Human liver organoids model progressive inflammatory and fibrotic injury in non-alcoholic fatty liver disease

Abstract

Chronic liver injury promotes inflammation, which can progress to fibrosis and cirrhosis, a major cause of mortality. Resolving the transcriptional changes orchestrating this transformation remains challenging. Here, we derive single-cell transcriptomic maps of progressive liver injury in human pluripotent stem cell (hPSC) - derived liver organoids (HLOs), modeling steatohepatitis with palmitic acid and fibrosis through TGF-β1 treatment. We observe that palmitic acid drives inflammation and non-alcoholic fatty liver disease (NAFLD) expression signatures, while TGF-β1 expands hepatic stellate-like populations, remodels cell cycle patterning, and induces extracellular matrix pathways. Analysis of receptor-ligand expression defines the induction of genes regulating Notch and fatty acid signaling with palmitic acid treatment, while the TGF-β1 response is shaped by the co-expression of COL1A1 and integrins to promote crosstalk between hepatocytes, cholangiocytes, and stellate cells. Finally, inflamed and fibrotic HLOs sequentially induce genes predicting disease progression in NAFLD. Our findings highlight HLOs as dynamic human in vitro systems to study evolving liver injury, providing a single-cell transcriptomic reference that will facilitate benchmarking future organoid-based liver injury models.

Introduction

Chronic liver injury promotes sustained inflammation, leading to liver fibrosis, which can progress to cirrhosis, a major cause of morbidity and mortality worldwide^{1, 2}. Non-alcoholic fatty liver disease (NAFLD) is among the most common causes of chronic liver disease and the most rapidly increasing indication for liver transplantation in the US^3–5. The majority of cases of NAFLD that lead to end-stage liver disease progress from simple steatosis to non-alcoholic steatohepatitis (NASH, inflammation due to steatosis) and then fibrosis^{1, 6}. While the progression from steatosis to fibrosis is well documented, there are currently few treatment options to disrupt this process other than weight loss. More broadly, there are no approved treatments available that target common inflammatory or fibrotic pathways arising from chronic liver injury, which could prevent the progression of NASH or other chronic liver diseases. Thus, the development of human cell-based models of liver inflammation and fibrosis to investigate molecular pathways and the activity of these pathways in individual cell types is critical for creating and testing new approaches to prevent liver failure.

Conceptually, the prerequisite for chronic liver injury is continued exposure to hepatotoxic agents, such as the saturated free fatty acid palmitic acid (PA)^{7, 8}, which accumulates in NAFLD^9–11. Oleic acid (OA) is another NAFLD-associated saturated free fatty acid¹² with protective^13–15 and toxic¹⁶ effects having been described. As a consequence of toxic agent accumulation, hepatocytes and other resident liver cell types may undergo apoptosis or activate pro-inflammatory signaling pathways leading to the secretion of cytokines such as tumor necrosis factor alpha (TNFα)^17–20. This subsequently triggers the activation of resident liver macrophages (Kupffer cells)²¹, the recruitment of systemic immune cells²², and their common secretion of cytokines, including transforming growth factor beta (TGF-β)²³. Finally, TGF-β activates hepatic stellate cells (HSCs) and promotes their differentiation towards myofibroblasts^{24, 25}, driving the acquisition of the fibrotic phenotype. Hallmarks of fibrosis include the up-regulation of transcripts COL1A1, COL3A1, ACTA2, TIMP1 and receptor-ligand pairs coding TNFS12/TNFRSF12A and PDGFB/PDGFR^{26, 27}, as well as the accumulation of myofibroblast-secreted collagen types I and III in the extracellular matrix (ECM) and an expansion of the ECM^{28, 29}.

Generation of human pluripotent stem cell (hPSC) - derived multi-lineage liver organoid systems (HLOs)^30–36 has allowed aspects of liver inflammation^{33, 37, 38} and fibrosis³⁹ to be modeled in individual systems in vitro. However, a multi-lineage liver organoid model effectively recapitulating evolving liver injury would further require the demonstration of specific transcriptional responses corresponding to different stages of liver injury. This would include the potential to create different degrees of liver injury displaying signatures indicative of the underlying injury mechanism and ideally recapitulate sequential phenotypes covering normal, inflammatory, and finally fibrotic stages. Single-cell RNA sequencing (scRNA-seq) has emerged as a key tool for benchmarking and deciphering such transcriptional dynamics, and efforts are underway to build single-cell atlases of multi-cellular organoids as promising 3D models to comprehensively evaluate their potential to recapitulate human diseases⁴⁰.

Here, we present an HLO model to profile multiple stages of liver injury, including PA-induced steatohepatitis and TGF-β1-induced fibrosis, providing a single-cell reference of the HLO injury landscape. We apply morphological, histological, and transcriptional analysis to evaluate the generation of liver disease phenotypes and optimal 3D culture conditions for this purpose. We compare the transcriptional landscape of differential injury conditions and demonstrate their ability to align along the sequential trajectory of liver fibrosis development. Finally, we apply clinically-validated liver fibrosis staging gene signatures that correctly classify the severity of liver injury in HLOs across disease progression of NAFLD.

Results

Human liver organoids recapitulate the transcriptional landscape of major cell types in the adult human liver

We first differentiated hPSCs into HLOs as previously described^{33, 34} (Fig. 1a). We confirmed loss of the canonical pluripotency genes SOX2, NANOG, and POU5F1⁴¹ by day 16 of differentiation and induction of the human liver genes ASGR1 and HNF4A (hepatocyte markers^{42, 43}), KRT19/CK19 and SOX9 (cholangiocyte markers^{44, 45}) as well as VIM and DES (HSC markers^{46, 47}) in day 21 HLOs (Fig. 1b). To evaluate HLO identity at the protein and histologic level, we performed hematoxylin and eosin (H&E) staining and immunohistochemistry (IHC) and found that HLOs at day 21 display an organized, sphero-luminal structure, and multiple cells express the nuclear hepatocyte marker CEBPα⁴⁸ (Fig. 1c, Supplementary Fig. 1). These results are in agreement with previous observations on the emergence of distinct liver cell types from at least two different lineages in HLOs after day 20³³.

Fig. 1 HLOs recapitulate major components of the adult human liver landscape.

a. Overview of the experimental design for this study. Human liver organoids (HLOs) are differentiated from human pluripotent stem cells (hPSCs), cultured in four different three-dimensional conditions, and tested for their respective potential to model liver injury (fibrosis induced by TGF-β1 and steatohepatitis by PA). HLOs acquiring an injured phenotype along with untreated HLOs are subjected to scRNA-seq and other readouts. b. Expression of pluripotency genes is reduced with HLO differentiation, while cell-type-specific adult liver markers are induced. Relative gene expression was measured by quantitative reverse transcriptase (qRT) PCR and normalized to housekeeping gene ACTB and displayed relative to hPSC controls (2^-ddCt). Comparison between hPSCs and HLOs from the same experiment were performed at day 16 or 21 as indicated. N = 6-8 individual experiments as indicated by circles. Mann-Whitney-U-statistics (two-tailed) hPSCs vs. HLOs: p-values^{day 16 HLOs} = 0.0024, Bonferroni-adjusted level of significance: 0.0125 (*); p-values^{day 21 HLOs} = 0.002165, Bonferroni- adjusted level of significance: 0.0071 (**). c. HLOs at day 21 show an organized, luminal structure and express adult liver proteins. Representative immunohistochemistry images of HLOs stained in hematoxylin & eosin (H&E), CEBPα and negative control with secondary antibody on day 21. Arrows indicate sites of nuclear protein staining. Scale bars indicate 50 µm, n = 3 individual experiments. d. Single-cell RNA sequencing (scRNA-seq) analysis of day 21 HLOs identifies the major cell types of the human liver across three different annotation strategies. UMAP plots show 17,898 cells from 10X scRNA-seq in day 21 HLOs (n = 3 technical replicates). Cell type annotations generated by three different methods are displayed: Literature-based annotation (top), database-annotation (center), and SingleCellNet⁵⁶ annotation comparing HLOs to human liver single-cell data⁵⁷ (bottom). e. Dotplot showing the expression of canonical cell-type-specific marker genes expected in the adult human liver across clusters. Cell types defined by single-cell data are displayed on the y-axis, and canonical genes (bottom) are sorted by cell type (top). The fraction of cells expressing a gene is indicated by the size of the circle, and the mean expression of a gene is indicated by color. Hierarchical clustering is represented by the dendrogram on the right. f. Enrichment of human WikiPathways 2019¹¹⁰ for three major cell types in HLOs. Terms have been shortened for readability, full lists are provided in Supplementary Fig. 3c.

To further validate the cellular composition of HLOs, we performed 10X scRNA-seq on day 21 and analyzed a total of 17,898 cells after quality control (QC) (Fig. 1d, Supplementary Fig. 2, Methods). Annotation of organoid cell populations can create difficulties since differentiating systems may contain transient cell states and retain premature features⁴⁹. To address these challenges, we initially evaluated three different annotation strategies⁵⁰. We first annotated cell clusters based on marker genes from the literature for fetal and adult liver cell types (Supplementary Table 1, Supplementary Fig. 2d-e), rendering five main clusters of hepatocyte-like, HSC-like, two cholangiocyte-like populations, and a small fraction of embryonic stem cell (ESC)-like cells expressing NANOG and POU5F1, but not SOX2 (Fig. 1d, top). In a second approach, we annotated cell clusters based on genes in cell-type marker databases^{51, 52}, composed from a wide range of scRNA-seq data, including one human liver study⁵³ (Supplementary Table 2, Supplementary Fig. 2d-e). This resulted in annotations closely related to the previous approach (Fig. 1d, center). Here, the cholangiocyte-like I cluster was annotated as hepatoblast (HB)-like, the common progenitor giving rise to cholangiocytes and hepatocytes^{54, 55}, since HB-markers KRT8 and KRT18 and cholangiocyte markers KRT19 and CLDN4 co-occurred across the top differentially expressed genes defining this cluster (Supplementary Table 3, Supplementary Fig. 2d). Increasing the clustering coarseness resolved a small cholangiocyte population with the HB annotation being largely preserved, suggesting that the majority of the previously cholangiocyte-like I labeled cells retain precursor features. In contrast, the literature based annotation preserved the cholangiocyte annotation throughout cluster coarseness levels, suggesting this method being potentially less sensitive discriminating cholangiocytes from their precursor states (Supplementary Fig. 2e). A fifth cluster (cholangiocyte-like II / smooth muscle cell-like) showed ambiguous annotations and thus could not be confidentially annotated by the two strategies. To complement this analysis with a third method, we applied SingleCellNet (SCN)⁵⁶, a random forest-based classifier, to compare HLOs to scRNA-seq data from fetal and adult liver tissue ⁵⁷. We chose this approach since SCN has previously been reported to yield high inter-dataset and inter-experimental accuracy for benchmarking data of human pancreatic cell populations⁵⁸, which developmentally are closely related to the liver⁵⁹. The SCN annotations for hepatocyte- and HSC-like cells are consistent with the previous methods, and the cholangiocyte-I/ HB-like cluster is mainly assigned to the normal adult liver tissue-specific KRT8+/KRT19+ population, labeled as bi-potent cells according to the terminology in the reference study⁵⁷ (Fig. 1d, bottom, Supplementary Fig. 3a-b). Overall, the attributions of cellular identities for the main clusters of hepatocyte- and HSC-like cells are overlapping across annotation strategies, and two out of three strategies indicated a cholangiocyte-precursor population with bipotent features constituting the third major cell type in day 21 HLOs. Based on this comparative analysis we decided to utilize the database annotation method for all subsequent exploratory analyses since it showed sensitivity to progenitor states at the cholangiocyte-HB interface and allows for the annotation of potentially emerging cell types beyond the repertoire of a single reference study.

To further ensure cellular identities we confirmed the expression of canonical marker genes for each cell type and found they overlap with the consensus annotations (Fig. 1e). EPCAM, CLDN4, KRT7 and KRT19 expression has been reported for adult cholangiocytes as well as hybrid (i.e. bi-potent) hepatic progenitors⁴⁴, and these genes are characteristic for the cholangiocyte I/HB-like cluster. These results support the consensus annotations from the previous strategies.

We next performed pathway enrichment analysis, revealing upregulated signatures for liver-characterizing metabolic processes and lipid metabolism in hepatocyte-like cells. HSC-like cells were enriched for pathways related to extracellular matrix and focal adhesion. Cholangiocyte-I/ HB-like cells showed enrichment for prostaglandin synthesis genes, in line with prostaglandins being involved in favoring liver fate decisions in early progenitors⁶⁰, as well as genes associated with epithelial-mesenchymal-transition (EMT) (Fig. 1f, Supplementary Fig. 3c). Together, these results indicate the presence of hepatocyte-, HSC- and emerging cholangiocyte-like populations in day 21 HLOs across multiple cell type annotation strategies.

Inducible liver injury phenotypes in HLOs upon TGF-β1 and fatty acid treatment

We next evaluated conditions to induce a fibrotic phenotype in the HLO model. TGF-β1 drives liver fibrogenesis in vivo^{61, 62}, and we treated HLOs with TGF-β1 (10 ng/ml and 25 ng/ml) in four different 3D culture systems. The final differentiation steps for HLOs are performed in Matrigel, and HLOs were either left in Matrigel or removed from Matrigel and cultured on i) ultra-low attachment plates (commercially available cell culture plates coated with a hydrogel layer to avoid attachment) as previously described³⁴, ii) 1% agarose coated plates, and iii) an orbital shaker (Fig. 2a).

Fig. 2 Inducible liver injury phenotypes in HLOs upon TGF-β1 and fatty acid treatment.

a. Overview of the four different 3D culture methods tested for HLO injury induction. b. HLOs undergo injury-specific morphologic changes when treated with TGF-β1 and oleic acid (OA) including contraction of the Matrigel dome with TGF-β1 treatment and darkening of color with each treatment when cultured in an orbital shaker (OS) after removal from Matrigel. Representative brightfield images of whole Matrigel domes (each 50 µl) with embedded HLOs (upper panel, scale bar 3 mm) and isolated HLOs cultured on an orbital shaker (lower panel, scale bar 0.1 mm) after five days of treatment with TGF-β1 (10, 25 ng/ml), OA (400 µM), and control conditions. c. Compaction assay for Matrigel cultured HLOs indicates a significant surface area reduction for HLOs treated with TGF-β1. Kruskal-Wallis test (two-tailed) followed by a post hoc Conover’s test with Bonferroni correction. Each color represents an individual experiment, and each circle represents a separate biological replicate. N = 3-4 individual experiments. d. H&E staining of HLOs cultured in an OS (upper row), Matrigel (middle row), and ultra-low attachment plate (ULA, lower row) after five days of treatment with TGF-β1 (10 or 25 ng/ml) display distinct structural changes. Representative H&E staining images showing HLO wall thickening in Matrigel- and ULA-TGF-β1 HLOs while hyaline-like intraorganoidal mass accumulation is present in OS-TGF-β1 HLOs. Scale bar 100 µm. N = 1-3 individual experiments. e. Representative thresholded images of Sirius red staining of HLOs cultured in four different 3D systems and stimulated with TGF-β1 (10 ng/ml) for five days showing collagen fiber deposition in Matrigel-cultured HLOs and OS-HLOs. Scale bars 100 µm. N ≥ 4 HLOs from one individual experiment. f. Sirius red staining quantification indicates significant collagen fiber deposition occurs only in Matrigel- and OS-cultured HLOs. Shown is the percentage of Sirius red positive tissue by total tissue normalized to the mean of the untreated control HLOs of the respective culture condition. Quantification is based on thresholded 200 x 200 µm squares around single HLOs. Kruskal-Wallis (two-tailed) test followed by a post hoc Conover’s test with Bonferroni correction. N ≥ 4 HLOs from one individual experiment. g. qPCR for COL1A1, TNFA and AFP transcripts indicates signficant induction of TNFA and COL1A1 only in HLOs incubated in an orbital shaker. Shown are relative mRNA levels normalized to ACTB for each individual culture condition (2^⁻ddCt). Kruskal-Wallis (two-tailed) test followed by a post hoc Conover’s test with Bonferroni correction comparing three treatment groups per culture method. P-values < 0.05 are indicated in the figure. Non-significant (ns) p-values: Matrigel: COL1A1 - control vs. TGF-β1^{10 & 25ng/ml}: p = 0.12, TNFA - Control vs. TGF-β1^{10 & 25ng/ml}: p = 0.94. 1% agarose: COL1A1 - control vs. TGF-β1^{10 & 25ng/ml}: p = 1.0, TNFA - Control vs. TGF-β1^{10 & 25ng/ml}: p = 0.7. ULA: COL1A1 - control vs. TGF-β1^{10 & 25ng/ml}: p = 0.97, TNFA - Control vs. TGF-β1^{10 & 25ng/ml}: p = 0.74. N ≥ 3 individual experiments. h. H&E staining of HLOs cultured on OS after five days of treatment with PA (500 µM) or OA (400 µM). Both culture methods display vacuole enrichment. Scale bar, 100 µm. i. TNFA expression is induced in OS-HLOs treated with PA (500 µM) for five days. Shown are qPCR results for relative mRNA levels of TNFA and COL1A1 normalized to ACTB. Mann-Whitney-U test (two-tailed) followed by Bonferroni correction (adjusted level of significance = 0.025). N = 4 individual experiments (represented by individual dot colors).

We observed morphological changes in HLOs cultured in Matrigel and on the orbital shaker after five days of treatment including HLO tissue consolidation, along with surface roughening in TGF-β1- treated HLOs (Fig. 2b, Supplementary Fig. 4a). For comparison, we also evaluated HLOs treated with OA, which did not demonstrate the same compaction in Matrigel as observed with TGF-β1 but was associated with a darker appearance on light microscopy when cultured in the orbital shaker, as previously described^{33, 34}. These results suggest HLOs are reacting in an injury-specific manner to the applied treatments.

We further quantified the contractile effect of TGF-β1 by culturing HLOs in Matrigel drops and measuring the Matrigel drop area after TGF-β1 application. This analysis demonstrates a significant reduction in droplet size with TGF-β1 treatment (Fig. 2c). To further investigate alterations in HLOs we stained for H&E and observed the intraluminal accumulation of hyaline, monomorph structures upon TGF-β1 treatment (Fig. 2d), providing further evidence for an injury response to TGF-β1 treatment.

We next performed Sirius red staining (Fig. 2e), a standard method for quantification of type I and III collagen deposition in liver fibrosis⁶³. We optimized an existing pipeline (Methods) for the computational quantification of Sirius red staining in the human liver for sections with low tissue amounts and options to select individual HLO areas of interest for the calculation of Sirius red percentage per tissue and per area separately (Methods, github.com/anjahess/sr_organoids). We utilized our pipeline to analyze HLOs cultured via the four previously described methods that were treated with TGF-β1 (10 ng/ml) for five days. We found a significant increase of collagen deposition only in Matrigel and orbital shaker-cultured HLOs (Fig. 2f), suggesting the latter culture methods lead to accumulation of extracellular type I and III collagen as observed in liver fibrosis.

We next evaluated the gene expression response across the four culture methods in HLOs when treated with TGF-β1. Canonical transcriptional changes in fibrosis include the induction of the alpha-1 subunit of type 1 collagen (COL1A1)²⁶ and TNFA⁶⁴, while TGF-β1 has been reported to reduce alpha-fetoprotein (AFP) levels in hepatic progenitor cells⁶⁵, foregut endoderm⁶⁶, and hepatocellular carcinoma (HCC) cell lines⁶⁷. We analyzed mRNA levels for the respective genes. We found that only HLOs cultured on an orbital shaker with TGF-β1 showed a significant increase in COL1A1 and TNFA expression (Fig. 2g), while AFP reduction was significant in HLOs cultured on Matrigel and 1% agarose, potentially reflecting different levels of TGF-β1-responsive progenitor-like cells across culture conditions. These results show that culturing in the orbital shaker provides the conditions under which HLOs demonstrate the most robust fibrotic and inflammatory response to TGF-β1.

We therefore focused on the orbital shaker method to screen for an inflammatory response to fatty acids. In H&E, both PA and OA treatments resulted in the increase of stain-free vacuoles as a correlate of lipid droplets (Fig. 2h, arrow heads). Treatment with PA was associated with an increase in TNFA mRNA levels, but COL1A1 mRNA and protein levels did not change significantly (Fig. 2i, Supplementary Fig. 4c). Treatment of HLOs with OA was not associated with an increase in TNFA or COL1A1 expression under the same conditions (Supplementary Fig. 4b) and did not lead to increased collagen deposition at the protein level (Supplementary Fig. 4c). These results show that under the observed treatment conditions, PA but not OA induced an inflammatory response and neither induced clear fibrotic injury.

Evaluating fibrotic and inflammatory responses at the single cell level

To dissect cell-type-specific transcriptional injury-response patterns, we treated HLOs cultured on an orbital shaker with either TGF-β1 or PA for five days (Fig. 3a) before performing scRNA-seq (n = 8 HLO samples, with replicates for each control and treatment condition). We analyzed 44,275 cells after QC (Methods), annotating cell types as described previously and identified hepatocyte-, cholangiocyte-, and HSC-like populations along with cells that more closely align with premature, cycling cells (MKI67+ progenitors), myoepithelial cells and HBs. Database and literature-based strategies overlapped in their annotation of cholangiocyte-like cells in the orbital shaker system, suggesting a less precursor-like state for this population compared to native HLOs at day 21 (Fig. 1, Supplementary Fig. 5d). Cell clusters were also identified near the HSC-like population with expression patterns associated with premature HSCs (PHSCs) and smooth muscle cells (SMC), the latter of which may represent a myofibroblast-like population (Fig. 3b, Supplementary Fig. 5-6). These results indicate a diversification of mesenchymal cell types in the HLO injury system suggestive of treatment-induced changes on the single-cell level.

Fig. 3 Evaluating fibrotic and inflammatory responses at the single cell level.

a. Brightfield images of day 25 HLOs after five days of treatment with TGF-β1 (10 ng/ml, upper panel) or PA (500 µM, lower panel) and respective controls. Scale bar indicates 50 µm. b. UMAP representations of 44,275 clustered cells from 8 HLO samples after five days of treatment with TGF-β1 (10 ng/ml) or PA (500 µM) and their respective controls. N = 2 replicates per condition, total n = 8, OS-cultured HLOs at day 25. c. TGF-β1 treatment induces fibrotic signatures while PA induces inflammatory signals. UMAP plots compare GO term¹⁰⁵ gene set score distributions for selected categories related to fibrosis and inflammation (labeled, left) for HLOs treated with TGF-β1 (10 ng/ml) and PA (500 µM) and their respective controls (labeled, top). Each condition contains two replicates, n-numbers as indicated. Increased expression is indicated by a shift from blue to red. Cell cluster annotations are provided in b. The scale bar in the lower right indicates the expression score for all plots except for immune response, which has a separate score. d. Cell cluster proportions as the fraction of total cells per sample are shown for each of the 8 individual replicates, color encodes the cell type as annotated in b., and cell types are listed in the order they are displayed. e. Dotplot showing canonical marker gene expressions for each cluster from d. Scaled mean expression is displayed, dot sizes correspond to the percentage of cells in each cluster expressing the marker gene. Canonical genes (bottom) are sorted by cell type (top). f. TGF-β1 treatment induces genes associated with fibrosis while PA induces genes associated with inflammation. Cell clusters were separated for expression analysis (left), and colors indicate the culture conditions for each cell. Differential expression (right) is displayed for selected genes associated with fibrosis and inflammation for TGF-β1 treatment, PA treatment, and controls. Differential expression is displayed as mean z-score calculated separately for each cell cluster. g. Top enriched Wiki pathways 2019¹¹⁰ for differential expressed genes in each treatment for each major HLO population as labeled in f. Terms have been shortened for readability, full lists are provided in Supplementary Fig. 8.

To understand which general injury gene expression signatures are activated with each treatment, we evaluated injury-response scores based on gene ontology terms⁶⁸. TGF-β1 treatment resulted in increased scores for fibroblast activation and smooth-muscle and contractile fiber activity along with alterations in cell cycle phase distributions, while these scores remained unchanged in the PA-treated HLOs. In contrast, PA-treated HLOs showed induction of immune response scores when compared to their controls and were enriched for chemokine-activity associated genes. Both treatments resulted in the reduction of liver-regeneration scores (Fig. 3c, Supplementary Table 4). These findings demonstrate the allocation of cells with induced inflammatory scores under PA treatment to the epithelial clusters (hepatocyte- and cholangiocyte-like cells) and the up-regulation of pro-fibrotic scores with TGF-β1 in mesenchymal clusters (HSC- and PHSC-like cells).

We then evaluated how cell type distributions change with each condition. TGF-β1-treated HLOs displayed an increase in HSC-related populations. HSC-like cells increased from less than 10% in controls (5.7% and 4.9%) to greater than 15% (22.7% and 16.5%) with TGF-β1 treatment, while PHSC-like cells increased from less than 10% (4.4 % and 7.5%) in the controls to around 70% (66.4% and 76.4%) with TGF-β1 treatment. In contrast, hepatocyte-like populations decreased by 10-fold from 19.6% to 1.9%, and a near-total loss of cycling MKI67⁺ progenitor was observed (Fig. 3d, Supplementary Table 5), in line with previously reported cell cycle arrest in TGF-β1-treated hepatocytes and other cell types^{69, 70}. While the analysis shows an expansion of HSC-like cells in response to TGF-β1, as observed in the development of liver fibrosis⁷¹, no significant alterations in cell population proportions were observed with PA treatment.

We also compared the distribution of cell types in the orbital shaker controls and the original HLO differentiation conditions (Fig. 1) and observed that culturing for five days in the orbital shaker was associated with a relative increase in hepatocyte- and cholangiocyte-like cells and a relative reduction in HSC-like cells (Supplementary Fig. 6c). These results suggest that the orbital shaker culturing method improves the ratio of epithelial to mesenchymal lineages resembling more closely the physiological proportions observed in the human liver^{26, 53, 57}.

PHSC-, SMC-, and myoepithelial-like cells cluster near HSCs (Fig. 3b), and these clusters share expression of many of the genes characteristic of HSCs (Fig. 3e). The cluster annotated as PHSCs is characterized by an HSC-like signature but does not express collagen genes at the level observed for HSC-like cells, instead marker genes of early progenitor cells CD59, EPCAM, ITGB1, PERP and KRT8 are among the differentially expressed genes (Supplementary Table 6). These cells also express CLDN4, KRT7 and KRT19 (Fig. 3e), genes that in the liver specify cholangiocytes. In the context of a general upregulation of immaturity markers in this cluster and the physiological presence of the respective transcripts in placental cells^72–74, this is in line with the acquisition of a premature phenotype of HSC-like cells in response to TGF-β1.

Differential gene expression analysis per cell cluster further displayed injury-specific signatures depending on the respective cell type. The fibrotic module of differentially expressed genes included COL1A1, TAGLN2 and TGFBI and was induced across TGF-β1 treated cell types, with HSC- and PHSC-like cells showing the strongest relative induction (Fig. 3f). The inflammatory module of differential genes included Interleukins IL1 and IL32 (the top up-regulated gene in NAFLD⁷⁵) and other canonical inflammatory genes including NFKBIA and CCL20 (Fig. 3f, Supplementary Tables 7-13), showing the strongest relative induction in hepatocyte- and cholangiocyte-like cells treated with PA, followed by HSC- and SMC-like populations (Supplementary Fig. 7). To contextualize differential gene expression more broadly, we performed cell-type resolved pathway analysis and found TGF-β1 hepatocyte-like cells enriching for TGF-β- and virus-associated pathways (being related to differentially expressed genes ITGB1, APP, CDKN2B, FN1, JUNB, SKIL (TGF-β) and ITGB1, GSN, NPC2, CLTB, HLA-C, CLTA, RHOB (virus), respectively) while the PHSC-like cells and cholangiocyte-like cells enriched for ECM- and cytoskeleton-associated pathways (Fig. 3g, Supplementary Fig. 8). Strikingly, PA-treated hepatocyte-, cholangiocyte-, and HSC-like cells displayed the nonalcoholic fatty liver disease (NAFLD) pathway among their top four ranked enriched pathways, associated with genes including NDUFA13, NDUFA11, UQCRQ, UQCR11 and cyclooxygenases COX7A2 and COX7C. These analyses show that TGF-β1 and PA each induce different patterns of gene expression across different cell types that mirror fibrotic injury and fatty liver-associated inflammation, respectively.

HLOs shift crosstalk towards HSC- and cholangiocyte-like cells inducing specific interactomes with TGF-β1 and PA treatment

To understand the cell-cell interactions in TGF-β1 and PA HLO injury models we utilized CellPhoneDB⁷⁶ and identified ligand-receptor pairs exclusive for each injury type. In TGF-β1-treated HLOs, receptor-ligand expression was observed involving TNFSF12 and PDGFB, which recapitulates findings from analyses on scar-associated macrophages in fibrotic human liver²⁶ (Fig. 4a, Supplementary Table 14). With PA treatment, receptor-ligand expression was observed that involves WNT7B, LGALS9 and CD55, reflecting the induction of previously described molecular mediators of cellular responses to injury and inflammation in the liver^77–79. We further identified ICAM1 and JAM2 containing ligand-receptor pairs shared between both liver injury entities. This general evaluation of the interactome reveals groups of receptor-ligand pairs changing in response to inflammatory and fibrotic injury in HLOs overlapping with previously reported molecular mediators in vivo.

Fig. 4 HLOs show increased crosstalk between HSC- and cholangiocyte-like cells inducing specific interactomes with TGF-β1 and PA treatment.

a. Venn diagram displays the intersecting and unique significant interacting groups from CellPhoneDB⁷⁶ analysis for day 25 OS-cultured HLOs treated with TGF-β1, PA, and their respective controls, n = 44,275 cells. b. Heatmaps showing the total number of interactions between cell clusters in day 25 OS-cultured HLOs treated with TGF-β1, PA, and their respective controls. Supplementary Fig. 6a provides cluster distributions across replicates. c. Dotplots showing all significant (p < 0.05) receptor-ligand pairs per cell-cell cluster pair from CellPhoneDB analysis for OS-cultured HLOs. Receptor-ligand pairs are shown on the x-axes, and cell clusters are shown on the y-axes. Mean expression is indicated by the size of the circle, and significant p-values are indicated by blue color. d. Differential receptor-ligand pair interactions with HSC-like cells in HLOs after exposure to TGF-β1. Dotplots showing differential COL1A1 receptor-ligand interactions per cell-cell cluster for OS-HLOs treated with TGF-β1 and corresponding controls. Mean expression (mean of all partner‘s individual average expression, Methods) is indicated by the size of the circle, and significant p-values are indicated by blue color. “Self“ indicates co-expression of receptor and ligand on the same cell types, which in this case is co-expression on HSC-like cells. e. Example of differential receptor-ligand pair interactions with cholangiocyte-like cells after exposure to TGF-β1. Dotplots showing differential EPHA4 receptor-ligand interactions per cell-cell cluster for OS-HLOs treated with TGF-β1 and corresponding controls. Mean expression is indicated by the size of the circle, and significant p-values are indicated by blue color. f. Differential receptor-ligand pair interactions in HLOs after exposure to PA. Dotplots showing significant hepatocyte-like-driven DLL1-NOTCH1-3 (left) and hepatocyte-like-driven FFAR2-FAM3C (right) receptor-ligand pair interactions in OS-HLOs treated with PA compared to their respective controls. Mean expression is indicated by the size of the circle, and significant p-values are indicated by blue color.

We assessed interactions between each cell type and found HSC-like cells to constitute the most interactive cluster based on receptor/ligand co-expression (Fig. 4b). In controls and following treatment with TGF-β1 and PA, HSC- and cholangiocyte-like cells showed the strongest interaction. The number of interactions also increased with TGF-β1 but not PA treatment. We also evaluated condition-specific patterns of receptor-ligand pair expression (Fig. 4c, Supplementary Fig. 9-10). This analysis allowed us to explore “hot-spots” of injury-specific interactions observed between HSC-like and cholangiocyte-, hepatocyte-, and other HSC-like cells. These interactions revealed new collagen-integrin pairs with TGF-β1 treatment. This includes multiple collagens that can interact with integrin α1β1 on cholangiocyte-like cells and a shift from collagen-integrin α1β1 receptor-ligand interactions between HSC- and hepatocyte-related cells to collagen-integrin α2β1 interactions (Fig. 4d). These findings suggest that TGF-β1 signaling induces expression of genes that can promote cross talk between HSCs and cholangiocytes and between HSCs and hepatocytes through integrin signaling.

Expression of members of the Ephrin receptor A (EPHA) family and its ligands, ephrines (EFN) characterizes a subpopulation of cholangiocytes with migration potential and has been linked to filopodia formation in cholangiocytes^80–81. We found that expression of receptor/ligand pairs for ephrin signaling are also regulated by TGF-β1. EPHA4 expression is paired with multiple ephrin ligands in cholangiocyte-like cells following TGF-β1 treatment (Fig. 4e), potentially relating to the oberserved surface alterations in these HLOs (Supplementary Fig. 4a). In addition, TGF-β1 treatment leads to expression of EPHB2 in cholangiocyte-like cells and expression of the gene encoding the ligand EFNA5 in HSC-like cells, indicating the presence of a receptor pathway previously reported to promote hepatic fibrogenesis in vivo^{82, 83}. These results suggest that TGF-β1 induces expression of genes in the ephrin pathway that can promote autocrine signaling in cholangiocytes and communication between HSCs and cholangiocytes.

With PA treatment, hepatocyte-like cells express the Notch-ligand DLL1 in conjunction with NOTCH1-3 in HSC- and cholangiocyte-like cells (Fig. 4f). We found PA treatment induced expression of free-fatty acid receptor FFAR2 on hepatocyte-like cells in relation to the ligand FAM3C in hepatocyte-like cells themselves, HSC-, and cholangiocyte-like cells. Together, these findings suggest that PA induces crosstalk between multiple cell types involving pathways linked to liver disease^84–89 and reveal potential molecular targets in fatty acid-induced liver injury.

Trajectory inference reconstructs the emergence of major HLO lineages and injury-specific terminal states

The development of liver fibrosis from steatohepatitis can be understood as a sequential process over time with fibrosis following steatohepatitis⁶. Therefore, we investigated the projection of our steatohepatitis model (PA) and our fibrosis model (TGF-β1) on a force-directed layout for generating continuous trajectories between discrete time points⁹⁰. We first visualized our previously annotated cell types across the trajectories (Fig. 5a, I.) before analyzing differential trajectories in control-, PA- and TGF-β1-treated HLOs (Fig. 5a, II.-IV.). PA-treated HLOs displayed a distinct trajectory towards hepatocyte-like cells (Fig. 5a, III., dashed circle). TGF-β1 treated MKI67+ progenitor-like cells related to the HSC- and SMC-like clusters mainly via the PHSC population (Fig. 5a, IV., dashed circle), while connections to HSC-like cells through the hepatocyte- and cholangiocyte-like subfractions shared among healthy and PA-treated HLOs were not observed, adding evidence that TGF-β1 not only changes cell type proportions favoring HSC-like progenitors, but also the developmental axis through which they emerge.

Fig. 5 Trajectory inference reconstructs the emergence of major HLO lineages and injury-specific terminal states.

a. Palantir results based on harmony force directed layout display the lineage relationships between control, inflammatory (PA-treated), and fibrotic (TGF-β1-treated) HLO cell populations, revealing increased pseudotime towards hepatocyte-, smooth muscle cell-, and HSC-like populations, while PHSC-like cells and PA-treated hepatocyte-like cells show lower pseudotime than their healthy counterparts and differentiation events specific to the injured phenotype. (I.) Harmony augmented affinity matrix-based force-directed layout colored by cell types corresponding to Fig. 3b. (II.) Control-HLO cells mapped to harmony force directed layout. (III.) PA-HLO cells mapped to harmony force directed layout, dashed circle (upper left) indicates clusters enriched relative to the control condition. (IV.) TGF-β1-treated HLO cells mapped to harmony force directed layout, dashed circle (lower right) indicates clusters enriched relative to the control condition. (V.-VII.) Palantir results include pseudotime (V., arrow heads indicate pseudotime maxima), terminal states (VI., white arrow head: MKI67-high early cell, colored arrow heads: terminal states), and regions of high differentiation potential (VII., arrow heads 1-4). n = 8 samples and 44,275 cells. b. MAGIC-imputed gene expression of canonical inflammation- and fibrosis-related genes overlaps with injury-specific terminal states. Fetal hepatocyte-related transcripts enrich along the trajectory towards the PA-treated hepatocyte-like cells. Projection of the imputed gene expression of transcripts related to inflammation (upper left), fibrosis (upper right), and developmental stages of fetal hepatocytes (lower left). c. Gene expression trends reveal the increase of inflammatory and pro-fibrotic transcripts along the pseudotime ordering towards injury-specific (PA and TGF-β1) terminal states. Heatmaps show imputed gene expression over Palantir pseudotime for canonical inflammation (II.) and fibrosis (III.) related genes for hepatocyte-like (left column) and HSC-like (right column) terminal states. Opposed are terminal states specific to injury conditions (PA-treated hepatocyte-like and TGF-β1-treated PHSC-like cells, indicated by red arrows). Imputed expression levels are indicated by color, and x-axes display pseudo-time. MKI67 expression, an indication of proliferation that marks early cells, is also quantified (I.) d. Inflammatory genes (left panel) plateau or decrease in TGF-β1-specific PHSC-like cells (brown line), while fibrosis-related (right panel) genes peak towards the end of pseudotime, suggesting a transition trough inflammatory states towards the fibrotic phenotype in these cells. PA-treated hepatocytes (red line) show a constant increase of inflammatory gene expression with only small increases in fibrotic signatures towards the end of pseudotime, suggesting an inflammatory terminal state with a beginning fibrotic phenotype for these cells. Gene expression trends of representative genes from 5c. across lineages and pseudotime. Pseudo-time is indicated on the x-axes and normalized expression is shown on the y-axes. Cell clusters are indicated by color (top) and can be identified in 5a (I). Pseudo-time is determined in relation to terminal states (5a. V.), and not all cell clusters reach a value of 1.

To further investigate the directionality of these trends, we overlaid Palantir⁹¹ selecting an MKI67-high progenitor-like cell as the early cell. As expected from the immature gene expression signature of the PHSCs, cells corresponding to this cluster aligned earlier along the pseudotime representation than their HSC counterparts. SMC- and hepatocyte-like cells corresponded to regions at the end of the pseudotime path (Fig. 5a, V., arrow heads). Unsupervised analysis inferred six terminal states (Fig. 5a, VI.) among which one hepatocyte-like cluster was dominated by PA-treated cells and the PHSC cluster represented mostly TGF-β1-treated cells. We next identified cells with high differentiation potential (Fig. 5a, VII.), showing enhanced activity in four defined regions. Cells with the highest differentiation potential were detected at the interface of premature cells identified as HB-like cells close to the hepatocyte-like terminal states and mesenchymal populations (Fig. 5a, VII., arrow head 1), potentially pointing at a hepatomesenchymal transition state previously described in normal development⁹². A second cluster of intermediate-high differentiation potential aligns with the early bidirectional fate decision towards either the hepatocyte-like or the mesenchymal, PHSC-like state (Fig. 5a, VII., arrow head 2). A third region of intermediate-high differential potential is present at the PHSC-like – HSC-like interface (Fig. 5a, VII., arrow head 3), and a fourth at the transition from HSCs towards SMC-like cells, whose transcriptional signature is closely related to myofibroblasts (Fig. 5a, VII., arrow head 4), consistent with a canonical feature of fibrosis pathology^{24, 71}.

To examine the differentiating regions as potential representations of healthy-to-injury transitions, we next applied MAGIC⁹³ to generate imputed data and visualize the expression of relevant genes along the differentiation trajectory with a focus on the acquisition of “pro-fibrotic” and “pro-inflammatory” gene expression signatures (Fig. 5b). Inflammatory transcripts IL32, NFKBIA, CCL20 and CXCL1 showed a bimodal expression signature increasing towards inflamed hepatocyte-like cells with PA treatment and PHSC-like cells with TGF-β1 treatment. The PA-treated hepatocyte terminal state showed enrichment for fetal hepatocyte transcripts AFP, APOM, APOA1 and APOB. We observed a targeted diffusion distribution of pro-fibrotic genes towards HSC-like cells, corresponding to the TGF-β1-treatment enriched trajectory via the PHSC-like cluster (e.g., COL1A1, TGFBI, IGFBP3). Specifically, myofibroblast-characterizing transcripts ACTA2 and TIMP1 increased towards the SMC- like cluster, again suggesting these cells could represent an HSC-derived myofibroblasts-like population. These transcripts are also characteristic for activated HSCs, a hallmark of fibrosis in vivo⁷¹, again suggesting the trajectory with enhanced differentiation potential connecting HSC- and myofibroblast-like populations (Fig. 5a, VII., arrow head 4) represents the activation of these HSC-like cells.

To better understand gene expression signatures accounting for deviating terminal states observed in injured HLOs (Fig 5a, V.-VI.), we next inferred gene expression trends and opposed hepatocyte-like and HSC-like cells with their injured counterparts (Fig. 5c, left: hepatocyte gene trends, right: HSC gene trends). As expected, MKI67 decreased towards all terminal states (Fig. 5c, I.). We observed a clear shift towards inflammatory (Fig. 5c, II.) and fibrotic genes (Fig. 5c, III.) along pseudotime in PA- specific hepatocyte-like cells and TGF-β1-specific PHSC-like cells (red arrows) when comparing them to their general population counterparts, which included the control cells (black arrows). A relative increase in inflammatory and fibrotic gene expression is evident in terminal states specific for PA and TGF-β1 treatment. The general population of hepatocyte-like cells used as a control does contain a mix of control and PA-treated cells, and the general population of HSC-like cell controls contain a mix of control and TGF-β1 treated cells (Fig. 5a), which could also reduce the differences observed between control and injury conditions. Projecting lineage-specific gene expression trends over pseudotime (Fig. 5d) allowed for the identification of inflammatory genes (Fig. 5d, left column) plateauing or slightly decreasing in TGF-β1-specific PHSC-like cells (brown line), while fibrosis-related genes (Fig. 5d, right column) peaked towards the end of pseudotime, suggesting a transition through inflammatory states towards the fibrotic phenotype in these cells. PA-treated hepatocyte-like cells (red line) showed a constant increase of inflammatory gene expression with only small increases in fibrotic signatures towards the end of pseudotime, suggesting an inflammatory terminal state with an emerging fibrotic phenotype for this population. Pseudotime is indicated on the x-axis, with a value of 1 representing the greatest pseudotime (SMCs). Other cell populations terminate before 1 based on pseudotime calculations for each terminal cell state. Together, these results support the presence of intermediate inflammatory stages on the path towards fibrosis.

Candidate gene signatures predicting disease progression in NAFLD progressively increase with treatment of PA and TGF-β1

We next investigated how changes in gene expression following treatment of HLOs with PA and TGF-β1 relate to the development of fibrosis in patients with NAFLD. We applied the 26 and 98 gene signatures established to predict fibrosis in NAFLD⁹⁴ to bulk gene expression in HLOs. This analysis revealed a significant increase in the score for both gene signatures with PA and TGF-β1 treatment, with the mean score increasing to the greatest extent with TGF-β1 treatment (Fig. 6a). We then plotted expression signatures for individual genes that demonstrated the greatest dynamic range in expression across clusters (Fig. 6b). Many of these genes show a trend of induction from control, to PA, to TGF-β1, including CXCL6, GSN, IL32 and TAGLN2, while other genes, such as S100A4 show the highest expression with PA treatment. This indicates the gradual induction of NAFLD disease progression scores in the HLO system.

Fig. 6 Gene signatures predicting fibrosis stages in NAFLD gradually increase from control HLOs to those treated with PA to those treated with TGF-β1.

a. Application of 26- and 98-gene signatures⁹⁴ to predict fibrosis stages across NAFLD to control and injured OS-HLOs. P-values derived from Kruskal-Wallis test (two-tailed) with post hoc Conover’s test followed by Bonferroni correction. Log scale; ns, non-significant. N = 8 samples, n = 44,275 cells. b. Dotplot showing scaled mean expression of cross-condition intersecting genes from the 98-signature score across treatment conditions. Genes displayed are expressed in at least 3 cells in each HLO sample and are present in each individual replicate. Dot sizes correspond to the percentage of cells expressing the respective gene in each condition, and the level of expression is indicated by color. Full differential expression data can be found in Supplementary Tables 7-13. c. Deconvolution of the scores from a. to individual cell types as annotated in Fig. 3b. P-values derived from Kruskal-Wallis test (two-tailed) followed by a post hoc Conover’s test with Bonferroni correction. N numbers are indicated below cell cluster names. Log scale, ns, non-significant.

We then investigated the 26 and 98 gene signatures at the single cell level in HLOs to understand the cell types responsible for each signature. The 26 gene signature was increased in hepatocyte-, smooth muscle-, myoepithelial-, cholangiocyte-, HSC-, and MKI67⁺ progenitor-like cells, suggesting these clusters all contribute to the fibrosis signature induced by PA or TGF-β1 treatment of HLOs. The 98-gene signature is significantly up-regulated in all TGF-β1 treated cell types, while PA treatment was associated with an increase in the 98 gene signature only in MKI67⁺ progenitor- and myoepithelial-like cell types (Fig. 6c). This again is in line with a gradual acquisition of fibrosis severity, where PA treatment does not induce the full spectrum of fibrosis-related genes observed with TGF-β1.

To further dissect the contribution of individual cell types to the score enrichment, we next plotted the relative expression of individual genes per cell cluster across treatment conditions (Supplementary Fig. 11). Hepatocyte-like cells in HLOs treated with TGF-β1 showed higher expression of BTG2, EHD4 and GSN, while PA induced IL32 and S100A4 in this population. HSC-like cells showed induction of COL4A1, COL4A2, COL5A1, PKM, VIM, and TPM4 with TGF-β1 treatment, while PA treatment resulted in reduced expression of these signature genes. Cholangiocyte-like cells showed an increased expression of signature genes CXCL6 and IL32 in both injury scenarios while MEAF6, GSN, S100A11, PGP, TMP4, VIM, and TAGLN2 showed higher expression only in the TGF-β1 condition. A gradual increase in expression levels from control to PA to TGF-β1 in PHSC-like cells was observed for CDC42SE1, CXCL6, GSN, PGP, PKM, S100A11, TAGLN2, TMEM51, TNFAIP8, and VOPP1. These results identify the individual cell types that contribute to the gene signature linked to disease progression in NAFLD.

Discussion

Here, we report that multicellular human liver organoids (HLOs) can model progression of steatohepatitis and liver fibrosis when maintained under specific 3D culture conditions. While studies have started to decipher the contribution of individual cell types to the response to liver fibrosis humans^{26, 95}, our understanding of how cell types interact during evolving liver injury, including their temporal order of activation, is limited. Systematic phenotypical, histological, and scRNA-seq analysis in the HLO system in this study indicates a high degree of specificity and temporal orchestration in the responses of individual cell populations to different injury types.

PA-treated HLOs as models of liver inflammation display altered phenotypes and induced TNFA expression. The general cell type distribution does not change with PA treatment, but gene scoring reveals enhanced signals for transcripts related to immune response and chemokine activity, while fibrosis-related scores are not induced. This is mainly driven by the hepatocyte- and cholangiocyte-like populations, followed by HSC-like cells, up-regulating not only the highly specific NAFLD cytokine IL32⁷⁵ but also canonical inflammatory pathways related to NFκB. All subpopulations except for the myoeptihelial-like cells additionally are enriched for NAFLD related transcripts including cyclooxygenase-coding genes COX6 and COX7 and NADH-ubiquinone oxidoreductase subunits NDUFA3, NDUFA11 and NDUFA13. Unique cell-cell cross talk defining interactions including FFAR2/FAM3C and DLL1/NOTCH1-3 are induced. While most inflamed HLO cells align on developmental trajectories with their respective controls, PA-treated hepatocyte-like cells deviate, giving rise to a distinct terminal state. Grading PA-treated HLOs with clinically validated NAFLD- fibrosis scores shows a moderate, but significant increase driven by gene expression across cellular subpopulations, including hepatocyte- and SMC-like (most likely myofibroblasts-reflecting) populations. Taken together, transcriptomic data generated from PA-treated HLOs capture an inflammatory transition at single-cell resolution that is driven primarily by hepatocyte-like cells.

TGF-β stimulation in the same HLO system triggers more pronounced changes in HLO morphology with increased collagen deposition and an expansion of the HSC-like population consistent with observations in chronic human liver disease, and scRNA-seq profiling revealed a premature transcriptional signature of the most significantly expanding HSC-subpopulation (PHSC), defined by the upregulation of CD59, EPCAM, ITGB1 and KRT8. TNFSF12 and PDGFB-driven receptor ligand interactions are induced as previously shown for human fibrotic livers²⁶, and cholangiocyte-like and HSC-like cells are the most interactive partners in the fibrotic HLO. PHSC-like cells emerge as a distinct terminal state on the liver injury trajectory. The progression of fibrosis is classified as more severe across all cell populations when applying the clinically validated gene signature score on TGF-β1 treated HLOs.

While many genes shaping the trajectories towards the injured terminal states overlap between PA- and TGF-β1 mediated injury, sequentially induced expression levels and the bifurcated anatomy of the pseudotime trajectory representation stress the different cell types (hepatocyte-versus PHSC-like cells) and transdifferentiation events that preserve the injury-response specificity in the HLO system and correspond well with in vivo observations finding maintained inflammatory hepatocyte activation preceding HSC-to-myofibroblast differentiation events in liver fibrosis development^{1, 2}. Taken together, these analyses suggest that PA and TGF-β1 treatment in HLOs recapitulate the sequentially evolving cell states of inflammatory and fibrotic liver injury established for NAFLD pathology. In addition, HLOs not only develop fibrosis-like pathologies, but are also capable of mirroring gradual transcriptome-wide injury responses corresponding to fibrosis severity classes in NAFLD patients. However, the current HLO system lacks immune cells, and it is possible that the introduction of cells such as macrophages would also promote the transition from inflammation to fibrosis.

To conclude, we present a comprehensive scRNA-seq evaluation of a stem-cell derived multi-cellular human liver organoid system capable of mimicking differential injury-responses in a clinical disease progression-like manner. We build on previous developments in human liver organoid models yielding multi-cellular systems demonstrating the feasibility of deriving injury scenarios of genetically-programmed, ADPKD-like fibrosis³⁹ and oleate-mediated steatohepatitis^{33, 37}. We explore multiple culture conditions and identify their contribution to generating physiological cell-type distributions towards the inducibility of multiple, stressor-specific liver injury responses. We utilize scRNA-seq to validate our multi-injury system and contribute to current efforts of establishing frameworks for the systematic benchmarking of HLOs as state-of-the art in vitro model systems overcoming previous limitations in 2D culture and non-human animal models⁴⁰.

Our multi-injury liver organoid system demonstrates the temporal emergence of transcriptionally altered cell populations in the course of liver injury and defines specific molecular pathways involved in the early cell-cell interactions preceding broad pro-fibrotic alterations of the HLO cellular landscape. This flexible, perturbable system will facilitate the study of liver injury development, enable stressor-specific testing of therapeutic candidates, and provide a baseline reference atlas for the benchmarking of novel liver injury organoid models in the future.

Methods

hPSC and HLO culture

H1 hPSC cells (WA01) were obtained from WiCell. The ESCRO approval was received from Massachusetts General Hospital. hPSCs were maintained in mTeSR medium (StemCell Technologies) on Matrigel (Corning, 354230) as previously described⁹⁶ and split with Accutase (Thermo Fisher Scientific) upon 80-90% confluence. For passage and thawing, Rho-associated kinase (ROCK)-Inhibitor (StemCell Technologies, Y-27632) was added to mTeSR at 10 µM. For the differentiation of human liver organoids (HLOs), cells were split at a ratio of 1:10 and cultured until ∼85% confluent before HLOs were differentiated as previously described^{33, 34}. On day 20, HLOs were either kept in Matrigel domes or isolated. To perform isolation, HLOs were incubated in DPBS(-/-) on ice for 15 minutes followed by repeated manual dissociation with a P1000 at 4°C followed by centrifugation at 190 x g for three minutes, removal of supernatant, and resuspension in 10 µL final medium per receiving well of a 24-well-plate. Isolated HLOs in media were either plated to i) 1% agarose coated plates ii) plates on an orbital shaker at 80 revolutios per minute, or iii) hydrogel plastics ultra low attachment surface 24-well plates (Corning, 3473).

Liver injury induction

Injury media solutions based on Hepatocyte Culture Medium (HCM, Lonza, CC-3198) were prepared to obtain solutions of TGF-β1 (10 ng/mL and 25 ng/mL, R&D Systems 240-B-002), palmitic acid (PA, 500 µM, Sigma Aldrich P0500), and oleic acid (OA, 400 and 800 µM, Sigma Aldrich O1383). HLO media was changed to injury media on day 20-21. Isolated HLOs were resuspended in the final injury media at the final step of the isolation process (see above) and kept in solution for five days. Controls for TGF-β1 were cultured in HCM. PA was dissolved into HCM with 10% BSA and 1% ethanol before dilution to final concentration in HCM. OA was dissolved into DPBS(-/-) with 12.5 mM NaOH and 1.67% BSA at 8 mM before dilution to final concentration in HCM. PA and OA controls were prepared accordingly, omitting the initial step of dissolving the active agent in the carrier solutions.

Contraction assay

Day 20-21 HLOs in 50 µL Matrigel domes cultured in HCM were exposed to TGF-β1 at 10 ng/mL and 25 ng/mL. Medium was changed every three days. On day 0 and 5 of incubation, images of plates with scale bars were taken. Images were loaded into ImageJ (version 1.53a), with the scale set based on the known distance. Subsequently, the circular HLO/matrix drop areas were manually selected and measured. All distances in mm² were normalized to the mean of the control areas in mm², and two-tailed Kruskal-Wallis- statistics were performed on all groups with a Conover post hoc test.

qPCR

Total RNA was isolated from HLOs to perform qPCR. Briefly, 1 mL medium was removed from 6 well plates. Matrigel domes containing HLOs were carefully detached from the wells with a cell scraper and transferred to a 1.5 mL tube including medium. If isolated HLOs served as starting material, the detachment step was not needed. Matrigel-embedded HLOs were left at 4°C for 15 minutes to dissolve the Matrigel. Then, HLOs were centrifuged for three min at ∼190 x g at 4°C (Matrigel-embedded HLOs) or at RT (non-matrigel HLOs or 2D cell layers). Supernatant was discarded, and the pellet was resuspended in 300 µL Trizol (Invitrogen, 15596026), rigorously vortexed or mechanically processed until fully dissociated, and finally incubated for 10 minutes at room temperature and then stored at −80°C. RNA was prepared via phenol-chloroform extraction. For cDNA synthesis, 200 or 500 ng RNA were reverse-transcribed using the iScript gDNA Clear cDNA Synthesis Kit (Biorad, 1725034), and no-RT-controls and mastermix controls were prepared. qPCR reactions were prepared from cDNA at 1:5 dilution with SYBR Green iTaq Universal SYBR Green Supermix (Biorad, 1725120) and qPCR primers at 10 µM in a total volume of 10 µL in a ≥ 40-cycle and melt curve reaction cycle r (Biorad, Base #CT009383, Optical Head #786BR2648). All biological samples were measured in technical triplicates.

qPCR data analysis

All gene Ct means were calculated from three technical replicates. Only samples with housekeeping Ct values below 35 were considered for analysis. Mean Ct values for housekeeping genes β-Actin (ACTB) or glyceraldehyde-3-phosphate dehydrogenase (GAPDH) were subtracted from target gene means to generate the ΔCt value. The control ΔCt value average was calculated from three biological replicates and subtracted from all experimental ΔCts to render ΔΔCt values. The log2 fold-change was calculated as 2^(⁻ΔΔCt).

Primer design and validation

Primers were designed using PrimerBlast. Briefly, transcripts of interest were selected by NCBI Reference Sequence ID and the ‘pick primers’ hyperlink function was used to import the cDNA sequence into the PrimerBlast mask. The following settings varied from the defaults: Exon junction span → primer must span an exon-exon junction, PCR-product size → 50-200, allowing primers to bind to variant transcripts of the same gene. Blast was performed against both, RefSeq mRNA (Homo sapiens) and RefSeq representative genomes (Homo sapiens) in order to obtain primers specific to transcript sequence and mRNA rather than gDNA. Primers with i) no additional match in RefSeq mRNA or at least > 4 mismatches and ii) no additional genomic hits or hits with at least > 800 bp product size, were selected for further validation. Next, primers were tested on human liver tissue RNA at a final concentration of 10 µM as described above, and 2% agarose gels of the primer product were prepared with a 100 bp ladder. Primers were selected for further experiments if they showed a single band in the expected size range. The band was cut, cDNA was extracted from agarose and sent for Sanger sequencing, and the product identity was assured by re-blasting. A primer was selected for experiments when the sequence matched the selected sequence of the initially selected transcripts (NCBI BLAST). Primer sequences are provided in Supplementary Table 15.

Histology and immunohistochemistry

Medium was removed from day 16, day 21 or day 25-26 HLOs treated for five days with pro-fibrotic substances, washed twice with warm DPBS(-/-) and fixed overnight at 4°C with 4% PFA in DPBS(-/-). PFA was removed and the HLOs were washed 2x with DPBS(-/-) before being transferred to 70% ethanol for storage and paraffin embedding. H&E and Sirius red staining was performed on cut and deparaffinized HLOs in line with standard protocols. Anti-human-CEBPα (Sigma Aldrich, HPA052734) and the secondary antibody (HPR anti-mouse) were used at a dilution of 1:200.

Organoid Sirius red quantification pipeline

The pipeline for HLO-specific Sirius red staining analysis is written in python (≥ version 3), ImageJ (version 1.53a) macro and bash and will be made publicly available upon release at https://github.com/anjahess/sirius_red. First, full slides were cleared from black artifacts. Then, for whole slide quantification, all image areas containing tissue were selected and saved for quantification. The Sirius red quantification module was adopted from the NIH-Image J macro “Quantifying Stained Liver Tissue“ (available at https://imagej.nih.gov/ij/docs/examples/stained-sections/index.html, requested 2020-11-14). Briefly, RGB stacks were generated from images, and the tissue containing area was defined and measured by the setAutoThreshold(ℌDefault stack”) function. The Sirius red staining threshold reached from minimum to maximum as retrieved from the setAutoThreshold and getThreshold(min, max) functions measured in the green channel of the RGB stack, and divided by an experiment-specifc brightness factor (0.95-1.3). For single HLO analysis, 200 x 200 µm selection squares were placed around HLOs in the interactive, user-supervised mode and resulting images were processed as described for whole slide scans. Finally, csv-formatted result tables were called from python (version 3.7) and the Sirius red stained area was calculated either per total area or per tissue area (measures as the auto-thresholded area in the blue channel). If normalization was performed, values were calculated as percentage of mean of all control samples of the respective experiment.

Single-cell RNA sequencing

For day 21 and 34 HLO single-cell RNA sequencing, HLOs underwent a passage step at day 16 and remained in HCM medium with 10% Matrigel on ultra-low attachment plates as previously described³⁴. At day 21 and 34, one well of a six well plate of free-floating HLOs was collected, washed 1X with warm DPBS(-/-) and briefly dissociated by incubation with 300 µL 0.05-0.25% Trypsin-EDTA for 10 minutes at 37°C (Corning, Ref. 25-052-CI RT). Day 34 and day 25-26 orbital shaker-cultured HLOs underwent an additional DBPS (-/-) wash and 10 minutes of trypsinization to allow complete HLO digestion. For each replicate, full single cell dissociation was confirmed manually by light microscopy. After a final DPBS(-/-) wash, HLOs were resuspended in DPBS(-/-), counted after Trypan blue staining with the TC20™ Automated Cell Counter according to a previously determined optimal gating of 7-20 µm and transferred on ice. Library preparation was performed on biological replicates with the highest viability counts. Libraries were prepared according to the 10X ChromiumSingle Cell 3ʹReagent Kits v3 instructions. The library was sequenced by paired-end sequencing on an Illumina® NextSeq 2000 P3 flow cell or the NovaSeq 6000 system according to the manufacturer’s recommendations.

Single-cell RNA sequencing analysis

Raw bcl files were converted to fastq by the command bcl2fastq --use-bases-mask Y26,I8,Y98. Fastq quality was assessed with multiqc⁹⁷. Fastq files were aligned to the GRCh38 genome with cellranger count, cellranger (=> version 3.0.2). Doublet scores were calculated with Scrublet⁹⁸ assuming 0.06 percent doublets at a calling threshold of 0.2-0.22 on the cellranger filtered count matrices (manually adjusted based on bimodal histogram distribution). Downstream analysis was performed with scanpy⁹⁹ (version 1.7.2, for functions abbreviated with “sc” from here on) in python (version 3.7). The following cellular quality control criteria were applied to each replicate after manual inspection of violin and scatter plots of the unfiltered inputs: (i) 100 features per cell, (ii) a maximum mitochondrial gene fraction of 50% (iii) genes-by-count ratio maximum of 8500. Transcripts were accepted if present in at least three cells. Subsequently, technical replicates and, when appropriate, replicates of the conditions to be compared were merged using the scanpy concatenate function with default settings. Concatenated datasets were total-count normalized (target sum 10,000), logarithmized (X=log(X+1), natural logarithm), and a total of 4000 highly variable genes were selected for downstream analysis. The data was scaled to a maximum value of 10.

Cell cycle scoring

Cell cycle scores (S-, G2M-Score) and cell cycle phase (S, G2M, G1) were assigned based on a previously published cell cycle defining gene list¹⁰⁰ with the sc.score_genes_cell_cycle function.

Dimensionality reduction, embedding, clustering, and marker gene identification

Normalized, log-transformed and scaled data were objected to calculation of principal component analysis (PCA) coordinates, loadings, and variance via sc.tl.pca. The neighborhood graph was calculated using the first 50 principal components and embedded utilizing the Uniform Manifold Approximation and Projection (UMAP)¹⁰¹ algorithm via sc.api.tl.umap. Clusters were identified with the Leiden algorithm¹⁰² at resolution 0.1 or 0.5. Coarsenesses lower than the default parameters were chosen to reproduce the underlying biology of approximately five cell types experimentally validated for HLOs³³. Finally, cluster-characterizing genes were defined using scanpy’s implementation of the Wilcoxon rank-sum test followed by Benjamini–Hochberg correction (sc.tl.rank_genes_group).

Batch correction

After merging and normalization, samples were manually controlled for batch effects by overlap inspection of PCA and UMAP plots. Batch effects were corrected with the scanpy implementation of Harmony¹⁰³ applying the sce.pp.harmony_integrate function followed by recalculation of the neighborhood graph and UMAP based on the Harmony representation. Even replicate distributions were assured by generation and inspection of barplots and UMAP plots indicating replicate proportions by cell cluster (Supplementary Figures 2,5-6).

Cell cluster annotation

Automated cluster annotation based on literature markers

Literature research was performed to assemble a set of specific marker genes for cell types potentially emerging in HLOs, including mature liver cells and progenitors (Supplementary Table 1). Wilcoxon Rank sum test derived marker gene lists (cf. above) were selected for the top 70 marker genes per cluster and forwarded to the gseapy.enrichr¹⁰⁴ function together with the literature reference list and the number of genes in the dataset was provided as background parameter. Results from enrichr with a p-value below 0.05 were assigned as the new cluster identity based on the number of significantly enriched marker genes. If no hits were found that met the p-value criteria, the label with the lowest adjusted p-value was chosen as the identity. No gene hits resulted in the label “unannotated“.

Automated annotation based on curated databases

The PanglaoDB⁵¹ marker gene set (version 27 March 2020) was downloaded and reduced to human transcripts and subgroups relevant in the context of human liver. Wilcoxon Rank sum test derived marker gene lists were selected and annotated as described above. One main cluster remained unannotated and showed enriched expression of CENPF, MKI67, UBE2C and TOP2A. Comparing against an additional reference, CellMarker⁵² (wget http://biocc.hrbmu.edu.cn/CellMarker/download/Human_cell_markers.txt), resulted in successful annotation as MKI67+ progenitor cells and the respective marker gene set complemented the reference for all subsequent annotations (Supplementary Table 2).

SingleCellNet annotation

Generation of a reference data object. For comparing HLO cells to human liver cells, human 10X scRNA-seq data were chosen as a reference⁵⁷. Respective cellranger count outputs were downloaded via: wget ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE156nnn/GSE156625/suppl/GSE156625%5FHCCFgenes%2Etsv%2Egz; wget ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE156nnn/GSE156625/suppl/GSE156625%5FHCCFmatrix%2Emtx%2Egz; wget ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE156nnn/GSE156625/suppl/GSE156625%5FHCCFbarcodes%2Etsv%2Egz and loaded into scanpy via sc.read_10x_mtx to obtain an AnnData object with raw count data. To add cell types defined by the authors, the public HCCF1F2.h5ad file of the respective dataset was downloaded from: https://data.mendeley.com/public-files/datasets/6wmzcskt6k/files/8ecdfe82-955a-40bd-b6b7-50edd67e9d31/file_downloaded and relevant metadata columns (“louvain”, “NormalvsTumor”) were extracted from the adata.obs dataframe, integrated into the previously generated AnnData object, and were added to the raw count data composing the new AnnData object. SingleCellNet classification of HLO-cell types based on human reference data. Using pysinglecellnet (version 0.1), the created AnnData object was defined as the reference dataset. A classifier was trained with the top 15 genes and gene pairs, respectively. Genes were not limited to highly variable genes. Subsequently, each HLO sample was classified. To this end, each cellranger filtered feature matrix raw output (matrix.mtx, features.tsv, barcodes.tsv) was loaded into scanpy as an individual query sample and underwent classification. The annotated cell type was stored as a categorical named “SCN_class” in the adata.obs slot from where it could be retrieved for downstream analyses.

Inflammatory gene and fibrosis scoring

To render gene lists for biological processes related to fibrotic and inflammatory injury scenarios, publicly available GO term^{68, 105} gene lists were browsed via Amigo2¹⁰⁶ filtering for Homo sapiens genes only and the “bioentity_label” column was exported. All GO terms are listed in Supplementary Table 4. Cell cycle scores were calculated as described above. Duplicate values were removed from lists prior to scoring. CYP450 superfamily scores were calculated based on a custom gene list and included CYP450 members and the transcript coding for the Cytochrome P450 Oxidoreductase (POR). Finally, the scanpy function sc.tl.score_genes using default options was used for scoring cells.

DGE and pathway enrichment analysis

Differential gene expression (DGE): DGE was performed on raw count data for each defined cluster separately based on batch-corrected datasets including all technical replicates of the respective time points or conditions that were compared (Control vs. TGF-β1 10 ng/mL vs. PA 500 µM). For each cluster, differentially expressed genes were rendered applying the Wilcoxon rank-sum test as described above, after exclusion of mitochondrial and ribosomal genes. Gene Set Analysis. For each cluster a hypergeometric enrichment test was performed on the top 150 differentially expressed genes per condition. Subsequently, the python implementation of GSEA/EnrichR¹⁰⁴ in gseapy (version 0.10.7) was applied to render enriched terms. Gene sets used were “GO_Biological_Process_2018“ and “WikiPathways_2019_Human“, the organism was set to “Human“.

CellPhoneDB cell-cell interactome

CellPhoneDB⁷⁶ (version 2.1.7) with rpy2 (version 3.0.5) was run on h5ad files containing log-transformed count data from OS-cultured HLOs treated with TGF-β1 (10ng/ml) or PA (500 µM) and their respective controls to infer cell-cell interactions. Metadata from previously harmony-integrated and cell-type annotated AnnData objects of the same conditions were exported to .txt files and provided as an input to the statistical_analysis function along with “gene_name“ set as an identifier for genes in the counts data. To generate the count network data and inputs for dotplots, the plot_heatmap_plot function was applied, and all plots were created in python (version 3.7) using seaborn¹⁰⁷. Mean expression indicated by dot size in all figures refers to CellPhoneDB output “means“, and is the mean of the respective two individual partner‘s average expression. To generate Venn diagrams displaying overlapping and distinct significant (p-value < 0.05) receptor-ligand interaction groups, the CellPhoneDB output file “pvalues” served as an input, containing p-values for the enrichment of the interacting ligand-receptor pair in each of the interacting pairs of cell types. The list was filtered for pairs with at least one p-value below 0.05 (meaning at least one significant interaction) for each condition. Next, the list of receptor-ligand pairs was grouped by the first interacting partner, which resulted in a total of 142 possible groups for each condition. The overlapping and distinct fractions were calculated (Supplementary Table 14), and Venn diagrams generated using pyvenn (Adam Labadorf, https://github.com/adamlabadorf/pyvenn).

Harmony timeseries forced layout

Concatenated, quality criteria filtered (cf. above) data from OS-HLOs controls (n = 4), OS-HLOs + TGF-β1 (n = 2), and OS-HLOs + PA (n = 2) were total-count normalized (10E4), log-transformed, and 5000 highly variable genes were selected. Subsequently, an augmented affinity matrix and force-directed layout was calculated with harmony_timeseries via the sce.tl.harmony_timeseries function with 1000 components.

Palantir trajectory inference for time course reconstruction of liver injury stages

Harmony augmented affinity matrices served as distances for Palantir⁹¹ trajectory analysis as inputted to sce.tl.palantir. The first cell from the MKI67+ progenitor-like cluster in the sorted adata.var indices was chosen as an early cell. No terminal or start cell was predefined. Palantir analysis was performed with 30 nearest-neighbors, 25 maximum iterations, scaling of components enabled and 1200 waypoints. Additional plots were generated with palantir (version 1.0.0).

Retrieval of human NAFLD signatures in injured HLOs

Previously published 26 and 98 gene signatures predicting fibrosis in NAFLD⁹⁴ were transferred to lists, and each cell from the integrated dataset of all orbital-shaker cultured HLOs (n = 4 controls, n = 2 PA, n = 2 TGF-β1, total 44,275 cells, quality control and processing as described above) obtained a score for the respective gene lists based on raw count data as described for other inflammatory and fibrosis-associated scores. Dot plots were generated with scanpy based on scaled count data for the entire dataset or for each cell type individually splitting into the three treatment groups. For statistical analysis, lists of numerical score values of all cells from each group (control, PA, TGF-β1), either in the full dataset or per cell cluster calculated with the leiden algorithm at resolution 0.5 and annotated with the database approach were forwarded to statstical testing described below.

Statistics and reproducibility

All statistical tests were performed in python (version 3.6) utilizing scipy.stats¹⁰⁸ (version 1.8.0) and scikit-posthocs¹⁰⁹ (version 0.6.7). Statistics comparing two groups were performed using the two-tailed Mann-Whitney U test followed by a Bonferroni correction. For three or more groups, the two-tailed Kruskal-Wallis non-parametric test was applied followed by a post hoc Conover’s test with Bonferroni correction. A difference of mean between groups was considered significant at a p-value below 0.05. N numbers for individual HLO samples in qPCR and immunohistochemistry experiments indicate individual experiments (individual differentiation cycles starting from day 0 hPSCs). For scRNA-seq, n numbers or samples refer to individual HLO single cell suspensions that arose from the same differentiation experiment if performed in replicate, e.g. for day 21 HLOs 3 replicates (n = 3) were sequenced, which in this case means from one HLO experiment, single cell suspensions from three different wells from the same treatment and time point were harvested and underwent individual library preps.

Data availability

Raw data for single-cell RNA sequencing are available at Gene Expression Omnibus (GEO) under the accession GSE207889 upon release.

Code availability

The following scripts are available upon request and will be made publicly available upon release: Sirius red quantification pipeline: https://github.com/anjahess/sirius_red ; Single-cell RNA sequencing scripts: https://github.com/anjahess/scrnaseq.

Supplementary information

Supplementary information includes 11 figures and 15 tables:

Supplementary Figure 1, Representative images of fixed HLOs, related to Figure 1c.

Supplementary Figure 2, Replicate distribution and marker gene expression from scRNA-seq, related to Figure 1d.

Supplementary Figure 3, SingleCellNet classification of HLO cells and analysis of WikiPathways, related to Figure 1d.

Supplementary Figure 4, Induction of fibrotic phenotypes in OS-cultured HLOs, related to Figure 2b, f, i.

Supplementary Figure 5, Marker gene expression and replicate distribution in 44,275 single cells from control and injured HLOs, related to Figure 3b.

Supplementary Figure 6, Contributions per replicate for TGF-β1 and PA treatment with respecitive controls, related to Figure 3b.

Supplementary Figure 7, Differential gene expression (DGE) for each major cell type after treatment with TGF-β1 and PA compared to controls and gene modules for fibrosis and inflammation, related to Figures 3e-f.

Supplementary Figure 8, Pathway enrichment across cell clusters, related to Figures 3g.

Supplementary Figure 9, Overview of differential cell-cell interactions in OS-HLOs upon TGF-β1 treatment, related to Figure 4c.

Supplementary Figure 10, Overview of differential cell-cell interactions in OS-HLOs upon PA treatment, related to Figure 4c.

Supplementary Figure 11, Expression of genes composing signatures predicting fibrosis stages in NAFLD, related to Figure 6c.

Supplementary Table 1, Literature-based marker genes, related to Figure 1.

Supplementary Table 2, Cell types for database annotation, related to Figures 1 and 2.

Supplementary Table 3, Top 100 cluster-defining genes in day 21 HLOs, related to Figure 1.

Supplementary Table 4, GO terms for scoring inflammation and fibrosis, related to Figure 3.

Supplementary Table 5, Cell type distributions for control, TGF-β1 and PA-treated HLOs, related to Figure 3.

Supplementary Table 6, Database annotation results for PHSC-like cells, related to Figure 3.

Supplementary Table 7, DGE for hepatocyte-like cells, related to Figure 3.

Supplementary Table 8, DGE for cholangiocyte-like cells, related to Figure 3.

Supplementary Table 9, DGE for HSC-like cells, related to Figure 3.

Supplementary Table 10, DGE for PHSC-like cells, related to Figure 3.

Supplementary Table 11, DGE for MKI67+ progenitor-like cells, related to Figure 3.

Supplementary Table 12, DGE for myoepithelial-like cells, related to Figure 3.

Supplementary Table 13, DGE for smooth muscle cell-like cells, related to Figure 3.

Supplementary Table 14, CellPhoneDB receptor-ligand pair groups, related to Figure 4.

Supplementary Table 15, Sequences of human qPCR primers.

Author contributions

A.H. and A.C.M. conceived and designed the study with additional assistance from S.D.G. Bench experiments were performed by A.H. and S.D.G. with A.B.S. and T.H. providing additional assistance. Computational analyses were performed by A.H. with assistance from R.R. The manuscript was written by A.H. and A.C.M with input from all other authors.

Competing interests

A.C.M. receives research funding from Boehringer Ingelheim, Bristol-Myers Squibb, and Glaxo Smith Klein for other projects and is also a consultant for Third Rock Ventures.

Supplementary information

Supplementary Figure 1. Representative images of fixed HLOs. Corresponding to Fig. 1c.

a. Panels showing individual HLOs from left to right: H&E, secondary antibody only (Control), anti-CEBPα (nuclear staining indicated by arrowheads). Scale bars 100 µm. b. Two HLOs from a., increased magnification for representative indication of nuclear CEBPα staining (arrowheads). Scale bars 50 µm.

Supplementary Figure 2. Replicate distribution and marker gene expression from scRNA-seq. Corresponding to Fig. 1d.

a. UMAP plot mapping 17,898 single cells from day 21 ULA-HLOs from Fig. 1 colored by replicate. b. Barplot indicating the replicate distribution as the proportion of total cells per cluster for each cluster calculated with the Leiden algorithm at resolution 0.1 and database annotation. c. UMAP plots are colored for the scaled expression of canonical cell type marker genes in HLOs after quality control (QC). d. Barplots showing cell type distributions across replicates as the proportion of each cell type per total cells per sample. Annotation approach (Methods) is indicated (from left to right: database-annotation utilizing PanglaoDB¹ and CellMarker² and statistical testing via GSEA-py enrichr³, literature-based marker sets and statistical testing via GSEA-py enrichr³, SingelCellNet⁵ using human liver scRNAseq data⁴ as a reference). Leiden resolution = 0.1. e. Barplots showing the cell type distibutions for database and literature annotations as described for e, this time at increased Leiden resolution of 0.5.

Supplementary Figure 3. SingleCellNet classification of HLO cells and analysis of WikiPathways⁶. Corresponding to Fig. 1d and f.

a. SingleCellNet (SCN)⁵ classifier output heatmap trained on 31 clusters from human liver scRNA-seq data⁴. Heatmap showing the score calculated by SCN for each cell of the reference data (x-axis) for each of the 31 possible clusters and the ‘random’ category (y-axis). b. Individual SCN⁵ results for each day 21 (D21) replicate. Heat maps showing classification score for each individual cell with the final SCN-annotation in the sample (x-axis) for each of the 31 possible clusters and the ‘random’ category (y-axis). Cluster labels of cell types accounting for < 5% of all cells were removed for readability. c. Enrichment of human GO-terms⁹ (left) and WikiPathways 2019⁶ (right) for the three major cell types in HLOs as labeled in main figure 1f. A maximum of ten significantly (p-value < 0.05) enriched terms was plotted. Leiden resolution = 0.1.

Supplementary Figure 4. Induction of fibrotic phenotypes in OS-cultured HLOs. Corresponding to Fig. 2b, f, i.

a. Enlarged representative brightfield images from Fig. 2, showing day 25 HLOs cultured on an orbital shaker (control, left) and treated with TGF-β1 (10 ng/ml, right). Arrowheads indicate areas of surface roughening. b. qPCR results for COL1A1 and TNFA. Shown are qPCR results for relative mRNA levels normalized to GAPDH (2^⁻ddCt). Kruskal Wallis test (two-tailed) followed by Conover’s post hoc test with Bonferroni correction. Ns, not significant. N ≥ 3 individual experiments. c. Sirius red quantification in HLOs treated with TGF-β1, PA, and OA. Shown is the percentage of Sirius red positive tissue by total tissue (normalized by the mean of the untreated control HLOs in each experiment). Quantification based on the whole slide Sirius red percentage mean (left), mean of averaged 200 x 200 µm regions of interest (ROIs) around single HLOs (center), or the Sirius red percentage per tissue for each HLO ROI (right). Kruskal-Wallis (two-tailed) test followed by a post hoc Conover’s test with Bonferroni correction. N ≥ 4 HLOs from one individual experiment (organoid ROIs), n ≥ 3 individual experiments (all plots).

Supplementary Figure 5. Marker gene expression and replicate distribution in 44,275 single cells from control and injured HLOs. Corresponding to Fig. 3b.

a. UMAP plot mapping 44,275 single cells from OS-HLOs treated with TGF-β1 (10 ng/ml, n = 2) or PA (500 µM, n = 2) and their respective controls (n = 4) colored by replicate. b. Barplot indicating the replicate distribution as the proportion of total cells per cluster for each cluster calculated with the Leiden algorithm at resolution 0.5, cluster labels as annotated by database approach (Methods). c. UMAP plots are colored for the scaled expression of canonical cell type marker genes expressed in HLOs after quality control (QC). d. Barplots showing cell type distributions across replicates as the proportion of each cell type per total cells per sample. Annotations generated by database annotation (left) and literature-based annotation (right), Leiden cluster resolution = 0.5.

Supplementary Figure 6. Contributions per replicate for TGF-β1 and PA treatment with respecitive controls. Corresponding to Fig. 3b.

a. Barplots indicating the contribution of each cell cluster (each in an individual color) per replicate, displayed as the fraction of total cells per sample. Distributions are shown for all OS-HLO conditions performed in duplicate: TGF-β1 (10 ng/ml, n = 2) or PA (500 µM, n = 2) and their respective controls (n = 2 for each condition). b. Barplots display the replicate contribution as the fraction of total cells per cluster for each cell cluster for all samples. c. Barplot shows the cell cluster proportions for control samples and indicates a relative reduction of HSC-like and an induction of hepatocyte- and cholangiocyte-like cells in OS-HLOs. Cell type contribution for each individual replicate is indicated as the fraction of total cells per sample for each cluster calculated with the Leiden algorithm at resolution 0.5 across harmony-integrated 94,753 single cells from 16 individual samples representing different HLO conditions and time points: day 21 ultra-low attachment (ULA-) cultured HLOs: n = 3 replicates and 17,854 cells, day 34 ULA-HLOs: n = 2 replicates and 16,670 cells, day 25 orbital shaker HLOs: n = 6 replicates and 30,569 cells, day 25 orbital shaker HLOs with treatments: n = 4 replicates and 25,609 cells, not shown, and one HLO reference: n = 1 replicate and 4,051 cells (Ouchi et al.)⁷.

Supplementary Figure 7. Differential gene expression (DGE) for each major cell type after treatment with TGF-β1 and PA compared to controls and gene modules for fibrosis and inflammation. Corresponding to Fig. 3e-f.

a. Heatmaps showing top differentially expressed genes based on Wilcoxon rank sum test calculated on raw count data for each cluster separately. Raw count data for each cell type was extracted, DGE was performed and expression was normalized and scaled again per cluster for visualization only. b. Differential expression is displayed for selected genes associated with fibrosis and inflammation for PA treatment, TGF-β1 treatment, and controls for all cell types indicated on the left. Differential expression is displayed as the mean z-score calculated separately for each cell cluster. Data are displayed for cell types not shown in Fig 3. Full lists of differentially expressed genes for each cell type are supplied in Supplementary Tables 7-13.

Supplementary Figure 8. WikiPathway 2019⁷ enrichment analysis across cell clusters. Corresponding to Fig. 3e and f.

Barplots show the negative decadic logarithm of adjusted p-values for Wiki-Pathway 2019⁶ terms enriched among differential genes for each condition for each cell cluster. Cut-off for plots is an adjusted p-value below 0.05.

Supplementary Figure 9. Overview of differential cell-cell interactions in OS-HLOs upon TGF-β1 treatment. Corresponding to Fig. 4c.

Dotplots showing all significant (p < 0.05) receptor-ligand pairs per cell-cell cluster pair from CellPhoneDB analysis for OS-cultured injured HLOs (red) and their respective controls (grey). Y-axes denote the receptor-ligand pair, each column (x-axis) represents a cell-cell pair. TGF-β1 (n = 2) and TGF-β1-control (n = 2) interactome. Only cell-cell pairs with significant interactions are displayed.

Supplementary Figure 10. Overview of differential cell-cell interactions in OS-HLOs upon PA treatment. Corresponding to Fig. 4c.

Dotplots showing all significant (p < 0.05) receptor-ligand pairs per cell-cell cluster pair from CellPhoneDB analysis for OS-cultured injured HLOs (red) and their respective controls (grey). Y-axes denote the receptor-ligand pair, each column (x-axis) represents a cell-cell pair. PA (n = 2) and PA-control (n = 2) interactome. Only cell-cell pairs with significant interactions are displayed.

Supplementary Figure 11. Expression of cell type- and injury-specific candidate genes composing signatures predicting fibrosis stages in NAFLD gradually change from healthy to PA-treated to TGF-β1-treated HLOs. Corresponding to

Fig. 6c. Dot plots of scaled expression values are displayed for genes from 98-gene signature⁸ to predict fibrosis stages in a cohort of 143 patients across different stages of NAFLD. Each panel shows the scaled expression (calculated previously on all clusters jointly) for the individual clusters among the database-annotated cell types corresponding to Fig. 3b. The gene list was reduced to genes present in each individual replicate (n = 4 controls, n = 2 PA, n = 2 TGF-β1, total 44,275 cells) and expressed in more than 3 cells. Dot sizes correspond to the fraction of cells in a group expressing the respective gene, and color indicates expression level.

Acknowledgments

The authors thank Ramnik Xavier for helpful discussions, Cristin McCabe and Jacques Deguine for assistance with data management, and Adam Slamin, Dan Dubinsky, and the Broad Genomics Platform for help with generation of single cell sequencing data. The authors also thank the Massachusetts General Hospital (MGH) NextGen Sequencing Core for additional assistance with generation of single cell sequencing data and Annika Gabriel for designing color palettes used in the manuscript. A.H. was supported by the Studienstiftung des deutschen Volkes. A.C.M. was supported through internal funding at MGH.

Abbreviations:

HLOs

Human liver organoids

OS

Orbital shaker – cultured HLOs

ULA

Ultra-low attachment plate – cultured HLOs

D21

Day 21 of differentiation

D25

Day 25 of differentiation

D34

Day 34 of differentiation

PA

Palmitic Acid at 500 µM

TGF-β1

TGF-β1 at 10 ng/µl