An autocrine signaling circuit in hepatic stellate cells underlies advanced fibrosis in nonalcoholic steatohepatitis.

Study design

In this study we leveraged recent developments in single nucleus RNA-seq and tissue clearing techniques to uncover signaling circuits that emerge in NASH that can serve as therapeutic targets for advanced fibrosis. We profiled HSC transcriptomes from both human and mouse NASH livers and identified a conserved autocrine signaling circuit that consists of 68 ligand-receptor pairs. Targeting one of these HSC autocrine receptors, NTRK3, reduced HSC activation in culture and reversed advanced fibrosis in vivo in the FAT-NASH model. All results in this study have been reproduced in independent cohorts of mice and in biological replicates for cell culture. For quantifications that can be subjected to human bias, samples were randomly coded and quantification was done in a blinded fashion. Use of human tissues for single nuclei RNA-seq was IRB-exempt in accordance with guidelines of the Mount Sinai Institutional Review Board, as there were no patient identifiers and the tissue was otherwise intended to be discarded.

Patient and clinical specimens

Samples collected at Mount Sinai Hospital, New York, were freshly resected liver samples collected at time of surgery with informed consent. Samples collected at University of California, San Diego were deidentified livers declined for transplantation obtained via Lifesharing OPO. Samples collected at Gordian Biotechnology were livers not suitable for transplantation from research consented donors were perfused and transported on ice to Gordian Biotechnology from regional hospitals after transplantable organ removal. Samples sections suitably sized for snRNA-seq (~4x4x4mm) were removed from at least 5mm below the outer surface of the liver, snap frozen in liquid nitrogen, and stored at −80°C until nuclei extraction.

Mice

The animal protocol was approved by the Institutional Animal Care and Use Committee (IACUC) at the Icahn School of Medicine at Mount Sinai, NY (IACUC-2018-0060). 6-week-old male and female C57BL/6J mice (housed separately) were purchased from Jackson Laboratories. Five mice per cage were housed in a Helicobacter-free room for 12 hours light - 12 hours dark cycle and weighed once weekly. In this study we used the FAT-NASH rodent model of human NASH, which is a model that we have extensively used and validated ( 12 ). This model reproducibly recapitulates the stages of progression of human NASH that include advanced fibrosis and tumors (hence the term FAT – NASH for “Fibrosis And Tumors”), along with changes in the microbiome ( 13 ). Carbon tetrachloride (CCl 4 ) was purchased from Sigma-Aldrich. CCl 4 was freshly dissolved in corn oil at a final concentration of 5% before injection. The final dose of pure CCl 4 was 0.2 μL/g of body weight of mice, delivered intraperitoneally once/week starting from initiation of the western diet/sugar water feeding and continued for a total period of 6, 12, or 24 weeks. Western diet containing 21.2% fat (42% Kcal), 41% sucrose and 1.25% cholesterol by weight was purchased from Envigo (Teklad Custom diet). Sugar water solution contained 18.9 g/L D-(+)-Glucose (Sigma-Aldrich) and 23.1 g/L D-(−)-Fructose (Sigma-Aldrich) dissolved in autoclaved water and filter sterilized. The diet and sugar water were replaced once weekly. For LOXO-195 in vivo studies, LOXO-195 (TargetMol) was first dissolved in DMSO to 41.6mg/mL, before corn oil was added to a final LOXO-195 concentration of 12.5mg/mL, 100μL of this drug solution or vehicle control was given to a 25g mouse to achieve a 50mg/kg dosage (weight adjusted daily). The dosing regimen was oral gavage twice a day (separated by 8 hours), 5 days a week for a total of 4 consecutive weeks. All mice were sacrificed at 72 hours after the last dose of LOXO-195 or vehicle control. At termination of study pieces of liver (unless used for tissue clearing), were formalin fixed for FFPE and directly frozen in Tissue-Tek OCT compound (ThermoFisher Scientific catalog no. 4583) over dry-ice then −80°C for subsequent immunohistochemistry. The remaining mouse livers were chopped into ~2 cm 3 pieces, flash frozen in liquid nitrogen, and stored at −80° C until nuclei extraction. Tissue clearing, staining, confocal imaging, and 3D image reconstruction Mice were anesthetized and perfused through the portal vein with 20 mL of PBS followed by 20 mL of 4% PFA/PBS, liver tissues were trimmed to ~2 mm 3 , cleared, and stained as described in the iDISCO protocol ( 18 ). A detailed protocol that accompanies the original iDISCO publication along with an updated list of validated antibodies for this method can be found at https://idisco.info/ . Antibodies used in this study were DESMIN (Abcam catalog no. ab15200, 1:400 dilution) and CD31 (R&D Systems catalog no. #AF3628, 1:400 dilution). Propidium iodide (Sigma catalog no. #P4170, 0.02 mg/mL final concentration) was used to stain DNA. Imaging was carried out on the Leica SP5 DMI confocal microscope and subsequently reconstructed using the IMARIS software at the Sinai Microscope CoRE. The IMARIS “surface” tool was used for segmenting the DESMIN staining using Surface Grain Size of 0.303 μm, diameter of largest sphere of 1.14 μm, and manual Threshold range from 13.4 to 168.4. The IMARIS “spot” tool was used to segment propidium iodide DNA staining into nuclei by setting the diameter of spots to 5μm, and these nuclei were then manually curated as belonging to HSC if surrounded by DESMIN signal in the overlay channel.

Generation of LX2-Cas9 cells and cell culture

LX-2 cells are a widely used, immortalized human HSC line previously characterized in our laboratory ( 29 ). Stable Cas9-expressing LX-2 cells were generated through lentiviral transduction of a Cas9-expressing plasmid (lentiCas9-Blast, Addgene 52962-LV). Cells were then selected based on blasticidin resistance, and homogenous Cas9-expressing LX-2 cells were isolated by single cell dilution cloning (available as LX-2 Cas9 line from Millipore-Sigima, SCC613). Scratch assay was performed in a 24-well plate by using a P200 pipet tip to scrape the LX-2 cell monolayer in a straight line. Debris were removed from the edge of the scratch by washing and replacing with fresh media. To obtain consistent field of view in imaging, markings were made on the plate using a fine tip marker. Plates were then imaged immediately after the scratch with a microscope using the phase-contrast objective at 10x magnification. Cells were returned to incubation and imaged again 24, 48, and 72 hours after initial scratch. Images acquired were analyzed quantitatively by using Photoshop. Images obtained at each time point and treatment condition were visually overlayed on top of each other and cropped to the same size. The perimeters and areas of the scratches were then delineated and quantified. For all cell culture assays, LX-2 cells were serum-starved overnight to quiesce and synchronize metabolic activity in serum-free DMEM (Thermo Fisher Scientific) supplemented with 0.1% BSA, without antibiotics at 37°C. All experiments were repeated independently 3 times, each performed in triplicates.

Immunoblotting

Flash frozen livers were homogenized using TissueLyser (Qiagen) in RIPA buffer containing protease inhibitors (Thermoscientific Pierce complete protease and phosphatase inhibitors (ThermoFisher, A32963 and 78420), 1 tablet/10 ml RIPA buffer). Bradford protein quantification was performed (Biorad, 5000006) for liver lysates. LX2 cells were directly collected in 2x Laemmli buffer (Biorad, 1610737). 50 μg liver protein or 10μL of LX2 lysate were analyzed by SDS-PAGE and immunoblotting with antibodies to hNTRK3 (Biotechne, AF373), mNTRK3 (Biotechne AF1404), αSMA (Abcam, ab5694), pERK (Cell Signaling, 4370S) and CALNEXIN (Abcam, ab75801). Picrosirius red staining and artificial intelligence-based collagen quantification in liver sections Liver was fixed in 10% formalin buffer and paraffin embedded liver tissues were sectioned using a 4 μm microtome. Slides containing tissue sections were baked at 60°C for 1 hour and rehydrated through xylene followed by graded ethanol (100%, 95%, 85% and 70%) into distilled water and processed for picrosirius red/Fast green staining. For collagen staining, rehydrated slides were stained for one hour in saturated picric acid with 0.1% Sirius Red (Direct Red-80; Sigma-Aldrich) followed by counterstain with 0.01% Fast Green (Sigma-Aldrich) for another hour. The slides were removed from the stain, rinsed in water, and rapidly dehydrated through graded ethanol (70%, 85%, 95%, and 100%) followed by xylene and finally placed on cover slips in Permount (ThermoFisher). Whole slides with stained sections were digitally scanned in an Aperio AT2 digital scanner (Leica Biosystems Inc.) at 40X (0.221 micron per pixel). FibroNest was used for the quantification of Sirius red-positive collagen fibers from each image, where Sirus-red positive pixels are used to detect collagen. The whole-tissue fibrosis phenotype was described for its collagen content and structure (12 traits), the morphometric traits of the segmented collagen fibers ( 12 ), and fibrosis architecture traits ( 7 ). The collective distribution of each trait were quantified with 7 statistical parameters (quantitative Fibrosis Traits, qFTs) to account for severity, progression, distortion, and variance for both fine and assembled collagens, resulting in a total of 315 qFTs. Principal qFTs were automatically detected using the mouse FAT-NASH fibrosis progression cohort consisting of 15 liver tissues (Chow mice n=5, 6 weeks FAT-NASH n=2, 12 weeks FAT-NASH mice n=5, 24 weeks FAT-NASH mice n=5, Fig. 3D , E )). Automatically selected principal qFTs were normalized and combined to form a continuous composite score for fibrosis severity in each phenotypic layer (collagen content, fibers morphometry, fibrosis architecture) and for the whole fibrosis phenotype (Ph-CFS). In addition, Collagen fibers were classified as “fine” or “assembled” based on the complexity of their skeleton, and the morphometric phenotypes can be quantified for each subgroup. The relative changes of each individual qFTs between animals and groups were visualized in the form of a heatmap.

Single nucleus RNA-seq and data processing

For snRNA-seq of mouse samples and human samples collected at Mount Sinai (CTRL 1-3, NASH 6-9), 40-60mg of total liver tissue was chopped on ice in 1mL of 0.03% TST buffer (146mM NaCl, 10mM Tris-HCl pH 7.5, 1mM CaCl 2 , 21mM MgCl 2 with freshly added 0.03% Tween-20, 0.01%BSA, and 0.4U/mL Ribolock RNase Inhibitor (Thermo Fisher Scientific FEREO0382) ( 30 ) with Tungsten Noyes Spring Scissors (FST 15514-12). The resulting nuclei suspension was then filtered through 40μm cell strainers (Thermo Fisher Scientific 352340) into a fresh 50mL conical tube on ice. 1 more mL of 0.03% TST was used to rinse the cell strainer and 3mL of ST buffer was added to the nuclei suspension followed by briefly mixing by gentle flicking. The nuclei suspensions were spun at 500xg for 5 minutes at 4°C in a swing-bucket centrifuge. The final nuclei pellet was suspended in 200μL of 0.4% BSA in PBS by pipetting 30 times with a regular p1000 tip. SnRNAseq datasets from 2 control mouse livers (“Chow_1” and “Chow_2) can be found at GEO accession # GSE212327 . For “NASH” and “NASH_HCC” samples, 20mg of liver from 3 FAT-NASH mice were pooled for each sample and processed together in an attempt to average out heterogeneity associated with the diseased liver. For snRNA-seq of human samples collected at Gordian (NASH 1-5), frozen human liver tissue (~60 mm 3 ) was lysed by Dounce homogenization (Kimble 2mL: 20 times with Pestle A over ~60 seconds) in 2mL lysis buffer (10mM Tris pH 7.0, 10mM NaCl, 3mM MgCl 2 , 0.05% Triton X-100, 0.13% RNAse Inhibitor (Enzymatics Y9240L), 0.25% Superase RNAse Inhibitor (Thermo Fisher Scientific AM2694) with 45μM Actinomycin D, then dripped through a 40μm cell strainer (Celltreat 229481) into a protein low-binding 15mL centrifuge tube and the filter was washed into the collection tube with another 2mL of lysis buffer. After 5 minutes total exposure time to lysis buffer, samples were spun for 2 minutes at 300xg at 4°C; the supernatant was aspirated from the top first to remove fat, then the nuclei-containing pellet was resuspended in 5mL of wash buffer (DPBS + 2% BSA + 0.13% Enzymatics RNAse Inhibitor + 0.25% Superase RNAse Inhibitor + 3μM Actinomycin D) and filtered again through a 35μm strainer into a 5mL FACS tube. Nuclei were sorted into a 1.7mL a protein low-binding microcentrifuge tube pre-coated with 900μL of wash buffer on a Biorad S3e FACS using FSC/SSC > FSC-H vs FSC-W > FL4 vs. SSC gating to eliminate debris. Nuclei were pelleted at 300xg for 3 minutes at 4°C in a swing-bucket centrifuge and resuspended in wash buffer for 10x capture using an LT200 pipettor. Nuclei preparations were processed by the Chromium 3′ Gene Expression V2 Kit (for mouse samples) or Chromium 3′ Gene Expression V3 Kit (for human samples) according to the manufacturer’s guidelines. Qubit 3 (Fisher Scientific) and 2100 Bioanalyzer (Agilent Technologies) were used for quality check of cDNA. Libraries were sequenced on the NovaSeq at Sinai or NextSeq 550 at Gordian. To generate a count matrix, the sequenced files from each independent sample were processed through 10X Genomics Cell Ranger software v6. The raw base call files were demultiplexed using Cell Ranger mkfastq pipeline to generate FASTQ files. Cellranger count pipeline was applied to the FASTQs to perform alignment against modified transcriptomes based on the mm10 and GRCh38 reference builds for mice and humans respectively that contain introns to increase mapping efficiency and the number of genes detected. Filtered feature-barcode matrices from Cell Ranger were subsequently run through a standard Seurat pipeline for quality control. Different QC parameters were used for snRNA-seq data generated using Chromium 3′ Gene Expression V2 (mouse samples) and V3 (human samples) Kits using Seurat. For snRNA-seq generated from mouse samples, we included nuclei that fit the following criteria: 1) total number of expressed genes greater than 300 and less than 6000, 2) < 5% of which annotated as mitochondrial, and 3) total counts greater than 500 and less than 15000. Bulk RNA-seq and data processing Bulk RNA-seq from LX2 cells was carried out by a commercial vendor (Novogene). Briefly, RNA-seq libraries from LX-2 cells were prepared with polyA capture. RNA-seq libraries were prepared according to the Illumina NovaSeq RNA sample preparation protocol. RNA pooled from 2 wells on 24-well plates was used to generate libraries, which were analyzed on an Agilent 2100 Bioanalyzer. 150 base-pair paired-end reads that passed quality control were aligned to reference genome GRCh38. After alignment, read counts were generated and analyzed using the DESeq2 for differential gene expression. Three biological replicates were adopted for all experiments. Dimensionality reduction, clustering, and visualization For mouse snRNA-seq datasets we normalized and transformed the filtered count matrix as per standard Seurat pipeline. We performed dimension reduction using the top 2000 highly variable genes, performed PCA on the top 30 principle components and clustered using Louvain community detection algorithm with resolution of 0.3. Mouse samples from different conditions were merged using SeuratWrappers/LIGER into a single dataset for downstream data analysis. Processing of human snRNA-seq datasets are shown in accompanying scripts from Jupyter notebook. We identified cluster-specific gene expression by differential gene expression of nuclei in the cluster versus all other nuclei using Wilcoxon rank sum test and manually assigned cell types based on top differentially expressed genes. We visualized the reduced dimensionality data using the same principal components as previously used for clustering and UMAP. For Gene Ontology analysis, differentially expressed genes between cell clusters (Wilcoxon rank sum test, Padj. < 0.05) were input into AmiGO2 web server ( http://amigo.geneontology.org/amigo ) and all significantly enriched “PANTHER GO-Slim Biological Processes)” pathways ( FDR < 0.05) with more than 1 gene reported. Predicting and visualizing cell-cell interactions Filtered and normalized feature-barcode matrices of mouse and human snRNA-seq datasets along with cell type annotation files were exported from Seurat/R and input into CellphoneDB/Python ( 31 ). Mouse gene names were converted to orthologous human gene names based on the Ensembl database, only genes with unique human orthologues were kept for downstream analysis. Default in put parameters were used for statistical analysis. For mouse datasets, significant autocrine receptor-ligand interactions (pvalue < 0.05) for all cell clusters are shown in Data file S4 . For human datasets, significant autocrine ligand-receptor interactions (pvalue < 0.05) for all cell clusters from CTRL or patients with NASH are shown in Data file S5 , respectively. Cytoscape was used visualize the number of ligand-receptor interactions between cell types, with the thickness of the connecting line directly proportional to the number of significant interactions between each pair of interacting cell types.

Statistical analysis

Results are shown as means ± standard deviation unless indicated otherwise. Statistical analysis was performed using Prism unless otherwise specified. Student’s t-test is used when comparing two groups with similar variance (F-test performed to compare variance, P > 0.05). Analysis involving multiple groups was performed using one-way or two-way ANOVA, followed by post hoc t test.

Supplementary Material Data file S1 Table S1. NASH patient information Data file S2 Table S2.

Differentially expressed genes in human NASH stellate cells

Data file S3 Table S3.

Differentially expressed genes in mouse NASH stellate cells

Data file S4 Table S4.

CellphoneDB predicted significant receptor-ligand interactions from CTRL mouse snRNA-seq data

Data file S5 Table S5. CellphoneDB predicted significant receptor-ligand interactions from 24 weeks FAT-NASH mouse snRNA-seq data Data file S6 Table S6.

CellphoneDB predicted significant receptor-ligand interactions from CTRL patient snRNA-seq data

Data file S7 Table S7. CellphoneDB predicted significant receptor-ligand interactions from NASH patient snRNA-seq data video S1 Movie S1. HSCs and endothelial cells in iDISCO cleared control livers imaged using confocal microscopy video S2 Movie S2. HSCs and endothelial cells in iDISCO cleared control livers imaged using lightsheet microscopy video S3 Movie S3. HSCs in iDISCO cleared chow livers imaged using confocal microscopy video S4 Movie S4. HSCs in iDISCO cleared 24 weeks FAT-NASH livers imaged using confocal microscopy Fig. S1 - Fig. S6 Fig. S1. Additional analyses from CTRL and NASH patient snRNAseq data Fig. S2. Additional analyses from CTRL and NASH mouse snRNAseq data Fig. S3 Overview of the liver clearing and 3D imaging pipeline Fig. S4. Effect of NTRK3 knockdown or pharmacological inhibition on LX2 cell viability or proliferation Fig. S5. Additional data from LOXO-195 treated FAT-NASH mice Fig. S6. Liver clearing and 3D imaging of DESMIN staining in vehicle or LOXO-195 treated FAT-NASH mice

Data and materials availability: All data are available in the main text or the supplementary materials . Jupyter notebooks and scripts for processing and integrating the human patient data are available on zenodo at DOI 10.5281/zenodo.7348604.are available a All count matrices and metadata for each sample are publicly available in the Gene Expression Omnibus ( http://www.ncbi.nlm.nih.gov/geo/ ) under data accession no. GSE212837 .