Finding individual gene controllers that may trigger cell and tissue renewal

Our underlying hypothesis is that perturbing a single TF can lead to downstream changes which shift a cell state from old/diseased back towards young/healthy. To develop our platform and test this approach, we utilized the classic cell aging model—passaged human neonatal dermal skin fibroblasts—to measure hallmarks of cellular replicative aging, wherein early, middle, and late passages reflect some aspects of young, middle-aged, and old cells (12). The first step of TRDP (Fig. 1A) is to perform global transcriptomics of the two cell states. Then, bioinformatics tools (Materials and Methods) are used to predict the genes regulated by each TF based on a combination of known TF binding motifs and their occurrence proximal to the promoters of each gene. TRDP compiles a list of TFs' target genes and identifies which were most significantly differentially expressed between the cell states. From this, TRDP outputs a ranked list of TFs which likely regulate the most gene expression differences between the two cell states. Next, we follow up on these top TF perturbations in vitro and in vivo with single cell RNA sequencing and phenotyping. To systematically perturb TFs in vitro, we used Perturb-seq (13, 14), which uses CRISPRa (CRA) for activation of a target gene or CRISPRi (CRI) for inhibition of a target gene in cells followed by single-cell RNA sequencing. Then, for further validation of these genes in vitro and in vivo, other tools can be used for genetic perturbations, such as an AAV virus delivery of an overexpression plasmid for in vivo expression. The final output of TRDP is a TF perturbation, or collection of perturbations, which lead to rejuvenation of an older cell or tissue back towards a younger state, as defined by gene expression and cellular phenotypes.

Through the computational pipeline of TRDP, we selected the top 200 candidate TFs to perturb in late passage fibroblasts, to try identifying the perturbations that reverse the gene expression in late passaged cells to that of early passaged cells (list of TFs in Dataset S1↗). Based on phenotyping we performed (data in Fig. 2), early passage cells were 0 to 20 population doublings (PD), middle passage cells were 21 to 30 PD, and late passage cells were over 31 PDs. We performed Perturb-seq screening of these top 200 TFs using CRISPRa and CRISPRi, followed by differential gene expression analysis in comparison to that of the baseline states (early vs late passage). Nontargeting sgRNA were used as controls and are noted as either CRA NT (for CRISPRa cell line with nontargeting sgRNA) or CRI NT (for CRISPRi cell line with nontargeting sgRNA). We ranked TF perturbations by those with the greatest global reversal of gene expression toward early passage cell state. We quantify the degree of reversal by R_rej, defined as the correlation between log fold change of gene expression of late passage cells vs. early passage cells and log fold change of late passage cells with a TF perturbation vs. late passage cells with a nontargeting control. TF perturbations with a significant negative R_rej indicated the perturbation reversed the gene expression changes due to replicative aging. We then ranked R_rej and selected the top hits (TFs with r =< −0.3 Table 1).

To have a global view of the downstream transcriptional response to the TF perturbations, we performed TF module analysis using the Scenic algorithm (17). TF modules are defined as groups of genes regulated by a TF, based on their covariation with the TF across single cells and the occurrence of the TF binding motif in their promoters (17) (Materials and Methods). We found that the top hit TF perturbations resulted in similar downstream gene expression changes at the level of TF modules, when viewed from a subset of the modules differentially expressed between late passage (PD32) and early passage (PD14) cells (Fig. 1B), even though the factors themselves originated from diverse upstream pathways. We observed similarities in the gene expression changes in these top TF perturbations as in other aging and rejuvenation models, including in young patients' human skin fibroblasts (15), various tissues from young mice (e.g., intestine, pancreas, marrow, and thymus) (16), and various tissues from old mice rejuvenated through heterochronic parabiosis (e.g., intestine, pancreas, marrow, lung, liver, and brain) (16) (Fig. 1C). Cell types in rejuvenated aged mice that exhibit this transcriptional signature are enriched for mitotically active cells, such as large intestine crypt stem cells, pancreas ductal cells, lung fibroblasts, and hematopoietic stem cells. These observations suggest that there could be a common set of gene expression changes, perhaps of proliferative cells, across species, cell types, and different rejuvenation methods.

It is clear that a major component of the downstream gene expression program is related to cell cycle, i.e., most of the top TF perturbations up-regulate cell cycle–related TF modules (e.g. several E2F family TF modules (Fig. 1B and SI Appendix, Fig. S1↗), and thus should boost cell cycle progression. This is confirmed by classifying single cells into different cell cycle phases and observing that most of the top TF perturbations decrease the percent of cells in G1 phase while increasing the percentage in S and G2/M phases (SI Appendix, Fig. S2↗). However, many top TF perturbations also reverse gene expression changes between late passaged and early passaged cells not obviously attributable to cell cycle redistribution, as shown by the R_rej calculated when restricted to a subset of cells in the same cell cycle phase (Dataset S2↗). E.g., when comparing the changes between late passage and early passage cells in G1 phase with those between TF perturbed cells and NT control cells in G1 phase, several top hit TF perturbations still have a significant R_rej. One example is EZH2 overexpression. Genes down regulated in late passage cells in G1 phase but reversed with EZH2 overexpression include a broad range of classes such as RNA processing, one carbon metabolism, protein synthesis, mitochondrial function, and stress response (Dataset S3↗).

Gene expression patterns at transcription module level are largely similar among the top hit TF perturbations when restricted to TF modules differentially expressed between early and late passaged cells, which is to be expected as we selected TF perturbations based on their similarity to early passage/late passage comparison. However, different TF perturbations do cause distinct gene expression patterns when viewed from all the TF modules generated from the single-cell data (17) (SI Appendix, Fig. S1↗). Some of these gene groups are unique to a given TF perturbation, indicating potential specific downstream phenotypes that may differentiate different TF perturbations.

We checked whether CRISPRa and CRISPRi of the same TF produced opposite gene expression changes, and we found that very often this is not the case (Dataset S4↗). While for some TFs (such as EZH2 and EGR1) overexpression and repression did lead to opposite global gene expression changes, for the majority of TFs there is no concordance. We speculate that some of these TFs are necessary but not sufficient for the expression of the target genes, while others may be sufficient but not necessary. Thus, perturbation in one direction may produce effect but not in the opposite direction (e.g., if a TF is sufficient but not necessary, overexpression will produce an effect while repression may not produce any effect).

To validate the top hits from the Perturb-seq screen, we performed comprehensive cellular and molecular phenotyping of various aging hallmarks (18) in late passage fibroblasts with a single TF targeted. Specifically, we focused cellular studies on four TF perturbations that reversed overall gene expression back to an earlier passage state: overexpression of EZH2 or E2F3 and repression of STAT3 or ZFX. E2F3 is largely involved in regulating cell cycle progression from G1 to S phase (19). EZH2 is a methyl-transferase best known for being a catalytic subunit of the polycomb repressive complex 2 (PRC2) (20) but has other roles outside the PRC2 complex (21 –24). STAT3 is a member of the STAT family that forms part of the JAK-STAT signaling cascade and plays an important function in immune/inflammatory response (25). ZFX is a relatively poorly understood TF (26), but it has links to stem cell renewal (27, 28). We selected these four single TF perturbations initially because they had a large effect on reversing aging gene expression, as seen from R_rej (Fig. 2A), and from the shift of the single cell distribution in the gene expression space towards early passage cells (SI Appendix, Fig. S3↗). They also came from distinct pathways and were relatively unexplored in terms of rejuvenation research.

We assayed a wide range of cell aging phenotypes, including proliferation, proteostasis, and mitochondrial function. Perturbations of any of these four factors in late passage fibroblasts led to slightly higher rates of cellular proliferation, ranging from 10 to 30% more on average, and decreased senescence-related gene expression (Fig. 2 B and C), bringing both back towards middle passage levels. These single TF perturbations also improved cellular proteostasis. Both proteasome gene expression and proteasome activity returned to a middle passage state, with up to ~40% increases (Fig. 2D). Mitochondrial gene expression and activity, both of which decrease with age and senescence, were significantly increased (Fig. 2E). The total lysosome staining decreased, once again returning to a middle passage state (Fig. 2F). Previous literature shows that the lysosomes of late passage cells gradually transform to residual bodies (containing undigested materials), and these bodies increase in number and size, reflecting degeneration of lysosomal function observed at older age (29, 30). As a comparison to these factors, we performed overexpression experiments of the Yamanaka factors (31) in late passage fibroblasts and validated increased mRNA expression by qPCR (SI Appendix, Fig. S4↗); note that OCT4 expression was too low in control cells to accurately measure changes in expression. The fold change caused by overexpressing the Yamanaka factors in cell proliferation, proteasome activity, mitochondria activity, and lysosome were comparable to those seen with our four single TF perturbations (Fig. 2G). Thus, single TF perturbations we identified led to amelioration of some replicative and cellular aging hallmarks comparable to the combined effect of the Yamanaka factors in an in vitro context.

These single TF perturbations did not generally cause significant changes in DNA damage, telomere length, methylation age, nor fibroblast identity related genes (SI Appendix, Figs. S5 and S6↗). The observation that the methylation age did not change with the TF perturbations is consistent with the previous report that cell senescence is unrelated to epigenetic clocks (32). We also found that the transcriptomic responses to the TF perturbations have no resemblance to those caused by cancerous transformations (Fig. 2H), arguing that these TF perturbations are not tumorigenic. These results indicate that these factors may target orthogonal pathways to factors related to stem cells or dedifferentiation. Thus, through TRDP, we identified TF perturbations which can each individually reverse aspects of passaged dermal fibroblast aging.

We note that there are some specificities in terms of phenotypes reversed by each TF perturbation. For example, E2F3 overexpression has the strongest effect on increasing proliferation (percent of KI67 positive cells, Fig. 2B), EZH2 overexpression is best at increasing mitochondria activity (Fig. 2E), while repression of STAT3 produced the strongest effect on increasing proteasome activity and lysosome function (Fig. 2 D and F). These differences are possibly caused by differences in the downstream transcriptional programs driven by the TF perturbations.

Thus far, we used TRDP to predict TF perturbations which could reverse gene expression—and some aspects of cellular aging—in passaged fibroblasts. Next, we asked whether a single TF perturbation validated in vitro could lead to reversal of aging phenotypes at the tissue level in vivo. We focused on the aged liver as it is both responsive to rejuvenation interventions, such as heterochronic parabiosis (1, 33), and amenable to temporally controlled and tissue-specific gene expression manipulations (34). We first analyzed bulk RNA sequencing data of young (3 mo) and aged (20 mo) mouse livers and looked at whether there were significant age-related differences in the expression of E2F3, EZH2, STAT3, and ZFX. EZH2 and E2F3 expression were significantly higher in young mice, STAT3 was significantly higher in old mice, and ZFX was not different (Fig. 3A). We elected to further investigate EZH2 given its relative better safety profile compared to E2F3 (35, 36) and fewer previous links to liver health unlike STAT3 (37).

We used a liver-specific viral-mediated overexpression approach, where young and aged mice were administered AAV8 encoding EZH2 or a control vector (EGFP control) via tail vein injection. Changes in gene expression and liver histopathology was assessed three weeks post injection (Fig. 3B). EZH2 was successfully overexpressed compared to control (log2fc = 2.9, or ~650% more), which resulted in the expression of thousands of genes being reversed back toward the young mouse state (Fig. 3C). The degree of correlation between aged/young and aged EZH2/aged control (R_rej = -0.42) is even stronger than that observed from the in vitro screening in human fibroblasts (R_rej = −0.36). We found that genes down-regulated in aged mouse livers and reversed by EZH2 overexpression (|log2fc| > 1.5 in both old/young and EZH_OE/control comparisons) encompassed a broad range of functions with enrichment in "transmembrane transporter" category. Genes up-regulated in aged mouse livers and reversed by EZH2 overexpression were enriched for "leukocyte mediated immunity" and "cardiac muscle cell action potential" (Dataset S5↗). These data indicate that EZH2 overexpression reversed age-related inflammation/immune response, loss of cell identify (possibly through epigenetic drift), and loss of transporter genes important for liver function.

EZH2 overexpression did not change general liver morphology within this 3-week period nor lead to any histological aberrations (Fig. 3D). Importantly, EZH2 overexpression did reverse hallmarks of liver health back to a young mouse state, with both a decrease in steatosis (Fig. 3E) and fibrosis (Fig. 3F) to at least half the level of the old mouse livers. Gene expression changes are concordant with rejuvenated histological phenotypes. For example, expression of lipid metabolism–related genes returned back to levels indicative of a more youthful state following EZH2 overexpression (SI Appendix, Fig. S7↗). Glucose tolerance tests (GTT) revealed that EZH2 overexpression in aged mice restored systemic glucose metabolism, as reflected by improved glucose clearance toward levels observed in young mice (Fig. 3G). EZH2 treatment also caused a decrease in inflammatory proteins, p-SAPK/JNK, but no change in p16 (SI Appendix, Fig. S8 A–D↗). Liver damage was not detected, observing no change in ALT activity as well as no differences in body weight (Fig. S8 E and F↗). Together, these data indicate that a TF perturbation like EZH2 overexpression identified by TRDP has the potential to be beneficial in vivo.

To assess potential cancer risk, we compared the liver transcriptome of aged mice with EZH2 overexpression with that from mouse liver cancer models. We analyzed several published datasets from independent experiments in which different treatments were used to cause liver cancer in mice (38). Our analysis indicates EZH2 overexpression did not share a gene expression signature with cancerous gene expression (Fig. 3H). Among the cancer datasets, there were clear conserved gene expression signatures, including up-regulation of genes related to stress response, cell adhesion/migration, regulation of cell population proliferation, and blood vessel development, and down-regulation of carboxylic acid metabolism and amino acid metabolism genes (Dataset S6↗). These signatures suggest a set of changes that are necessary for the growth/migration and metabolic rewiring of the cancer cells. EZH2 overexpression mouse liver profiles have no resemblance to these gene expression signatures. This, together with the histology and functional data, argues that EZH2 overexpression is not simply pro-oncogenic.