Spatial analysis reveals targetable macrophage-mediated mechanisms of immune evasion in hepatocellular carcinoma minimal residual disease

Lea Lemaitre; Nia Adeniji; Akanksha Suresh; Reshma Reguram; Josephine Zhang; Jangho Park; Amit Reddy; Alexandro E Trevino; Aaron T Mayer; Anja Deutzmann; Aida S Hansen; Ling Tong; Vinodhini Arjunan; Neeraja Kambham; Brendan C Visser; Monica M Dua; C Andrew Bonham; Nishita Kothary; H Blaize D’Angio; Ryan Preska; Yanay Rosen; James Zou; Vivek Charu; Dean W Felsher; Renumathy Dhanasekaran

doi:10.1038/s43018-024-00828-8

. Author manuscript; available in PMC: 2025 Sep 20.

Published in final edited form as: Nat Cancer. 2024 Sep 20;5(10):1534–1556. doi: 10.1038/s43018-024-00828-8

Spatial analysis reveals targetable macrophage-mediated mechanisms of immune evasion in hepatocellular carcinoma minimal residual disease

Lea Lemaitre ^1,⁹, Nia Adeniji ^1,⁹, Akanksha Suresh ^1,⁹, Reshma Reguram ¹, Josephine Zhang ¹, Jangho Park ¹, Amit Reddy ¹, Alexandro E Trevino ², Aaron T Mayer ², Anja Deutzmann ³, Aida S Hansen ⁴, Ling Tong ³, Vinodhini Arjunan ¹, Neeraja Kambham ⁵, Brendan C Visser ⁶, Monica M Dua ⁶, C Andrew Bonham ⁶, Nishita Kothary ⁷, H Blaize D’Angio ², Ryan Preska ², Yanay Rosen ⁸, James Zou ⁸, Vivek Charu ⁵, Dean W Felsher ^3,^✉, Renumathy Dhanasekaran ^1,^✉

PMCID: PMC12258040 NIHMSID: NIHMS2094406 PMID: 39304772

Abstract

Hepatocellular carcinoma (HCC) frequently recurs from minimal residual disease (MRD), which persists after therapy. Here, we identified mechanisms of persistence of residual tumor cells using post-chemoembolization human HCC (n = 108 patients, 1.07 million cells) and a transgenic mouse model of MRD. Through single-cell high-plex cytometric imaging, we identified a spatial neighborhood within which PD-L1 + M2-like macrophages interact with stem-like tumor cells, correlating with CD8⁺ T cell exhaustion and poor survival. Further, through spatial transcriptomics of residual HCC, we showed that macrophage-derived TGFβ1 mediates the persistence of stem-like tumor cells. Last, we demonstrate that combined blockade of Pdl1 and Tgfβ excluded immunosuppressive macrophages, recruited activated CD8⁺ T cells and eliminated residual stem-like tumor cells in two mouse models: a transgenic model of MRD and a syngeneic orthotopic model of doxorubicin-resistant HCC. Thus, our spatial analyses reveal that PD-L1⁺ macrophages sustain MRD by activating the TGFβ pathway in stem-like cancer cells and targeting this interaction may prevent HCC recurrence from MRD.

HCC is associated with high rates of recurrence, which contribute to a grim 5-year survival rate below 20% (ref. 1). In patients with nonmetastatic HCC who respond well to transarterial chemoembolization (TACE), most but not all tumor cells are eliminated, leaving behind a small fraction of residual tumor cells. Even after prolonged latency, these residual tumor cells can lead to recurrence, which is generally unresponsive to therapy and contributes to poor survival^2–5. Thus, deciphering how residual tumor cells evade post-therapy immune surveillance is crucial for preventing recurrence and enhancing outcomes.

Residual tumor cells may employ two major mechanisms to persist after therapy and instigate recurrence: cell-intrinsic or immune-based mechanisms. First, recurrence can arise from a subset of residual cells that could function as stem-like tumor cells, characterized by unique attributes such as chemoresistance and dormancy^6–8. Previously, we have reported a transgenic mouse model of HCC dormancy with such persistence of stem cell-like residual tumor cells⁹. Second, dormancy and recurrence can be promoted by protumor macrophages^10–12. Our previous findings support this, as we observed an M2-like macrophage population in mouse models of HCC that facilitates immune cell evasion¹³ and metastasis¹⁴. While tumor and immune factors may both drive HCC recurrence, their interplay remains unclear.

We aimed to evaluate the interplay between cell-intrinsic and immune mechanisms contributing to the persistence of residual tumor cells. To do this, we analyzed both human samples of post-TACE residual human HCC and mouse models of MRD. First, we generated a spatial map of residual human HCC, which showed that interactions between stem-like tumor cells and protumor macrophages, mediated by the TGFβ pathway, were linked to CD8⁺ T cell exhaustion. Second, we employed three distinct yet complementary mouse models: two transgenic models mimicking MRD (TRE-MYC-Cebpb-tTA⁹ and TRE-MYC/Twist1/Luc-Cebpb-tTA¹⁴), where oncogene inactivation caused tumor regression but allowed undetectable MRD to persist and lead to recurrence, and a syngeneic orthotopic allograft model mimicking chemoresistance. In these mouse models, we showed that targeting the TGFβ and PD-L1 pathways eliminated residual tumor cells and prevented recurrence.

Results

Single-cell spatial atlas reveals heterogeneity in human HCC

To elucidate the spatial organization of tumor and immune cells in post-chemoembolization residual HCC, we constructed a comprehensive single-cell spatial map of human HCC. We microdissected areas of viable HCC, constructed formalin-fixed paraffin-embedded (FFPE) tissue microarrays (TMA) and used co-detection by indexing (CODEX)¹⁵ to analyze the spatial architecture. We imaged a total of 1.07 million cells from 108 HCC samples with multiplex staining using 41 antibodies targeting tumor, immune, stromal and functional markers (Fig. 1a, Supplementary Fig. 1 and Supplementary Table 1). Through cellular segmentation, marker quantification, normalization and unsupervised clustering, we identified 11 major cell types within HCC (Fig. 1b and Extended Data Fig. 1a). Thus, by utilizing FFPE-optimized CODEX, we achieved highly multiplexed single-cell marker visualization and cellular phenotyping in human HCC.

We used canonical marker expression in CODEX to further subset three major cell types: tumor cells, macrophages and T cells (Fig. 1c,d). First, among the tumor cells we identified three discrete subsets. Two were stem cell-like, cytokeratin 19 (CK19)⁺ (6.7%, 34,436 cells) and epithelial cellular adhesion molecule (EpCAM)⁺ (5.2%, 26,622 cells) tumor cells (Fig. 1c,d). They expressed stemness, prosurvival and mesenchymal markers (Extended Data Fig. 1b). A third subset of tumor cells expressed the immune checkpoint PD-L1 (3.8%, 19,737 cells). The heterogeneity observed among these tumor cell subsets aligned with previous reports from single-marker studies^16–18. Second, among the host immune cells, we identified the CD68⁺ tumor-associated macrophages (TAMs) as the most common immune cell subset (14.5%, 154,899 cells). Three subsets of TAMs were identified: CD206⁺ M2-like (CD206⁺/CD163⁺/HLA-DR⁻/PD-L1⁻), PD-L1⁺ M2-like (PD-L1⁺/CD11b⁺/CD206⁻/HLA-DR⁻) and HLA-DR⁺ M1-like (HLA-DR⁺/S100A4⁺/CD206⁻/PD-L1⁻) (Fig. 1c,d). The expression of other macrophage markers aligned with this classification (Extended Data Fig. 1c). Third, we defined three subsets among the tumor-infiltrating CD8⁺ T cells (17,647 cells, 0.01%): exhausted (PD-1⁺/TIM3⁺/CD44⁻), effector (CD45RO⁺/CD44⁺/PD-1⁻/TIM3⁻) or memory (CD45RO⁺/CD44⁻/PD-1⁻) CD8⁺ T cells (Fig. 1c,d). Thus, through CODEX analysis we classified discrete subsets of tumor cells, macrophages and CD8⁺ T cells (Fig. 1d and Supplementary Table 2).

To validate our tumor and immune cell classification we used three approaches. First, we overlaid our unsupervised cell-calling on immunostained images and confirmed the accuracy of classification based on marker expression (Fig. 1e). Second, we demonstrated a consistent proportion of immune and stromal cell types across various HCC clinical subgroups, indicating the reliability of the classification, with NASH HCC notably displaying increased CD8⁺ T cell exhaustion as previously reported¹⁹ (Extended Data Fig. 1d–f). Third, we noted that the observed subset identification and proportion of macrophage and T cell subsets are similar to previous reports^20–22; however, we did note a higher proportion of neutrophils, possibly attributable to their underrepresentation in transcriptome-based deconvolution analyses^23,24 (Extended Data Fig. 2a). Thus, our CODEX analysis provides an accurate and quantitative single-cell spatial profile of 20 major cell types in human HCC.

Residual human HCC has enrichment of immunosuppressive cells

Next, we used the above single-cell phenotyping to identify differences in tumor and immune cell subsets that would distinguish residual and primary HCC. The above 108 samples were composed of two cohorts- residual HCC (n = 55; 427,126 cells) and primary HCC (n = 53; 638,158 cells). Residual HCC was derived from patients with HCC who had undergone transarterial chemoembolization (mean of 2.42 cycles; s.e.m. = 1.32) and retained post-treatment viable HCC in the liver explant. The control group was derived from patients with treatment-naive primary HCC. The two groups were matched for key variables, including age, sex, race, ethnicity, comorbidities, grade, microvascular invasion, alphafetoprotein (AFP), clinical tumor stage and pathological tumor stage, but not matched for etiology and cirrhosis (Supplementary Table 3). We could measure all 20 cell types in both residual and primary HCC. Thus, we could distinguish residual and primary HCC through their distinct distribution of cellular subtypes (Fig. 2a,b).

Fig. 2 | — a, Immune and tumor cell proportions in residual HCCs from chemoembolization-treated liver explants (residual HCC, n = 55 patients) compared to primary resected primary HCC (n = 53 patients). b, H&E and CODEX imaging representations of a residual and primary HCC tumor core. In each CODEX image, eight canonical markers (DAPI, CD31, CD4, CD15, CD68, PanCK, CD8 and αSMA) are overlaid onto the image in different colors. c, Volcano plot showing the differential enrichment of tumor and immune cell populations (with statistical significance) in residual (n = 55 patients) versus primary HCCs (n = 53 patients). Two-tailed unpaired t-test shown with Bonferroni correction for multiple comparisons was used. d, Proportion of PD-L1⁺ macrophages, PD-L1⁺ tumor cells, mast cells and exhausted T cells in residual (n = 55 patients) versus primary HCCs (n = 53 patients). Box indicates 25th–75th percentile, whiskers show 5th–95th percentile and the line shows the median e, Representative Voronoi plot of cores showing PD-L1⁺ macrophage, PD-L1⁺ tumor cells, mast cells and exhausted T cells density in residual (n = 55 patients) versus primary HCCs (n = 53 patients). f, Comparison of normalized expression of stemness marker CD44 and CK19 marker in tumor cells from residual (n = 217,105 cells) and primary HCC (n = 297,608 cells) (both Padj value < 2.2 × 10⁻³⁰⁸). Two-tailed unpaired t-test shown with Bonferroni correction for multiple comparisons. The curve represents the probability density function of the normalized expression levels, with the peak indicating the mode. g, Comparison of normalized expression of CD11b and podoplanin in PD-L1⁺ macrophages from residual (n = 47,605 cells) and primary HCC (n = 6,108 cells) (both Padj value < 2.2 × 10⁻³⁰⁸). Two-tailed unpaired t-test shown with Bonferroni correction for multiple comparisons. The curve represents the probability density function of the normalized expression levels, with the peak indicating the mode. h, Comparison of normalized expression of PD-1 and TIM3 in T cells and NK cells from residual (T cells, n = 16,549 cells; NK cells n = 11,701 cells) and primary HCC (T cells, n = 45,494 cells; NK cells n = 2,111 cells) (both Padj value < 2.2 × 10⁻³⁰⁸). Two-tailed unpaired t-test shown with Bonferroni correction for multiple comparisons. The curve represents the probability density function of the normalized expression levels, with the peak indicating the mode. Statistical significance was assessed by an unpaired, two-tailed t-test, Bonferroni adjustment was used for P values. CD, cluster of differentiation; αSMA, smooth muscle alpha actin; BCL2, B cell lymphoma 2; Sig, significant.

We found higher infiltration of immunosuppressive tumor and immune cell subsets in residual HCC than in primary HCC. Specifically, PD-L1⁺ macrophages (Padj = 2.2 × 10^−6,), PD-L1⁺ tumor cells (Padj = 3.4 × 10⁻⁵), mast cells (Padj = 3.1 × 10⁻⁴), exhausted CD8 T cells (Padj = 6.6 × 10⁻⁴) and natural killer (NK) cells (Padj = 1.8 × 10⁻³) were more abundant in residual HCC (Fig. 2c–e and Extended Data Fig. 2b). Conversely, neutrophils (Padj = 1.4 × 10⁻⁴) and fibroblasts (Padj = 1.5 × 10⁻²) were relatively depleted in residual HCC compared to primary HCC (Fig. 2c and Extended Data Fig. 2c). These observations remained consistent within the subgroup of HCCs arising in a cirrhotic background (Extended Data Fig. 2d). Moreover, we show that M2-like macrophage enrichment was specific to the presence of therapy-resistant residual tumors post-TACE, rather than treatment effects, as evidenced by higher macrophage levels in TACE-refractory HCC²⁵ compared to TACE-exposed peritumoral cirrhotic liver tissue (Extended Data Fig. 2e–g) or sites of post-TACE complete response (Extended Data Fig. 2h), with results being consistent across different TACE methods (Extended Data Fig. 2i). Hence, residual HCC showed a more protumor redistribution of specific immune cells than primary HCC.

In addition to changes in cellular distribution, we observed four differences in cellular phenotype between residual and primary HCC. First, tumor cells in residual HCC showed a more stem-like (CD44^high, CK19^high) phenotype (Fig. 2f). Second, the abundant PD-L1⁺ macrophages within the residual HCC were more likely to be monocyte-derived (CD11b^high) and migratory (podoplanin^high) (Fig. 2g). Third, the CD8⁺ T and NK cells infiltrating residual HCCs were more exhausted, with higher PD-1 and TIM3 expression (Fig. 2h). Thus, residual HCC seems to be characterized by cancer stemness, enrichment of protumor macrophages and decreased immune surveillance.

Spatial interactions in residual HCC drive immune evasion

We hypothesized that the differences observed between residual and primary HCC in both the tumor and immune cellular subtypes may be related to cellular interactions. To test this, we integrated single-cell phenotype data with spatial coordinates to analyze interactions between nonhomotypic cells in direct contact with each other, which we defined as within a 25-μm radius. Based on the above results, we focused on interactions among tumor cells, macrophages and CD8⁺ T cells.

We first compared the interactions of tumor cells in residual versus primary HCC (Fig. 3a, Extended Data Fig. 3a). The most significant change was seen in tumor cell interactions with PD-L1⁺ macrophages (Fig. 3b). In residual HCC, all four types of tumor cells we identified had more frequent direct interactions with PD-L1⁺ macrophages compared to primary HCC (Fig. 3b). This was true, whether tumors had high or low levels of PD-L1⁺ macrophages (Extended Data Fig. 3b). In contrast, tumor cells did not differentially interact with exhausted CD8⁺ T cells, despite the increased presence of the latter in residual HCC (Fig. 3c). Among the interactions of tumor cells, those between EpCAM⁺ tumor cells and either M2-like PD-L1⁺ or CD206⁺ macrophages were associated with poor recurrence-free survival in residual HCC (Fig. 3d). We validated this finding in The Cancer Genome Atlas (TCGA) cohort by showing that overexpression of genes representing the interaction between stem-like cancer cells (EPCAM, KRT19 and CD44) and M2-like macrophages (CD68, CD274, MRC1 and CD163) was indeed associated with a significantly poorer recurrence-free survival in HCC (P = 5.2 × 10⁻⁵) (Extended Data Fig. 3c). Further, by employing three-dimensional (3D) co-culture tumoroid in vitro experiments, we showed that co-culture of HCC cancer cells with M2-like macrophages induced more cancer stemness than co-culture with M1-like macrophages (Extended Data Fig. 3d). Thus, in residual HCC, stem-like tumor cells spatially interact with M2-like PD-L1⁺ macrophages, correlating with worse clinical outcomes.

Next, we focused on the interactions of macrophages (Extended Data Fig. 3e). M2-like PD-L1⁺ macrophages were found to more frequently directly interact with exhausted CD8⁺ T cells (P = 1.11×10⁻³) (Fig. 3e) but not effector (P = 0.17) or memory CD8⁺ T cells (P = 0.38) in residual HCC. These interactions of PD-L1⁺ macrophages in residual HCC seemed to occur predominantly within fibrovascular bundles. PD-L1⁺ macrophages interacted more frequently with both fibroblasts (Padj = 2.99 × 10⁻³) and endothelial cells (P = 7.7 × 10⁻³) in residual than primary HCC (Fig. 3f). Similarly, we found that there was closer spatial proximity between PD-L1⁺ macrophages and fibroblasts (44.99 versus 179.75 μm; P = 0.027) or endothelial cells (36.2 versus 68.5 μm; P = 0.004) in residual than primary HCC. This was not observed for the other two subsets of HLA-DR⁺ or CD206⁺ macrophages and fibroblasts (42.2 versus 38.6 μm P = 0.329; 30.3 versus 31.9, P = 0.197, respectively). Thus, in residual HCC, PD-L1⁺ macrophages seem to directly interact with, and may result in the exhaustion of CD8⁺ T cells within fibrovascular bundles.

We next examined interactions that did not occur by direct contact but could still be mediated indirectly by intermediary cells. To first establish the range of indirect influence of a central cell, we developed a regression model (Supplementary Fig. 2). We found that the variance in a given central cell’s expression of Ki67 and BCL2 could be explained by evaluating an interaction radius ranging from 25–100 μm (Fig. 3g and Extended Data Fig. 3f). Based on this, we evaluated indirect interactions between nonhomotypic cells lying between 25 μm and 100 μm of a given cell. We found that in residual versus primary HCC, all four tumor cell subsets more frequently indirectly interacted with PD-L1⁺ macrophages (Fig. 3h and Extended Data Fig. 3g,h); however, different from what we observed for direct interactions, we found that in residual HCC, the tumor cells did indirectly interact more frequently with exhausted CD8⁺ T cells (Fig. 3i). This raises the possibility that PD-L1⁺ macrophages could serve as an intermediary facilitating indirect interactions between tumor cells and exhausted CD8⁺ T cells. Thus, residual HCC exhibits remodeling in cellular interactions between stem-like tumor cells, PD-L1⁺ macrophages, and exhausted CD8⁺ T cells which could be responsible for changes in immune surveillance, as we explain below.

Spatial neighborhoods reprogram immune cells in residual HCC

We wondered whether the remodeling of cell–cell interactions in residual HCC had a higher-order spatial organization. To test this, we defined ‘cellular neighborhoods (CNs)’ aiming to capture the intricate spatial arrangements within HCC as opposed to viewing them just as uniform sheets of cells. To determine these spatial CNs, we employed a previously reported approach that defined neighborhoods by clustering individual cells and their neighbors to identify broad patterns of spatial organization²⁶ (Fig. 4a). We identified nine distinct neighborhoods across all the HCC tissues (Fig. 4b). Among these, four were tumor-cell-dominant, four were immune-cell-dominant and one was of mixed cell population (termed ‘other’, of unclear significance) (Fig. 4b). Thus, through CN analysis, we delineated higher-order spatial organization in HCC.

To confirm that we were identifying real spatial structures, we used three approaches. First, we overlaid the neighborhoods with hematoxylin and eosin (H&E)-stained sections and fluorescent images, confirming the accurate recapitulation of known spatial structures such as fibrovascular bundles and lymphoid infiltrates (Extended Data Fig. 4a). Second, we observed the presence of these CNs in tumors across all stages, grades and etiologies, suggesting that their presence was a shared feature in the microenvironment (Extended Data Fig. 4b). Third, we confirmed that the canonical marker of the dominant cell within each CN was indeed overexpressed in its respective CN (Extended Data Fig. 4c). Thus, the microenvironment of HCC seems to be organized into spatial CNs.

We examined whether the distribution of the identified spatial CNs was different between residual and primary HCCs (Fig. 4c). Residual HCCs exhibited a higher prevalence of two neighborhoods, the M2-like macrophage immune CN (P = 4 × 10⁻⁵) and the vascular inflammatory tumor CN (P = 0.02). In contrast, the CNs with abundant antitumor immune cells, the innate immune CN (P = 8.5 × 10⁻⁷) and T cell immune CN (P = 0.012), were less frequent in residual HCC (Fig. 4c). These findings show that in residual versus primary HCC there is a protumor restructuring of the microenvironment.

We further examined the M2-like macrophage immune CN enriched in residual HCC. We hypothesized this CN may influence the cellular phenotype of tumor cells and T cells residing within it. The tumor cells residing in the M2-macrophage CN (17,046 cells, 7.4%) exhibited a more cancer stem-like (CK19^high/EpCAM^high/CD44^high), mesenchymal (vimentin^high) and prosurvival (BCL2^high and Ki67^high) phenotype than the tumor cells within another CN abundant in tumor cells but devoid of M2-like macrophages, the pauci-immune tumor CN (109,391 cells, 47.2%) (Fig. 4d). Additionally, T cells within the M2-macrophage CN (4,299 cells, 25.9%) exhibited a more exhausted phenotype (PD-1^high, TIM3^high, CD44^low and CD45RO^low) than the T cells residing within another CN enriched in T cells (8,345 cells, 50.3%) (Fig. 4d). The observations at the cellular level were also true at the tissue level, where residual HCC with higher M2-like macrophage CN had greater infiltration of EpCAM⁺ tumor cells and exhausted CD8⁺ T cells (Extended Data Fig. 4d). Thus, M2-like macrophages seem to promote stemness and CD8⁺ T cell exhaustion within specific spatial CNs rather than across the entire tumor.

Next, we investigated the phenotype of macrophages and T cells residing within the two stem-like tumor cell CNs in residual HCC. Macrophages residing within the two stem-like tumor CNs were more likely to be M2-like (CD206^high and CD163^high), and CD8⁺ T cells were more likely to be exhausted (PD-1^high and TIM3^high), than in the other two nonstem-like tumor CNs (Fig. 4d and Extended Data Fig. 4e). Additionally, the abundance of EpCAM⁺ tumor CNs was associated with poor recurrence-free survival in residual HCC (P = 0.02, hazard ratio (HR) 5.0) (Fig. 4e). Taken together, these data indicate that the spatial organization into CNs is distinct in residual HCC than primary HCC. Specifically, the M2-macrophage CN seems to promote spatially constrained stemness and CD8⁺ T cell exhaustion. This suggests that such spatial organization into M2-macrophage CN may serve as a mechanism by which residual tumor cells evade CD8⁺ T cell surveillance.

To experimentally test our observations from post-TACE residual human HCC, we conducted in vitro experiments to investigate how resistance to doxorubicin, the most commonly used chemotherapy agent in TACE, reprograms cancer cells and macrophages. Employing single-cell RNA sequencing (scRNA-seq), we analyzed 3D heterotypic tumoroids (n = 31,058 cells) composed of macrophages co-cultured with either doxorubicin-resistant or control HCC cells (Fig. 4f). In the doxorubicin-resistant tumoroids, cancer cells demonstrated a stem-cell-like phenotype with higher expression of stemness, chemoresistance and cytokines, which can drive M2-like macrophage polarization (Extended Data Fig. 5a,b). Among the macrophages, a specific macrophage cluster (C3), enriched in doxorubicin-resistant tumoroids, showed a distinct M2-like phenotype, in contrast to the less-abundant C4 cluster (CXCR4/LPLP^high; STAT1/ISG15^low) (Fig. 4g,h). Additionally, immunofluorescence (IF) analysis confirmed that macrophages co-cultured with doxorubicin-resistant cancer cells exhibited PD-L1 overexpression compared to those co-cultured with control cells (P = 1.6 × 10⁻⁶) (Fig. 4i). In concordance, IF and flow cytometry analysis confirmed that monocyte-derived macrophages from patients with HCC treated with conditioned medium from doxorubicin-resistant patient-derived tumoroids exhibited significantly higher PD-L1 expression compared to those treated with conditioned medium from control tumoroids (P = 1.6 × 10⁻⁶) (Fig. 4j and Extended Data Fig. 5c) or with doxorubicin alone (Extended Data Fig. 5d). Overall, our findings demonstrate that doxorubicin-resistant cancer cells not only adopt a more stem-cell-like phenotype but also significantly influence macrophage polarization toward an M2-like phenotype with elevated PD-L1 expression.

TGFβ pathway activation drives persistence of residual tumor

To investigate the transcriptional changes within the tumor cell and macrophage neighborhoods identified by CODEX analysis, we used spatially resolved transcriptomics. We separately quantified the expression of 1,812 genes within tumor cell (pan cytokeratin (panCK)⁺, CD45⁻/CD68⁻) and macrophage (panCK⁻, CD45⁺/CD68⁺) areas of interest (AOIs) using NanoString GeoMx DSP (Fig. 5a,b). A subset of the samples which were analyzed by CODEX and exhibited clearly defined tumor and macrophage areas were included in spatial transcriptomics; total of 105 AOIs (tumor cell (n = 52) and macrophage (n = 53); patient details in Supplementary Table 4). The robustness of AOI classification was confirmed by canonical gene expression of each AOI (Fig. 5c). We examined pathways activated within the AOIs of residual HCC compared to nontumorous liver AOIs. Residual HCC tumor cell AOIs displayed upregulated immunosuppressive interleukin (IL)-10 pathway (P = 4.7 × 10⁻⁴⁷) and PD-L1 pathway (P = 2.79 × 10⁻³⁷) (Fig. 5d and Supplementary Table 5). On the other hand, macrophage AOIs showed upregulated angiogenesis (P = 3.75 × 10⁻¹⁰) and invasiveness (P = 8.33 × 10⁻⁹) pathways (Supplementary Table 6) (Fig. 5d). Thus, using spatial transcriptomics, we revealed distinct yet complementary protumor pathway activation within the tumor cell and macrophage AOIs of residual HCC.

Fig. 5 | — a, Workflow of NanoString spatial transcriptomics analysis. The expression of targeted transcriptomes of cancer and immune-related genes were quantified in tumor cell and macrophage AOIs. b, Representative immunofluorescent ROIs showing tumor cell and macrophage-enriched AOI based on the expression of panCK, CD45, CD68 and DAPI. c, PCA and volcano plot showing differential expression of genes in the tumor cell (n = 52 AOIs) and macrophage (n = 53 AOI). Welch’s two-tailed unpaired t-test was performed on log₂-transformed normalized count data and the Benjamini–Yekutieli false discovery rate correction was applied. d, Molecular pathways activated in tumor cells (n = 19 AOI versus n = 11 AOI) and macrophage AOIs (n = 18 AOI versus n = 9 AOI) of residual HCC compared to nontumorous liver samples. e, Schematic showing derivation of tumor cell and macrophage signatures from the spatial transcriptome data of primary and residual HCC. f, Kaplan–Meier curve showing the prognostic significance of tumor cell and macrophage signatures applied to a validation cohort of human HCC (n = 340 patients). A log-rank test was used to compare the groups. g, Upstream regulators of transcriptional regulators of gene expression changes in the tumor cell and macrophage AOIs of residual HCC compared to nontumorous liver samples. h, Plot comparing *TGFB1* gene expression levels in tumor cells (n = 19 AOIs) and macrophage AOIs (n = 18 AOI) of residual HCC. Data are presented as mean ± s.e.m. A two-tailed unpaired t-test was used. i, Correlation plot showing coexpression of *SMAD2* and *CD274 (PD-L1)* in residual HCC (n = 38 AOI) but not primary HCC (n = 48 AOI). A two-sided Spearman test was used to correlate expressions. j, Representative fluorescent mRNA FISH images along with quantification of *TGFB1* mRNA expression in primary (n = 51 patients) and residual HCC (n = 46 patients). Representative fluorescent mRNA FISH images along with quantification of *TGFB1 and CD68* mRNA expression in primary (n = 31 patients) and residual HCC (n = 30 patients). Data are presented as mean ± s.e.m. A two-tailed unpaired t-test was used. rHCC, residual hepatocellular carcinoma; TGFB1, transcription growth factor β1.

To examine whether transcriptional changes within tumor cell or macrophage AOIs were prognostic, we established two gene signatures enriched in the respective AOIs of residual, but not primary HCC (Fig. 5e). We applied the residual HCC tumor cell and macrophage signatures to stratify the independent cohort of human TCGA HCC (n = 340). Enrichment of the macrophage signature was associated with poor overall- and recurrence-free survival on multivariable analysis of TCGA HCC, adjusting for age, sex, and tumor stage (P = 8.2 × 10⁻⁴, HR 1.9 (1.3–3.1)) (Fig. 5f and Extended Data Fig. 5e). In contrast, the tumor cell signature did not show significant associations with overall or recurrence-free survival (P = 0.628, HR 1.1 (0.8–1.6)) (Fig. 5f and Extended Data Fig. 5e). This suggests that macrophages influence the trajectory of residual HCC toward recurrence and poor prognosis.

Next, we assessed upstream regulators of the molecular pathways activated within tumor cell and macrophage AOIs. The receptor TGFBR1/2 kinase pathway was the top upstream regulator of the transcriptional changes in the tumor cell AOIs (P = 3.1 × 10⁻⁴⁴, z-score 2.5), whereas its corresponding ligand, TGFB1 was the top upstream regulator of transcriptional changes in the macrophage AOIs of residual HCC (P = 2.7 × 19⁻⁹, z-score 1.2) (Fig. 5g, Extended Data Fig. 5f). Moreover, TGFB1 was expressed at a higher level in the macrophage AOIs than tumor AOIs of residual HCC (Fig. 5h). The TGFB1 pathway effector gene SMAD2 positively correlated with CD274 (PD-L1) within residual, but not primary HCC (Fig. 5i). Further, we used messenger RNA fluorescence in situ hybridization (FISH) to show that TGFB1 mRNA expression and the proportion of TGFB1-expressing cells that were CD68⁺ macrophages were higher in residual than primary HCC (Fig. 5j). Collectively, our CODEX and spatial transcriptomics analyses highlight the central role of macrophage-mediated immunosuppressive pathways in residual tumor cell persistence, suggesting that PD-L1⁺ macrophages, may activate the TGFβ1 pathway in residual tumor cells.

Transgenic mouse model shows recurrence can arise from MRD

Based on our analysis of residual human HCC, we hypothesized that PD-L1⁺ macrophages promote persistence of residual disease by concurrently activating the TGFβ1 pathway in residual tumor cells, and also driving CD8⁺ T cell exhaustion. We next sought to directly test this hypothesis using transgenic mouse models of MRD in HCC.

To do this, we employed hepatocyte-specific transgenic mouse models of MYC-driven HCC, which express key HCC-specific genes and proteins, and show concordance with gene expression patterns in multiple human HCC datasets (Extended Data Fig. 6a–c). The non-metastatic MYC-HCC model has been previously reported to retain intrahepatic dormant residual tumor cells upon oncogene inactivation and tumor regression⁹. To more closely mimic human HCC and simulate disseminated residual tumors, we additionally employed the MYC/Twist1 model of metastatic HCC^9,14 (MT-HCC). We now show that MT-HCC retains small clusters of hyperchromatic, nonproliferative (phospho histone 3⁻), nonapoptotic (cleaved casp3⁻) viable residual tumor cells, both within the liver and at extrahepatic sites, upon oncogene inactivation (Fig. 6a,b and Extended Data Fig. 6d). Notably, these areas or residual HCC are radiologically undetectable (Fig. 6b), thus fulfilling criteria to be referred to as minimal residual disease (MRD)²⁷. Thus, our models effectively mimic the persistence of residual tumor cells post-tumor regression, as observed in post-TACE human residual HCC.

Fig. 6 | — a, Schematic illustrates the process of oncogene activation triggering tumor progression, subsequent inactivation leading to tumor regression and then reactivation promoting recurrence. Representative H&E and IHC for MYC show the temporal evolution of tumor progression and tumor regression in MYC/Twist1 HCC in the liver and lungs. b, Representative bioluminescence imaging, MRI liver and CT lungs show regression of tumors upon oncogene inactivation and recurrence upon oncogene reactivation. MRD-bearing mice do not show any radiological evidence of tumor. c, Overall survival of MYC-HCC (n = 17 mice) and MT-HCC (n = 21 mice) mice with primary tumor from oncogene activation (MYC-HCC n = 12, MT-HCC n = 16 mice) or with recurrent tumor from oncogene reactivation in MRD-bearing mice (MYC-HCC n = 5, MT-HCC n = 5 mice). A log-rank test was used to compare the survival in the Kaplan–Meier analysis. d, Intravital microscopy (MYC-HCC n = 3, MT-HCC n = 3 mice) and subcutaneous transplant allograft model (MYC-HCC n = 10, MT-HCC n = 10 mice) confirms persistence of a small proportion of viable residual tumor cells in perivascular niches, which serve as a source of recurrence upon oncogene reactivation. e, Cross-species transcriptome analysis of a 45-gene signature enriched in residual murine HCC (n = 6 mice) and not primary (n = 10 mice) or recurrent HCC (n = 3 mice) was able to predict survival in an independent human HCC cohort (n = 372 patients). A log-rank test was used to compare survival between Kaplan–Meier curves. f, Experimental scheme for scRNA-seq of microdissected areas containing in vivo MRD. Graph-based clustering of scRNA-seq represented as a Uniform Manifold Approximation and Projection (UMAP), with each color representing the correspondingly named cell type (n = 15,300 cells). The residual tumor cells (n = 355 cells) are highlighted in a red box. g, Overlap with cell types and enriched pathways in residual tumor cells (n = 355 cells) compared to hepatocyte clusters 1 (3,607 cells) and 2 (2,144 cells) in scRNA-seq of MRD. Bar graphs indicate −log(P value) from hypergeometric tests to assess significance of pathway enrichment. h, Expression of *Tgfb1* and *Tgfbr1* in primary (MYC-HCC n = 5, MT-HCC n = 5 mice), residual HCC (MYC-HCC n = 3, MT-HCC n = 3 mice) and recurrent HCC (MYC-HCC n = 3 mice). Data are presented as mean ± s.e.m. A two-tailed unpaired t-test was used to compare the mean between the groups. Horizontal bar plot showing activation of Tgfb1 pathway as upstream regulator of transcriptional changes in MYC-HCC and MT-HCC MRD and MYC pathway inactivation in MRD. i, Expression of Tgfbr1 by IF staining of primary cancer cell lines derived from MYC-HCC upon MYC inactivation and reactivation. Representative IF images and quantification of phosphorylation of smad2/3 primary cancer cell lines derived from MYC-HCC upon MYC inactivation. Mean and s.e.m. are shown (n = 12 biological replicates each). A two-tailed unpaired t-test was used to compare the mean between the groups. IHC, immunohistochemistry; BLI, bioluminescence imaging; IVM, intravital microscopy; Prim, primary; Res, residual; Rec, recurrent.

We evaluated whether the MYC-HCC and MT-HCC model are comparable to post-chemoembolization residual human HCC in four ways. First, we used time-series analysis to show that oncogene reactivation led to the recurrence of tumors from residual tumor cells within the regressed tumor bed, recapitulating the clinical phenomenon of recurrence (Fig. 6a,b). These recurrent tumors had shorter latency and poorer survival rates, pointing to their origin from MRD (Fig. 6c). Second, we unambiguously traced the origin of recurrent tumors to MRD using a combination of subcutaneous transplantation studies and intravital microscopy to track live, residual RFP-labeled MT-HCC cells (Fig. 6d). Third, we found that the overexpression of a gene signature from mouse MRD was associated with poor overall (P = 5.4 × 10⁻⁴, HR 1.8) and recurrence-free survival (P = 0.03, HR 1.4) in the TCGA cohort of human HCC (Fig. 6e). Fourth, we confirmed the translational relevance of our model to the context of chronic inflammation by demonstrating that tumor regression and MRD-driven recurrence also occur in the background of NASH induced by a high-fat diet. (Extended Data Fig. 6e–g). These results validate using MYC-HCC and MT-HCC mouse models to study MRD-driven recurrence in human HCC and to test adjuvant therapies.

Residual tumor cells in mouse MRD demonstrate cancer stemness

In our analysis of human HCC, we found that residual tumor cells exhibit cancer stem cell-like features. We characterized the in vivo residual tumor cells in our mouse models by performing scRNA-seq of a microdissected tumor bed containing MRD (n = 15,300 cells) (Fig. 6f). A subset of cells (n = 355, 0.02%) clustered with hepatocytes but differentially expressed several genes compared to the two main hepatocyte clusters (586 differentially expressed genes; 460 up and 126 down) (Fig. 6f and Supplementary Table 7). Gene set enrichment analysis revealed that these cells were transcriptionally analogous to embryonic stem cell lines (Fig. 6g). Additionally, gene signatures associated with progenitor or cancer stem cells were enriched in the residual tumor cells (Fig. 6g). Particularly, a gene signature representing liver cancer stem cell CK19⁺ tumors associated with human HCC recurrence²⁸ was strongly upregulated in this cluster (Padj = 1.1 × 10⁻¹¹) (Fig. 6g). Further, the residual tumor cells in the liver and metastatic sites expressed liver cancer stem cell markers CD133 and CK19 (Extended Data Fig. 6h,i). Thus, as seen in human post-chemoembolization residual tumor cells, the tumor cells in MRD in our mouse models exhibit cancer stemness.

Tgfβ pathway is activated in residual tumor cells in mouse models of MRD

Next, we found that the TGFβ pathway, identified as a key regulator of residual human HCC, was also activated in mouse MRD using three approaches. First, we show that the TGFβ pathway was the top activated upstream regulator of the transcriptional changes in MRD of both MYC-HCC (Padj = 1.8 × 10⁻⁵⁰; enrichment score (ES) 6.6) and MT-HCC (Padj = 7.9 × 10⁻⁴⁹; ES 6.2) upon oncogene inactivation (Fig. 6h). Second, we corroborated that the TGFβ pathway activation was specifically confined to a subset of stem-like tumor cells in vitro using scRNA-seq (Extended Data Fig. 7a–e). Third, we found that residual tumor cells in vitro have increased phosphorylation of smad2/3 and reversible overexpression of the receptor Tgfbr1 but did not express Tgfb1 (Fig. 6h,i). Thus, in our mouse model of MRD, stem-like residual tumor cells show TGFβ pathway activation.

Mechanisms of immune evasion by residual tumor cells

We investigated if in our mouse models, as suggested by human residual HCC, macrophages within the MRD niche contribute to TGFβ activation in residual tumor cells (Fig. 5g). Single-cell analysis of the in vivo mouse MRD demonstrated three distinct macrophage subsets (Fig. 7a). Two macrophage subsets had a proinflammatory M1-like phenotype with higher expression of genes such as Il1r1, Cxcl1, Ido2 and Nlrp6 (Fig. 7b). One of the macrophage clusters demonstrated an immunosuppressive phenotype (n = 1,061 cells) with overexpression Cd274 (Pdl1) and enrichment of M2-like macrophage²⁹ and monocyte-derived macrophage gene signatures (979 differentially expressed genes; 669 up and 310 down) (Fig. 7b,c and Supplementary Table 8). Moreover, this M2-like macrophage cluster expressed higher levels of Tgfb1 (Padj = 3.95 × 10⁻⁷). Thus, Pdl1^high macrophages seem to be the source for the Tgfβ pathway activation in residual tumor cells in our mouse MRD.

Fig. 7 | — a, Graph-based clustering of scRNA-seq of MRD represented as UMAP with each color representing the correspondingly named cell type (n = 15,300 cells). The three subsets of macrophages are highlighted in a red box. The M2-like macrophage cluster (C6) (1,061 cells) is shown separately in the UMAP below. b, Volcano plot showing the differentially expressed genes between the M2-like C6 cluster (1,061 cells) of macrophages and the C5 (1,097 cells) and C9 cluster (681 cells) of M1-like macrophages (Padj < 0.05) are shown in green if underexpressed or red if overexpressed in the M2-like macrophage cluster. DESeq2 was used in differential expression analysis to calculate fold changes and P values. c, Heatmap showing the enrichment of gene signature of M2-like macrophage polarization and monocyte-derived macrophage differentiation in the M2-like C6 cluster of macrophages (1,061 cells) than the C5 (1,097 cells) and C9 cluster (681 cells) of M1-like macrophages. d, Representative H&E and IF images of CD133⁺ stem-like cancer cells and Pdl1⁺ macrophages near MRD (<200 μm) compared to liver distant from MRD (n = 8 areas, each group). Data are presented as mean ± s.e.m. Bar plot shows comparison for mean PD-L1⁺ macrophages between groups using two-sided unpaired t-tests. e, Representative H&E and IF images of Tgfβ1 expression on Pdl1⁺ macrophages near MRD (<200 μm) compared to liver distant from MRD (n = 9 areas, each group). Data are presented as mean ± s.e.m. Bar plot shows comparison for mean Tgfβ1⁺/PD-L1⁺ macrophages between groups using two-sided unpaired t-tests. f, Schematic showing collection of conditioned medium from primary or residual HCC cells from MYC-HCC to treat macrophages in vitro. Bar plot shows comparison for mean concentration of cytokines secreted by residual tumor cells (n = 6 biologically independent samples) and primary tumor cells (n = 6 biologically independent samples) using two-sided unpaired t-tests. Data are presented as mean ± s.e.m. g, Macrophages are treated with conditioned medium (CM) from negative control (n = 2 biologically independent samples), primary (n = 3 biologically independent samples) or residual tumor cells (n = 3 biologically independent samples) in vitro. Bar plot shows comparison for phagocytosis between CM from primary or residual tumor cell groups using two-sided unpaired t-tests. Data are presented as mean ± s.e.m. h, Bar plots showing comparison of mean CD4⁺ T, CD8⁺ T and exhausted CD8⁺ T cells in livers with MRD (n = 4 mice) versus control non-MRD livers (n = 4 mice) using two-sided unpaired t-tests. Representative flow cytometry plots of exhausted CD8⁺ T cells based on TIM3 and PD-1 expression. Data are presented as mean ± s.e.m. Ctrl, control.

We found that Pdl1^high macrophages were enriched in MRD and not in non-MRD-bearing areas of the liver and also found that they closely interacted with stem-like residual tumor cells, both in the liver (Fig. 7d) and the lungs (Extended Data Fig. 8a,b). Notably, these Pdl1^high macrophages were confirmed to coexpress Tgfβ1 (Fig. 7e). Further using a multiplex ELISA assay, we show that residual tumor cells, and not primary tumor cells, secrete four cytokines, Cxcl5, Il6, Il22 and Il18, which are known to promote M2-like macrophage polarization and increase PD-L1 expression^30–33 (Fig. 7f). Also, macrophages exposed to the medium from residual tumor cells showed more impaired phagocytosis than those exposed to the medium from primary tumor cells (Fig. 7g). Thus, residual tumor cells in our mouse models seem to use a cytokine-mediated mechanism to recruit Tgfβ1-secreting nonphagocytic Pdl1⁺ macrophages to MRD.

We examined immune cell functionality within MRD using flow cytometry. While infiltration of other innate immune cells (Extended Data Fig. 8c) CD4⁺ T and CD8⁺ T cells (Fig. 7h), including naive, central memory or effector phenotypes (Extended Data Fig. 8d), were similar between MRD-bearing and control livers (Supplementary Fig. 3), we observed a significant enrichment of exhausted (Pd1⁺/Tim3⁺) CD8⁺ T cells in MRD-bearing liver (Fig. 7h). Thus, as we see in residual human HCC, in our mouse model of MRD we observed exhaustion of CD8⁺ T cells, suggesting that this is a mechanism of impaired immune surveillance.

Inhibition of Tgfβr1 and Pdl1 blocks HCC recurrence from MRD

Our analysis of human residual HCC and mouse MRD suggests that targeting the Tgfβ1 pathway and Pdl1^high macrophages could eliminate residual tumor cells and prevent tumor recurrence. To experimentally address this, we examined in our residual MYC-HCC or MT-HCC-bearing transgenic mice, whether a Tgfbr inhibitor (TgfbrI) or anti-Pdl1 antibody, either alone or in combination, prevented recurrence (Fig. 8a). We confirmed the absence of macroscopic tumors by magnetic resonance imaging (MRI) imaging before randomizing transgenic mice with MRD to the four treatment groups (Fig. 8a). Upon oncogene ractivation, mice receiving combined anti TgfbrI and Pdl1 therapy showed an eight-fold reduction in recurrent tumor burden and a fivefold reduction in recurrence rate, compared to controls (both P < 0.05; Fig. 8b,c). These differences were also significant when compared to monotherapy with TgfbrI or anti-Pdl1 (all P < 0.05; Fig. 8b,c). Combination anti-TgfbrI and Pdl1 therapy-treated livers demonstrated areas of higher CD4⁺ and CD8⁺ T cell recruitment than either control or monotherapies (all P < 0.05; Fig. 8d), Moreover, CD4⁺ T and CD8⁺ T cells of mice treated with combination therapy were more activated (CD69⁺/CD44⁺) than the control group (Fig. 8e). The combined inhibition of Tgfbr1 and Pdl1 in a subcutaneous allograft model improved recurrence-free survival, confirming its effectiveness in preventing recurrence from MRD (Extended Data Fig. 9a–d). Additionally, we found that combined inhibition of Tgfbr1 and Pdl1 lacked efficacy in preventing the progression of primary MT-HCC, suggesting mechanistic specificity to MRD (Extended Data Fig. 9e,f). Thus, combined blockade of Tgfβ1 pathway and Pdl1 promoted recruitment and activation of T cell response and prevented tumor recurrence from MRD.

Fig. 8 | — a, Experimental scheme for treatment of oncogene-deprived MRD-bearing mice with control antibody or monotherapy with Tgfbr inhibitor (TgfbrI) or anti-Pdl1 or their combination. Treatment is followed by oncogene reactivation to induce tumor recurrence. b, Representative gross images of recurrent tumor burden upon oncogene reactivation in the liver of mice treated with control antibody or monotherapy with TgfbrI or anti-Pdl1 inhibitor or their combination. c, Quantification of recurrent tumor burden upon oncogene activation in the liver of MYC-HCC and MT-HCC mice treated with control antibody (n = 11 mice) or monotherapy with Tgfbr inhibitor (n = 9 mice) or anti-Pdl1 inhibitor (n = 9 mice) or their combination (n = 16 mice). Plots compare the mean between the groups with two-sided unpaired t-tests. Box shows 25th–75th percentiles, whiskers show 5th–95th percentiles and the line represents the median. d, Representative images and quantification from H&E staining and IHC staining for CD4 and CD8 in recurrent tumors in mice treated with control antibody (n = 5 mice) or monotherapy with TgfbrI (n = 5 mice) or anti-Pdl1 inhibitor (n = 5 mice) or their combination (n = 5 mice). Plots compare the mean between the groups with two-sided unpaired t-tests. Box shows 25th–75th percentiles, whiskers show 5th–95th percentiles and the line represents the median. e, Representative flow cytometry images and quantification of activated CD4⁺ and CD8⁺ T cells, which are CD69⁺/CD44^high in the recurrent tumor in the liver and lungs of MYC-HCC and MYC/Twist1 HCC mice treated with control antibody (n = 3 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n = 4 mice). Bar plots compare the mean between the groups with two-sided unpaired t-tests. Data are presented as mean ± s.e.m. f, Experimental scheme for treatment of oncogene-deprived residual HCC-bearing mice with control antibody or Tgfbr1 or Pdl1 inhibitors or combination therapy with TgfbrI and anti-Pdl1 inhibitor. Residual tumor niches are then evaluated for immune response. g, Quantification of IHC staining for CD8 and PD-L1 in residual HCC niche in mice treated with control antibody (n = 5 mice) or Tgfbr1 (n = 5 mice) or Pdl1 inhibitors (n = 5 mice) or combination therapy with Tgfbr1 and Pdl1 inhibitors (n = 5 mice). Bar plots compare the mean between the groups with two-sided unpaired t-tests. Data are presented as mean ± s.e.m. h, Representative H&E staining and IHC staining for CD8 and PD-L1 in MRD (red arrows in gross image and black boxes in H&E) in mice treated with control antibody (n = 5 mice) or Tgfbr1 (n = 5 mice) or Pdl1 inhibitors (n = 5 mice) or combination therapy with Tgfbr1 and Pdl1 inhibitors (n = 5 mice).

To determine whether blocking Tgfβ1 and Pdl1 eliminated MRD, we administered either control or TgfbrI or anti-Pdl1 antibodies or their combination to mice bearing residual HCC and euthanized mice immediately after the completion of treatment, without reactivating the oncogenes (Fig. 8f). Transgenic mice treated with the combination therapy exhibited increased recruitment of CD8⁺ T cells and decreased Pdl1⁺ macrophages within MRD than treatment with control or the respective monotherapies (Fig. 8g,h). We also saw an increased number of apoptotic cells in MRD with combination therapy than control or monotherapies (Extended Data Fig. 9g). Flow cytometry analysis confirmed that CD8⁺ T cells in mice treated with combination anti TgfbrI/Pdl1 therapy were more activated (CD69⁺/CD44⁺) than with control (Extended Data Fig. 9h and Supplementary Fig. 4). Hence, combined blockade of Tgfβ and Pdl1 decreased immunosuppressive macrophage infiltration, elicited a robust CD8⁺ T cell response, eliminated MRD and prevented HCC recurrence.

We note that our transgenic mouse models of MRD recapitulate the natural history of residual tumor cell persistence and HCC recurrence, but they do not involve persistence of residual tumors after exposure to therapeutic agents such as doxorubicin, as is used in TACE for human HCC³⁴. Hence, we complemented our MRD mouse model by developing a immunocompetent mouse model of syngeneic orthotopic allografts in wild-type C57/BL6 mice with Hep 53.4 HCC cells that underwent in vitro selection to be either doxorubicin-resistant or control (Extended Data Fig. 10a). Treatment with either control or combined TgfbrI and anti-Pdl1 antibodies in the doxorubicin-resistant orthotopic allografts resulted in a significant reduction in liver tumor burden (P = 0.01) (Extended Data Fig. 10b,c). Immune analysis revealed increased CD8⁺ T cell infiltration, elimination of PD-L1⁺/CD206⁺ M2-like macrophages and increased infiltration of CD86⁺/MHCII⁺ M1-like macrophages (Extended Data Fig. 10d,e). In contrast, control allografts (Extended Data Fig. 10f) showed no significant response to this treatment (Extended Data Fig. 10b,c,g and Supplementary Fig. 5). These observations confirm the findings from our transgenic MRD model, highlighting the complementary nature of both models and underscoring the therapeutic potential of dual Tgfβ1 and Pdl1 blockade in eliminating both oncogene-deprived stem-like MRD and doxorubicin-resistant murine HCC.

Discussion

We combined spatial analysis of both human clinical samples and transgenic mouse models of HCC to determine the mechanism of persistence of residual disease in HCC and suggest a possible therapeutic approach to improve clinical outcome (Extended Data Fig. 10h). We identified that interactions between stem-like tumor cells and immunosuppressive PD-L1⁺ macrophages within spatially constrained neighborhoods, in both human and mouse residual HCC, were linked to CD8⁺ T cell exhaustion. Further, macrophage-mediated TGFβ pathway activation within residual tumor cells was found to enable persistence of both human and mouse MRD. Finally, we provide preclinical evidence that combined inhibition of PD-L1⁺ macrophage-mediated TGFβ activation eliminates MRD in mouse HCC and prevents recurrence. Our results provide mechanistic insight that suggests a therapeutic strategy for eliminating residual disease to prevent cancer recurrence.

We harness the power of multiple complementary spatial biology technologies to draw mechanistic insights into residual HCC. Notably, residual disease that persists after therapy is a poorly characterized entity, especially in human solid tumors. Previous studies have employed scRNA-seq of primary HCC^35–37 or CODEX of treatment-naive HCC^38,39; however, the spatial organization of post-TACE residual HCC is not known. Our spatial mapping provides two key insights. First, we show that spatial interactions of cancer stem cells and protumor macrophages, not just their abundance as previously suggested^40–42, are critical drivers of recurrence from residual HCC. Second, we reveal a spatial restructuring of residual HCC into neighborhoods enriched with M2-like macrophages. Within these spatially constrained areas, M2-like macrophages promote CD8⁺ T cell exhaustion, thus facilitating immune evasion by stem-like tumor cells. This interplay adds a spatial dimension to the known role in PD-L1⁺ macrophages in cancer progression^43–45. Thus, our spatial analysis identified tumor cell and host immune cell interactions that could enable evasion of immune surveillance and eventual tumor recurrence of HCC.

We studied MRD mechanisms in HCC using transgenic mouse models^9,14, which, in contrast to previous studies using xenograft models in immunocompromised hosts^46–48, allow for in situ investigation of MRD with an intact immune system. A signature derived from these transgenic MRD models was indicative of shorter recurrence-free survival in human HCC, demonstrating translational relevance. Using these models, we identify that stem-like tumor cells reversibly overexpress Tgfbr1 and spatially interact with the Tgfβ1-secreting Pdl1⁺ macrophages abundant in MRD. Further, blocking the Tgfβ pathway alone was not effective in preventing HCC recurrence. Rather, the concordant elimination of Pdl1⁺ macrophages, which are a source of Tgfb1, was required to induce a robust T cell response and eliminate residual tumor cells. Additionally, we validated the efficacy of this combination in a mouse model of doxorubicin-resistant HCC. Combined blockade of TGFβ1 and PD-L1 has been trialed in primary cancers^49–51, yielding inconsistent results⁵² and exhibiting toxicity with extended use. Our study suggests that targeting MRD through brief adjuvant therapy can potentially mitigate these limitations and prevent HCC recurrence.

Our study presents a comprehensive single-cell spatial map of post-chemoembolization residual human HCC, highlighting critical interactions between stem-like tumor cells, M2-like macrophages and exhausted CD8⁺ T cells, insights possible only through the preservation of spatial context. Moreover, using a transgenic mouse model for disseminated MRD we demonstrate that the insights gained from spatial analysis can indeed be actionable, thus guiding us to target Tgfβ and Pdl1 to eliminate MRD in HCC. Thus, our results suggest a new adjuvant therapeutic strategy for reducing recurrence in HCC and improving patient outcomes.

Methods

Ethical approval

All procedures and methods were conducted in compliance with federal, state and Stanford University guidelines. Tissue collection from patients with HCC was approved by the Stanford Institutional Review Board (no. 28374). Written patient consent was obtained as per Institutional Review Board approval for the applicable components of the study. Animal experiments were approved by Stanford’s Animal Protocol and Laboratory Animal Care (APLAC) and adhere to the USDA Animal Welfare Act and PHS Policy on Humane Care and Use of Laboratory Animals.

Patient cohort selection

In this study, we identified two cohorts of patients diagnosed with HCC who met the following inclusion criteria. For cohort 1 (residual HCC), patients with a confirmed diagnosis of HCC who received bridging therapy with TACE, subsequently underwent liver transplantation and had viable residual HCC in the explanted liver tissue, were selected. Additional criteria included availability of sufficient tissue in FFPE blocks. We carefully selected those nodules which had been targeted for TACE and ensured adequate samples for our downstream analysis. Exclusion criteria included a history of other malignancies, receipt of resection before liver transplantation, receipt of radiation therapy before transplant or evidence of metastatic cancer. All patients in this cohort (n = 55) had received doxorubicin-based TACE (total procedures n = 116), a majority of which was DEB-TACE with doxorubicin-eluting beads (n = 102, 89%) and a smaller proportion received conventional TACE (n = 14, 11%), none had received bland TACE. A consistent team of experienced interventional radiologists at a single institution performed all TACE procedures. Patients undergoing TACE underwent follow-up computed tomography (CT) or MRI scans 8–12 weeks post-procedure to assess therapeutic response, which was evaluated using the modified Response Evaluation Criteria in Solid Tumors (mRECIST) system⁵³. Decisions regarding additional TACE treatments were made during multidisciplinary tumor board meetings. For the control group (primary HCC), patients who underwent surgical resection for HCC and had not received any locoregional therapy or systemic therapy before resection were selected. Additionally, these patients did not have any other form of cancer or metastatic tumors. In both groups, we collected comprehensive clinical and pathological data, including patient demographics, tumor characteristics, details of the treatments received and outcomes following the treatment. Patients were followed until death or until August 2023.

Construction of tissue microarrays

All tumor tissues were processed uniformly in the Stanford clinical pathology laboratory. H&E-stained sections from each FFPE block were carefully reviewed by the pathologist and areas of viable tumor or nontumorous liver selected. We generated a total of five TMAs with 1.5-mm diameter cores which were assembled using a TMA Grand Master automated tissue microarrayer.

CODEX panel development and staining

CODEX multiplex staining and analysis.

A 41-plex custom CODEX antibody panel was developed and validated (Enable Medicine) for ultra-high-plex imaging utilizing purified, carrier-free antibodies conjugated to unique DNA oligonucleotide barcodes (Akoya Biosciences) (Supplementary Table 1). Image processing and analysis were performed as described previously²⁶.

CODEX antibodies were validated on FFPE tonsil sections, and staining patterns were confirmed via comparison with online databases (The Human Protein Atlas, www.proteinatlas.org; Pathology Outlines, www.pathologyoutlines.com) and the published literature. Tissue microarrays containing FFPE biopsies from the cohort described in this study were sectioned at 5 μm and placed on 15 × 15-mm glass coverslips (Electron Microscopy Sciences, 72204-01) precoated with poly-l-lysine (Sigma, P8920). Coverslip staining was performed by Enable Medicine.

Briefly, FFPE tissue sections on coverslips were pretreated by heating on a slide warmer for 25 min at 55 °C. Tissue deparaffinization and hydration were next performed by incubating the FFPE tissue sections on coverslips for 5 min each following a solvent series (Histochoice Clearing Agent, Histochoice Clearing Agent, 100% ethanol, 100% ethanol, 90% ethanol, 70% ethanol, 50% ethanol, 30% ethanol, ddH₂0 then ddH₂0). Antigen retrieval was performed in 0.01 M citrate buffer at high pressure. The tissue was washed and equilibrated before staining for 3 h at room temperature with the 41-plex CODEX antibody cocktail in a staining buffer containing blocking solution (Akoya Biosciences). Post-staining, the tissues were washed and fixed in 1.6% PFA, followed by an ice-cold methanol incubation. After washing, the final tissue fix was performed using Fixative reagent (Akoya Biosciences). FFPE tissues on coverslips were stored in a six-well plate containing the storage buffer at 4 °C until CODEX acquisition.

CODEX multiplexed imaging and processing.

Stained coverslips were mounted onto the CODEX stage plate v.2 (Akoya) and secured onto the stage of a BZ-X810 inverted fluorescence microscope (Keyence). Reporter plates were prepared by adding fluorescently labeled oligonucleotides (Atto550, Cy5, AF750) made up in a reporter stock solution of nuclease free water, 10× CODEX buffer, assay reagent and nuclear stain to a black Corning 96-well plate. Automated image acquisition of tissue regions was performed at Enable Medicine using a CFI Plan Apo λ ×20/0.75 objective (Nikon) and fluidics exchange managed via the CODEX instrument and CODEX Instrument Manager software (v.1.29.3.6, Akoya Biosciences), according to the manufacturer’s instructions, with slight modifications. Raw fluorescent TIFF image files were processed, deconvolved and background subtracted utilizing the Enable Processor Pipeline (Enable Medicine) and antibody staining was visually assessed for each biomarker and tissue region using the Enable Visualizer (Enable BIOS, Enable Medicine). OME-TIFF hyper stacks were segmented based on 4,6-diamidino-2-phenylindole (DAPI) stain, pixel intensities were quantified, and spatial fluorescence compensation was performed, which generated comma-separated value and flow cytometry standard files for downstream analysis.

CODEX data analysis.

All analyses were run in R v.4.0.5 unless otherwise indicated. R functions are specified using the following notation: ‘<package_name>::<function_name>’.

Cell clustering.

For cell clustering and cell neighborhood analysis, data from 108 cores were analyzed. Possible batch effects were addressed by performing an inverse hyperbolic sine transform (‘base::asinh’) on cell expression values for every marker, in every region of interest (ROI). Next, normalized values were z-scaled across both cells and markers. To cluster cells, dimensionality reduction was first performed on scaled expression values using principal-component analysisA (PCA) with 20 components (‘stats::prcomp’). Next, a k-nearest-neighbor graph was constructed to build a similarity network between cells in principal-component space (‘dbscan::kNN’, k = 30). Finally, cells were clustered using the Leiden graph clustering algorithm (‘igraph::cluster_leiden’, cluster_resolution = 1.0).

Cell populations were defined using iterative unsupervised clustering using subsets of the full markers-by-cells expression matrix. An initial set of coarse cell clusters was first defined by unsupervised clustering on all cells that passed quality control (QC) and the major cell lineage markers in the panel: CD20 (B cell), CD15 (neutrophil), CD68 (macrophage), FoxP3 (T_reg), CD31 (endothelial cell), CD56 (NK cell), CD8/CD4/CD45/CD3e (T cell), CD117 (mast cell), CD11c (dendritic cell), PanCK (tumor) and αSMA (fibroblast). Clustering parameters, including granularity, nearest-neighbor number and marker subsets, were optimized by assessing clustering results visually, overlaid on images (Enable Medicine Visualizer); by manually examining the distribution of expression values in each cluster; and by quantifying cluster purity using the silhouette score. Next, cell subtypes were defined by subclustering of major cell categories. The full marker-by-cell expression matrix was subsetted with the following criteria: tumor cells: PanCK, EpCAM, CK19 and PD-L1; macrophages: HLA-DR, PD-L1 and CD206; and T cells: CD4, PD-1, TIM3, CD44 and CD45RO. Other canonical markers of these cell types were evaluated to further confirm the cell definitions. In each case, subclustered cells were examined visually for proper expression of lineage and subtype markers.

Construction of a spatial cellular interaction graph.

To perform spatial analyses on the data, we first constructed a spatial nearest-neighbor graph. Cell coordinates were derived by taking the centroid of each segmented cell nucleus relative to the corner of the ROI. A fixed radius neighbor algorithm was next used on these coordinates (‘dbscan::fixedrad’, r = 25 or r = 100). This graph thus represents, for each cell, its closest neighbors in two-dimensional space within the specified fixed radius. Each cell was assessed for its interaction with neighboring cells based on their proximity. Direct interactions were defined as those occurring between cells within a 25-μm radius. Indirect interactions were considered for cells located between 25 and 100 μm from each other. Homotypic interactions, where a cell interacted with another of the same type, were excluded from the analysis. For each sample, we further calculated the proportion of each cell type engaged in interactions with other distinct cell types. This was determined for each central cell by dividing the number of interacting cells of another specific type by the total number of that central cell type present in the sample. This normalized the interaction data for varying cell densities across regions.

Cell neighborhood analysis.

To define CNs, the number of neighbors of each cell type was counted, resulting in a matrix of cells by cell clusters, with each row representing a cell, each column representing a cell annotation (cell type) from the clustering above and each value representing the count of neighbors of the given annotation. The neighbor cell proportion was computed for each row. The resulting matrix was clustered using k-means clustering (‘stats::kmeans’), where the optimal k was determined empirically by maximizing the silhouette score metric (‘cluster::silhouette’). Each cluster was defined as a CN. Thus, each cell was given both a cell type annotation, which depends only on the cell’s own marker expression and a cell neighborhood annotation, which depends on the cell type and the identities of its nearest neighbors. To compare CNs between patient cohorts, we determined the proportion of cells in each ROI belonging to each CN. Proportions were transformed using the inverse hyperbolic tangent (‘base::asinh’) and split by cohort. We then performed pairwise t-tests (‘stat::t.test’) on the transformed proportions, comparing each CN between primary and residual HCC cohorts. The resulting P values were corrected for multiple testing by the Bonferroni method (‘stat::p.adjust’, method = ‘Bonferroni’).

NanoString digital spatial profiling

We used the published experimental methods for the NanoString GeoMx analysis⁵⁴. The GeoMx Digital Spatial Profiling instrument from NanoString Technologies was used for IF imaging. We used 5-μm sections of the FFPE human HCC TMAs for H&E and Cancer Transcriptome Atlas (CTA) processing. CTA processing involved baking slides at 60 °C, deparaffinization, antigen retrieval, proteinase K digestion, hybridization to RNA probes, washing to remove off-target probes and counterstaining with morphology markers. The morphology markers used were anti-panCK-Alexa Fluor 532, anti-CD45-Alexa Fluor 594 and anti-CD68-Alexa Fluor 647. The GeoMx Digital Spatial Profiling instrument from NanoString Technologies was used for IF imaging, ROI selection, AOI segmentation and spatially indexed barcode cleavage and collection. The tissue microarrays underwent staining with four markers, including panCK, CD45, CD68 and DAPI. These slides were imaged on the GeoMX platform, which functions in part as a fluorescent slide scanner. ROIs; (n = 12 per slide) were selected based on the visualization markers, using a custom-designed web-based control program (NanoString). Both H&E slides and multiplex IF-stained slides were visually examined in collaboration with a pathologist. ROIs were selected based on criteria such as sufficient cellularity and the absence of artifacts. Within each ROI, two specific AOIs were identified: one enriched in epithelial cells (PanCK⁺/CD45⁻/CD68⁻) and the other in macrophages (PanCK⁻/CD45/CD68⁻). After AOIs were chosen, the GeoMX platform utilized an automatically controlled UV laser to illuminate each AOI in turn, specifically cleaving oligonucleotide tags within the AOI but not in the surrounding tissue. A microcapillary collection system then collected the liberated oligonucleotides from each region and plated them into an individual well on a microtiter plate. This process was repeated in turn for each AOI. After AOI collection was complete, oligonucleotides were hybridized to complementary NanoString counting beads and counted using an nCounter analysis platform (NanoString).

Library preparation involved PCR amplification and sequencing on a NovaSeq S2 to achieve a minimum sequencing depth of 150–200 reads per μm² of illumination area. Digital counts from barcodes corresponding to gene probes were normalized using internal spike-in controls to account for system variation. Subsequently, these counts were further normalized to the area of their respective compartments. Differential gene expression and gene set enrichment analysis were performed in the NanoString GeoMx web portal. A Benjamini–Hochberg correction was used to decrease the false discovery rate. PCA was conducted in the Omics Explorer software from QluCore (Lund, Sweden, v.3.7)). Upstream regulators of transcription were discovered using the Ingenuity Pathway Analysis software from QIAGEN Digital Insights (v.94302991).

Multiplex fluorescent in situ hybridization

The RNAscope (Advanced Cell Diagnostics) fluorescent in situ hybridization technology designed primarily for use on FFPE sections was used to detect TGFB1 and CD68 mRNA in human HCC tissue microarrays. The procedure was performed as described previously⁵⁵. Post-fixed TMA sections were subjected to staining using the RNAscope assay protocol, employing the TGFB1 and CD68 primary target probe. Simultaneously, positive and negative in-house control probes were incorporated within each run. The stained sections were scanned on a fluorescence microscope (BZ-X800) from Keyence Corporation of America and quantified using QuPath analysis software⁵⁶.

In vitro 3D co-culture and scRNA-seq

Macrophage polarization for the in vitro co-culture experiment was performed as previously described⁵⁷. In brief, the human THP1 monocytes were cultured in RPMI medium with 10% heat-inactivated fetal bovine serum. THP1 monocytes were differentiated into macrophages by 24 h incubation with 150 nM phorbol 12-myristate 13-acetate (PMA, Sigma, P8139) followed by 24 h incubation in RPMI medium. Macrophages were polarized in M1 macrophages by incubation with 20 ng ml⁻¹ IFNγ (R&D system, 285-IF) and 10 pg ml⁻¹ LPS (Sigma, 8630) for 48 h. Macrophage M2 polarization was obtained by a 48-h incubation with 20 ng ml⁻¹ IL-4 (R&D Systems, 204-IL) and 20 ng ml⁻¹ IL-13 (R&D Systems, 213-ILB). After staining the macrophages with Cell Trace Violet (C34557), 3,000 cells of M1 and M2 type cells were added separately to an ultra-low attachment 96-well plate containing 3,000 Huh7 cells in each well. The cells were centrifuged at 1,400 rpm for 3 min following which they were co-cultured at 37 °C for 72 h. IF was performed on the resultant spheroids. Tubulin antibody (CST, 5335S) was used as a common marker for all the cells and preconjugated AF488 CK19 (ab87014) antibody was used to identify the stem cells in the spheroids.

In this study, we evaluated how doxorubicin-resistant cancer cells reprogrammed macrophages. The human HCC cell line Huh7 (Gift from the Torok laboratory) and SNU449 (CRL-2234, ATCC) were cultured in DMEM, 10% v/v serum, 2 mM l-glutamine and antibiotics in a humidified atmosphere of 5% CO₂ at 37 °C. Cells were tested negative for Mycoplasma contamination. Cells were cultured in the presence of increasing concentrations of doxorubicin for 96 h to determine a half-maximum inhibitory concentration (IC₅₀) of 1 μM. Our determined IC₅₀ for doxorubicin in HCC cell lines were consistent with previous in vitro studies^58–60 and mirrored the concentrations observed in human HCC explants following TACE⁶¹. Clones of cells that remained viable after doxorubicin exposure or cells exposed to control treatment for 96 h were counted (2,500 cells per well) and co-cultured with macrophages (THP1, TIB-202 and ATCC) (2,500 cells) in 3D spheroids created in 96-well low attachment plates. After 96 h of co-culture, the heterotypic 3D tumoroids from the two experimental conditions were pooled separately as single-cell suspensions followed by scRNA-seq using a BD Rhapsody Single-Cell Analysis System⁶² to investigate molecular changes in both the cancer cells and macrophages. A total of 34,474 passed initial QC and were sequenced. Analysis of scRNA-seq data was performed using the Seurat platform (v.5 of our Seurat toolkit)⁶³.

Human monocyte-derived macrophages and patient-derived 3D tumoroids

Peripheral blood samples were collected from patients with HCC and used to isolate peripheral blood mononuclear cells, which were then sorted into monocytes using CD14⁺ beads (130-097-052, Miltenyi). These monocytes were cultured in six-well plates with 10 ng ml⁻¹ of M-CSF (300-25, Peprotech) and GM-CSF (G5035 Sigma) for 5 days to generate monocyte-derived macrophages (MoMs), as described before⁶⁴. MoMs were subsequently treated with doxorubicin (130 ng ml⁻¹, AAJ64000MF, Fisher Scientific) or control for 48 h and then the MoMs were trypsinized, stained and analyzed by flow cytometry. In parallel, these MoMs were cultured in a eight-well chamber slide (177402 Lab-Tek) and IF was performed for PD-L1 (ab205921 Abcam).

In parallel, tumor samples from patients with HCC who underwent resection were dissociated into single cells and cultured in tumoroid medium until forming spheres in a 96-well plate, as described previously⁶⁵. Upon sphere formation, patient-derived organoids were treated with doxorubicin (130 ng ml⁻¹) or control. After 72 h, the conditioned medium from the patient-derived organoid cultures, treated with control or doxorubicin, was collected and added to the MoMs. Following an additional 48 h of culture with conditioned medium, the MoMs were trypsinized, stained and analyzed by flow cytometry.

Transgenic mice and in vivo treatment

Animals were housed in a pathogen-free environment at Stanford University and all procedures were performed in accordance with Stanford’s APLAC protocols. Mice were generally maintained on a 12-h dark–light cycle at 70 °F (±2 °F) and 50% (±20%) humidity. LAP-tTA/tet-O-MYC and LAP-tTA/TRE-Twist1/Luc transgenic lines were used, as previously described^9,14. Only male mice of this strain carry the tet-O-MYC transgene and develop liver cancer upon doxycycline withdrawal. Littermate controls were used for all experiments. Mice were administered weekly doses of 0.1 mg ml⁻¹ doxycycline (Sigma) in drinking water during mating and until 4 weeks of age. At 4 weeks, mice were taken off doxycycline. Mice were screened for tumors via MRI at approximately 2–3 months of age, at which time they developed tumors between 50–150 mm³. Stanford University’s APLAC allows a maximal tumor diameter of 1.70 cm, and this was not exceeded. Once liver tumors and/or lung metastases were confirmed, mice were placed on doxycycline to induce tumor regression. MRI was used to confirm complete tumor regression. To induce tumor recurrence, mice were taken off doxycycline and monitored closely. To evaluate the effectiveness of therapies in preventing tumor recurrence, two weeks following complete radiologic tumor resolution, mice were enrolled into one of four treatment groups and treated for 2 weeks (Fig. 8a). Control rat IgG (BioXCell) and anti-PD-L1 (clone 10F.9G2, BioXCell) antibodies were given intraperitoneally (i.p.) (100 μg per mouse) every other day. The Tgfbr1 inhibitor SB431542 was dosed 10 mg kg⁻¹ (dissolved in dimethylsulfoxide) daily i.p. as previously described⁶⁶. At the end of treatment, mice were taken off doxycycline and monitored for tumor recurrence with MRI. Humane end points for early euthanasia included signs of pain, labored breathing or the inability to move and stand normally. Other early euthanasia criteria were a total body weight loss greater than 20% or more than 10% body weight loss in a single week. Mice were killed by gradual carbon dioxide displacement and subsequent confirmatory cervical dislocation.

Establishment and treatment of immunocompetent syngeneic allograft models of HCC

The murine HCC cell line Hep 53.4 (ref. 67) (gift the Mark Yarchoan Laboratory) derived from HCC arising in the C57BL/6 background was used to establish orthotopic syngeneic allografts. The cells were cultured in DMEM, 10% v/v serum, 2 mM l-glutamine and antibiotics in a humidified atmosphere of 5% CO₂ at 37 °C. Cells were tested as negative for Mycoplasma contamination. Cells were cultured in the presence of increasing concentrations of doxorubicin for 96 h to determine an IC₅₀ of 1 μM. Clones of cells that remained viable after doxorubicin treatment or cells exposed to control treatment for 96 h were used to create orthotopic allografts. Female C57BL/6J mice at 6–8 weeks old were purchased from The Jackson Laboratories and maintained in accordance with the APLAC protocol. A total of 3 × 10⁶ cells in 40 μl Matrigel (Corning) and a phosphate-buffered saline (PBS) mixture per mouse were injected orthotopically into the left lobe of the livers of mice as previously described^68,69. Two weeks after engraftment, mice were enrolled into two treatment groups and treated for 2 weeks (Extended Data Fig. 10a). Control rat IgG (BioXCell) and anti-PD-L1 (clone 10F.9G2, BioXCell) antibodies were given i.p. (100 μg per mouse) every other day. The Tgfbr1 inhibitor SB431542 was dosed 10 mg kg⁻¹ (dissolved in dimethylsulfoxide) daily i.p. as previously described⁶⁶. At the end of treatment, mice were killed and evaluated for treatment response.

Animal magnetic resonance imaging

MRI was performed using a 7T small animal MRI (Agilent conversion) with a 40-mm Varian Millipede RF coil (ExtendMR) at the Stanford Small Animal Imaging Facility as previously described^70,71. In brief, animals were anesthetized with 1–3% isoflurane and placed into the MRI scanner containing a 40-mm Varian Millipede RF coil (ExtendMR). ParaVision (PV6.01) was used to acquire the DICOM images and tumor volumes were quantified from images using Osirix image processing software (Osirix, UCLA).

Intravital microscopy

The window chambers were prepared following previously described methods⁷². Under sterile conditions and with appropriate anesthesia, the titanium window frames were implanted onto the back of NSG mice, creating a 1-cm diameter hole in the dorsal skin-fold flap while preserving the integrity of the opposing dermis, fascial plane and vasculature. RFP-labeled primary HCC cells derived from MYC-HCC and MYC/Twist1 HCC tumors were injected near a major vessel between the fascia and dermis. The window-chamber mice were anesthetized using 1.0–2.0% isoflurane in oxygen flow. Intravital images were obtained using an A1 MP+ Multiphoton microscope (Nikon) with a ×10 water immersion objective. Throughout the imaging process, the mouse’s temperature was maintained at 37 °C with a warm plate. Large-field imaging combined with blood vessel imaging (Dextran labeling) was used as a ‘position mark’ to focus on the same imaging region and obtain tumor images on different days.

Whole-transcriptome sequencing of mouse tumors

RNA sequencing of MYC-HCC was performed at the Beijing Genomics Institute using their BGIseq 500 platform single-end 150-bp, 20 million reads per sample. MYC-activated tumors (MYC-On, MT-On, n = 10), tumors 7 days after MYC inactivation (MYC-Off, MT-Off, n = 6), and tumors that recurred upon MYC reactivation (n = 3) underwent whole-transcriptome sequencing. The gene expression level was quantified by a software package called RSEM. DESeq software was used to perform differential expression analysis. Ingenuity Pathway Analysis (QIAGEN) was used to perform functional pathway and similarity analysis.

scRNA-seq of mouse liver

MRD in MYC/Twist-HCC liver.

We identified an area of the liver that contained MRD based on the presence of scar of tumor regression on the liver surface. We microdissected this area of the liver and isolated nuclei from mouse liver tissue and constructed 3′ single-cell gene expression libraries (Next GEM v.3.1) using the 10x Genomics Chromium system. The library was sequenced on Illumina NovaSeq (PE150). After sequencing, clean reads were analyzed with mouse reference genome mm10-2020-A using Cell Ranger v.5.0. A total of 15,300 cells passed QC metrics, had an average of 24,020 reads per cell and median of 1,646 genes per cell.

Residual tumor cells from in vitro studies.

MYC was inactivated in a clonal population of primary cells derived from MYC-HCC for 1 week. MYC inactivation was confirmed by immunoblotting and IF. Around 90% of the cells died over the period of 1 week; the residual viable cells were submitted for single-cell sequencing. The 10x barcoding and complementary DNA synthesis were performed using 10x chromium 3′ scRNA-seq V2 chemistry according to the manufacturer’s instructions. The final libraries were sequenced with the Illumina Hiseq2500 according to recommended specifications. A total of 4,367 cells passed our QC metrics, had an average of 93,743 reads per cell and a median of 1,347 genes per cell.

Data were analyzed by ROSALIND (https://rosalind.onramp.bio/), with a HyperScale architecture developed by ROSALIND. Quality scores were assessed using FastQC. Cell Ranger (v.5.0) was used to align reads to the Mus musculus genome-build GRCm38, count UMIs, call cell barcodes and perform clustering. Individual sample reads were normalized via relative log expression using the DESeq2 R library. Read distribution percentages, violin plots, identity heatmaps and sample MDS plots were generated as part of the QC step using RSeQC. DESeq2 was also used to calculate fold changes and P values and perform optional covariate correction. Clustering of genes for the final heatmap of differentially expressed genes (fold change ≥2, Padj value < 0.05) was carried out using the Partitioning Around Medoids method using the fpc R library. Hypergeometric distribution was used to analyze the enrichment of pathways, Gene Ontology, domain structure and other ontologies.

Immunohistochemistry

Tissues were fixed in 10% paraformaldehyde and embedded in paraffin for sectioning. Sections were deparaffinized by incubation in xylene and rehydrated by sequential incubation in 100%, 95%, 80%, 60% ethanol and deionized water. Antigen retrieval was performed in a prewarmed container using the Dako pH 6.1 Target Retrieval Solution from Agilent Technologies. The sections were treated with 3% hydrogen peroxide to quench endogenous peroxidase activity. The sections were covered in the Dako Serum-Free Protein Block from Agilent Technologies. The sections were incubated with primary antibody overnight at 4 °C in a humid chamber. Details of the antibodies used and dilutions are provided in Supplementary Table 9. Subsequently, the sections were incubated with biotinylated anti-mouse (1:300 dilution, Vector Laboratories) or biotinylated anti-rabbit (1:300 dilution, Vector Laboratories) for 30 min at room temperature. Sections were placed for 30 min at room temperature in an ABC reagent (1:300 dilution, Vectastain ABC kit, Vector Laboratories). Sections were developed using 3,3′-diaminobenzidine, counterstained with hematoxylin and mounted with Permount.

Immunofluorescence

For paraffin sections, tissues were fixed in 10% paraformaldehyde and embedded in paraffin for sectioning. The sections were deparaffinized by incubation in xylene and rehydrated by sequential incubation in 100%, 95%, 80% and 60% ethanol and deionized water. For frozen sections, tissues were frozen in Tissue-Tek O.C.T. Compound (Sakura Finetek). The sections were equilibrated to room temperature in 1× PBS and then fixed with 4% methanol-free formaldehyde. The sections were covered in 100% methanol at −20 °C for 10 min. The sections were treated with 3% hydrogen peroxide. The slides were incubated in the Immunofluorescence Blocking Buffer from Cell Signaling Technologies. The sections were incubated with primary antibody overnight at 4 °C in a humid chamber. Subsequently, the sections were incubated with fluorescent-labeled secondary anti-mouse or anti-rabbit antibodies (1:300 dilution, Abcam) for 30 min at room temperature. The sections were mounted with DAPI. The sections were scanned on the All-in-One Fluorescence Microscope (BZ-X800) from Keyence Corporation of America and quantified on ImageJ software (National Institutes of Health; NIH).

Flow cytometry

Liver tissues were collected from treated mice and processed into single-cell suspensions through mechanical disaggregation and enzymatic digestion with a collagenase solution. The obtained cells were then stained with fluorochrome-conjugated antibodies specific to murine immune cell markers. Following staining, cells were analyzed using flow cytometry to identify and quantify distinct immune cell populations based on their fluorescent characteristics. LIVE/DEAD Fixable near-IR (Thermo Fisher Scientific, L34975) was used for live/dead discrimination. Data were analyzed using FlowJo (v.10.10), with cell populations identified through gating strategies that discriminated based on size (forward scatter) and granularity (side scatter). Details of the antibodies used and dilutions are provided in Supplementary Table 9.

Macrophage phagocytosis assay

Conditional HCC cell lines were derived from LAP-tTA and Tet-O-MYC or MYC/Twist1 mice. Cells were grown in DMEM (Invitrogen), supplemented with 10% FBS (Invitrogen) and cultured at 37 °C in a humidified incubator with 5% CO₂. Cell lines were confirmed to be negative for Mycoplasma contamination. Raw 264.7 (macrophage) cell suspensions from ATCC (TIB-71) were prepared with a concentration of about 3 million cells per ml. Cell suspension was pipetted onto a plate in 100-μl aliquots. Conditioned medium from MYC-HCC cells and conditioned medium from residual MYC-HCC cells (created through chronic treatment with 0.01% doxycycline) were collected. The negative control for the conditioned medium was DMEM (Invitrogen), supplemented with 10% FBS (Invitrogen) and 0.01% doxycycline solution. The cell suspensions were treated with the conditioned medium for 2 h. Then, the phagocytosis assay was performed (Phagocytosis Assay kit, ab234054) from Abcam). The reagents were prepared according to the package instructions. Zymosan slurry was added to the plated cells and the mixture was incubated at a controlled temperature of 37 °C with 5% CO₂ for 2 h. The cells were washed in an ice-cold phagocytosis assay buffer. Scanning was completed on the All-in-One Fluorescence Microscope (BZ-X800) from Keyence Corporation of America and quantification was performed on ImageJ software (NIH).

Luminex assay

Conditioned medium from MYC-HCC cells (derived from a conditional MYC transgenic mouse model of liver cancer) and conditioned medium from residual MYC-HCC cells (created through chronic treatment with 0.01% doxycycline) were collected and frozen at −80 °C. A Luminex assay using a 48-plex mouse cytokine array was performed by the Human Immune Monitoring Center, Stanford University. Concentration data was analyzed and visualized in GraphPad Prism (v.9.5.1).

Statistical analysis and reproducibility.

All statistical tests were performed in R (v.4.05 and 4.3.1). All statistical tests were two-sided unless otherwise stated. Differences between groups were analyzed using a Student’s t-test or one-way analysis of variance. The Benjamini–Hochberg method was used for adjusting P values. A chi-squared test was used to compare categorical variables. Kaplan–Meier analysis with a log-rank test was performed for survival analysis. All graphs are presented as the mean ± s.e.m. An adjusted P value <0.05 was considered to be significant. No statistical method was used to predetermine sample size and no data were excluded. Data normalization was performed before analysis but this was not formally tested, we have shown data distribution in the plots as individual data points. If not stated otherwise, the experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment. To ensure the reproducibility of our results, all experiments were conducted with adequate replicates. Specific details on the number of replicates for each experiment are provided in the respective methods sections.

Extended Data

Extended Data Fig. 3 | — a. Heatmap showing different patterns of direct interactions between the four tumor cell subtypes in primary (n=53 patients) vs residual HCC (n=55 patients). Two-tailed unpaired t-test was used to compare the groups. * Indicates pAdj value<0.05. b. Dot plot showing mean frequency of interaction between tumor cell subtypes and PDL1+ macrophages in primary and residual HCC stratified by median frequency of PDL1+ macrophages in tumor. Two-tailed unpaired t-test was used to compare the groups. * Indicates pAdj value<0.05. c. Kaplan Meir curve showing survival analysis of human HCC TCGA cohort (n=372 patients) classified based on expression of genes related to M2-like macrophages and cancer stem cells. Log rank test used. d. IF analysis of human HCC cell lines Huh7 co-cultured in 3D tumoroids with THP1 macrophages polarized to M1-like (n=6 tumoroids) or M2-like (n=8 tumoroids) macrophages. Two-tailed unpaired t-test was used to compare the groups. Bar plot shows quantification of CK19 expression on the 3D tumoroids. Data are presented as mean values +/− SEM. e. Heatmap of direct interactions between immune cells identified by CODEX in primary (n=53 patients) and residual HCC (n=55 patients). f. Large neighborhood sizes demonstrate reduced, but still significant spatial autocorrelation for Ki67 and BCL2 markers. Spatial autocorrelation is calculated using Geary’s C statistic. The spatial autocorrelation for each marker is calculated independently for all tumor regions (n=1.07 million cells) and reported as C’ = 1 - C. The maximal spatial autocorrelation possible is 1. Box: 25–75 percentile, whiskers: 5–95 percentile, line: median. g. Alluvial plot of statistically significant indirect tumor cell subtype-immune cell interactions in primary (n=53 patients) vs residual HCCs (n=55 patients). h. Heatmap showing different patterns of indirect (25-100um) interactions of tumor cell subtypes with immune cells in primary (n=53 patients) vs residual HCC (n=55 patients). Two-tailed unpaired t-test was used to compare the groups. * Indicates pAdj value<0.05. Abbreviations: HCC- hepatocellular carcinoma, PDL1- programmed cell death ligand 1, TCGA- the cancer genome atlas project, M2- Type 2 Macrophage, IF- Immunofluorescence, 3D- 3 Dimensional, THP1- , M1- Type 1 Macrophage, CK- cytokeratin, CODEX- co-detection by indexing, CD- cluster of differentiation, EPCAM- epithelial cell adhesion molecule, HLA-DR - major histocompatibility complex II cell surface receptor, NK- natural killer, Eff- Effector, Exh- Exhausted, Mem- Memory, DC- Dendritic cell, EC-, Macro- Macrophage, CSC- Cancer Stem cell, Mac- Macrophage, BCL2- B cell lymphoma 2.

Extended Data Fig. 4 | — a. Cross-matching of Voronoi plots showing spatial neighborhoods and the associated H&E and CODEX immunofluorescence staining images representations of the same tumor core. The canonical markers of key cells represented in each neighborhood are shown. b. Stacked bar chart comparing relative proportions of all 9 cellular neighborhoods identified by CODEX by HCC AJCC Stage (I n=58 vs II-III n=50 patients), grade (1 n=40, 2 n=57, and 3 n=9 patients), and etiologies (NASH n=28 patients, ALD n=11 patients, HepC n=35 patients, and HepB n=21 patients). c. Boxplot comparisons of CK19, EpCAM, CD3, aSMA, CD15, and CD68 expression amongst the identified cellular neighborhoods (CK19+Tumor CN=33708 cells, EpCAM+ Tumor CN=28190, Fibroinflammatory immune CN= 140,273, Innate immune CN= 53,912, M2- macrophage Immune CN= 188,407, Other= 6,130, T cell Immune CN=51,987, Pauci-immune tumor CN=310,977, Vascular inflammatory Tumor CN=251,581). Box: 25–75 percentile, whiskers: 5–95 percentile, line: median. d. Comparison of proportion of EpCAM+ tumor cells and exhausted CD8 T cells in residual HCC stratified by median frequency of M2-macrophage CN (Low n=17; High n=38 patients). e. Comparison of specific canonical marker expression on macrophage and T cells in the EpCAM+ tumor CN (macrophages n=272 cells, T cells n=13 cells), CK19+ tumor CN (macrophages n=1201 cells, T cells n=354 cells) compared to vascular inflammatory tumor CN (macrophages n=16143 cells, T cells n=1275 cells) in residual HCC. Abbreviations: H&E-hematoxylin and eosin, HCC- hepatocellular carcinoma, CODEX- co-detection by indexing, NASH-nonalcoholic steatohepatitis, ALD- alcoholic liver disease, HepC- hepatitis C, HepB- hepatitis B, CK- cytokeratin, CD- cluster of differentiation, EPCAM- epithelial cell adhesion molecule, aSMA- alpha smooth muscle actin, CN- Cellular neighborhood, PD1- programmed cell death protein 1, TIM3- T cell immunoglobulin and mucin domain-containing protein 3, Macro- macrophages, DAPI-4’,6-diamidino-2-phenylindole, PanCK- Pan-cytokeratin, ECad- E cadherin.

Extended Data Fig. 5 | — a. Characterizing the HCC cancer cell clusters C0 cluster between doxorubicin-resistant (DoxR) (n=4320 cells) and control samples (n=3627 cells). Gene set expression analysis shows enrichment of HCC progenitor signature and doxorubicin resistance signature in the cancer cells of the DoxR samples. b. Dot plot shows mean expression of key differentially expressed genes in C0 and C5 clusters between control (C0 n=3627 cells; C5 n=161 cells) and DoxR (C0 n=4320 cells; C5 n=57 cells) samples. c. Flow cytometry of monocyte-derived macrophages (MoM) from HCC peripheral blood mononuclear cells (PBMCs) were cultured for 48 hours with conditioned media from patient-derived organoid (PDO) which were either treated with doxorubicin (MoM+ DoxoR PDO CM; n=11,753 cells) or control (MoM+ ctrl PDO CM; n=2089 cells), d. Flow cytometry of monocyte-derived macrophages (MoM) from HCC peripheral blood mononuclear cells (PBMCs) were treated with doxorubicin-containing media (MoM+Doxorubicin, n=7559 cells) or doxorubicin-free control media (MoM+ctrl media, n=15,546 cells). This experiment was repeated using MoM from patients with HCC which were then treated with doxorubicin (n=3 patients) or control (n=3 patients). Two-tailed unpaired t-test used to compare the proportion of PDL1+ MoM on flow cytometry. e. Recurrence-free survival predicted by the macrophage and tumor cell signatures derived from the spatial transcriptomic analysis. Kaplan Meir analysis with log rank test was performed (n=372 patients, unique biological samples). f. TGFBR1/2 pathway displayed as an upstream regulator of the network of transcriptional changes in the tumor cell AOI and the corresponding ligand TGFB1 in the macrophage AOI of residual HCC. Upstream analysis in Ingenuity Pathway Analysis was performed using a two-tailed Fisher’s Exact Test to identify likely upstream regulators based on differential gene expression data. Abbreviations: HCC- hepatocellular carcinoma; DoxR- Doxorubicin-Resistant; PDO- Patient-derived organoid; CM- Conditioned media; Ctrl- Control; MoM- Monocyte-derived macrophage; PBMCs- Peripheral blood mononuclear cells, PDL1- programmed cell death ligand 1, TGFB1 transforming growth factor beta 1; Res- Residual; Tum- Tumor; HR- hazard ratio. AOI- Area of interest.

Extended Data Fig. 6 | — a. Representative H&E images and IHC for HCC-specific proteins glutamine synthetase, HNF4A and stem cell marker CK19. This was performed on multiple tumor samples (MYC-HCC=5, MT-HCC=5 mice). b. Expression of HCC-specific genes in primary (MYC-HCC=5, MT-HCC=5 mice) compared to WT FVB mouse liver (n=2 mice). Data are presented as mean values +/− SEM. c. Heatmap shows similarity between MYC-HCC=5, MT-HCC=5 and five human HCC transcriptome datasets including the TCGA cohort; the top 5 upregulated and downregulated genes across the different datasets are shown d. Representative H&E and IHC images showing phospho histone 3 (pH3) and cleaved caspase 3 (cC3) staining in primary HCC, minimal residual disease (MRD), and recurrent HCC in MYC-HCC and MYC-Twist1 HCC livers. Bar plots comparing quantification of phospho histone 3, and cleaved caspase 3, in primary tumors, MRD, and recurrent tumors in the liver of MYC-HCC. This was repeated n=5 mice in each group. Data are presented as mean values +/− SEM. e. Experimental scheme to induce NASH in MT-HCC. Gross liver images, H&E, and trichrome staining demonstrating induction of NASH, fibrosis, and HCC in this model. Experiments were performed on 3–5 animals in each group. f. Confirming induction of NASH by demonstrating hepatic inflammation, obesity, hyperlipidemia with high-fat diet (HFD) (c1 chow=5, HFD=3; c2-5 chow=4; HFD=4 mice). Data are presented as mean values +/− SEM. Two-sided unpaired t-tests were used for comparison. g. Confirming reversible residual HCC upon oncogene inactivation in MT-HCC with diet-induced NASH. Experiments were performed on 5 mice on HFD diet. h. Representative H&E and IHC images showing persistence of CD133+ stem-like cancer cells in perivascular niches of oncogene inactivated MRD in the liver, lungs and subcutaneous spaces. Experiments were performed on 5 mice in each group. i. Representative H&E and IHC images showing persistence of CK19+ stem-like cancer cells in perivascular niches of MRD in the subcutaneous spaces. Bar plots comparing quantification of CK19 and CD133 in primary tumors, MRD, and recurrent tumors in the liver of MYC-HCC. Experiments were performed on 5 mice in each group. Data are presented as mean values +/− SEM Abbreviations: H&E-hematoxylin and eosin, IHC- Immunohistochemistry, HCC- hepatocellular carcinoma, HNF4A- Hepatocyte nuclear factor-4 alpha, CK- cytokeratin, MYC- HCC- MYC-driven hepatocellular carcinoma, MT-HCC- MYC/Twist hepatocellular carcinoma, WT FVB- wild-type FVB, TCGA- the cancer genome atlas project, pH3- phospho histone 3, cC3- cleaved caspase 3, MRD- Minimal residual disease, NASH Nonalcoholic steatohepatitis, HFD- High-fat diet, ALT-Alanine aminotransferase, LDL- Low density lipoprotein, CD- cluster of differentiation.

Extended Data Fig. 7 | — a. Experimental scheme for single cell sequencing on MYC inactivation residual tumor cells (n=4367 cells). b. Two major clusters upon MYC inactivation, stem cell-like cluster (n=1184 cells) and non-stem-like cluster (n=3183 cells). c. Heatmap shows differentially expressed genes between the stem cell-like cluster and non-stem-like clusters. d. Activation of Tgfβ1 pathway in the stem cell-like cluster of residual tumor cells. e. Top molecular pathways activated in the stem cell-like cluster. *Abbreviations: TGFB1 transforming growth factor beta 1, HCC- hepatocellular carcinoma, scRNA-seq-single cell RNA sequencing*.

Extended Data Fig. 8 | — a. Comparison of macrophages in the spatial proximity of MRD (n=6 mice) versus areas without MRD (n=6 mice) in the lungs of MYC/Twist1 mice. Data are presented as mean values +/− SEM. b. Representative images demonstrating spatial interactions of stem-like cancer cells and macrophages in the spatial proximity of MRD in the lungs and liver of MYC/Twist1 mice. c. Comparison of immune cell subsets between livers without MRD ctrl Liver (n=4 mice) and livers with MRD (n=4 mice). Bar plots compared using two-tailed unpaired t-tests. Data are presented as mean values +/− SEM. d. Comparison of CD8 T cell subsets between livers without MRD ctrl Liver (n=4 mice) and livers with MRD (n=4 mice). Bar plots compared using two-tailed unpaired t-tests. Data are presented as mean values +/− SEM. *Abbreviations: MRD- Minimal residual disease, ctrl- control, CD- cluster of differentiation, H&E-hematoxylin and eosin, DAPI-4’,6-diamidino-2-phenylindole, HCC- hepatocellular carcinoma*.

Extended Data Fig. 9 | — a. Experimental scheme for treatment of oncogene-deprived subcutaneous MRD-bearing mice with control antibody or combination therapy with Tgfbr1 and Pdl1 inhibitors. Treatment is followed by oncogene reactivation to induce tumor recurrence. Kaplan–Meier analysis with log rank test was performed. b. Kaplan–Meier curves show time to recurrence in oncogene-deprived subcutaneous MRD-bearing mice with control antibody (n=4 mice) or combination therapy with Tgfbr1 and Pdl1 inhibitors (n=5 mice). c. Experimental scheme for treatment of oncogene-deprived subcutaneous MRD-bearing mice with control antibody or combination therapy with TgfbrI and anti-Pdl1 inhibitor. Residual tumor niches are then evaluated at the end of treatment. d. Macroscopic and microscopic evaluation of subcutaneous MRD sites shows elimination of residual tumor cells in mice treated with combination therapy with TgfbrI and Pdl1 inhibitors (n=4 mice) than control (n=4 mice). e. Experimental scheme for treatment of oncogene-activated primary MT-HCC with control antibody or combination therapy with TgfbrI and anti-Pdl1 inhibitor. Liver tumor burden is assessed at the end of treatment. f. Quantification of liver tumor burden, gross images, and H&E images of primary MT-HCC treated with control antibody (n=4 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n=4 mice). Data are presented as mean values +/− SEM. Two-tailed unpaired t-test was used to compare the groups. g. Apoptosis measured by cleaved caspase 3+ cells in liver MRD of MYC-HCC mice treated with control (n=5 mice) versus Tgfbr1 (n=5 mice) or Pdl1 inhibitors (n=5 mice) or combination therapy with Tgfbr1 and Pdl1 inhibitors (n=5 mice). Data are presented as mean values +/− SEM. Two-tailed unpaired t-test was used to compare the groups. h. Flow cytometry quantification of activated CD4 and CD8 T cells which are CD69+/CD44^high in MRD of mice treated with control antibody (n=3 mice) or combination therapy with Tgfbr1 and Pdl1 inhibitors (n=3 mice). Bar plots compare the mean between the groups with unpaired t-tests. Data are presented as mean values +/− SEM. Two-tailed unpaired t-test was used to compare the groups. Abbreviations: HCC- hepatocellular carcinoma, PDL1- programmed cell death ligand 1, TGFBr1 transforming growth factor receptor beta 1, H&E-hematoxylin and eosin, Ctrl- control, MRD- Minimal residual disease, MT-HCC- MYC/Twist hepatocellular carcinoma, CD- cluster of differentiation.

Extended Data Fig. 10 | — a. Schematic showing the establishment of mouse model of control (n=5 mice) or doxorubicin-resistant syngeneic orthotopic allografts (n=11 mice) which were then treated with control antibody (n=6 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n=10 mice). b.Representative gross images, H&E and IHC images of control or doxorubicin-resistant orthotopic HCC-bearing mice treated with control antibody (n=6 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n=10 mice). IHC for CD8 T cells shows in the bottom panel. c. Quantification of tumor burden and CD8T cells in the liver of control or doxorubicin-resistant orthotopic HCC-bearing mice treated with control antibody (n=6 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n=10 mice). Bar plots compare the mean between the groups with two-sided unpaired t-tests. Data are presented as mean values +/− SEM. d. Flow cytometry-based quantification of tumor-infiltrating leukocytes, T cells, M2-like (PDL1+ or CD206+) and M1-like macrophage (CD86+ or MHCII+) subsets in doxorubicin-resistant orthotopic HCC-bearing mice treated with control antibody (n=4 mice) or combination therapy with TgfbrI and anti-Pdl1 inhibitor (n=4 mice). Bar plots compare the mean between the groups two-sided with unpaired t-tests. Data are presented as mean values +/− SEM. e. Representative flow cytometry images and quantification of M1-like and M2-like macrophages within the tumor in the doxorubicin-resistant orthotopic HCC-bearing mice treated with control antibody or combination therapy with TgfbrI and anti-Pdl1 inhibitor. f. Hep 53.4 HCC cell lines are treated with control or doxorubicin for 96 hours and viable cells are selected for orthotopic implantation into mouse liver, confirmed by MRI. g. Flow cytometry analysis shows no difference in leukocyte or T cell or macrophage infiltration in WT mice bearing control HCC treated with either control antibody (n=2 mice) of combined inhibition of Tgfbr1 and anti-Pdl1 antibody (n=3 mice). Data are presented as mean values +/− SEM. h. Schematic summarizing the main findings- i. The spatial organization of post-TACE residual HCC is unveiled through integrated analysis of single-cell spatial profiling employing CODEX and GeoMx spatial transcriptomics of human HCC. ii. Mouse model of minimal residual disease (MRD) reveals that TGFβ1 derived from PDL1+ macrophages enable the persistence of residual stem-like tumor cells and induces exhaustion in CD8T cells. iii. In two mouse models of MRD, we target the TGFβ pathway and PDL1, eliminating residual tumor cells and preventing HCC recurrence. *Abbreviations: HCC- hepatocellular carcinoma, Tgfb -transforming growth factor beta, H&E-hematoxylin and eosin, Ctrl-control, Res-residual, MRD- Minimal residual disease, DoxR- doxorubicin-resistant*.

Supplementary Material

Supplemental Figures

NIHMS2094406-supplement-Supplemental_Figures.pdf^{(8.9MB, pdf)}

Supplementary Tables

NIHMS2094406-supplement-Supplementary_Tables.xlsx^{(443.2KB, xlsx)}

Source Data Main Fig

NIHMS2094406-supplement-Source_Data_Main_Fig.xlsx^{(145.4KB, xlsx)}

Source Data Ext Fig

NIHMS2094406-supplement-Source_Data_Ext_Fig.xlsx^{(78KB, xlsx)}

The online version contains supplementary material available at https://doi.org/10.1038/s43018-024-00828-8.

Acknowledgements

R.D. acknowledges NIH grant CA222676 from the National Cancer Institute, American College of Gastroenterology Junior Faculty Career Development Grant and Cancer League Award. D.F. acknowledges NIH grant CA208735 and CA253180 from the National Cancer Institute. P. Chu helped with mouse histology services.

Footnotes

Reporting summary

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

Competing interests

A.T.M., A.E.T., R.P. and H.B.D. are employees of Enable Medicine. All other authors declare no competing interests.

Extended data is available for this paper at https://doi.org/10.1038/s43018-024-00828-8.

Data availability

Source data for all figures have been provided as Source Data files. All other data supporting the findings of this study are available from the corresponding author on reasonable request. CODEX data (the raw data matrix), which includes cell-level information on normalized expression for all markers, centroid coordinates and the assigned cell type and cohort for each cell, have been deposited in Figshare (ref. 73). In addition, if raw images are sought, we will make them available at https://app.enablemedicine.com/portal/visualizer upon request to the corresponding author (dhanaser@stanford.edu) within 10 days of receiving the request. Nanostring data: the gene expression data matrix, which includes sample-level information on normalized expression for all genes, and the assigned AOI type and cohort for each sample, has been deposited in Figshare (ref. 74). The mouse single-cell and whole-transcriptome sequencing data generated in this study are publicly available in the National Center for Biotechnology Information Gene Expression Omnibus database under accession codes GSE243176, GSE242746, GSE242745 and GSE242743. Publicly available TCGA liver cancer data were used for analysis from https://www.cbioportal.org/study/summary?id=lihc_tcga_pan_can_atlas_2018. Source data are provided with this paper.

Code availability

No custom code or algorithms were generated in this manuscript.

References

1.Lee Y-T et al. The mortality and overall survival trends of primary liver cancer in the United States. J. Natl. Cancer Inst 113, 1531–1541 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Lei J et al. Response to transarterial chemoembolization may serve as selection criteria for hepatocellular carcinoma liver transplantation. Oncotarget 8, 91328–91342 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Kim DJ et al. Recurrence of hepatocellular carcinoma: importance of mRECIST response to chemoembolization and tumor size. Am. J. Transplant 14, 1383–1390 (2014). [DOI] [PubMed] [Google Scholar]
4.Agopian VG et al. Impact of pretransplant bridging locoregional therapy for patients with hepatocellular carcinoma within Milan criteria undergoing liver transplantation: analysis of 3601 patients from the US multicenter HCC transplant consortium. Ann. Surg 266, 525–535 (2017). [DOI] [PubMed] [Google Scholar]
5.Adeniji N et al. Impact of bridging locoregional therapies for hepatocellular carcinoma on post-transplant clinical outcome. Clin. Transplant 34, e14128 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Ghajar CM Metastasis prevention by targeting the dormant niche. Nat. Rev. Cancer 15, 238–247 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Lee TK-W, Guan X-Y & Ma S Cancer stem cells in hepatocellular carcinoma - from origin to clinical implications. Nat. Rev. Gastroenterol. Hepatol 19, 26–44 (2022). [DOI] [PubMed] [Google Scholar]
8.Price TT et al. Dormant breast cancer micrometastases reside in specific bone marrow niches that regulate their transit to and from bone. Sci. Transl. Med 8, 340ra73 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Shachaf CM et al. MYC inactivation uncovers pluripotent differentiation and tumour dormancy in hepatocellular cancer. Nature 431, 1112–1117 (2004). [DOI] [PubMed] [Google Scholar]
10.Ding W et al. Clinicopathologic and prognostic significance of tumor-associated macrophages in patients with hepatocellular carcinoma: a meta-analysis. PLoS ONE 14, e0223971 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
11.Zhu X-D et al. High expression of macrophage colony-stimulating factor in peritumoral liver tissue is associated with poor survival after curative resection of hepatocellular carcinoma. J. Clin. Oncol 26, 2707–2716 (2008). [DOI] [PubMed] [Google Scholar]
12.Borriello L et al. Primary tumor associated macrophages activate programs of invasion and dormancy in disseminating tumor cells. Nat. Commun 13, 626 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Dhanasekaran R et al. MYC overexpression drives immune evasion in hepatocellular carcinoma that is reversible through restoration of proinflammatory macrophages. Cancer Res. 83, 626–640 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Dhanasekaran R et al. MYC and Twist1 cooperate to drive metastasis by eliciting crosstalk between cancer and innate immunity. eLife 10.7554/elife.50731 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Black S et al. CODEX multiplexed tissue imaging with DNA-conjugated antibodies. Nat. Protoc 16, 3802–3835 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Uenishi T et al. Cytokeratin 19 expression in hepatocellular carcinoma predicts early postoperative recurrence. Cancer Sci. 94, 851–857 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Khosla R et al. EpCAM⁺ liver cancer stem-like cells exhibiting autocrine wnt signaling potentially originate in cirrhotic patients. Stem Cells Transl. Med 6, 807–818 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Liu C-Q et al. Expression patterns of programmed death ligand 1 correlate with different microenvironments and patient prognosis in hepatocellular carcinoma. Br. J. Cancer 119, 80–88 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Pfister D et al. NASH limits anti-tumour surveillance in immunotherapy-treated HCC. Nature 592, 450–456 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Cancer Genome Atlas Research Network. Comprehensive and integrative genomic characterization of hepatocellular carcinoma. Cell 169, 1327–1341.e23 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Guo Y et al. The heterogeneity of immune cell infiltration landscape and its immunotherapeutic implications in hepatocellular carcinoma. Front. Immunol 13, 861525 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Yu S et al. Tumor-infiltrating immune cells in hepatocellular carcinoma: Tregs is correlated with poor overall survival. PLoS ONE 15, e0231003 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Veglia F et al. Analysis of classical neutrophils and polymorphonuclear myeloid-derived suppressor cells in cancer patients and tumor-bearing mice. J. Exp. Med 218, e20201803 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Salcher S et al. High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell 40, 1503–1520.e8 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Singal AG et al. AASLD practice guidance on prevention, diagnosis, and treatment of hepatocellular carcinoma. Hepatology 78, 1922–1965 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Schürch CM et al. Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell 183, 838 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Dasari A, Grothey A & Kopetz S Circulating tumor DNA-defined minimal residual disease in solid tumors: opportunities to accelerate the development of adjuvant therapies. J. Clin. Oncol 36, JCO2018789032 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Villanueva A et al. Combining clinical, pathology, and gene expression data to predict recurrence of hepatocellular carcinoma. Gastroenterology 140, 1501–12.e2 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Pello OM et al. Role of c-MYC in alternative activation of human macrophages and tumor-associated macrophage biology. Blood 119, 411–421 (2012). [DOI] [PubMed] [Google Scholar]
30.Li Z et al. Cancer-associated fibroblasts promote PD-L1 expression in mice cancer cells via secreting CXCL5. Int. J. Cancer 145, 1946–1957 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Zhang W et al. IL-6 promotes PD-L1 expression in monocytes and macrophages by decreasing protein tyrosine phosphatase receptor type O expression in human hepatocellular carcinoma. J. Immunother. Cancer 8, e000285 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Xi X et al. Interleukin-22 promotes PD-L1 expression via STAT3 in colon cancer cells. Oncol. Lett 22, 716 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Kobori T et al. Interleukin-18 amplifies macrophage polarization and morphological alteration, leading to excessive angiogenesis. Front. Immunol 9, 334 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Patel M & Pillai A Management of intermediate-stage hepatocellular carcinoma: systemic versus locoregional therapy. Surg. Oncol. Clin. N. Am 33, 159–172 (2024). [DOI] [PubMed] [Google Scholar]
35.Lu Y et al. A single-cell atlas of the multicellular ecosystem of primary and metastatic hepatocellular carcinoma. Nat. Commun 13, 4594 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Ma L et al. Single-cell atlas of tumor cell evolution in response to therapy in hepatocellular carcinoma and intrahepatic cholangiocarcinoma. J. Hepatol 75, 1397–1408 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Ma L et al. Tumor cell biodiversity drives microenvironmental reprogramming in liver cancer. Cancer Cell 36, 418–430.e6 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Ruf B et al. Tumor-associated macrophages trigger MAIT cell dysfunction at the HCC invasive margin. Cell 186, 3686–3705.e32 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Maestri E et al. Spatial proximity of tumor-immune interactions predicts patient outcome in hepatocellular carcinoma. Hepatology 10.1097/hep.0000000000000600 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Govaere O et al. Keratin 19: a key role player in the invasion of human hepatocellular carcinomas. Gut 63, 674–685 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Zhang J et al. Overexpression of Epcam and CD133 correlates with poor prognosis in dual-phenotype hepatocellular carcinoma. J. Cancer 11, 3400–3406 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Noh C-K et al. EpCAM as a predictive marker of tumor recurrence and survival in patients who underwent surgical resection for hepatocellular carcinoma. Anticancer Res. 38, 4101–4109 (2018). [DOI] [PubMed] [Google Scholar]
43.Zhou C et al. Disruption of SLFN11 deficiency-induced CCL2 signaling and macrophage M2 polarization potentiates anti-PD-1 therapy efficacy in hepatocellular carcinoma. Gastroenterology 164, 1261–1278 (2023). [DOI] [PubMed] [Google Scholar]
44.Murai H et al. Multiomics identifies the link between intratumor steatosis and the exhausted tumor immune microenvironment in hepatocellular carcinoma. Hepatology 77, 77–91 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Li X et al. Targeting of tumour-infiltrating macrophages via CCL2/CCR2 signalling as a therapeutic strategy against hepatocellular carcinoma. Gut 66, 157–167 (2017). [DOI] [PubMed] [Google Scholar]
46.Martin-Padura I et al. Residual dormant cancer stem-cell foci are responsible for tumor relapse after antiangiogenic metronomic therapy in hepatocellular carcinoma xenografts. Lab. Invest 92, 952–966 (2012). [DOI] [PubMed] [Google Scholar]
47.Haraguchi N et al. CD13 is a therapeutic target in human liver cancer stem cells. J. Clin. Invest 120, 3326–3339 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Schrader J et al. Matrix stiffness modulates proliferation, chemotherapeutic response, and dormancy in hepatocellular carcinoma cells. Hepatology 53, 1192–1205 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Ravi R et al. Bifunctional immune checkpoint-targeted antibody-ligand traps that simultaneously disable TGFβ enhance the efficacy of cancer immunotherapy. Nat. Commun 9, 741 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Mariathasan S et al. TGFβ attenuates tumour response to PD-L1 blockade by contributing to exclusion of T cells. Nature 554, 544–548 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Lan Y et al. Enhanced preclinical antitumor activity of M7824, a bifunctional fusion protein simultaneously targeting PD-L1 and TGF-β. Sci. Transl. Med 10, eaan5488 (2018). [DOI] [PubMed] [Google Scholar]
52.Metropulos AE, Munshi HG & Principe DR The difficulty in translating the preclinical success of combined TGFβ and immune checkpoint inhibition to clinical trial. eBioMedicine 86, 104380 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
53.Lencioni R & Llovet JM Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin. Liver Dis 30, 52–60 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
54.Merritt CR et al. Multiplex digital spatial profiling of proteins and RNA in fixed tissue. Nat. Biotechnol 38, 586–599 (2020). [DOI] [PubMed] [Google Scholar]
55.Cosi C, Millar M, Beltran M, Sherry L & Gatti-McArthur S Quantitative analysis of RNAscope staining for expression in mouse brain tissue as a measure of neuronal activation. MethodsX 8, 101348 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
56.Bankhead P et al. QuPath: open source software for digital pathology image analysis. Sci. Rep 7, 16878 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Genin M, Clement F, Fattaccioli A, Raes M & Michiels C M1 and M2 macrophages derived from THP-1 cells differentially modulate the response of cancer cells to etoposide. BMC Cancer 15, 577 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Gambari R, Hau DK-P, Wong W-Y & Chui C-H Sensitization of Hep3B hepatoma cells to cisplatin and doxorubicin by corilagin. Phytother. Res 28, 781–783 (2014). [DOI] [PubMed] [Google Scholar]
59.Hu Q-D et al. NSC 74859 enhances doxorubicin cytotoxicity via inhibition of epithelial-mesenchymal transition in hepatocellular carcinoma cells. Cancer Lett. 325, 207–213 (2012). [DOI] [PubMed] [Google Scholar]
60.Buschauer S et al. Hepatocellular carcinoma cells surviving doxorubicin treatment exhibit increased migratory potential and resistance to doxorubicin re-treatment in vitro. Oncol. Lett 15 10.3892/ol.2018.7887 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
61.Namur J et al. Embolization of hepatocellular carcinoma with drug-eluting beads: doxorubicin tissue concentration and distribution in patient liver explants. J. Hepatol 55, 1332–1338 (2011). [DOI] [PubMed] [Google Scholar]
62.Ulbrich J, Lopez-Salmeron V & Gerrard I BD Rhapsody^™ single-cell analysis system workflow: from sample to multimodal single-cell sequencing data. Methods Mol. Biol 2584, 29–56 (2023). [DOI] [PubMed] [Google Scholar]
63.Hao Y et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat. Biotechnol 10.1038/s41587-023-01767-y (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Nielsen MC, Andersen MN & Møller HJ Monocyte isolation techniques significantly impact the phenotype of both isolated monocytes and derived macrophages in vitro. Immunology 159, 63–74 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Wu PV & Nusse R 3D culture of primary patient-derived hepatoblastoma tumoroids. Methods Mol. Biol 2544, 259–267 (2022). [DOI] [PubMed] [Google Scholar]
66.Wang SK, Xue Y & Cepko CL Microglia modulation by TGF-β1 protects cones in mouse models of retinal degeneration. J. Clin. Invest 130, 4360–4369 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Kress S et al. p53 mutations are absent from carcinogen-induced mouse liver tumors but occur in cell lines established from these tumors. Mol. Carcinog 6, 148–158 (1992). [DOI] [PubMed] [Google Scholar]
68.Brown ZJ, Heinrich B & Greten TF Establishment of orthotopic liver tumors by surgical intrahepatic tumor injection in mice with underlying non-alcoholic fatty liver disease. Methods Protoc. 1, 21 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Zabransky DJ et al. Profiling of syngeneic mouse HCC tumor models as a framework to understand anti-PD-1 sensitive tumor microenvironments. Hepatology 77, 1566–1579 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Lai I et al. Lipid nanoparticles that deliver IL-12 messenger RNA suppress tumorigenesis in MYC oncogene-driven hepatocellular carcinoma. J. Immunother. Cancer 6, 125 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
71.Dhanasekaran R et al. MYC ASO impedes tumorigenesis and elicits oncogene addiction in autochthonous transgenic mouse models of HCC and RCC. Mol. Ther. Nucleic Acids 21, 850–859 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Palmer GM et al. In vivo optical molecular imaging and analysis in mice using dorsal window chamber models applied to hypoxia, vasculature and fluorescent reporters. Nat. Protoc 6, 1355–1366 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Dhanasekaran R Spatial analysis reveals targetable macrophage-mediated mechanisms of immune evasion in hepatocellular carcinoma minimal residual disease. Figshare 10.6084/m9.figshare.26539345 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Dhanasekaran R Nanostring_GeoMx_HCC_9_2_24.csv. Figshare 10.6084/m9.figshare.26898886.v1 (2024). [DOI] [Google Scholar]

Associated Data

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

Supplementary Materials

Supplemental Figures

NIHMS2094406-supplement-Supplemental_Figures.pdf^{(8.9MB, pdf)}

Supplementary Tables

NIHMS2094406-supplement-Supplementary_Tables.xlsx^{(443.2KB, xlsx)}

Source Data Main Fig

NIHMS2094406-supplement-Source_Data_Main_Fig.xlsx^{(145.4KB, xlsx)}

Source Data Ext Fig

NIHMS2094406-supplement-Source_Data_Ext_Fig.xlsx^{(78KB, xlsx)}

Data Availability Statement

[R1] 1.Lee Y-T et al. The mortality and overall survival trends of primary liver cancer in the United States. J. Natl. Cancer Inst 113, 1531–1541 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R2] 2.Lei J et al. Response to transarterial chemoembolization may serve as selection criteria for hepatocellular carcinoma liver transplantation. Oncotarget 8, 91328–91342 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R3] 3.Kim DJ et al. Recurrence of hepatocellular carcinoma: importance of mRECIST response to chemoembolization and tumor size. Am. J. Transplant 14, 1383–1390 (2014). [DOI] [PubMed] [Google Scholar]

[R4] 4.Agopian VG et al. Impact of pretransplant bridging locoregional therapy for patients with hepatocellular carcinoma within Milan criteria undergoing liver transplantation: analysis of 3601 patients from the US multicenter HCC transplant consortium. Ann. Surg 266, 525–535 (2017). [DOI] [PubMed] [Google Scholar]

[R5] 5.Adeniji N et al. Impact of bridging locoregional therapies for hepatocellular carcinoma on post-transplant clinical outcome. Clin. Transplant 34, e14128 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R6] 6.Ghajar CM Metastasis prevention by targeting the dormant niche. Nat. Rev. Cancer 15, 238–247 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R7] 7.Lee TK-W, Guan X-Y & Ma S Cancer stem cells in hepatocellular carcinoma - from origin to clinical implications. Nat. Rev. Gastroenterol. Hepatol 19, 26–44 (2022). [DOI] [PubMed] [Google Scholar]

[R8] 8.Price TT et al. Dormant breast cancer micrometastases reside in specific bone marrow niches that regulate their transit to and from bone. Sci. Transl. Med 8, 340ra73 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R9] 9.Shachaf CM et al. MYC inactivation uncovers pluripotent differentiation and tumour dormancy in hepatocellular cancer. Nature 431, 1112–1117 (2004). [DOI] [PubMed] [Google Scholar]

[R10] 10.Ding W et al. Clinicopathologic and prognostic significance of tumor-associated macrophages in patients with hepatocellular carcinoma: a meta-analysis. PLoS ONE 14, e0223971 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]

[R11] 11.Zhu X-D et al. High expression of macrophage colony-stimulating factor in peritumoral liver tissue is associated with poor survival after curative resection of hepatocellular carcinoma. J. Clin. Oncol 26, 2707–2716 (2008). [DOI] [PubMed] [Google Scholar]

[R12] 12.Borriello L et al. Primary tumor associated macrophages activate programs of invasion and dormancy in disseminating tumor cells. Nat. Commun 13, 626 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R13] 13.Dhanasekaran R et al. MYC overexpression drives immune evasion in hepatocellular carcinoma that is reversible through restoration of proinflammatory macrophages. Cancer Res. 83, 626–640 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R14] 14.Dhanasekaran R et al. MYC and Twist1 cooperate to drive metastasis by eliciting crosstalk between cancer and innate immunity. eLife 10.7554/elife.50731 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R15] 15.Black S et al. CODEX multiplexed tissue imaging with DNA-conjugated antibodies. Nat. Protoc 16, 3802–3835 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] 16.Uenishi T et al. Cytokeratin 19 expression in hepatocellular carcinoma predicts early postoperative recurrence. Cancer Sci. 94, 851–857 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R17] 17.Khosla R et al. EpCAM⁺ liver cancer stem-like cells exhibiting autocrine wnt signaling potentially originate in cirrhotic patients. Stem Cells Transl. Med 6, 807–818 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.Liu C-Q et al. Expression patterns of programmed death ligand 1 correlate with different microenvironments and patient prognosis in hepatocellular carcinoma. Br. J. Cancer 119, 80–88 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] 19.Pfister D et al. NASH limits anti-tumour surveillance in immunotherapy-treated HCC. Nature 592, 450–456 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] 20.Cancer Genome Atlas Research Network. Comprehensive and integrative genomic characterization of hepatocellular carcinoma. Cell 169, 1327–1341.e23 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R21] 21.Guo Y et al. The heterogeneity of immune cell infiltration landscape and its immunotherapeutic implications in hepatocellular carcinoma. Front. Immunol 13, 861525 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R22] 22.Yu S et al. Tumor-infiltrating immune cells in hepatocellular carcinoma: Tregs is correlated with poor overall survival. PLoS ONE 15, e0231003 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.Veglia F et al. Analysis of classical neutrophils and polymorphonuclear myeloid-derived suppressor cells in cancer patients and tumor-bearing mice. J. Exp. Med 218, e20201803 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] 24.Salcher S et al. High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell 40, 1503–1520.e8 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R25] 25.Singal AG et al. AASLD practice guidance on prevention, diagnosis, and treatment of hepatocellular carcinoma. Hepatology 78, 1922–1965 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R26] 26.Schürch CM et al. Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell 183, 838 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] 27.Dasari A, Grothey A & Kopetz S Circulating tumor DNA-defined minimal residual disease in solid tumors: opportunities to accelerate the development of adjuvant therapies. J. Clin. Oncol 36, JCO2018789032 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R28] 28.Villanueva A et al. Combining clinical, pathology, and gene expression data to predict recurrence of hepatocellular carcinoma. Gastroenterology 140, 1501–12.e2 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.Pello OM et al. Role of c-MYC in alternative activation of human macrophages and tumor-associated macrophage biology. Blood 119, 411–421 (2012). [DOI] [PubMed] [Google Scholar]

[R30] 30.Li Z et al. Cancer-associated fibroblasts promote PD-L1 expression in mice cancer cells via secreting CXCL5. Int. J. Cancer 145, 1946–1957 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] 31.Zhang W et al. IL-6 promotes PD-L1 expression in monocytes and macrophages by decreasing protein tyrosine phosphatase receptor type O expression in human hepatocellular carcinoma. J. Immunother. Cancer 8, e000285 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R32] 32.Xi X et al. Interleukin-22 promotes PD-L1 expression via STAT3 in colon cancer cells. Oncol. Lett 22, 716 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] 33.Kobori T et al. Interleukin-18 amplifies macrophage polarization and morphological alteration, leading to excessive angiogenesis. Front. Immunol 9, 334 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R34] 34.Patel M & Pillai A Management of intermediate-stage hepatocellular carcinoma: systemic versus locoregional therapy. Surg. Oncol. Clin. N. Am 33, 159–172 (2024). [DOI] [PubMed] [Google Scholar]

[R35] 35.Lu Y et al. A single-cell atlas of the multicellular ecosystem of primary and metastatic hepatocellular carcinoma. Nat. Commun 13, 4594 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] 36.Ma L et al. Single-cell atlas of tumor cell evolution in response to therapy in hepatocellular carcinoma and intrahepatic cholangiocarcinoma. J. Hepatol 75, 1397–1408 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R37] 37.Ma L et al. Tumor cell biodiversity drives microenvironmental reprogramming in liver cancer. Cancer Cell 36, 418–430.e6 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R38] 38.Ruf B et al. Tumor-associated macrophages trigger MAIT cell dysfunction at the HCC invasive margin. Cell 186, 3686–3705.e32 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.Maestri E et al. Spatial proximity of tumor-immune interactions predicts patient outcome in hepatocellular carcinoma. Hepatology 10.1097/hep.0000000000000600 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R40] 40.Govaere O et al. Keratin 19: a key role player in the invasion of human hepatocellular carcinomas. Gut 63, 674–685 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R41] 41.Zhang J et al. Overexpression of Epcam and CD133 correlates with poor prognosis in dual-phenotype hepatocellular carcinoma. J. Cancer 11, 3400–3406 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R42] 42.Noh C-K et al. EpCAM as a predictive marker of tumor recurrence and survival in patients who underwent surgical resection for hepatocellular carcinoma. Anticancer Res. 38, 4101–4109 (2018). [DOI] [PubMed] [Google Scholar]

[R43] 43.Zhou C et al. Disruption of SLFN11 deficiency-induced CCL2 signaling and macrophage M2 polarization potentiates anti-PD-1 therapy efficacy in hepatocellular carcinoma. Gastroenterology 164, 1261–1278 (2023). [DOI] [PubMed] [Google Scholar]

[R44] 44.Murai H et al. Multiomics identifies the link between intratumor steatosis and the exhausted tumor immune microenvironment in hepatocellular carcinoma. Hepatology 77, 77–91 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R45] 45.Li X et al. Targeting of tumour-infiltrating macrophages via CCL2/CCR2 signalling as a therapeutic strategy against hepatocellular carcinoma. Gut 66, 157–167 (2017). [DOI] [PubMed] [Google Scholar]

[R46] 46.Martin-Padura I et al. Residual dormant cancer stem-cell foci are responsible for tumor relapse after antiangiogenic metronomic therapy in hepatocellular carcinoma xenografts. Lab. Invest 92, 952–966 (2012). [DOI] [PubMed] [Google Scholar]

[R47] 47.Haraguchi N et al. CD13 is a therapeutic target in human liver cancer stem cells. J. Clin. Invest 120, 3326–3339 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R48] 48.Schrader J et al. Matrix stiffness modulates proliferation, chemotherapeutic response, and dormancy in hepatocellular carcinoma cells. Hepatology 53, 1192–1205 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R49] 49.Ravi R et al. Bifunctional immune checkpoint-targeted antibody-ligand traps that simultaneously disable TGFβ enhance the efficacy of cancer immunotherapy. Nat. Commun 9, 741 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R50] 50.Mariathasan S et al. TGFβ attenuates tumour response to PD-L1 blockade by contributing to exclusion of T cells. Nature 554, 544–548 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R51] 51.Lan Y et al. Enhanced preclinical antitumor activity of M7824, a bifunctional fusion protein simultaneously targeting PD-L1 and TGF-β. Sci. Transl. Med 10, eaan5488 (2018). [DOI] [PubMed] [Google Scholar]

[R52] 52.Metropulos AE, Munshi HG & Principe DR The difficulty in translating the preclinical success of combined TGFβ and immune checkpoint inhibition to clinical trial. eBioMedicine 86, 104380 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R53] 53.Lencioni R & Llovet JM Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin. Liver Dis 30, 52–60 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R54] 54.Merritt CR et al. Multiplex digital spatial profiling of proteins and RNA in fixed tissue. Nat. Biotechnol 38, 586–599 (2020). [DOI] [PubMed] [Google Scholar]

[R55] 55.Cosi C, Millar M, Beltran M, Sherry L & Gatti-McArthur S Quantitative analysis of RNAscope staining for expression in mouse brain tissue as a measure of neuronal activation. MethodsX 8, 101348 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R56] 56.Bankhead P et al. QuPath: open source software for digital pathology image analysis. Sci. Rep 7, 16878 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R57] 57.Genin M, Clement F, Fattaccioli A, Raes M & Michiels C M1 and M2 macrophages derived from THP-1 cells differentially modulate the response of cancer cells to etoposide. BMC Cancer 15, 577 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R58] 58.Gambari R, Hau DK-P, Wong W-Y & Chui C-H Sensitization of Hep3B hepatoma cells to cisplatin and doxorubicin by corilagin. Phytother. Res 28, 781–783 (2014). [DOI] [PubMed] [Google Scholar]

[R59] 59.Hu Q-D et al. NSC 74859 enhances doxorubicin cytotoxicity via inhibition of epithelial-mesenchymal transition in hepatocellular carcinoma cells. Cancer Lett. 325, 207–213 (2012). [DOI] [PubMed] [Google Scholar]

[R60] 60.Buschauer S et al. Hepatocellular carcinoma cells surviving doxorubicin treatment exhibit increased migratory potential and resistance to doxorubicin re-treatment in vitro. Oncol. Lett 15 10.3892/ol.2018.7887 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R61] 61.Namur J et al. Embolization of hepatocellular carcinoma with drug-eluting beads: doxorubicin tissue concentration and distribution in patient liver explants. J. Hepatol 55, 1332–1338 (2011). [DOI] [PubMed] [Google Scholar]

[R62] 62.Ulbrich J, Lopez-Salmeron V & Gerrard I BD Rhapsody^™ single-cell analysis system workflow: from sample to multimodal single-cell sequencing data. Methods Mol. Biol 2584, 29–56 (2023). [DOI] [PubMed] [Google Scholar]

[R63] 63.Hao Y et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat. Biotechnol 10.1038/s41587-023-01767-y (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R64] 64.Nielsen MC, Andersen MN & Møller HJ Monocyte isolation techniques significantly impact the phenotype of both isolated monocytes and derived macrophages in vitro. Immunology 159, 63–74 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R65] 65.Wu PV & Nusse R 3D culture of primary patient-derived hepatoblastoma tumoroids. Methods Mol. Biol 2544, 259–267 (2022). [DOI] [PubMed] [Google Scholar]

[R66] 66.Wang SK, Xue Y & Cepko CL Microglia modulation by TGF-β1 protects cones in mouse models of retinal degeneration. J. Clin. Invest 130, 4360–4369 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R67] 67.Kress S et al. p53 mutations are absent from carcinogen-induced mouse liver tumors but occur in cell lines established from these tumors. Mol. Carcinog 6, 148–158 (1992). [DOI] [PubMed] [Google Scholar]

[R68] 68.Brown ZJ, Heinrich B & Greten TF Establishment of orthotopic liver tumors by surgical intrahepatic tumor injection in mice with underlying non-alcoholic fatty liver disease. Methods Protoc. 1, 21 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R69] 69.Zabransky DJ et al. Profiling of syngeneic mouse HCC tumor models as a framework to understand anti-PD-1 sensitive tumor microenvironments. Hepatology 77, 1566–1579 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R70] 70.Lai I et al. Lipid nanoparticles that deliver IL-12 messenger RNA suppress tumorigenesis in MYC oncogene-driven hepatocellular carcinoma. J. Immunother. Cancer 6, 125 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R71] 71.Dhanasekaran R et al. MYC ASO impedes tumorigenesis and elicits oncogene addiction in autochthonous transgenic mouse models of HCC and RCC. Mol. Ther. Nucleic Acids 21, 850–859 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R72] 72.Palmer GM et al. In vivo optical molecular imaging and analysis in mice using dorsal window chamber models applied to hypoxia, vasculature and fluorescent reporters. Nat. Protoc 6, 1355–1366 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R73] 73.Dhanasekaran R Spatial analysis reveals targetable macrophage-mediated mechanisms of immune evasion in hepatocellular carcinoma minimal residual disease. Figshare 10.6084/m9.figshare.26539345 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R74] 74.Dhanasekaran R Nanostring_GeoMx_HCC_9_2_24.csv. Figshare 10.6084/m9.figshare.26898886.v1 (2024). [DOI] [Google Scholar]

PERMALINK

Spatial analysis reveals targetable macrophage-mediated mechanisms of immune evasion in hepatocellular carcinoma minimal residual disease

Lea Lemaitre

Nia Adeniji

Akanksha Suresh

Reshma Reguram

Josephine Zhang

Jangho Park

Amit Reddy

Alexandro E Trevino

Aaron T Mayer

Anja Deutzmann

Aida S Hansen

Ling Tong

Vinodhini Arjunan

Neeraja Kambham

Brendan C Visser

Monica M Dua

C Andrew Bonham

Nishita Kothary

H Blaize D’Angio

Ryan Preska

Yanay Rosen

James Zou

Vivek Charu

Dean W Felsher

Renumathy Dhanasekaran

Abstract

Results

Single-cell spatial atlas reveals heterogeneity in human HCC

Fig. 1 |. Single-cell spatial atlas of the tumor immune microenvironment in HCC.

Residual human HCC has enrichment of immunosuppressive cells

Fig. 2 |. Immunosuppressive cell enrichment and tissue remodeling in residual HCC.

Spatial interactions in residual HCC drive immune evasion

Fig. 3 |. Remodeling of spatial cellular interactions in residual HCC.

Spatial neighborhoods reprogram immune cells in residual HCC

Fig. 4 |. Synchronized spatial remodeling of cellular neighborhoods in residual HCC.

TGFβ pathway activation drives persistence of residual tumor

Fig. 5 |. TGFβ pathway activation promotes persistence of residual tumor cells in HCC.

Transgenic mouse model shows recurrence can arise from MRD

Fig. 6 |. Recurrence of HCC arises from MRD in a transgenic mouse model of HCC.

Residual tumor cells in mouse MRD demonstrate cancer stemness

Tgfβ pathway is activated in residual tumor cells in mouse models of MRD

Mechanisms of immune evasion by residual tumor cells

Fig. 7 |. Mechanisms of immune evasion by cancer stem cells in the mouse MRD niche.

Inhibition of Tgfβr1 and Pdl1 blocks HCC recurrence from MRD

Fig. 8 |. Blockade of Tgfβr1 and Pdl1 prevents recurrence from MRD in mouse HCC.

Discussion

Methods

Ethical approval

Patient cohort selection

Construction of tissue microarrays

CODEX panel development and staining

CODEX multiplex staining and analysis.

CODEX multiplexed imaging and processing.

CODEX data analysis.

Cell clustering.

Construction of a spatial cellular interaction graph.

Cell neighborhood analysis.

NanoString digital spatial profiling

Multiplex fluorescent in situ hybridization

In vitro 3D co-culture and scRNA-seq

Human monocyte-derived macrophages and patient-derived 3D tumoroids

Transgenic mice and in vivo treatment

Establishment and treatment of immunocompetent syngeneic allograft models of HCC

Animal magnetic resonance imaging

Intravital microscopy

Whole-transcriptome sequencing of mouse tumors

scRNA-seq of mouse liver

MRD in MYC/Twist-HCC liver.

Residual tumor cells from in vitro studies.

Immunohistochemistry

Immunofluorescence

Flow cytometry

Macrophage phagocytosis assay

Luminex assay

Statistical analysis and reproducibility.

Extended Data

Extended Data Fig. 1 |. Identification, validation and relative proportion of major cell types using CODEX.

Extended Data Fig. 2 |. Comparison of immune cell distribution in residual HCC.