A spatial atlas of the healthy human liver from live donors Authors: Oran Yakubovsky, Keren Bahar Halpern, Sapir Shir, Roy Novoselsky, Amichay Afriat, Adi Egozi, Tal Barkai, Yotam Harnik, Rouven Hoefflin, Yael Korem Kohanim, Ofra Golani, Inna Goliand, Yoseph Addadi, Merav Kedmi, Hadas Keren-Shaul, Liat Fellus-Alyagor, Lena Prichislov, Dana Hirsch, Chen Mayer, Ron Pery, Niv Pencovich, Timucin Taner, Ido Nachmany, Shalev Itzkovitz Published: 2026 ABSTRACT AbstractReconstructing gene expression atlases for human tissues is challenging due to limited access to healthy samples from live individuals. Neurologically deceased donors often show ischaemic changes, and tissues near diseased regions may have altered gene expression1,2. The liver, with its unique regenerative capacity, allows analysis from live healthy donors. Here, using spatial transcriptomics (Visium, Visium HD3, multiplexed error-robust fluorescence in situ hybridization (MERFISH)4 and PhenoCycler imaging5) and single-nucleus RNA sequencing6, we analysed 16 liver samples: 8 from young live healthy donors and 8 from individuals with liver pathology, sampling ‘adjacent normal’ tissue. Livers from live healthy donors displayed significant gene expression differences compared with the adjacent normal tissues from individuals with liver pathology. Hepatocytes and non-parenchymal cells exhibited marked zonation along the porto–central axis of the liver lobules, with key functions being pericentrally shifted compared to mice and other mammals. Our atlas identified dynamic programmes in early steatotic hepatocytes, including a decline in nuclear-encoded mitochondrial proteins and a compensatory increase in mitochondria-encoded transcripts. This study presents a spatial gene expression reference for the healthy human liver and insights into hepatocyte changes in early steatosis.Subject terms: Liver, RNA sequencing Main The liver is a multitasking organ that maintains nutrient homeostasis, produces circulating proteins, secretes bile acids and detoxifies foreign substances. These functions are carried out by hepatocytes, the main parenchymal cells, which are organized into hexagonal liver lobules measuring 0.5–1 mm in diameter. Lobules are polarized by blood flowing from portal veins to central veins, creating gradients of oxygen and nutrients. In response to this graded microenvironment, key hepatic functions are spatially segregated across lobule layers, a phenomenon known as liver zonation7. Many liver diseases, including metabolic-associated steatotic liver disease8, autoimmune hepatitis9, cholangitis10 and parasite infections11,12, exhibit zonated pathological patterns. High-resolution spatial expression maps of human hepatocytes are therefore essential for understanding liver pathology. Although recent human liver cell atlases have provided major insights13–22, limited access to truly healthy human liver tissue remains a key challenge. The selection of appropriate healthy reference tissues is critical for human cell atlas studies23. Human liver cell atlases have been reconstructed either from neurologically deceased donors (NDDs)14,24 or from histologically healthy-appearing tissues adjacent to liver pathologies16,20. Although commonly treated as healthy references, both sources present limitations1,2. Livers from NDDs experience metabolic disturbances caused by prolonged mechanical ventilation, non-enteral nutrition and medications, all of which can alter hepatic gene expression prior to tissue collection. In addition, post-mortem samples are subject to variable ischaemia times, leading to transcript-dependent RNA degradation that can explain up to 40% of variability in RNA-sequencing (RNA-seq) data2. Similarly, tissues adjacent to liver lesions may be affected by local or systemic disease, as liver gene expression can change even in patients with extra-hepatic tumours25. To establish a truly healthy human liver spatial transcriptomics atlas, we analysed samples from live human donors (LHDs), who underwent rigorous medical screening to ensure a healthy liver baseline prior to donation. We used 10X Visium spatial transcriptomics3, MERFISH4, single-nucleus RNA-seq (snRNA-seq)6 and PhenoCycler imaging5 to reconstruct zonation in LHD livers and compared it with adjacent normal tissue, revealing substantial expression differences. We also generated spatial atlases of other mammalian livers to identify conserved and species-specific zonation features. Because LHD grafts are considered transplantable with up to 20% steatosis, we analysed a broad range of steatosis levels in otherwise healthy livers. By integrating spatial transcriptomics with machine learning-based pixel classification of haematoxylin and eosin (H&E) images, we stratified zonal hepatocytes by lipid content and tracked gene expression changes during lipid accumulation. This work provides a resource for human liver biology and reveals spatially dependent hepatocyte changes in early steatosis. LHD livers are transcriptomically unique To assemble a detailed gene expression atlas of the human liver, we collected specimens from two procedure types: live organ donor hepatectomy, in which a portion of the liver is surgically removed from an LHD prior to transplantation, and liver resection, in which tissue is removed owing to a liver pathology (Fig. 1a and Methods). We also performed spatial transcriptomics of livers from other mammals to identify shared and unique zonated expression patterns (Fig. 1a). Adjacent normal samples showed substantial heterogeneity in age, body mass index, smoking status and medication history, whereas LHD samples were highly homogeneous (Fig. 1b). We categorized the samples according to their histological diagnosis. Our cohort consisted of eight LHDs, three of whom exhibited varying degrees of mild steatosis and eight adjacent normal sample donors, five of whom exhibited varying degrees of steatosis, two with portal fibrosis, and one patient who underwent associating liver partition and portal vein ligation for staged hepatectomy18 (Fig. 1a,b). We performed 10X Visium spatial transcriptomics on fresh-frozen samples, obtaining the transcriptomic signatures of 49,092 spots, with a median of 3,008 spots per patient and 3,393 unique molecular identifiers (UMIs) per spot (Extended Data Fig. 1a–d). We filtered outlier spots and retained 44,243 high-quality spots (Methods). A spatial expression atlas of human and mammalian livers. a, Study design. Liver tissue gene expression was measured using 10X Visium spatial transcriptomics (human, n = 16; mouse, n = 6; boar, n = 2; domesticated pig, n = 3; cow, n = 2), Visium HD (human, n = 3), MERFISH (human, n = 2) and snRNA-seq (human, n = 4). Diagram created in BioRender. Barkai, T. (2026) https://biorender.com/uke1h6i. b, Clinical characteristics of patients from whom liver samples were collected. ALPPS, associating liver partition and portal vein ligation for staged hepatectomy; CCA, cholangiocarcinoma; CRLM, colorectal liver metastases; HCC, hepatocellular carcinoma. c, Heat map of hierarchical clustering of pseudo-bulk transcriptomic profiles. Analysis was done on highly expressed genes. Representative genes are shown on the right. Expression units are z-scores of log10-transformed UMI-sum normalized expression. d, Principal component analysis (PCA) of pseudo-bulk transcriptomic profiles. Data in c,d based on 323 genes with a maximal expression across patients above 5 × 10−4 (Methods). e, MA plot showing the gene expression ratios between the LHD samples and the adjacent normal samples. Genes with maximal expression above 5 × 10−5, q-value below 0.25 and fold change ratio above 1.5 are highlighted in colour. q-values calculated using the Benjamini–Hochberg procedure. Rank-sum P values were calculated for all genes that had maximal expression above 5 × 10−5 in all 16 patients. Selected markers for distinct immune functions and classic hepatocyte functional genes are highlighted with a circle. Source data Live organ donor livers differ from adjacent normal liver samples. a. H&E images of tissues from Visium experiments. The capture area within the fiducial frame is 6.5×6.5 mm. Samples labeled with ‘M’ correspond to liver tissues obtained from LHDs. Samples labeled with ‘P’ correspond to adjacent normal liver tissues resected due to underlying pathological conditions. b–d. Quality metrics of Visium data, n = 16 patients examined in 4 independent experiments. Patient order matches that in Fig. 1b. Circles in c-d are medians, gray line span interquartile ranges. e. Gene set enrichment analysis of the differences between LHD liver samples and adjacent normal liver samples. f. Heatmap of hierarchical clustering of pseudo-bulk transcriptomic profiles of LHD, adjacent normal (black - this study, gray – Matchett et al.20) and NDD (Visium data from Andrews et al.14). Analysis done on genes showing differential expression between LHDs and adjacent normal livers (expression>2e-4, q-value < 0.05 and ratio above 1.5). Genes mentioned in the main text are in bold. Expression units are Z-scores of log10 UMI-sum normalized expression. g-l. Multilinear regression to assess the impact of covariates on gene expression variability in the liver samples studied highlights the impact of sample type (LHD vs. adjacent normal liver). Pseudobulk Visium gene expression data was log-normalized and standardized across samples. A linear regression model including the covariates of sample type, histology, age and sex was assessed over the set of genes with maximal expression above 1e-4. Analysis performed on all LHDs and on the steatotic adjacent normal liver samples P6, P7, P14, P18. Volcano plots show the respective regression coefficients and raw p-values for sample type (g), histology (h), age(i) and sex (j). Both the overall significance (-log10(geometric mean of all p-values), k) and number of genes with p-value < 0.01, l, highlight the higher impact of sample type (LHD vs. Adjacent normal) followed by histology, with a more minor impact of sex and age. P-values in g-l are t-statistic p-values of the two-sided hypothesis tests. Source data Pseudo-bulk and differential gene expression analyses revealed clear clustering by clinical context, most prominently separating LHD livers from adjacent normal livers (Fig. 1c–e and Supplementary Table 1). Adjacent normal livers, both from this study and a recent Visium-based report20, showed reduced expression of genes involved in canonical hepatocyte functions, alongside increased expression of immune-related genes (Extended Data Fig. 1e,f). Reduced hepatocyte functions included bile acid and lipid metabolism, genes encoding urea cycle enzymes (ASS1) and cytochrome P450 enzymes (CYP2E1 and CYP2C8) and the gluconeogenic gene PCK2 (Fig. 1c,e). Increased immune gene expression included the calprotectin genes S100A8 and S100A9, Kupffer cell genes C1QA and C1QB, plasma cell genes IGHA1 and IGHG3, and SAA1, SAA2, CRP and TIMP1. Analysis of an NDD sample14 (Methods) revealed increase in stress (IFITM2, IFITM3 and SAA1,2) and hypoxia-related programmes (LDHA and GAPDH), with partial reduction of hepatocyte gene expression (Extended Data Fig. 1f). Linear mixed regression modelling showed that sample type (LHD versus adjacent normal) was the dominant driver of gene expression variability, exceeding effects of age, sex or histology (Extended Data Fig. 1g–l). These results highlight marked transcriptional differences between LHD, adjacent normal and NDD livers, motivating focused analysis of LHD livers to define healthy liver zonation. Zonation in the healthy human liver To assemble a spatial reference frame, we inferred the position of each Visium spot along the porto–central axis using zonated landmark gene expression. We first validated the consistent pericentral expression of CYP2E126,27 in hepatocytes by comparing spots near or distant from portal triads, as annotated by a certified pathologist (Extended Data Fig. 2a). We then identified the 20 genes most correlated or anticorrelated with CYP2E1 and defined them as pericentral and periportal landmark gene sets. The normalized summed expression of these sets was used to assign a percentile-based zone index to each spot (Fig. 2a and Methods). Spots were grouped by lobule zone, and expression was averaged per zone. Of 1,724 hepatocyte-specific genes (Methods), 1,141 showed significant expression differences (q < 0.25), indicating widespread zonation (Fig. 2b,c). In-situ validation using MERFISH for zonation profiles computed from spatial transcriptomics. a. Violin plots showing the spatial distribution of CYP2E1 across pericentral and periportal spots labeled by a certified pathologist. n ≥ 5 portal areas and n ≥ 5 central areas per sample, annotated by a pathologist. Kruskal-Wallis p < 1.94e-12 for all samples. Circles are medians, gray boxes span interquartile ranges. b–c. MERFISH image of patient M5 (b) and M8 (c) showing selected zonated genes. Blowouts of selected regions are shown on the right. d. mRNA zonation profiles for representative genes. Profiles are means over the 2 patients in MERFISH (red, M5 and M8) and 5 patients in Visium (black, M4, M5, M6, M7,M8), normalized to their maximal expression level across zones, patches show the SEM. e. Spearman correlation between the portal bias (log2portal expressioncentral expression) in the MERFISH and Visium methods. Shown are highly expressed genes (maximal expression above 5e-5). Selected genes’ names are shown. P-values are two-sided. Source data Zonation of hepatocyte gene expression. a, Spatial transcriptomics Visium 10X liver slides, coloured by spot-inferred zone index (Methods) in a range from yellow (portal) to blue (central). Spots predominantly composed of fibrotic tissue are represented in grey (Methods). b, Heat map of the mRNA zonation profiles based on the Visium 10X data. In total, 1,724 highly expressed hepatocyte-specific genes with a gene zonation score (Methods) are shown. Profiles are normalized to their maximal value across zones across all healthy, non-steatotic liver samples (M4–M8; Fig. 1b). Profiles are sorted by the zonation profiles centres of mass (Methods). c, mRNA zonation profiles of representative hepatocyte-specific genes in b. Data are means (line) and s.e.m. (shaded region) across zonation profiles of all healthy, non-steatotic liver samples. To validate our reconstructed zonation profiles and increase spatial resolution, we applied two complementary spatial transcriptomics methods: MERFISH and Visium HD. We performed multiplexed single-molecule RNA imaging of 500 genes using the MERSCOPE platform (Extended Data Fig. 2b,c, Methods and Supplementary Table 3). Zonation profiles showed strong agreement between MERFISH and Visium data across two patients (M5 and M8; Spearman R = 0.49, P = 2 × 10−7 and R = 0.57, P = 5 × 10−324, respectively; Extended Data Fig. 2d,e). As an additional validation, we applied Visium HD to formalin-fixed, paraffin-embedded (FFPE) sections from three patients (M1, M2 and M6; Extended Data Fig. 3 and Supplementary Table 4). Visium HD, which relies on targeted probe hybridization, as opposed to the polyA capture method of Visium, yielded highly concordant zonation profiles (Spearman R = 0.7, P < 10−30; Extended Data Fig. 3b,c). We further validated zonation using highly multiplexed immunofluorescence (PhenoCycler5,28) targeting 48 markers in LHD and adjacent normal livers (Supplementary Table 5). The zonation atlas is accessible via our online web app (https://shalevapps.weizmann.ac.il/liver_app/). We next turned to analyse the functional significance of zonation in the healthy human liver. Visium HD validates zonation reconstruction with higher spatial resolution. a. H&E images and overlayed inferred lobule layers for three LHD patients (M1, M2, M6) using 8um*8um spot resolution. Layers computed for all spots with more than 20 UMIs. Boxed region highlighted on the right column with three example genes – the pericentral GLUL (green), the mid-lobule gene KLKB1 (red) and the periportal gene SDS (blue). Images colored by inferred 8um spot lobule layers (left). Scale bars are 200um for M1, M2 and 40um for M6. Shown are all 8*8um spots with more than 20 UMIs. CV – central vein. b. Max-normalized average zonation profiles inferred from Visium (black) and Visium HD (blue) for the three patients. Patches are SEM. c. Zonation profiles strongly overlap between the two measurement techniques (Spearman correlation R = 0.7, two-sided p < 1e-30, computed over all genes with maximal expression above 5e-5 in both datasets). Source data Reconstructed zonation profiles revealed that pericentral hepatocytes exhibit increased expression of xenobiotic metabolism genes such as CYP2E1, bile acid biosynthesis genes including CYP27A1, and lipid biosynthesis genes such as FASN and MLXIPL (Fig. 2b,c). By contrast, periportal hepatocytes showed higher expression of immune-related genes, including CRP and SAA1, as well as the gluconeogenic genes PCK1 and ALDOB. Notably, the mitochondrial gluconeogenesis gene PCK2, which is human-specific, displayed pericentral zonation, challenging the prevailing view that gluconeogenesis is exclusively periportal in mammals7,26. The main hepatic glucose transporter GLUT2 (encoded by SLC2A2) was also pericentrally zonated (Fig. 2b,c). Immunohistochemistry confirmed pericentral enrichment of PCK2 and GLUT2 proteins (Extended Data Fig. 4). Discordant zonation profiles between human and mice. a,c – immunohistochemistry validations for the respective genes based on the human protein atlas68. b,d – Visium reconstructed zonation profiles of the patients analyzed in our study – M prefix denotes the LHD patients, P prefix denotes the adjacent normal livers. Zonation profiles normalized to the maximum across zones. e – Immunofluorescence validation for the pericentral zonation of GLUT2 (green) in LHD liver (M5, left) and the periportal zonation in mice (right). f – Immunofluorescence validation for the pericentral zonation of DPP4 (green) in LHD liver (M5, left) and the non- zonated expression in mice (right). red – GS (glutamine synthetase) antibody. In a,c,e,f PP – periportal, PC- pericentral. Scale bars – 50 um. Mouse images representative of 2 repeats, Human images representative of 3 repeats. We applied gene set enrichment analysis29 (GSEA) to identify strong pericentral enrichment of multiple hepatic functions, including xenobiotic metabolism, fatty acid metabolism and peroxisomal pathways (Extended Data Fig. 5a). In addition, we observed pronounced zonation of signalling ligands and receptors30 (Extended Data Fig. 5b,c). Pericentral zones were enriched for genes encoding WNT pathway components, including the ligands WNT2 and RSPO3 and the RSPO3 receptor LGR531–33, whereas periportal zones showed increased expression of genes encoding ligands (JAG1 and JAG2) and receptors (NOTCH3) of the Notch pathway, a pathway that is central to liver development, regeneration and disease34. Zonation was also evident at the level of transcription factors35 (Extended Data Fig. 5d). Pericentral hepatocytes showed enrichment of the aryl hydrocarbon receptor AHR, an activator of cytochrome P450 xenobiotic metabolism genes36, whereas periportal hepatocytes expressed higher levels of the Notch-activated transcription factor genes HES1, HEY1 and HEYL. Notably, HNF4A37, a master hepatic transcription factor that is periportal in mice, exhibited strong pericentral zonation in humans (Extended Data Fig. 5d, q-value = 8.5 × 10−14). Our analysis further identified genes with peak expression in mid-lobule zones (Extended Data Fig. 5e,f), including HSD17B13, C6, KLKB1, LIPC, HGD and the proteoglycan gene SDC1 (encoding CD138), which we validated using PhenoCycler antibody staining (Extended Data Fig. 5f). Canonical mouse mid-lobule genes IGFBP2 and HAMP26 were not enriched in the human mid-lobule. Although absolute expression levels differed between LHD and adjacent normal livers (Fig. 1c–e), overall zonation trends were highly correlated (Spearman R = 0.82, P < 10−320; Supplementary Table 6). Pathway enrichment gene zonation analysis in healthy LHD liver samples. a. Gene Set Enrichment Analysis (GSEA)29 of the sorted gene zonation scores, calculated as the linear combination of the zone index with the highest gene expression and the gene’s center of mass. The analysis utilized the KEGG69 and Hallmark70 gene sets in healthy, non-steatotic LHD liver samples (n = 5). b–d. Zonation patterns of ligands (b), receptors (c), and transcription factors (d) for highly expressed highly variable genes (q-value less than 0.05, expression more than 3e-6, dynamic range (ratio of maximal to minimal zonation value) more than 0.5) were analyzed in healthy, non-steatotic LHD liver samples. The cumulative significance of gene zonation across all patients (n = 5) was determined using Fisher’s method, combining the zonation bias p-values of each gene into a chi-squared distribution. Profiles are normalized by their maximum and sorted by centers of mass. e. The human liver includes non-monotonically zonated genes. Shown are the top 90 genes sorted by their mid-lobule score (ratio of maximal expression in layers 4–6 to maximal expression in other layers). Analysis performed on genes with maximal expression above 2e-5. Zonation profiles of bold font genes are shown on the right. Lines are means and patches are standard errors of the means over the non-steatotic LHD liver samples (n = 5). All analyses based on the Visium dataset. f. PhenoCycler image of patient M3 showing the mid-lobule zonation of SDC1 (CD138, red). PANCK in cyan stains portal cholangiocytes, aSMA (ACTA2) stains periportal and pericentral fibroblasts. Scale bar – 100 µm. PC – pericentral, PP – periportal. Source data Zonation differences in humans and mice Our zonation analysis revealed that several key hepatocyte functions and transcription factors are pericentrally enriched in humans, in contrast to patterns reported in mice, and that the identities of mid-lobular genes differ between species. To systematically compare zonation, we analysed a published mouse spatial transcriptomics dataset38 (n = 6; Fig. 3a). Correlation between portal biases in mice and humans, defined as log2-transformed expression ratios between periportal and pericentral zones, was unexpectedly low (R = 0.15). Although some genes, including pericentral RHBG, SLC13A3 and CYP1A2, or periportal HAL and ASS1, showed conserved zonation, many exhibited pronounced species-specific differences (Fig. 3b). Notably, murine mid-lobule genes IGFBP2 and HAMP were periportally zonated in humans, whereas VNN1, THBD and INMT were pericentral in mice but periportal in humans. Major functional shifts in humans included pericentral enrichment of bile acid metabolism, glycolysis, gluconeogenesis and fatty acid metabolism (Fig. 3b,c). Lipid metabolism genes such as FASN, APOA5, GPAM, ACSL5 and DPP4 (Extended Data Fig. 4f) were strongly pericentral in humans but periportal or weakly zonated in mice. The urea cycle is an important hepatic process by which toxic ammonia is converted into urea that is excreted in the urine. With the exception of the periportal ASS1, key urea cycle enzymes, including NAGS, CPS1, OTC and ASL, shifted pericentrally in humans. The pericentral expression of OTC could be related to the pericentral zonation of HNF4A, an upstream regulator of this gene39 (Fig. 3b and Extended Data Fig. 5d). In addition, key Kupffer cell markers (CD5L, CD163 and C1QB) were periportal in mice but pericentral in humans (Fig. 3b). Marker-based analyses (Methods) further revealed discordant zonation of lipid-associated macrophages (LAMs) and T cells, with more periportal enrichment in humans (Fig. 3d). Distinct and shared zonation programmes across mammals. a, Humans (y axis) and mice (x axis) show only weak correlation between portal bias (log2 of the ratio between expression in the most periportal and pericentral layers). Included are intersecting genes with maximal zonal expression less than 10−5. (Spearman correlation R = 0.15, two-sided P value = 5.18 × 10−25). b, Representative mRNA zonation profiles for humans (blue) and mice (red). Lines represent mean and shaded regions indicate s.e.m. over the individual repeats within each species, normalized to the maximum across zones. PC, pericentral; PP, periportal. c, GSEA of the differences in portal bias between humans and mice. EMT, epithelial–mesenchymal transition; FDR, false discovery rate; H, Hallmark pathways; K, KEGG pathways; NF-κB, nuclear factor-κB; TNF, tumour necrosis factor; UV response DN, ultraviolet radiation response, downregulated genes. d, Portal biases (log2portalexpressioncentralexpression) of the summed zonation profile of cell-type-specific marker genes (Methods), highlighting discordances between mouse and human (labelled in red). e–g, Visium measurements of liver sections of wild boar (e), pig (f) and cow (g), coloured by spot-inferred lobule layer (blue, pericentral; yellow, periportal). h, Hepatocyte-specific genes show pericentral shifts in their zonation profiles in humans compared with mice and other larger mammals. Box plot showing the portal bias (log2portalexpressioncentralexpression) of highly expressed hepatocyte-specific genes (Methods). The P value is the maximum rank-sum test P values between human and each of the other four species individually. Dots represent the mean log2-transformed portal bias of individual samples. Red lines are medians of the log2-transformed portal bias of the UMI-summed weighted zonation profiles (Methods), and boxes span the interquartile range. Human zonation data includes all non-steatotic LHD liver samples (M4–M8), n = 5 independent samples. Human icon in a,b,h created in BioRender. Barkai, T. (2026) https://biorender.com/ckgl5k9. Animal icons in b,e–h created in BioRender. Barkai, T. (2026) https://biorender.com/uke1h6i. Source data Mice differ from humans in having higher metabolic rates and smaller liver lobules40. To test whether the pericentral shifts observed in human hepatic functions are shared with larger mammals with more similar metabolic rates and lobule sizes to humans, we generated new spatial transcriptomics datasets from wild boar (n = 2; Fig. 3e), domestic pig (n = 3; Fig. 3f) and cow (n = 2, Fig. 3g) (Fig. 3e–g). We performed 10X Visium spatial transcriptomics on fresh-frozen samples (Extended Data Fig. 6 and Methods). Spot zonation was determined using landmark gene expression in cow and mouse, and spatial distance from portal fibrotic septa in pig and boar (Methods). For hepatocyte genes with 1:1 orthologues across all species (n = 12,410; Supplementary Table 7), we computed periportal-to-pericentral log2-transformed expression ratios. Across species, hepatic genes were significantly more pericentrally biased in humans (Fig. 3h; one-tailed rank-sum P < 1.45 × 10−6 for all species). These results indicate enhanced pericentral functional specialization and diversity in human hepatocytes compared with other mammals. Non-human mammalian Visium samples. a. H&E images of tissues from spatial transcriptomics experiments. The capture area within the fiducial frame is 6.5 ×6.5 mm. ‘C’ - cow, ‘P’ - wild boar. ‘PD’ - domesticated pig. b–d. Quality metrics of Visium data, n = 7 subjects of 3 mammalian species examined in 3 independent experiments. Circles in c-d are medians, gray boxes span interquartile ranges. e. Zonation profiles of representative key genes with consistent zonation patterns across the mammals studied. Representative genes that are pericentral (top row) or periportal (bottom row) across species. Lines are means of the max-normalized zonation profiles, patches are standard errors of the means (Methods). Animal icons in a,e created in BioRender; Barkai, T. https://biorender.com/uke1h6i (2026); human icon in e created in BioRender; Barkai, T. https://biorender.com/ckgl5k9 (2026). Source data In contrast to the pronounced pericentral zonation of key hepatocyte genes in humans, we also found universally zonated genes that showed similar trends across organisms. These included the pericentral GLUL and ADH4, and the periportal ALB and GLS2 (Extended Data Fig. 6e). GLS2 encodes glutaminase, which breaks down glutamine into glutamate, whereas GLUL encodes glutamine synthetase, which re-synthesizes glutamine from glutamate. The periportal expression of GLS2 and pericentral expression of GLUL matches the need to recycle this amino acid to ensure balanced representation in the liver entry and exit7. The WNT pathway constitutes a major regulator of liver zonation in mice41–45. Our ligand analysis indicated that the key WNT ligands WNT2 and RSPO3 are pericentrally zonated in humans (Extended Data Fig. 5b). To assess overall WNT pathway zonation, we reconstructed mean zonation profiles of WNT-activated and WNT-inhibited genes43. We found that WNT-activated and inhibited genes were pericentrally and periportally zonated, respectively in all analysed species (Extended Data Fig. 7a,b). The RSPO ligands and the matching LGR5 receptor, critical for maintaining liver zonation32 were pericentrally zonated in all species (Extended Data Fig. 7c). The WNT pathway therefore seems to be an evolutionary conserved regulator of liver zonation. Zonation of Wnt pathway components. a–b. Cross-species zonation profiles of Wnt-activated (a) and Wnt-inhibited (b) genes. Lines/patches are relative means/SEMs over the max-normalized zonation profiles of the genes within each set. Gene sets are the most Wnt-up-regulated and Wnt-down-regulated genes analyzed in Gougelet et al. (Methods)43. Profiles are normalized to their maximal expression across zones. blue-human, red-murine, cyan-cow, pink-boar, purple-domesticated pig. Human n = 7, mouse n = 6, cow n = 2, boar n = 2, domestic pig n = 3. c. Max-normalized zonation profiles of major Wnt pathway components71. Included for each data only genes with maximal zonal expression above 1e-6, genes for each species sorted by center of mass. Human icon created in BioRender; Barkai, T. https://biorender.com/ckgl5k9 (2026); animal icons created in BioRender; Barkai, T. https://biorender.com/uke1h6i (2026). Zonation of liver NPCs Non-parenchymal cells (NPCs) have critical roles in mediating liver function7,33. These include endothelial cells that line the major blood vessels (liver vascular endothelial cells (LVECs)) and the sinusoids (liver sinusoidal endothelial cells (LSECs)), Kupffer cells, the liver’s resident macrophages and stromal cells, which include portal myofibroblasts, vascular smooth muscle cells (VSMCs) and hepatic stellate cells. In mice, LSECs and fibroblasts show strong zonated expression46,47. Our ligand and receptor zonation analysis revealed zonated morphogens, such as the pericentral WNT and RSPO ligands (Extended Data Fig. 5b). We next sought to disentangle the specific zonal signatures of the key human liver NPCs. To resolve zonal programmes of human liver NPCs, we performed snRNA-seq on 4 LHD livers, generating an atlas of more than 67,000 nuclei with broad representation of LSECs, Kupffer cells and fibroblasts (Fig. 4a,b). The dataset showed high capture efficiency (mean 8,020 UMIs per nucleus) but lacked spatial resolution owing to tissue dissociation. Zonation of liver NPCs. a, Single-cell atlas of cells from patients M5, M6, M7 and M8. UMAP, uniform manifold approximation and projection. b, Markers of cell types. c–f, Single-cell atlas of highly zonated genes for each of the four main cellular lineages in the liver: LSECs (c), Kupffer cells (d), fibroblasts (e) and hepatocytes (f). Shown are the 80 most significantly zonated genes for each cell type out of the genes with expression greater than 5 × 10−4 and q-value < 0.01. g, Representative zonated ligands and matching receptors for pairs of cell types. Lines indicate the mean and shaded regions represent the s.e.m. over the cells in each lobule layer. Icons created in BioRender. Itzkovitz, S. (2026) https://BioRender.com/nu556sy. Source data To recover spatial information, we used the Visium HD dataset to extract zonated, highly expressed NPC landmark genes26,46, enabling inference of the zone of origin for each nucleus in the snRNA-seq atlas. We developed a segmentation and annotation pipeline using Visium HD H&E images, applying Stardist for cell segmentation, cell expansion without overlap, and a random forest classifier to distinguish hepatocytes from NPCs (Extended Data Fig. 8a and Methods). Assigning 2 μm × 2 μm Visium HD bins to enclosing cells yielded a spatially resolved single-cell atlas of hepatocytes and NPCs for two LHD livers (M2 and M6), comprising 75,949 cells (Extended Data Fig. 8b–d). Although mRNA capture was low (mean of 630 UMIs per cell), spatial coordinates enabled precise lobule zone assignment and landmark gene extraction (Extended Data Fig. 8e). Zonation of cell-type markers showed periportal enrichment of LVECs, T cells, VSMCs, type 2 conventional dendritic cells (cDC2), B cells and cholangiocytes, whereas fibroblasts, hepatocytes, Kupffer cells and LSECs spanned all zones (Extended Data Fig. 8f), motivating further analysis of these four cell types. Extraction of NPC landmark genes from Visium HD experiments. a. Pipeline for single cell segmentation and classification based on H&E images. Nuclei are detected (1.), cells are expanded and classified into hepatocytes and NPC morphologies based on object classification (Methods). (2.), single cell expression is obtained by aggregating the spots included in each cell (3). Figure taken from patient M2, scale bars are 20um in 1-2 and 5um in 3. b–d. UMAP of segmented single cells from M2 and M6 colored by morphology class (b), Leiden clusters (c) and lobule layer (d). Lobule layer was the median layer of all cells within a diameter of 50um, assigned based on the Visium landmark genes. e. Zonated landmark genes based on the Visium HD NPC analyses (Methods). f. Cell type-specific markers extracted from the snRNAseq atlas (Methods) summed over the Visium-measured zonal expression, demonstrating the biased periportal localization of T-cells, B-cells, VSMC, cDC2, LVEC and cholangiocytes. Visium zonation profiles reconstructed based on all 8 LHD patients. We summed landmark gene expression for each cell type in the snRNA-seq atlas and assigned individual nuclei to eight lobule layers (Methods). Averaging expression within each layer yielded zonated expression programmes (Fig. 4c–f, Extended Data Fig. 9a,b and Supplementary Table 8). We focused on ligands and matching receptors with correlated zonation patterns, computing Spearman correlations between ligand profiles in one cell type and receptor profiles in another, reasoning that concordant zonation may support zone-specific intercellular communication (Fig. 4g, Extended Data Fig. 9c and Supplementary Table 9). We restricted the analysis to hepatocytes, Kupffer cells, LSECs and fibroblasts, the four cell types spanning the parenchyma (Extended Data Fig. 8f). Identified interactions included pericentral WNT2 from LSECs with WNT receptors on hepatocytes (FZD5) and fibroblasts (FZD1), and pericentral RSPO3 from fibroblasts and LSECs with hepatocyte receptors LGR5 and LGR4. Additional interactions suggested signalling from pericentral fibroblasts to hepatocytes and LSECs via BMP9 (encoded by GDF2) and BMP10 through ACVR1 and ENG, respectively, as well as multiple zonated NPC–NPC interactions involving pathways such as Notch and Ephrin. Our NPC zonation profiles can be explored using the dedicated web app (https://shalevapps.weizmann.ac.il/liver_app/), enabling systematic exploration of interactions of the major zonal cell types in the human liver. NPC zonation and zonal interactions between liver cell types. a. Reconstructed fibroblast zonation profiles for VIM (periportal, red), TNC (pericentral, green) and the bimodal ACTA2 (aSMA, yellow). Profiles are normalized to their maximal expression across zones, patches are SEM. b. PhenoCycler image of patient M3 demonstrating the periportal (PP) expression of VIM, pericentral (PC) expression of TNC and bimodal expression of aSMA. Shown are the combined channels (left), as well as the pericentrally-biased TNC (center) and aSMA (right) stanings. Scale bar – 100 µm. Images are representative of the 6 patients that were measured with PhenoCycler. c. Examples of correlated zonation profiles between ligands and matching receptors across liver cell types. For each pair of cells shown are the spatially resolved snRNAseq zonation profiles of ligands (left) and matching receptors (right). All pairs have Spearman correlations of ligand and receptor zonation profiles that exceed 0.5 and expression above 1e-4 of both ligands and receptors in their respective cell types. White horizontal lines separate pericentral and periportal interactions. Icons in c created in BioRender. Itzkovitz, S. https://BioRender.com/7gr66bw (2026). Functional changes in early steatosis Metabolic-associated steatotic liver disease (MASLD) is characterized by abnormal lipid accumulation in hepatocytes, affecting around 25% of the population48,49. The disease progresses through stages of increasing steatosis, inflammation and fibrosis50,51. Hepatic lipid accumulation arises from multiple mechanisms, including increased circulating fatty acids, increased lipid import or synthesis, and reduced lipid export or breakdown. MASLD exhibits strong zonation, with lipid accumulation initiating predominantly in the pericentral zone52,53. Our cohort included four patients with varying steatosis severity (M1, M2, M3 and P6; Fig. 5a–d). We applied machine learning algorithms to quantify the amounts of lipids per spot (Fig. 5a–c, Supplementary Table 10 and Methods). Our patients exhibited pronounced pericentral bias in lipid content (Extended Data Fig. 10a,b). Our spatially resolved atlas enabled exploring functional changes in pericentral hepatocytes at increasing lipid levels. To this end, we ranked pericentral spots by their binned lipid contents and averaged gene expression across ranges of hepatic lipid accumulation (Fig. 5e–g and Supplementary Table 11). We performed k-means clustering of the average dynamic expression programmes across patients. We found that hepatocytes show a decrease in expression of genes involved in lipid influx or synthesis, such as the fatty acid transporter SLC27A3, the triglyceride synthesis enzyme DGAT1 and the cholesterol absorption gene NPC1L1. Conversely, hepatocytes increased the expression of genes involved in beta-oxidation (DIO1 and HSD17B4) and lipid secretion (ABCB4). We further analysed genes that exhibited consistent changes across all patients. We observed a global decrease in expression of the lipid droplet stabilizing enzyme gene HSD17B13, and an increase in ELOVL5, previously associated with a decrease in triglyceride content54 (Fig. 5g). These results highlight a potential compensatory adaptation of hepatocytes tuned for minimization of lipid accumulation in these early steatotic stages. Our analysis also identified intriguing changes in hepatic transcription factors such as KLF10 and ONECUT2 (Fig. 5f,g) and a global decline in nuclear-encoded mitochondrial proteins (for example, the respiratory genes NDUFA9, NDUFS4 and NDUFS8; Fig. 5f), yet a compensatory increase in mitochondria-encoded RNA, such as MT-ND1 and MT-ND4L (Fig. 5g). Dynamic gene expression changes in steatotic hepatocytes. a, H&E-stained Visium slide demonstrating the centrally biased zonation of lipid droplets (sample M1). Image representative of n = 4 steatotic donors (M1, M2, M3 and P6). b, Overlaid spots coloured by their inferred zone index for the same image in a. c, Overlaid colour-coded pixel classification results for the same image as in a and b. Scale bars in a–c are 100 µm. d, Distributions of lipid areas in pericentral spots (n = 4) according to the pixel classification in c. Horizontal lines demarcate per cent quintile thresholds used to annotate spots into five bins. e, k-means clustering (k = 4) of the highly expressed significant genes based on expression in the increasing lipid bins (no lipid bin and five bins with increasing lipid content). Expression is averaged over the lipid bin expression profiles of the four individual patients. Included are genes with maximal expression above 5 × 10−5 and dynamic range above 0.5. Selected genes shown on the left, white horizontal lines demarcate the different clusters. f,g, Selected genes that are globally decreasing (f) or increasing (g) with lipid content. Blue lines are means over the spots in each lipid content bin, shaded regions represent s.e.m. Coloured lines show the profiles in the four individual patients analysed. In f,g, M2 and M3 do not contain lipids with the highest lipid bin sizes. Profiles were smoothed with a running window of three bins. h, Computational pipeline for segmenting individual lipid droplets in the new Visium HD data of patient M1. i, Genes that exhibit higher expression in hepatocytes with large lipid droplets in the Visium HD experiment also show increased expression in the Visium spots with higher lipid content (Spearman correlation R = 0.21, two-sided P = 3.33 × 10−3). Source data Lipid droplet enriched spots are pericentrally zonated. a. Boxplots showing the spots lipid content (in percentage from the spot’s area) along the porto-central axis. P values are Kruskal-Wallis tests between all zones. Red lines are medians, boxes span interquartile ranges. A total of n = 1,806 (M2), n = 2,721 (M3), n = 3,361 (P6), and n = 3,992 (M1) spots were included in the analysis. b. Mean lipid content across lobule layers. Averages taken over the lipid content of all patients each normalized to its maximum across layers. Patches are SEM. c. Zonation reconstruction of mouse Visium data from chow-fed mice and Choline-deficient L-amino acid defined high fat diet (CDAA-HFD). d. Comparison of the log2 ratios of expression between the steatotic and non-steatotic pericentral spots in each of the human patients’ livers and the log2 ratios of expression between pericentral spots in the CDAA-HFD and chow-fed mice. Red highlights the only patient for which expression changes are highly correlated – the severe steatosis patient P17. Shown are the Spearman correlations, p-values are two-sided. Source data To further validate changes in steatotic hepatocytes, we analysed the Visium HD dataset from steatotic donor M1. We performed artificial intelligence-guided segmentation of individual lipid droplets (Fig. 5h) and compared gene expression between Visium HD 2-μm spots within small versus large droplets. Owing to intracellular RNA diffusion, RNA accumulated within each droplet reflects the expression profile of the enclosing hepatocyte. Differential expression between large and small droplets significantly correlated with changes identified in the Visium dataset (Fig. 5i, R = 0.21, P = 3.33 × 10−3). We further compared expression ratios between pericentral steatotic and non-steatotic hepatocytes in human donors and with changes in pericentral gene expression in steatotic and non-steatotic mice, based on a published Visium study using a choline-deficient L-amino acid-defined high-fat diet38 (Extended Data Fig. 10b,c and Methods). Gene expression changes were weakly correlated with mice, except for severe steatosis patient P17 (Extended Data Fig. 10d), demonstrating that our analysed cohort exhibits early steatosis that preceded steato-hepatitis. Our analyses provide a unique resource for exploration of the early states in hepatic steatosis in the human context, and can be explored using our online web app (https://shalevapps.weizmann.ac.il/liver_app/). Discussion Our study provides a spatial expression blueprint of the healthy human liver and reveals that livers from LHDs are transcriptomically distinct from adjacent normal livers commonly used as controls. LHD livers show higher expression of hepatic genes and lower expression of immune-related genes. Together with their greater clinical and physiological uniformity, this makes LHDs an optimal cohort for reconstructing the healthy human liver blueprint. We identified hepatocyte zonation features that are unique to humans compared with mice and other mammals. Notably, human pericentral hepatocytes perform diverse hepatic functions that are periportal in other species, including gluconeogenic components (PCK2) and key urea cycle enzymes (ASL, NAGS and OTC). We also found that HNF4A, periportally zonated in mice, is significantly pericentrally zonated in humans (Fig. 3b). Similarly, SLC2A2, encoding the glucose transporter GLUT2, is also pericentrally shifted in humans (Fig. 2c, Extended Data Fig. 4 and Supplementary Tables 2 and 4). This pericentral zonation may be in line with the ‘glucose paradox’ observation that hepatic glycogen stores are replenished directly from glucose in pericentral hepatocytes, yet from lactate through gluconeogenesis in periportal hepatocytes55. Consistently, our analysis showed higher periportal expression of LDHB, which converts lactate into pyruvate, and a non-zonated expression of LDHA, which converts pyruvate to lactate (Supplementary Table 2). Finally, lipogenic genes that are periportal in mice are pericentrally zonated in humans, with implications for modelling liver steatosis progression. Zonation differences and pericentral shifts in humans were observed not only relative to mice but also compared with other large mammals, including boars, pigs and cows. This was unexpected, as unlike mice, these species have metabolic rates and lobule sizes that are more closely resemble those of humans, and might therefore be expected to exhibit more similar zonation patterns. Moreover, all examined species shared a pericentral WNT niche (Extended Data Fig. 7), indicating that differences in WNT signalling alone cannot account for the human-specific shifts. The human pericentral zonation shifts might be shaped by additional factors beyond lobule geometry and conserved WNT patterning. One possibility is altered metabolite gradients along the porto–central axis, which arise from species-specific differences in zonal uptake and secretion by hepatocytes. Consistent with this idea, we identified pronounced interspecies differences in the zonation of key nutrient transporters and metabolic enzymes, including FASN and GLUT2 (Fig. 3b and Extended Data Fig. 4e). Future work may clarify how nutrient and hormonal gradients interact with canonical patterning pathways to generate the distinctive zonation architecture of the human liver. Our study revealed complex zonal interactions among key hepatic cell types, including LSECs, Kupffer cells and fibroblasts. Several ligands, such as RSPO3 and DLL1, were expressed across multiple cell types, suggesting potential functional redundancy (Supplementary Table 8). We identified fibroblast-expressed BMP9 (encoded by GDF2) and BMP10 as pericentrally zonated morphogens, with corresponding pericentral BMP receptors on hepatocytes and LSECs. This is consistent with a recent mouse study showing that BMP signalling from hepatic stellate cells can act upstream of endothelial WNT ligand expression56. The zonal interaction maps we generated (Extended Data Fig. 9 and Supplementary Table 9) provide a resource for targeted investigation of hepatic cell circuits in liver health and disease57. We also observed a species-specific discordance in Kupffer cell zonation. While Kupffer cells are periportally enriched in mice58, they showed a pericentral bias in humans. This shift may relate to their role in clearing dying hepatocytes, which may undergo higher turnover in the pericentral zones in humans owing to their higher metabolic activity compared with other species. Our spatially resolved expression atlas enabled analysis of hepatic gene expression programmes along the course of early steatosis, in the background of healthy individuals. Our findings of increased expression of mitochondria-encoded RNA is in line with previous work that described an increase in the number of mitochondria in early steatosis59. This increase may represent an adaptive response to limit lipid accumulation through enhanced beta-oxidation. Notably, the expression levels of nuclear-encoded mitochondrial components reduced with elevated lipid content, highlighting a potential imbalance that may contribute to hepatic failure in clearing excess lipids. Our analysis of steatosis focused on relatively early stages, before the commencement of hepatic inflammation (steato-hepatitis), a condition that has been explored in depth in several recent publications16,60. Our combined analysis of spatial transcriptomics and lipid content, quantified based on machine learning of H&E images, could be applied to other organs. For example, it could be used to explore expression differences in adipocytes depending on their lipid content. Our spatial expression atlas forms a blueprint of the healthy liver state, which will form a reference for detailed analysis of zonal changes in diverse liver pathologies such as viral hepatitis, fibrosis, drug-induced liver injuries and autoimmune hepatitis. Methods Experimental methods Human sample preparation Human samples and clinical data were obtained from patients undergoing either hepatectomy surgery due to a hepatic pathology or a donor hepatectomy for a living donor liver donation (Fig. 1b). In the hepatectomies that were performed due to liver pathologies, a sample was taken at least 5 cm from the closest margin of the pathology. In the hepatectomies that were performed as donor hepatectomy for a living donor liver donation, a liver biopsy of 1 cm³ was obtained immediately upon abdominal opening, prior to mobilization of the liver and initiation of resection. All hepatectomies due to liver pathologies were performed at Sheba Medical Center, Israel with approval from the Sheba Medical Center Helsinki committee. All donor hepatectomies for a living donor liver donation were performed at Mayo Clinic, Rochester, MN, with approval from the Mayo Clinic IRB committee. Informed consent was obtained from all patients, and all experiments followed all the Helsinki committee guidelines and regulations. Samples were taken only in cases where no pathology was observed in the sampling area during pre-operative and intraoperative evaluations. In all cases ~1 cm of hepatic parenchyma distant from the liver capsule area were resected. Tissues were gently washed in PBS and were embedded in OCT for Visium and MERFISH or fresh-frozen for Visium HD. Non-human sample preparation Whole washed pig (Sus scrofa domesticus) livers were bought from a butcher at Kibbutz Lahav C.R.O., Israel, which sells meat products from surplus animals. Specimens were adult males (5.5 months, ~85 kg) and an adult female (5.5 months, ~75 kg) raised under farming conditions. Washed cattle (Bos taurus) liver sections were donated by a butcher in Haifa, Israel. Specimens were adult males (17 months, ~670 kg) raised under farming conditions. Wild Boar (Sus scrofa) liver samples were taken from the cadavers of animals euthanized as part of population control regulated hunting done by the Israel Nature and Parks Authority (INPA). The sampling permit was acquired beforehand under INPA regulation. The researchers did not influence the time, location or manner of death; they were solely permitted to take samples from the cadavers. Age and weight estimation of the specimens were evaluated by trained INPA rangers and a veterinarian who participated in the hunting expedition. The boars’ ages were approximately 1–2 years, and they weighed approximately 60 kg. Liver tissues were gently washed in PBS and were embedded in OCT for spatial transcriptomics (Fig. 3b–d and Extended Data Fig. 6). 10X Visium spatial transcriptomics Fresh-frozen OCT embedded blocks were cryo-sectioned into 10-µm-thick slices and carefully placed on a Visium Spatial Gene Expression slide. Each slide contained 4 slices from 4 different patients, total of 4 Visium slides were used for the 16 human samples. Slides were fixed with pre-chilled methanol at −20 °C, stained for H&E as described in the Visium user guide and were captured using a Leica DMi8 widefield inverted microscope equipped with a Leica DFC7000T colour camera and an HC PL APO 20×/0.80 DRY objective (506530, Leica Microsystems). Permeabilization time was set according to the 10X Tissue optimization protocol, resulting in 24 min for all 16 human and 7 non-human mammalian samples. After tissue permeabilization and transcript capture, libraries were prepared according to the Visium Spatial Gene Expression User guide with 14–17 cycles of PCR for cDNA amplification. Four Libraries for each slide were pooled and loaded at a concentration of 400 pM to a SP100 flow cell and sequenced on Illumina NovaSeq 6000. The Fastq reads generated from the sequencer were preprocessed by 10X Genomics space-ranger software (version 2.0.1) which included spatial de-barcoding, read-alignment to hg38 and UMI-generation. Further preprocessing was done in MATLAB R2022a. 10X Visium HD spatial transcriptomics Fresh-frozen tissue samples were processed into FFPE blocks. Sections of 5 μm thickness were mounted onto Visium HD slides and deparaffinized according to the Visium HD Spatial Gene Expression User Guide. H&E staining was performed to visualize tissue morphology, and high-resolution brightfield images were captured using a Leica DMi8 widefield inverted microscope equipped with a Leica DFC7000T colour camera and an HC PL APO 20×/0.80 DRY objective (506530, Leica Microsystems). Following imaging, sections were destained and underwent de-crosslinking as described in the Visium HD FFPE Tissue Preparation Handbook. Three specific probes for each targeted gene were added to the deparaffinized, destained, and decrosslinked tissues. MERFISH spatial transcriptomics Liver OCT blocks were sectioned at width 10 μm. Sections were mounted on MERSCOPE glass slides (Vizgen) and were fixed with 4% paraformaldehyde pre-warmed to 47 °C for 30 min, rinsed, permeabilized, and preserved in sterile 70% ethanol. The procedures of probe hybridization, gel embedding, and tissue clearing were performed according to the MERSCOPE FFPE sample preparation user guide, using the predesigned pan cancer pathways 500-gene panel. In brief, tissues were dried for 1 h in room temperature to improve slide adherence. Then, tissues were photobleached for 3 h using the Vizgen photobleacher, followed by a de-crosslinking step, gel embedding and tissue clearing for 3 days. Next, sections were incubated with MERFISH probe mix for 48 h at 37 °C. After probe hybridization, the samples were labelled with DAPI and poly T and placed in the MERSCOPE flow chamber for imaging. The raw image files obtained were processed via the MERlin image analysis pipeline61. Cellpose software version 1.x was used to segment cells from the nuclear DAPI signal. The decoded RNA molecules were then partitioned into individual cells to generate single-cell count matrices. Immunofluorescence staining Immunofluorescence was performed on 5-µm paraffin-embedded sections using using the BOND RXm research detection platform (Leica, DS9455). Tissue sections were deparaffinized using BOND dewaxing solution. Antigen retrieval was carried out in heat induced antigen retrieval HIER1 (pH 6) for 10 min. Subsequently, sections were permeabilized and blocked for non-specific binding in 20% normal horse serum (Vector labs, S-2000) and 0.1% PBST. Primary antibodies against GS (Abcam, ab64613 and ab73593; 1:100), GLUT2 (Abcam, ab54460; 1:100) and DPP4 (a combination of AF1180 and TA500733, 1:100) were diluted in the 2% normal horse serum and 0.1% PBST and incubated for 30 min. After three washes, sections were incubated for 30 min at room temperature with secondary antibodies and Hoechst (invitrogen, H3570) in PBS. The secondary antibodies used were: Cy3-conjugated Donkey anti-Mouse IgG (Jackson ImmunoResearch, 715-165-151; 1:100), Cy5-conjugated Donkey anti-Rabbit IgG (Jackson ImmunoResearch, 711-175-152; 1:100) and Cy3 Donkey anti Goat (Jackson ImmunoResearch, 705-165-147, 1:100). Slides were washed three times in PBS, mounted with aqua poly mount (polysciences, 1860620), Imaging was done using the Phenoimager HT system (Akoya Biosciences) at 20× magnification. PhenoCycler antibodies and imaging Our antibody panel consisted of DAPI nuclear staining and 48 antibodies targeting proteins expressed in diverse immune, stromal and epithelial liver cell types: CD45, CD3, CD69, CD90, CD11c, Ki67, CD8, CD4, HLA-DR (MHCII), FN1, CD31, VIM, PDPN, EGFR, CD279, SOX4, CD14, CD138, PDGFRA, CA9, GLUT1, CD44, CHI3L1, PanCK, CD163, NDRG1, APOD, SOX2, APOE, MT1MMP, S100B, SNAP25, aSMA, CD19, SOX9, TNC, CD24, CD83, SPARC, CD1c, ATRX, POSTN, VCAN, COL1, CD15, TNR, DCLAMP and HSPG2. Staining was performed on two LHD livers (M1, M3) and four adjacent normal livers (P2, P6, P17 and P4). Each antibody was validated using conventional immunofluorescence on control tissues (liver, duodenum and glioma) for which matched Visium spatial transcriptomic data were available. PhenoCycler Fusion runs were performed according to the manufacturer’s protocol. In brief, fresh-frozen sections were fixed in 4% paraformaldehyde for 20 min, followed by blocking with 4% bovine serum albumin and 0.25% Triton X-100 for 30 min. Primary antibodies were incubated overnight at 4 °C, and secondary antibodies for 2 h at room temperature. All antibodies, including pre-conjugated reagents from Akoya, were further validated on control samples through multiple PhenoCycler runs to optimize antibody concentrations and exposure settings for optimal signal-to-noise ratio. Single-nucleus isolation form FFPE tissue samples Three 25-µm-thick tissue sections were washed with 3 ml of Xylene for 10 min to remove paraffin. Subsequently, they were rehydrated via a sequence of 1 ml ethanol washes, each lasting 30 s: 2 rounds in 100% ethanol, then 70%, and finally 50% ethanol. Next, the tissue sections were washed with nuclease-free water and finally with PBS. Tissue disruption was performed using the 37C_FFPE_1 gentleMACS in 2 ml of digestion mix composed of 1 mg ml−1 Liberase TM (5401119001, Roche), in RPMI1640. The separated nuclei were filtered through a 40-μm mesh, double-rinsed with 1× PBS and centrifuged at 850g at 40 °C for 5 min. They were then resuspended in 500 μl Quench Buffer B (10X genomics PN2001300). The concluding nuclear count was performed using the Denovix cell counter. Single-nucleus RNA profiling from FFPE tissue samples Gene expression libraries were prepared using the GEM-X Flex Gene Expression Human (1000793 10X Genomics), in accordance with the user guide (CG000787 revision B). Additional optimizations in the pre-amplification and Indexing cycles tailored for nuclei derived from FFPE samples. Over 300,000 nuclei per sample underwent a 20-h hybridization with the BC01-BC04 probe sets. This was followed by three washes using the Post-Hyb Buffer as recommended in the guide. Post-hybridization, nuclei were resuspended in the Post-Hyb Resuspension Buffer, counted, combined to achieve target of 20,000 cells per sample, and loaded onto Chromium X chip for capture, adhering to the guide’s procedures. An eight-cycle pre-amplification and nine-cycle indexing PCR was conducted. Library final concentration of 150 pM with the addition of 15 pM from PhiX library was loaded on NovaX (Illumina) sequencing machine aiming at 20,000 reads per cell with the following setting: Read1, 28 bp; Index1, 10 bp; Index2, 10 bp; Read2, 90 bp. Raw reads were de-multiplexed using Bcl-conver (v4.2.4) and aligned using Cellranger multi v9.0.0. GRCh38 (build 2024-A, 10X Genomics) and Chromium Human Transcriptome Probe Set v1.1.0 references were used for probe read mapping. Mouse immunofluorescence Mouse experiments were approved by the Institutional Animal Care and Use Committee of the Weizmann Institute of Science (approval number 05220824–2) and performed in accordance with institutional guidelines. Two-month-old C57BL/6 male mice, housed under regular 12 h:12 h light:dark cycle and fed ad libitum were used in the experiments. Mice were euthanized by cervical dislocation and tissues were collected and fixed in 4% paraformaldehyde in PBS in 4 °C. Livers were incubated for 24 h, transferred to 70% Ethanol, embedded in paraffin, and stored at room temperature before staining. Computational methods Genome alignment Human samples were aligned to genome assembly hg38. Domesticated pig samples were aligned to genome assembly Sscrofa11.1. Cattle samples were aligned to genome assembly BTauARS-UCD1.2. Boar samples were aligned to genome assembly Sscrofa11.1. Spatial transcriptomics preprocessing To remove outlier spots, for each patient, spots with log10 of total UMI counts lower than 3 s.d. from the mean log10 of total UMI counts across all spots, or mitochondrial fraction over 5 s.d. from mitochondrial fraction mean across all spots, were filtered out. Portal spots with prominent fibrosis, identified histologically based on the H&E staining, were manually marked and excluded from the analyses (Fig. 2a, patients M8, P7, P17, P2 and P21). In addition, spots located in areas where the tissue anatomy integrity was unclear were manually annotated and filtered out. The remaining spots were used for analysis. To correct for high abundant transcript diffusion into neighbouring spots51, we calculated for each gene the mean UMI count in spots within the fiducial frame that were not under the tissue to compute a background vector of expression. Next, the background vector was subtracted from each spot under the tissue, and negative values were set to zero. Finally, for the remaining spots, the raw data was normalized to relative counts per spot by dividing by the sum of all genes. Pseudo-bulk analyses of human samples Mitochondrial (^MT-) and ribosomal (^RPL, ^RPS) genes were removed from the variable gene list to minimize batch effect contributions to the analysis. Background subtracted counts were normalized, and a mean expression level was computed for each gene in every patient, creating a pseudo-bulk value for each gene. A z-score was calculated for each gene across patients, and hierarchical clustering (Fig. 1c) and PCA (Fig. 1d) were performed on the resulting z-scores for all genes with maximal normalized expression above 5 × 10−4. Spearman correlation distance and average linkage were used for hierarchical clustering. Comparison with NDD liver (Extended Data Fig. 1f) was performed on Visium data from Andrews et al.14. Pseudo-bulk was computed as the mean of spots annotated as hepatocytes and that further showed ALB expression above 0.003 of summed UMIs. GSEA29 was performed using version 3.0 with default parameters using the gene sets Hallmark and KEGG. Human portal–central axis spot annotation For Visium, we first extracted for each gene the spearman correlation coefficient between the gene’s expression level and CYP2E1 expression level over all spots. The 20 genes that were most correlated or anticorrelated with CYP2E1 over all spots were defined as our pericentral and periportal landmark gene sets (pericentral landmark gene set: CYP3A4, ADH1B, CYP1A2, CYP2E1, APOA2, APOC1, ADH4, ADH1A, APOH, AMBP, GSTA2, ADH1C, SLCO1B3, AOX1, APOA5, DCXR, RBP4, OAT, CYP2C19, GC; periportal landmark gene set: SERPINA1, APOA1, ALB, C7, NNMT, HAMP, ALDOB, ASS1, CYP2A7, MGP, A2M, FXYD2, CCL21, HAL, IGFBP2, SDS, AQP1, CYP2A6, FBLN1, PTGDS). Those landmark gene sets’ normalized summed expression levels enabled assigning a percentile-categorized zonation score to each spot, computed as the sum of periportal expression divided by the sum of periportal expression plus the sum of pericentral expression (sumofperiportalexpressionsumofperiportalexpression+sumofpericentralexpression). This zonation score was then binned into eight zones with equal percentiles, and a zone index was assigned to each spot. To correct for spurious annotations, we applied a median filter by re-assigning the zone of each spot as the median zone of all spots surrounding it (relative distance of under 150 µm; Fig. 2a). Subsequently, the zonation bias P value of each gene was computed using the Kruskal–Wallis test, followed by Benjamini–Hochberg correction for multiple hypotheses. Centre of mass was calculated for every orthologous gene per sample as the weighted average of the zones26,28. Gene zonation scores were calculated as the linear combination of the zonal layer with the highest gene expression and the gene’s centre of mass (Fig. 2). Gene zonation scores were non-defined for genes that exhibited low expression in our Visium datasets. For the MERFISH analysis, we used NOTCH3 as a periportal landmark gene and reconstructed cell zonation using the distance from a NOTCH3 positive cell. For the Visium HD analysis, we worked on the 8-μm spot size data. Non-parenchymal spots, identified histologically based on the H&E staining, were manually marked and excluded from the analyses, as were spots at the section periphery and spot with fewer than 50 UMIs. Zonation score was initially assigned to each spot using the same landmark gene method used for the Visium slides and binned into eight zones based on equal percentiles over the analysed spots. Subsequently, zone values were assigned the median zone of 100 μm ×100 μm areas. Spots were binned and averaged, yielding the Visium HD-reconstructed zonation profiles. Human hepatocyte marker gene identification To identify genes that serve as markers for hepatocytes, we utilized the single-cell atlas generated by Massalha et al.17. We defined marker genes as those with a normalized expression level higher than 10−5 and with a ratio of expression in hepatocytes to their maximum expression in all other cell types greater than 0.5. Since hepatocytes contain more than tenfold total mRNA than non-parenchymal liver cells, this value ensures that the majority of transcripts for such genes in each spot are likely to originate from hepatocytes46. We identified 1,724 hepatocyte marker genes (Figs. 2b,c and 3h). Cross-species zonation reconstruction To remove outlier spots, for each non-human sample spot with log10 of total UMI counts that are either higher or lower than 1.5 standard deviations from the mean log10 of total UMI counts across all spots were filtered out. For the non-human livers, each spot’s zone was determined based on landmark gene expression in cow and mouse26 and based on distances from the portal fibrotic septa in pigs and boars; Spots were then binned into 6 distinct zones, by equally proportioned percentiles and the assigned zone was median filtered using spots closer than 150 µm. Zonation profiles per species were computed as the weighted means of the zonation profiles of individual samples, weighted by the median number of UMIs per spot. When comparing zonation trends of specific gene subsets (Fig. 3h), the ratios between the expression levels in the most periportal and pericentral zones were computed. Only hepatocyte-specific genes that were highly expressed in at least one non-human species and in humans (maximum expression above 10−5 in each) were analysed. Cross-species orthologous genes analysis To allow for a common ground analysis, we sub-setted our data to genes that show 1:1 orthologue across all species (1 orthologue per gene per species). In total, 12,410 human, mouse, pig, boar and cattle orthologous genes were found using the Ensembl biomart tool55 (Supplementary Table 7) 1,249 hepatocyte-specific genes with orthologues across all species genes were found (Fig. 3h). Only human patients with steatosis level below 10% were used (n = 7, M2, M3, M4, M5, M6, M7 and M8). For the analysis of WNT-activated and WNT-repressed genes we utilized the data from a published WNT-hyperactivating liver-specific Apc-knockout model mice43 and used the 20 genes that were most WNT-upregulated and the 20 genes that were most WNT-downregulated. When comparing the zonal trends of NPCs (Fig. 4) we considered the subset of genes that were expressed in the respective cell type at a level above 5 × 10−5 and threefold higher levels than the maximal expression in any other cell type. For human liver we used the pseudo-bulk data of Massalha et al.17, for mouse we used the data of Ben-Moshe et al.62. Liver cell atlas using Visium HD Visium-HD was performed on liver tissue sections from two donors, M2 and M6. Single-nucleus segmentation was performed on the H&E image using the Stardist algorithm63 implemented within QuPath, utilizing the pre-trained H&E model he_heavy_augment. Segmented nuclei were expanded by up to 6 µm or until contacting neighbouring nuclei. Each segmented object was classified as either hepatocyte or NPC using a random forest object classifier trained on morphological features. The 2 µm × 2 µm Visium HD bins were next assigned to the segmented cells if their centroids resided within the cell area. Each cell received classification based on object classifier and the gene expression values that equal the sum of UMIs of the enclosed bins. For downstream single-cell analysis, cells with fewer than 200 UMIs were excluded. To generate a single-cell uniform manifold approximation and projection (UMAP), the data from both patients were combined. Gene expression was log-normalized, and highly variable genes (HVGs) were identified using thresholds of min_mean = 0.095 and min_disp = 0.5. z-scored HVG expression values were concatenated with a weighted classification binary output vector from the object classifier, which encoded the hepatocyte versus NPC label. PCA was applied, 25 components were selected and batch correction was performed using the Harmony algorithm (Harmonypy 0.0.10). The Harmony-corrected PCs were used to build the neighbourhood graph and UMAP visualization, using Scanpy (version 1.10.0) standard functions (n_neighbors = 20, resolution = 1.7). The Visium HD single-cell atlas was used to extract landmark genes for each major cell type: hepatocytes, LSECs, Kupffer cells and fibroblasts. To this end, the periportal and pericentral landmark genes from the Visium experiments were first used to assign a zonation score, which was then divided into three equal percentile bins. Next, the zonation profiles of each cell type were computed as the mean expression of all cells in the specific bin. The centres of mass were computed for each gene and sorted, and the top and bottom 20 genes with the most extreme centres of masses were selected as the cell-type-specific portal or central landmark gene set respectively. This calculation was performed over all genes for hepatocytes, and only for genes with maximal zonal expression above 10−4 and twofold higher expression in the respective cell type compared to the maximal expression in all other liver cell types, based on the human liver cell atlas of Massalha et al.17. Liver cell atlas using snRNA-seq snRNA-seq data were generated from liver tissues of four individuals (M5, M6, M7 and M8) and combined into a single AnnData object. Cells with mitochondrial transcript fractions exceeding 40% were excluded. To reduce ambient RNA background, we performed background subtraction by estimating gene-wise mean expression in low-complexity cells (150–400 total UMIs) and subtracting this background from all cells, followed by transforming negative values to zero. All downstream analysis was performed on log-transformed expression data. To retain high-quality cells, we selected those with total UMI counts between 500 and 50,000 and with up to 7,000 detected genes. HVGs were identified using thresholds of min_mean = 0.095, max_mean = 6, and min_disp = 0.5. To refine the annotation, we removed likely contaminant non-hepatocyte cells with ALB expression greater than 0.01. A second round of HVG selection, PCA, Harmony batch correction, UMAP and Leiden clustering was then performed using 50 principal components, 15 neighbours, a resolution of 1.0 and a UMAP minimum distance of 0.5. Only clusters with more than five cells in each of the four patients were retained. Final cell-type annotations were curated based on canonical marker gene expression. The lobule layer of the single nuclei was inferred based on the cell-type-specific landmark genes that were extracted from the Visium HD single-cell atlas. For each of the four main cell types, the periportal and pericentral landmark gene sets’ normalized summed expression levels enabled assigning a percentile-categorized zonation score to each cell, computed as the sum of periportal expression divided by the sum of periportal expression plus the sum of pericentral expression (sumofperiportalexpressionsumofperiportalexpression+sumofpericentralexpression). This zonation score was binned into eight zones with equal percentiles, and a zone index was assigned to each cell. The zonation profiles were computed as the means across all cells belonging to the same zone. To examine the zonation of liver cell types (Fig. 4c–g and Extended Data Fig. 9), cell-type-specific markers were extracted as genes with mean expression higher than 10−4 and 20-fold higher expression in the respective snRNA-seq atlas cluster compared to the maximal expression in the other clusters, with the exception of hepatocytes, for which twofold higher expression threshold was used. Ligand–receptor and transcription factor analysis The database of Ramilowsky et al.30 was used as reference for ligands and matching receptors. Spearman correlations were computed between the zonation profiles of all ligands and matching receptors in each of the 16 pairs of cell types in the snRNA-seq dataset (Extended Data Fig. 9, Supplementary Table 9). Transcription factors are based on ref. 35. Lipid droplet quantification Analysis of gene expression changes in early steatosis was performed on four out of the seven donors who exhibited steatosis, and for which image quality was sufficient to reliably extract lipid content values (M1, M2, M3 and P6). To extract the lipid percentage of each spot in the H&E attained Visium slide, an artificial neural network pixel classifier was trained on the H&E images using the QuPath software v.4.064. The pixel classifier was trained to identify three histologic features: fat droplet, non-fatty liver tissue and the slide background. For each slide, the pixel classifier was trained on multiple fatty droplets and non-fatty tissue from the same H&E image (Fig. 5a), with the following parameters: resolution: ‘very high’, channels: ‘Hematoxylin and Eosin’, scales: ‘1,2,4’, features: all 12 available features (the default is Gaussian). The spot locations were uploaded from the Visium output, and the QuPath measurement tools was used to quantify the number of fat droplet pixels and non-fatty liver tissue pixels in each spot (Fig. 5c). Comparison to mouse model of steato-hepatitis Visium spots in the mouse data from Zhou et al.38 were assigned to one of 8 lobule layers following the procedure described in ‘Human portal–central axis spot annotation’, using the mouse landmark genes from Droin et al.27. For each mouse, pericentral expression was defined as the mean expression in the two most pericentral layers. The log2 of the ratio between the mean pericentral expression in the steatotic and non-steatotic mice was computed for each gene, and correlated with the log2 of the respective ratio between the pericentral spot expression in each of the steatotic human liver samples and the mean of the pericentral spot expression in all non-steatotic human samples. Ethics statement All work was conducted in compliance with the principles outlined in the Declaration of Helsinki. Human experiments were approved by the Sheba Medical Center Helsinki committee (Helsinki approval number 866521SMC) and by the Mayo Center Institutional Review Board number 22-008260. Informed consent was obtained from all participants. Mouse experiments were approved by the Institutional Animal Care and Use Committee of the Weizmann Institute of Science (approval number 05220824–2). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-026-10377-y. Supplementary information Supplementary Tables 1–11 Reporting Summary Peer Review file Source data Source Data Fig. 1 Source Data Fig. 3 Source Data Fig. 4 Source Data Fig. 5 Source Data Extended Data Fig. 1 Source Data Extended Data Fig. 2 Source Data Extended Data Fig. 3 Source Data Extended Data Fig. 5 Source Data Extended Data Fig. 6 Source Data Extended Data Fig. 10 Acknowledgements The authors thank R. Lapid, R. King, G. Kahila Bar-Gal, B. Rotblat, M. Zachut and J. Shpirer for their help in sourcing the non-human samples. S.I. is supported by the Helen and Martin Kimmel Award for Innovative Investigation, the Yad Abraham Research Center for Cancer Diagnostics and Therapy, the Moross Integrated Cancer Center, the Minerva Stiftung grant, a Weizmann-Sheba grant, the Israel Science Foundation grants no. 908/21 and 3663/21, the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme grant no. 768956 and a grant from the Ministry of Innovation, Science and Technology, Israel. Y.K.K. is supported by the JSMF Postdoctoral Fellowship in Understanding Dynamic and Multi-scale Systems. Extended data figures and tables Author contributions O.Y., K.B.H. and S.I. conceived the study. I.N., T.T., N.P. and R.P. operated on the patients and provided samples. O.Y., K.B.H., T.B. and Y.H. collected and processed human samples. A.A. and O.Y. collected and processed non-human samples. S.I., O.Y., A.E., A.A., R.N., S.S. and Y.K.K. performed computational analysis. Visium experiments were performed under the guidance of M.K., H.K.-S., Y.H. and I.G. MERFISH experiments were performed under the guidance of L.F.-A. and D.H. O.Y. and R.H. performed the PhenoCycler experiments. R.N. and L.F.-A. performed immunofluorescence experiments. K.B.H. performed the snRNA-seq experiments. L.P., O.G. and Y.A. contributed to microscopy imaging and pixel classification development. C.M. contributed to the histology slide evaluation. O.Y. and S.I. wrote the manuscript. All the authors discussed the results and commented on the manuscript. Peer review Peer review information Nature thanks Kok Hao Chen, Bruce Wang who co-reviewed with Elise Lelou, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer review reports are available. Data availability Porto–central gene expression zonation profiles and gene expression changes along the steatosis gradient can be browsed using our web app (https://shalevapps.weizmann.ac.il/liver_app/). Spatial transcriptomics data of the human and non-human samples that were generated in this study can be downloaded from Zenodo, containing preprocessed 10X raw count tables, spot barcodes, tissue images and loupe browser files of all the human and non-human samples (https://zenodo.org/records/17735506 (ref. 65)). PhenoCycler data can be found at https://zenodo.org/records/17735558 (ref. 66). Matlab datasets required for GitHub codes can be found at https://zenodo.org/records/17735587 (ref. 67). Mouse Visium data was downloaded from the Gene Expression Omnibus (GEO) under accession GSE248077. Adjacent normal liver Visium data used in Extended Data Fig. 1f was downloaded from GEO accession GSE223559. NDD data used in Extended Data Fig. 1f was downloaded from GEO accession GSE240429. Source data are provided with this paper. Code availability Custom codes were central to the conclusions of the paper. The code is available on GitHub at https://github.com/OranYak/Human-liver. Matlab codes use some of the data supplied in the Zenodo repository https://zenodo.org/records/17735587 (ref. 67). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data is available for this paper at 10.1038/s41586-026-10377-y. Supplementary information The online version contains supplementary material available at 10.1038/s41586-026-10377-y. References Associated Data Supplementary Materials Supplementary Tables 1–11 Reporting Summary Peer Review file Source Data Fig. 1 Source Data Fig. 3 Source Data Fig. 4 Source Data Fig. 5 Source Data Extended Data Fig. 1 Source Data Extended Data Fig. 2 Source Data Extended Data Fig. 3 Source Data Extended Data Fig. 5 Source Data Extended Data Fig. 6 Source Data Extended Data Fig. 10 Data Availability Statement Porto–central gene expression zonation profiles and gene expression changes along the steatosis gradient can be browsed using our web app (https://shalevapps.weizmann.ac.il/liver_app/). Spatial transcriptomics data of the human and non-human samples that were generated in this study can be downloaded from Zenodo, containing preprocessed 10X raw count tables, spot barcodes, tissue images and loupe browser files of all the human and non-human samples (https://zenodo.org/records/17735506 (ref. 65)). PhenoCycler data can be found at https://zenodo.org/records/17735558 (ref. 66). Matlab datasets required for GitHub codes can be found at https://zenodo.org/records/17735587 (ref. 67). Mouse Visium data was downloaded from the Gene Expression Omnibus (GEO) under accession GSE248077. Adjacent normal liver Visium data used in Extended Data Fig. 1f was downloaded from GEO accession GSE223559. NDD data used in Extended Data Fig. 1f was downloaded from GEO accession GSE240429. Source data are provided with this paper. Custom codes were central to the conclusions of the paper. The code is available on GitHub at https://github.com/OranYak/Human-liver. Matlab codes use some of the data supplied in the Zenodo repository https://zenodo.org/records/17735587 (ref. 67).