The transcriptome of hepatic fibrosis revealed by single cell RNA sequencing

Signaling programs of hepatic disease have historically been studied by comparing bulk diseased and healthy liver tissue. As tissues contain heterogeneous cell mixtures, signals from less represented populations can be dampened however single cell RNA sequencing (scRNAseq) circumvents this. In revealing the transcriptomes of individual cells, scRNAseq has exposed the complex division of labor in normal hepatic homeostasis (1) and led to conceptualization of the “fibrotic niche,” (2) regions of scarring composed of scar-associated cells, extracellular matrix and fibrotic programs. In a recent issue of Nature, Ramachandran and colleagues use scRNAseq to resolve the cellular transcriptomes of healthy and cirrhotic human livers, revealing previously unidentified cell types, cell-cell interactions and fibrotic programs which comprise the human liver fibrotic niche (3). Additionally, in a parallel Cell Reports paper, the same group has verified hepatic stellate cell (HSC) subpopulation zonation within the murine fibrotic niche using scRNAseq (4).

scRNAseq has emerged as a powerful methodology to link genotype with phenotype and reveal transcriptional culprits of diseases. Starting from a range of platforms such as smaller scale laser capture microdissection to larger scale flow activated cell sorting (FACS) and microfluidics, RNAseq applied to individual cells can be used to identify cellular subpopulations, unfold the history of a cell’s transcriptome (pseudotime reconstruction) and reconstruct gene regulatory networks (5). When utilizing full transcripts (as opposed to 5’ or 3’ tails), scRNAseq can demonstrate splice variants, RNA edits and rare RNA isoforms (e.g. circular RNA), as well as detect transcriptome changes secondary to allelic expression, revealing downstream impacts of single nucleotide polymorphisms and mutations (5).

Ramachandran and colleagues employ scRNAseq to report a single cell atlas of non-parenchymal cells (NPC) from 5 healthy and 5 cirrhotic human livers. Utilizing FACS, NPCs were separated into CD45+ (resident/non-resident leukocytes) and CD45- (non-leukocyte) fractions. Over 100,000 cells (66,135 deemed liver resident) were sequenced with microfluidics-based scRNAseq, clustered using principal component analysis and identified using known ontogeny markers. This analysis revealed 21 cell populations from 10 lineages, all of which were represented in healthy and cirrhotic livers. Subpopulations expanded within the cirrhotic liver were identified after comparative analysis of cirrhotic and control livers (3).

Using topographical in silico trajectory and in vitro phenotype analyses, Ramachandran and colleagues reveal that three cell subpopulations, two previously unrecognized, interact spatiotemporally to drive fibrosis within the human liver’s fibrotic niche. Endothelial cells expressing ACKR1 (a regulator of chemokine bioavailability which promotes leukocyte recruitment) and PLVAP (a membrane protein which structurally regulates microvascular permeability) were significantly expanded within the fibrotic niche when compared to healthy livers. Also in the niche were circulatory-derived CD9+ macrophages expressing TREM2, a cell surface receptor most studied in neurodegeneration which yields a host of downstream immune responses. Finally, collagen-producing Mes(3) cells expressing PDGFRA (a tyrosine kinase receptor increasingly implicated in the profibrotic response of HSCs) were expanded in cirrhotic livers and found to be derived from cluster Mes(2). The authors performed pseudotemporal ordering and RNA velocity analysis to determine that the Mes(2) cell subpopulation, which is present in cirrhotic and healthy livers, gives rise to Mes(3) cells and offer this as the human analog to the well-studied process of murine HSC to myofibroblast transdifferentiation. Ligand-receptor modeling using the signaling repository CellPhoneDB revealed an in silico interactome validated with in vitro studies. Scar-associated endothelial cells signaling via non-canonical Notch and macrophages via PDGFRA promoted Mes(3) collagen production and proliferation, respectively (3). In a complementary study, this group further orders the topographical distribution of HSCs into subpopulations within the fibrotic niche of the carbon tetrachloride murine model of centrilobular fibrosis using scRNAseq, finding central vein-associated HSCs (expressing ADAMTSL2 and RSPO3) produce more collagen than portal vein-associated HSCs (expressing NGFR and ITGB3) (4). Whether portal vein-associated HSCs produce the majority of collagen in periportal models of liver injury remains to be determined.

A major strength of Ramachandran et al. was the comprehensive approach tackling temporal and cross-species challenges in regards to the discovery of TREM2+CD9+ scar-associated macrophages. As cirrhosis represents the end-stage of longstanding programs, the ephemeral programs regulating a cell’s transition from quiescence to pro-fibrotic cannot be directly inferred from study of healthy and cirrhotic livers. The authors verified scar-associated TREM2+ macrophage expansion was also present in non-cirrhotic livers along a continuum of disease severity in non-alcohol related steatohepatitis (NASH). Likewise, Xiong et al. (6) solidify the importance of TREM+CD9+ cells in NASH in their recent scRNAseq study of healthy and NASH livers. Both of these studies found TREM2+ macrophages have a molecular signature intermediate to Kupffer cells and monocytes however work remains to determine the exact lineage origin of these unique disease-associated cells.

Limitations of many recent scRNAseq studies, including Ramachandran et al., stem from the relative infancy of the field. The power to resolve a single cell’s transcriptome stands in stark comparison to the heterogeneity in methodology and noise created in the process. No gold standard exists for how to most effectively harness this technology. Thus, as with evolving methodologies inherent to technological revolution, the field is faced with challenges. Given variability in cell isolation techniques, scRNAseq library preparation and downstream bioinformatics platforms, there may be loss of rare but relevant signals. Indeed, researchers must acknowledge the error of misinterpreting technical dropout (RNA not detected by the sequencer due to technical reasons) as biological dropout (the gene is not expressed) and heed the risk of “dimension reduction” during data analysis which in turn may shrink signals of “important biological information” (5). As Ramachandran and colleagues chose to exclude cells expressing <300 genes and genes expressed in < 3 cells from their analysis, one wonders how inclusion of these data would shape their conclusions. On the contrary, amplification of technical noise and stochastic transcription may lead researchers astray to study inflated, less significant pathways (5).

The field urgently needs a universal approach to sort, denoise and interpret the big data generated by scRNAseq studies and to place these data within the framework of the molecular signature of the cell. Artificial intelligence (AI), i.e. deep learning neural networks, have emerged including BERMUDA (PMID: 31405383), DCA (deep count autoencoder) (PMID: 30674886), SAUCIE (PMID: 31591579) and VPAC (PMID: 31074382) with goals to provide batch effect correction, reduce noise and/or improve clustering for subpopulation identification within and between scRNAseq datasets. The more input available to AI, the more accurate the outcome. As such, it is vital for developers of machine learning to collaborate and for researchers to share exact methodologies and raw scRNAseq data. Lastly, as RNA represents one component of a cell’s molecular profile, integration of scRNAseq data with proteome and epigenome data yielding a multimodal single cell profile as reviewed (7) will bring additional challenges and opportunities for investigators.

In summary, Ramachandran and colleagues validate the longstanding importance of transdifferentiated HSCs as central mediators of hepatic fibrosis and present circulation-derived TREM2+ CD9+ macrophages as key regulators of the fibrotic niche now validated in a second human scRNAseq study (6). The resolution of the fibrotic niche is possible by studying hepatic injury in context of the whole organ, including its connections to circulatory and immune systems. This begs the question – should the frontier of disease study encompass the entire organism (e.g. whole human scRNAseq)? As the Tabula Muris Consortium reports an atlas with scRNAseq data from 20 adult mouse organs (8), a whole human atlas is conceivable. Within the field of liver disease, Ramachandran and colleagues bring us closer to understanding the mechanism of cirrhosis with elucidation of the hepatic fibrotic niche, an area where drug development can be strategized. With the convergence of AI, scRNAseq and multimodal single cell profiling among collaborative researchers, the future will bring a new cellular understanding of liver disease with higher resolution than ever.

Abbreviations:

ACKR1

atypical chemokine receptor 1

AI

artificial intelligence

FACS

flow activated cell sorting

HSC

hepatic stellate cell

NASH

non-alcohol related steatohepatitis

NPC

non-parenchymal cell

PDGFRA

platelet-derived growth factor receptor alpha

PLVAP

plasmalemma vesicle-associated protein

scRNAseq

single cell RNA sequencing

TREM2

triggering receptor expressed on myeloid cells 2