Bjorkstrom, N. K., Ljunggren, H. G. & Michaelsson, J. Emerging insights into natural killer cells in human peripheral tissues. Nat. Rev. Immunol.16, 310–320 (2016). [DOI] [PubMed] [Google Scholar]

[CR23] 23.Casanova-Acebes, M. et al. Tissue-resident macrophages provide a pro-tumorigenic niche to early NSCLC cells. Nature595, 578–584 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Niehrs, C. The complex world of WNT receptor signalling. Nat. Rev. Mol. Cell Biol.13, 767–779 (2012). [DOI] [PubMed] [Google Scholar]

[CR25] 25.Erazo-Oliveras, A., Munoz-Vega, M., Salinas, M. L., Wang, X. & Chapkin, R. S. Dysregulation of cellular membrane homeostasis as a crucial modulator of cancer risk. FEBS J.291, 1299–1352 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Zhao, Y. et al. Model-informed clinical pharmacology profile of a novel fixed-dose combination of nivolumab and relatlimab in adult and adolescent patients with solid tumors. Clin. Cancer Res.: Off. J. Am. Assoc. Cancer Res.10.1158/1078-0432.CCR-23-2396 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Curtis, C. et al. The genomic and transcriptomic architecture of 2000 breast tumours reveals novel subgroups. Nature486, 346–352 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell184, 3573–3587.e3529 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Guo, W. et al. scCancer: a package for automated processing of single-cell RNA-seq data in cancer. Brief. Bioinform10.1093/bib/bbaa127 (2021). [DOI] [PubMed] [Google Scholar]

[CR30] 30.Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods16, 1289–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Hao, Y. et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat. Biotechnol.42, 293–304 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Ravi, V. M. et al. Spatially resolved multi-omics deciphers bidirectional tumor-host interdependence in glioblastoma. Cancer Cell40, 639–655.e613 (2022). [DOI] [PubMed] [Google Scholar]

[CR33] 33.Wu, T. et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innov. (Camb.)2, 100141 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst.1, 417–425 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR35] 35.Kerseviciute, I. & Gordevicius, J. aPEAR: an R package for autonomous visualization of pathway enrichment networks. Bioinformatics10.1093/bioinformatics/btad672 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Street, K. et al. Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics19, 477 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Efremova, M., Vento-Tormo, M., Teichmann, S. A. & Vento-Tormo, R. CellPhoneDB: inferring cell-cell communication from combined expression of multi-subunit ligand-receptor complexes. Nat. Protoc.15, 1484–1506 (2020). [DOI] [PubMed] [Google Scholar]

[CR38] 38.Alquicira-Hernandez, J., Sathe, A., Ji, H. P., Nguyen, Q. & Powell, J. E. scPred: accurate supervised method for cell-type classification from single-cell RNA-seq data. Genome Biol.20, 264 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Shi, L. et al. A tumor-suppressive circular RNA mediates uncanonical integrin degradation by the proteasome in liver cancer. Sci. Adv.10.1126/sciadv.abe5043 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Natural killer cell-mediated immunosurveillance modulates liver cancer evolution through cancer stemness enhancement and lipid metabolism reprogramming

Liang Shi

Boqiang Liu

Ruya Sun

Jing He

Hao Shen

Weiqi Li

Yi Wang

Weijun Zhao

Chenqi Jin

Binghan Jin

Yuanshi Tian

Lingfeng Ma

Lidan Hou

Mingyu Chen

Jiaying Shen

Bo Shen

Xiao Liang

Hong Yu

Yifan Wang

Di Wang

Xiujun Cai

Abstract

Introduction

Results

Spatial distribution of NK cells positively correlates with early transcriptome evolution of liver cancer

Fig. 1. Spatial Distribution of NK Cells Positively Correlates with Early Transcriptome Evolution of Liver Cancer.

NK cell mediated-immune pressure induces liver cancer evolution in murine models

Fig. 2. NK Cell Mediated-Immune Pressure Induces Liver Cancer Evolution in Murine Models.

Alterations of liver cancer cell state mediated by NK cells lead to evolution

Fig. 3. Alterations of Liver Cancer Cell State Mediated by NK Cells Lead to Evolution.

Liver cancer patients undergo NK cell-mediated tumor evolution

Table 1.

Enhancement of cancer stemness is a prominent feature of NK cell-mediated tumor evolution

Fig. 4. Enhancement of Cancer Stemness is a Prominent Feature of NK Cell-Mediated Tumor Evolution.

Cholesterol metabolism mediates liver cancer evolution induced by NK cells

Fig. 5. Cholesterol Metabolism Mediates Liver Cancer Evolution Induced by NK Cells.

LXR activation enhances the efficacy of Anti-LAG-3 immunotherapy in HCC

Fig. 6. LXR Activation Enhances the Efficacy of Anti-LAG-3 Immunotherapy in HCC.

Discussion

Methods

Patients

Mice

Materials (antibodies, primers and etc.)

Spatiotemporal orthotopic tumor evolution model in immune-humanized mice

Sequential orthotopic tumor passage model

In vitro tumor evolution model

Cell culture

Tumor vaccination murine model

In vitro cytotoxicity assay

NK cell depletion

Scanning Electron Microscopy (SEM) Imaging

CRISPR/Cas9 whole-genome knockout library screening

Cholesterol measurement by Amplex Red assay

Quantitative real-time PCR (RT-qPCR)

Bulk RNA sequencing

scRNA sequencing and snRNA sequencing

Library construction

Read alignment and quality control

scRNA-seq data processing

snRNA-seq processing and annotation

Spatial transcriptome sequencing

Library construction

Read alignment

Region annotation and quality control

Data integration and clustering

ST analyses

DEG analyses and pathway enrichment

Calculation of signature score

Metabolism analyses

Survival analyses

Trajectory analyses

Cell-cell communication analysis

Cell state classifier

Lentiviral transduction

H&E staining

Immunohistochemistry

Immunofluorescence

Statistics and reproducibility