Transcription factor fluctuations underlie cell-to-cell variability in a signaling pathway response

Abstract

Stochastic differences among clonal cells can initiate cell fate decisions in development or cause cell-to-cell differences in the responses to drugs or extracellular ligands. One hypothesis is that some of this phenotypic variability is caused by stochastic fluctuations in the activities of transcription factors (TFs). We tested this hypothesis in NIH3T3-CG cells using the response to Hedgehog signaling as a model cellular response. Here, we present evidence for the existence of distinct fast- and slow-responding substates in NIH3T3-CG cells. These two substates have distinct expression profiles, and fluctuations in the Prrx1 TF underlie some of the differences in expression and responsiveness between fast and slow cells. Our results show that fluctuations in TFs can contribute to cell-to-cell differences in Hedgehog signaling.

Introduction

Stochastic processes cause fluctuations in gene expression between single cells (McAdams and Arkin 1997; Raser 2004) (Elowitz 2002; Ozbudak et al. 2002; Raj et al. 2006) (Blake et al. 2003). These fluctuations can produce phenotypic variability between identical cells. For example, cell-to-cell fluctuations of SCA1 help determine the lineage along which hematopoietic stem cells will differentiate (Chang et al. 2008), and cell-to-cell fluctuations of EGFR and AXL underlie the resistance of rare melanoma cells to chemotherapy (Shaffer et al. 2017, 2018). Because such molecular fluctuations can have important phenotypic consequences, a key question is what are the molecular mechanisms underlying cell-to-cell variability within clonal cell populations?

Several mechanisms underlying cell-to-cell variability in gene expression. Some variables that can cause differences between genetically identical cells include cell-cycle stage (Zopf et al. 2013), cellular volume(Kempe et al. 2015; Padovan-Merhar et al. 2015), mitochondrial content (das Neves et al. 2010), ribosome numbers (Guido et al. 2007), and cell state (Kiviet et al. 2014; Iwamoto, Shindo, and Takahashi 2016; Topolewski et al. 2022). However, there are likely additional mechanisms controlling cell-to-cell variability in gene expression, and the advent of single-cell genomic (Trapnell 2015; Eling, Morgan, and Marioni 2019) may provide approaches for identifying these mechanisms.

Stochastic fluctuations in the activities of transcription factors (TFs) are also hypothesized to be an important source of cell-to-cell differences among genetically identical cells (Raj et al. 2006; Senecal et al. 2014; Das et al. 2017). A key prediction of this hypothesis is that the expression variability of certain TFs, or that of their target genes, will correlate with phenotypic differences between clonal cells. Cells that stochastically express high amounts of a TF's target genes might behave differently from clonal cells expressing lower amounts of the same genes. This hypothesis has been tested in several systems. Stochastic levels of TFs underlie cell fate decisions in Drosophila (Wernet et al. 2006), C. elegans (Attner et al. 2019), and bacteria (Samoilov, Price, Arkin 2006) (Süel et al. 2006). Here, we tested whether TF fluctuations might underlie the variable response of mammalian cells to extracellular signaling by Hedgehog.

We use the cellular response to extracellular Hedgehog as a model for stochastic phenotypic variability. Hedgehog signaling polarizes developing tissues through a signaling pathway that results in the activation of the Gli family of TFs (Briscoe and Thérond 2013; Lee, Zhao, and Ingham 2016; Kong, Siebold, and Rohatgi 2019). Hedgehog signaling can therefore be measured using cells carrying a genome-integrated Gli-responsive reporter gene (Pusapati et al. 2018). However, we found that even among clonal cells, treatment with Hedgehog results in significant cell-to-cell variation in the activation of a Gli-responsive reporter gene. We therefore attempted to identify TFs whose fluctuations might account for these cell-to-cell differences in Hedgehog responsiveness using single-cell RNA sequencing (scRNA-seq). We identified fast- and slow-responding subsets of cells characterized by distinct expression profiles that were caused, in part, by stochastic fluctuations in the Prrx1 TF. We found that over-expression of Prrx1 was sufficient to speed up the response to Sonic Hedgehog agonist (SAG). Our results support the hypothesis that TF fluctuations can underlie cell-to-cell phenotypic variation.

Materials and methods

Cell culture

We grew NIH/3T3-CG cells and their derivatives in DMEM media supplemented with sodium pyruvate, 10% Bovine Serum (Gibco 16170078), and 1% Penicillin–Streptomycin. We grew the cells in an incubator maintained at 37° C with 5% CO2.

RNA-sequencing experiments

We generated scRNA-seq libraries using the 10× Single Cell 3′ Reagent Kits v3.1 (10× genomics 1000269). We first released the cells from the cell culture flasks by adding Trypsin 0.25%. We then prepared a cell suspension by following manufacturer's instructions and targeting a final capture of 2,000 cells for every sample. We then proceeded with library construction as described in the 10× protocol. We sequenced the final libraries at the DSIL and MGI at Washington University in St. Louis. Bulk RNA sequencing was performed using the services of Novogene corporation. Briefly, we extracted total RNA from cells, performed initial QC and shipped it to Novogene where library preparation and sequencing were performed targeting 20 million reads per sample.

RNA-sequencing analysis

We obtained counts for each transcript in each cell by using Cell Ranger version 3.1.0 (https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/using/count). We analyzed the cell by count matrix using the standard analysis workflow of Seurat version 3.1.5. (Butler et al. 2018; Stuart et al. 2019) For quality control, we kept cells that had at-least 10,000 UMIs per cell, had <20% of UMIs mapped to mitochondrial genes, and had between 2,000 and 8,000 genes covered. After QC, for the initial single-cell sequencing experiment, we were left with transcriptome measurements for 1,192 cells for the untreated time-point, 797 cells at 17 h and 899 cells at 30 h. For the second scRNA-seq experiment using the inducible Prrx1 cell-line, under the no-Dox condition, we were left with 4,532 cells at the untreated time-point, 3,726 cells at 19 h and 2,645 cells at 30 h after QC. Under the Dox condition, after QC, we were left with 2,362 cells at the untreated time-point, 3,014 cells at 19 h and 2,645 cells at 30 h.

We performed dimension reduction of single cells using Uniform Manifold Approximation and Projection (UMAP) (McInnes, Healy, and Melville 2018). We identified differentially expressed genes between clusters using the FindMarkers function in Seurat. We identified pathways that were enriched in the set of differentially expressed genes using the Gene Ontology (GO) web resource (Ashburner et al. 2000; Mi et al. 2019; Gene Ontology Consortium 2021).

We performed trajectory analysis using Monocle3 version 1.0.0 (Trapnell et al. 2014; Qiu et al. 2017) and Slingshot version 1.4.0 (Street et al. 2018). The Monocle workflow consists of a series of steps. The first step involves pre-processing of the data and dimensionality reduction. We started with the cell by gene matrix that was pre-processed using Seurat and the UMAP dimension reduction results for trajectory analysis. We next identified clusters of the cells using the cluster_cells method of Monocle. We then used the learn_graph method to learn a trajectory graph and ordered the cells according to pseudotime using the order_cells method. We were able to infer directionality on the Monocle trajectory using the time point information that the cells come from (each time point is a different library on the 10× platform), the arrows in Fig. 1 go from the earlier time-point to later-time-points as the cells progress through the SAG treatment. For trajectory inference using Slingshot, we used the same cell by gene matrix pre-processed using Seurat, UMAP dimension reduction results and passed it to the slingshot method. We did not specify the start and end clusters for either Monocle or Slingshot. More details regarding the parameters used for the various programs are available in the R notebooks made available in the Zenodo repository listed below.

Fig. 1.

Cells show variability in response to Hedgehog stimulation. (A) Cells were treated with Hedgehog agonist SAG for 0, 17, and 30 h and collected for single-cell RNA sequencing and flow cytometry. (B) Flow cytometry results at 17 and 30 h indicate that cells show variability in their timing of response, not all same cells respond at the same time. (C) Cells clustered with a clear temporal progression of Hedgehog response when colored by sample time-point. (D) Cells were scored for Hedgehog pathway response by looking at mRNA of four canonical Hedgehog response genes, responder cells cluster in one part of the graph. Response trajectories were inferred using trajectory analysis software. (E) Unsupervised clustering of cells is shown. Cells take two different trajectories to respond to Hedgehog pathway activity, one path is fast responding consisting of only cells at 0 h, the other response is from slow responders consisting of cells at 0, 17, and 30 h.

We computed the Hedgehog pathway activity score for each cell by looking at the mRNA expression of four genes—the GFP reporter, Gli1, Gli2, and Ptch1. For each of the four genes in each cell, we assigned a score of 1 if the gene is detected above background expression level and 0 if not. We then sum over all four genes to obtain a score in the range 0–4 for each cell.

For bulk RNA-sequencing analysis, we pseudo-aligned the reads to the mouse reference cDNA using Kallisto version 0.43.0 (Bray et al. 2016). We used the counts from Kallisto to identify differentially expressed genes using DESEQ2 version 1.26.0 (Love, Huber, and Anders 2014). We used a false-discovery rate of 0.05 to annotate genes as differentially expressed across conditions.

Enrichment calculation

We performed enrichment analysis of the genes that are perturbed when a TF is overexpressed using a hypergeometric test. We detect ∼14,000 genes in our single-cell data, out of which there are 300 genes in the fast-responder gene signature. Using a hypergeometric test, for each TF overexpression experiment, we ask if the overlap of the 300 genes with the number of genes that change expression upon TF overexpression is a statistically significant enrichment under the null hypothesis of no enrichment. We indicate statistically significant enrichments in the figures with an asterisk. We used the hypergeom function in the scipy.stats Python package to compute the enrichment P-values.

Hedgehog assay

We performed the Hedgehog assay on NIH/3T3-CG cells as previously reported in Pusapati et al. Briefly, we grew cells to confluence in media containing regular amounts (10%) of serum. We then switched the cells to low serum media (0.5% serum) overnight, and treated the cells with the Hedgehog pathway agonist SAG (Tocris Biosciences #4366) at 100 nM concentration. We measured the amount of fluorescence in the cells on a flow cytometer after allowing for the appropriate time period of response depending on the experiment. For the initial experiment in Fig. 1, we collected cells at 17 and 30 h. For the Prrx1 overexpression single-cell experiment, we collected cells at 19 and 30 h. For both these experiments, we treated cells that didn’t receive SAG as untreated cells or time-point zero. The time course treatments were done in a staggered manner so that all the cells could be harvested at the same time for scRNA-seq library preparation (Supplementary Fig. 4). This enabled us to process all the cells in one batch to minimize experimental variability for scRNA-seq.

Flow cytometry

We performed the flow cytometry experiments on a Beckman Coulter Cytoflex S instrument. We performed QC based on manufacturer provided QC beads prior to every experiment. We used the same cytometer gain settings for all experiments. To prepare cells for flow cytometry, we first released cells from culture wells by adding Trypsin and then added appropriate volume of culture media to neutralize Trypsin. We gated cells based on forward scatter and side scatter and measured GFP intensity on the FITC channel. We used a negative control sample (no SAG added) to identify the cutoff for the GFP intensity and measured the proportion of responders as the percentage of cells above this cutoff under different experimental conditions.

Plasmid transfection

To transfect the plasmids overexpressing the TFs, we used Lipofectamine 3000 (Invitrogen L3000001) reagent. We used 2.5 µg of DNA per transfection reaction, on a six-well plate, and followed manufacturer’s instructions for amounts of Lipofectamine and P3000 reagents. We measured transfection efficiency on a flow cytometer using the fluorescence of a reporter gene on the transfected plasmid or a control plasmid transfected in parallel. If we observed sufficient transfection efficiency (>50% of cells express transfected plasmid), we extracted total RNA, 24 h post transfection, from the cells and performed bulk RNA sequencing. The control plasmid used is a mCherry reporter gene driven by a cytomegalovirus (CMV) promoter.

Lentiviral generation and transduction

We chose the canonical coding transcripts and sequences for all the genes from UniProt (UniProt 2021). We then cloned the protein coding regions of Prrx1, Snai1, and Srebf2 to the pINDUCER21 Dox-inducible lentiviral vector (Meerbrey et al. 2011) using the services of Genscript (Addgene #46948, Supplementary Figs. 6 and 7). We used the pINDUCER21 system since it allows us to control the level of over-expression of a gene with precision by adding different levels of Doxycycline (Dox) to the cell growth media. This construct also has a fluorophore (miRFP670) on it which allows us to isolate cells that contain the integrated construct. The Hope Center at Washington University in St. Louis generated high-titer lentiviruses using the constructs. Detailed lentivirus generation protocol used by the Center is described in their publication (Li et al. 2012).

We used the generated virus to transduce the NIH3T3-CG cells and used 4 µg/ml polybrene to maximize transduction efficiency. After 24 h of cell-growth in the transduction media, we replaced the media with fresh regular media. We used a fluorescent marker of integration to sort single cells (miRFP670) using a Sony SH800 cell-sorter and grew out single-cell clones that we evaluated for TF induction.

We evaluated single-cell clones based on induction levels assessed by qPCR. For induction of clones, we used Dox at a concentration of 500 ng/ml. For TF induction followed by RNA-seq experiments, cells were treated with Dox for a period of 24 h to induce TF expression. We identified one clone for each TF that offered a good level of inducibility (Supplementary Fig. 8). We then used these clonal cell lines as a model to study the effect of TF overexpression on the Hedgehog assay.

To perform the Hedgehog assay, we grew the cells in two different growth conditions—in one condition, we added Dox to the media and in the other condition, we omitted Dox. We then grew the cells to confluence, for 24 h, and added SAG to both populations of cells to initiate the Hedgehog pathway response. We used flow cytometry to determine the effect of the overexpression of these TFs on the response to Hedgehog stimulation by measuring the GFP fluorescence of the reporter gene on the FITC channel.

qPCR run and analysis

We first extracted total RNA from cells using the Qiagen RNEasy kit (Qiagen 74004). We generated cDNA from the total RNA using the RDRT reagent (Sigma RDRT-100RXN) by following manufacturer's protocol. We then mixed SYBR green PCR master mix (Applied Biosystems 4301955) with 2 µl of cDNA from the previous step, water and primers to set up a standard qPCR run on the QuantStudio instrument (Applied Biosystems). For the TF induction experiments using the transduced cell-lines, we used the no-Dox sample as the baseline for computing the delta Ct value. We used HPRT primers to normalize as a within sample control for TF expression (Supplementary Table 3). We analyzed the results of the QuantStudio run using the Design and Analysis 2 software from Thermo Fisher.

Results

Variability in the Hedgehog response

We first asked whether clonal cells growing in the same environment show variability in their responses to Hedgehog signaling. To address this question, we used NIH3T3-CG cells, which are an established model of Hedgehog signaling (Pusapati et al. 2018; Kinnebrew et al. 2019). Treating these cells with SAG activates the pathway, resulting in elevated activity of Gli TFs (Briscoe and Thérond 2013; Lee, Zhao, and Ingham 2016; Kong, Siebold, and Rohatgi 2019). This increased activity of Gli is read out by a genome-integrated GFP reporter gene regulated by eight Gli binding sites (Supplementary Fig. 1). To assess variability in the Hedgehog response we treated NIH3T3-CG cells with SAG and monitored the response of the GFP reporter gene by flow cytometry (Fig. 1a). Although monocultures of NIH3T3-CG cells are clonal and are grown in the same controlled environment, we detected significant cell-to-cell variability in reporter gene expression in response to SAG (Fig. 1b). For example, at 30 h post SAG treatment only 39% of cells had activated the reporter gene, whereas by 92 h most cells responded (Supplementary Fig. 2). We observed similar variability in multiple experimental replicates grown on different days and derived from different single-cell clones (Supplementary Fig. 3). The flow cytometry profiles do not support a bimodal distribution, but rather a highly skewed unimodal distribution in which cells in the right tail of the distribution are responding rapidly to SAG [Hartigan's dip test, D = 0.0043, P = 0.99 (Hartigan and Hartigan 1985)]. What accounts for the difference between fast- and slow-responding cells in monocultures with no genetic or environmental variation?

Fast- and slow-responding NIH3T3-CG cells

We asked whether cells that respond quickly to SAG derive from untreated cells with expression profiles that are distinct from the slow-responding cells. We performed scRNA-seq on cells at 0, 17, and 30 h after the addition of SAG. We chose 30 h because we observe the maximal cell-to-cell variability in response at this time-point. We then included the 17 h time-point because it is roughly halfway between the start of the response and 30 h. (Methods).

We visualized cells at the three time-points using UMAP (Fig. 1c) and performed unsupervised clustering of cells. On the same UMAP plot, we colored the cells for Hedgehog response using four Hedgehog response genes (Methods) and observed a region of the UMAP where the responding cells reside (Fig. 1d). We observed similar results when we visualized cells using Principal Components Analysis (Supplementary Fig. 5), but UMAP highlighted the local differences between cells better and captured all the variation in two dimensions.

Next, we used Monocle to perform trajectory analysis (Qiu et al. 2017; Trapnell et al. 2014) to identify the cell states from which slow and fast responders derive. We observed two distinct trajectories after SAG treatment that lead towards cells expressing genes indicative of the Hedgehog response (Fig. 1d). Fast-responding cells start in cluster 0 and follow a trajectory that leads into cells that express Hedgehog responsive genes at 17 and 30 h (Fig. 1e). Slow-responding cells start in cluster 1 and many of these cells remain in the same cluster at 17 and 30 h, while other slow-responding cells follow a trajectory that leads towards, but does not reach Hedgehog responding cells. We interpret this result to mean that untreated NIH3T3-CG cells consist of two distinct subpopulations, one that is primed to respond quickly to SAG (fast-responders—cells in cluster 0 of Fig. 1e) and another that takes longer to respond (slow responders—cells in cluster 1 of Fig. 1e). Very few untreated cells lie outside these two clusters. We confirmed these results using a second trajectory analysis software, Slingshot (Street et al. 2018), which also identified two trajectories, one starting from the fast responders and another from the slow responders, leading to the Hedgehog responsive cells. Thus, global differences in mRNA expression profiles define fast- and slow-responding sub-states in NIH3t3-GC cells.

One hypothesis that might explain cell-to-cell differences in the timing of the Hedgehog response between the two subpopulations would be transient activation of Hedgehog signaling in the absence of ligand. Cells that are in the midst of transiently activating the pathway would then respond faster to extracellular ligand than the remaining cells. To test this hypothesis, we looked for subsets of cells that expressed targets of Hedgehog signaling. In the unstimulated population, we did not observe individual cells with activation of Hedgehog target genes (Supplementary Fig. 7). Thus, Hedgehog signaling is regulated tightly enough that stochastic activation of the pathway in the absence of ligand is unlikely to account for the cell-to-cell differences in cellular response after the addition of ligand.

Differences in cell cycle state do not fully explain fast- and slow-responding cells

We next asked whether differences in the cell-cycle state of cells might explain the differences between the fast- and slow-responding cells. We serum-starve cells prior to Hedgehog treatment to synchronize the cells in G1. Because this synchronization is not perfect, we used the expression profiles of serum-starved untreated cells to assign them to different phases of the cell cycle and then asked whether fast-responding cells were enriched for any specific phase of the cell-cycle. We observed a 1.5-fold enrichment of G1 cells in the fast responders compared to the slow responders (74.3% of fast responders vs 50.4% in the slow responders) (Supplementary Fig. 8). While we found a statistically significant enrichment, the fact that half of slow-responding cells were in G1 suggests that being in G1 is not sufficient for a cell to respond faster.

To validate the Seurat cell cycle measurements, we also measured the percentage of cells in different phases of the cell cycle using Propidium Iodide staining after 24 and 48 h of serum starvation (Supplementary Fig.9a). The computed percent of cells in different phases of the cell cycle after 24 h of serum starvation agrees with the Seurat estimated cell-cycle fractions. We also tested whether longer serum starvation would increase synchrony and reduce variability in the Hedgehog response. We observe similar amounts of variability in the Hedgehog response after 48 h of serum starvation as compared to 24 h of serum starvation (Supplementary Fig.9b). We conclude that cell-cycle differences do not fully explain the differences between fast- and slow-responding cells.

Differentially expressed TFs and pathways in the fast-responder cells

We identified differentially expressed genes between untreated cells defined as either fast or slow responding (Methods). We focused on the top 300 genes, ranked by fold-change, that were statistically significantly different between the fast and slow populations (Supplementary File S1). We use these 300 genes in subsequent analyses as markers of the fast responder state. Among the 300 genes, 37 genes are TFs (Supplementary Table 1). We decided to focus on the TFs in the differentially expressed genes because we hypothesize that fluctuations in TF activities account for the differences between fast and slow cells. Stochastic differences in the activities of TFs can be caused by fluctuations in their expression, localization, or post-translational modification. However, in this study, use the mRNA expression levels of TFs as proxies for their activities. While this measurement of activity is convenient and allows us to track TFs at single-cell resolution, it necessarily does not account for post-transcriptional or post-translational regulation of TF activities.

We also asked if any biological pathways were enriched among differentially expressed genes between the fast- and slow-responding cells through a GO analysis (Methods). We observed that the cholesterol biosynthesis pathway is significantly enriched in the set of genes differentially expressed between the two groups (Supplementary Table 2, Enrichment = 28.38, P = 1.7e-4, n = 5 out of 13 genes). Cholesterol boosts Hedgehog signaling by serving as a cofactor for Smoothened (Luchetti et al. 2016; Huang et al. 2018, 2016; Kinnebrew et al. 2019; Radhakrishnan, Rohatgi, and Siebold 2020). Increased cholesterol biosynthesis may therefore contribute to the phenotype of fast-responding cells.

Slow-responding cells pass through the fast state before responding

Slow-responding cells might respond more slowly because they first have to enter the fast-responding state from their current state or because they follow a distinct set of cellular states that lead to the Hedgehog response. We distinguished between these two possibilities by examining the set of genes that come on in the slow responders at later time points. Specifically, we examined cells from hour 17 and hour 30 in the slow-responding cluster (cluster 1 in Fig. 1d) and asked what genes are differentially expressed compared to untreated cells in the same cluster. When the slower responders respond, we observe that they turn on the gene signature of the fast-responding cells. At 17 h, these cells express 32 of the 300 fast responder genes (total number of DE genes = 117), and at 30 h, they express 48 of the 300 fast responder genes (total number of DE genes = 196), these are both statistically significant enrichments over random expectation (Hypergeometric test; P = 7.9e-27, 1.8e-37). We interpret this result to mean that the slower responders take longer to follow the same trajectory as fast-responding cells rather than following a different trajectory.

Overexpressing candidate TFs partially recreates the fast responder gene signature

Our results suggest that fluctuations in a small set of TFs and/or the upregulation of cholesterol biosynthesis pathway causes the differences between fast- and slow-responding cells. To test this hypothesis, we over-expressed these TFs and the cholesterol biosynthesis pathway and measured the bulk expression profiles of the resulting cells. We focused on five TFs based on their expression in the fast responder cells and their annotated functions as developmental TFs (Supplementary Table 1): Jun, Egr1, Prrx1, Snai1, and Srebf2 (Berge et al. 1998) (Carver et al. 2001). We chose the TF Srebf2, the main regulator of the cholesterol biosynthesis pathway (Horton et al. 1998), because this pathway is upregulated in the fast responder cells (Supplementary Table 2). The expression of these differentially expressed TFs co-localize with the fast responder cell populations (Supplementary Fig. 10). We designed plasmids encoding the coding regions of Jun, Egr1, Prrx1, Srebf2, and Snai1 under the control of a strong CMV promoter, and with a cotranscribed mCherry gene to measure transfection efficiency (Supplementary Figs. 11 and 12). Separately, we used a plasmid containing an mCherry reporter gene driven by a CMV promoter as a control to measure whether mCherry expression has any effect on the Hedgehog response. We transfected these plasmids into NIH3T3-CG cells and performed bulk RNA-seq. We identified genes that are differentially expressed upon transfection of each of the TFs but not after the control transfection (Supplementary Files 2–6).

Overexpression of four of the five TFs (Prrx1, Snai1, Jun, and Srebf2) individually led to misexpression of a significant number of the fast responder genes (Fig. 2; Hypergeometric test: P = 9.7e-4, 6.3e-7, 1.2e-10, 1.9e-16), whereas Egr1 overexpression did not (P = 0.13). The union of genes targeted by all four of these TFs (Prrx1, Snai1, Jun, and Srebf2) resulted in 115 out of the 300 fast responder gene signature. This number is smaller than the naive sum of the individual target genes and indicates overlap in the sets of target genes regulated by the TFs. Overall, we interpret this result to mean that each TF independently regulates a subset of the fast-responder signature.

Fig. 2.

Overexpressing transcription factors partially recreates the fast responder gene expression signature. We overexpressed five transcription factors, one at a time, Prrx1 (A), Snai1 (B), Srebf2 (C), Jun (D), and Egr1 (E) using a plasmid transfection assay followed by bulk RNA sequencing. We computed the overlap between the genes that change expression upon transcription factor overexpression (compared to control transfection, n = 3 replicates each) and the 300 genes in the fast-responder gene signature identified using the single-cell data. The (*) indicates a statistically significant enrichment using a hypergeometric test.

Prrx1 is a regulator of the fast responder state

We next asked whether the overexpression of any of these TFs was sufficient to increase the fraction of cells that respond to the Hedgehog signaling pathway. To test this prediction, we engineered cells, using lentiviral vectors (Methods, Supplementary Figs. 13 and 14), carrying Dox-inducible (Supplementary Fig. 15) versions of Prrx1, Srebf2, and Snai1 integrated into their genomes. This allowed us to compare the fraction of cells that respond to SAG when a TF is either induced or uninduced. For these experiments, we chose Srebf2 since it is the main regulator of the cholesterol biosynthesis pathway, which is the top differentially enriched pathway in the fast-responding cells (Supplementary Table 2) and showed a statistically significant enrichment. Additionally, we chose Prrx1 and Snai since they showed strong enrichment for competent genes in the transfection assay and because of their roles as developmental TFs.

Inducing Prrx1 expression resulted in more cells responding to the Hedgehog pathway (Fig. 3a, Supplementary Figs. 16 and 17). In the +Dox condition, at 32 h post SAG treatment, 87.9% of cells become GFP+, whereas in the -Dox condition 65.4% of cells become GFP+. We did not observe a difference between the two groups with the induction of Snai1 (Supplementary Fig. 18). Strong induction of Srebf2 was toxic to cells, and at induction levels that were not toxic, we did not observe an increase in the response to SAG (Supplementary Fig. 19). From these results, we conclude that Prrx1 plays a role in the regulation of the Hedgehog response.

Fig. 3.

Inducing Prrx1 expression makes more cells fast-responders. (A) We induced Prrx1 expression in engineered cells by adding Doxycycline prior to performing the Hedgehog assay. We compared the Hedgehog response of induced cells to cells that were not induced using flow cytometry 32 h post SAG treatment (n = 3 replicates each). (B) We identified the genes that change expression when Prrx1 is overexpressed by bulk RNA sequencing. We then computed the overlap of these genes with the 300 genes in the fast responder signature. The (*) indicates a statistically significant enrichment using a hypergeometric test. Error bars shown are standard error above and below the mean.

Bulk RNA-seq profiles from Prrx1-induced cells (Supplementary File 7) revealed a stronger overlap with fast-responding genes than in the plasmid overexpressed cells (Fig. 3b; n = 74 out of 300, P = 4e-34, enrichment = 5.37 fold). This stronger response is likely because in transiently transfected cells, only a subset of the cells are expressing Prrx1, whereas in the lentiviral transduced cells, all cells are overexpressing Prrx1. Induction of Prrx1 results in significant enrichment of genes involved in the cholesterol biosynthesis pathway (adjusted P = 0.025, enrichment = 10.42, n = 4 out of 13 genes in the pathway). Prrx1 induction also results in the upregulation of Gli2, the primary effector TF of Hedgehog signaling (Briscoe and Thérond 2013; Lee, Zhao, and Ingham 2016; Kong, Siebold, and Rohatgi 2019). Gli2 is regulated post-translationally by Hedgehog signaling, so the upregulation of Gli2 in uninduced cells may result in a larger pool of Gli2 protein to activate when the pathway is activated. We observe slightly leaky expression of Prrx1 in the transduced clone with a fold change of 1.5× relative to wild type clones even in the absence of Dox, which explains the higher response of these cells compared to wild-type cells in Fig. 1. Our results indicate that overexpression of Prrx1 can recreate a substantial part of the fast responder signature, which we interpret as the reason more of these cells respond to Hedgehog signaling.

Most cells activate the fast responder gene signature upon Prrx1 induction

We next examined the trajectory taken by cells overexpressing Prrx1. We generated scRNA-seq profiles of Prrx1-induced and uninduced cells followed by stimulation with SAG (Fig. 4a). We started with the cells carrying the genome-integrated inducible Prrx1 cassette described in the previous section. We grew cells in the absence or presence of Dox to induce Prrx1 expression and measured the Hedgehog response by flow cytometry. We again observed a larger response in Prrx1-induced cells than in uninduced cells (Supplementary Fig. 20). We then collected cells at three time points after SAG treatment for both Prrx1-induced and uninduced cells and performed scRNA-seq.

Fig. 4.

Prrx1 induction results in faster response trajectory. (A) Cells were grown in the absence and presence of Doxycycline for 24 h, to induce Prrx1 expression, and treated with the Hedgehog pathway agonist for 0 (Untreated), 19, and 30 h. (B) The clustering of the cells from the three time points across the two conditions. (C) The inferred trajectory of Hedgehog response. Cells were scored and colored for Hedgehog pathway response by looking at mRNA expression of four canonical Hedgehog response genes. (D) The unsupervised clustering of the cells using Seurat. Fast responders and slow responders predicted using a cell-type classifier are shown.

Prrx1-induced cells follow a faster and stronger response trajectory compared to their uninduced counterparts. When Prrx1 is not induced, cells from each time-point again cluster separately, suggesting a steady progression of response to SAG through time (Fig. 4b). By contrast, when we induced Prrx1 expression with Dox, cells at the 19 and 30 h time point cluster together, suggesting a faster response to SAG induction (Fig. 4b). Overlaying the Hedgehog response on these clusters supports this interpretation as Prrx1-induced cells at 19 h show similar levels of the Hedgehog response as induced cells at 30 h (Fig. 4c). However, even at 30 h induced and uninduced cells do not cluster together, which suggests that Prrx1 induction results in a stronger response to Hedgehog in cells that do respond. Taken together, these results show that the induction of Prrx1 causes more cells to respond to SAG, and that those cells that do respond, respond faster and stronger than cells that do not express Prrx1.

The trajectory analysis also revealed differences between induced and uninduced cells in the early response to SAG that are consistent with the idea that Prrx1 causes cells to adopt fast-responding expression profiles. Uninduced cells again showed distinct converging trajectories from slow- to fast-responding states (clusters 4 and 5 in Fig. 4d), but Prrx1-induced cells respond primarily along the fast trajectory. In addition, the Prrx1-induced cells start from a position along the fast trajectory that is closer to the later time points than the uninduced cells (clusters 9, 10, and 11 in Fig. 4d), which suggests that they are “further along” the fast trajectory than uninduced cells even before stimulation with SAG.

We next asked whether inducing Prrx1 creates more fast-responding cells. To detect fast-responding cells in this experiment, we used the model-based cell-type classifier Garnett (Pliner, Shendure, and Trapnell 2019). We trained the classifier to learn the features of fast- and slow-responding cells using the cells from the untreated condition shown in Fig. 1. We then used the classifier to classify untreated cells grown in the +Dox and -Dox conditions as fast responders and slow responders. In the untreated cells grown in -Dox condition, the classifier classified 19% of cells as fast responders, 48% as slow responders, and 33% of the cells as unknown. The fast responders in this group separate from the slow responders in the UMAP plots and are slightly ahead in the response trajectory even before SAG treatment (Supplementary Fig. 21a and b). In contrast, in the untreated cells from the +Dox condition, the classifier classified 91% of cells as fast-responders, 2% slow-responders, and 7% as unknown (Supplementary Fig. 21c and d). Thus, in the presence of Dox, the majority of cells display the fast-responder expression signature. We infer from this result that inducing Prrx1 generates a fast responder cell-state which makes more cells respond to the Hedgehog agonist, as early as 19 h.

Discussion

Fluctuations in TFs produce transient changes in gene expression that can generate phenotypic differences among clonal cells in the same environment. Consistent with this hypothesis, we showed that distinct gene expression profiles define fast- and slow-responding states among clonal NIH3T3-CG cells, and that a small number of TFs accounts for a substantial fraction of the expression differences that define the two states. The idea that these expression profiles cause differences in the response to Hedgehog signaling is supported by the observation that overexpression of Prrx1 is sufficient to generate part of the fast-responding expression signature and drive a faster and stronger response to Hedgehog in a larger fraction of cells. However, because Prrx1 only accounts for part of the fast-responding expression signature, there must be fluctuations in other TFs (Sigal et al. 2006), or other types of signaling molecules, that generate the fast-responding cell state. Much deeper sequencing of unstimulated cells might reveal groups of differentially expressed target genes that could indicate which TFs and signaling molecules are involved, much the same way in which coherent differential expression of the cholesterol biosynthesis pathway suggested the involvement of Srebf2 in the fast-responding state. Alternatively, if the transient activation of non-TFs (e.g. receptors, kinases) (Colman-Lerner et al. 2005) underlies the fast-responding state, then it will be necessary to identify specific changes in gene expression that are diagnostic of changes in the activities of such genes.

We observe parallels between our model underlying variability in the Hedgehog response and other previously described systems where fluctuations of a few key molecules underlie cell-to-cell variability in phenotypic outcomes. For example, differences in the levels of a few apoptotic regulator proteins prior to drug treatment determines how fast cells die in the presence of an apoptosis inducing ligand (Spencer et al. 2009). Fluctuations of a few resistance genes in cancer cells results in cell states that are resistant to cancer drugs (Shaffer et al. 2020; Emert et al. 2021), though the exact mechanisms underlying the emergence of such cell states remains unknown and is hypothesized to involve the fluctuation of multiple upstream TFs. In stem cells, fluctuations of a key TF Nanog affect whether cells differentiate or remain in the pluripotent state (Miyanari and Torres-Padilla 2012; Torres-Padilla and Chambers 2014). A recent study using fluorescence microscopy determined that most of the variability in the JAK-STAT signaling response can be attributed to fluctuations in the molecular content of cells (Topolewski et al. 2022). Fluctuations in the amount of TFs between cells have been hypothesized to underlie disease where some cells in a population are unable to express the TF above a required threshold (Cook, Gerber, and Tapscott 1998; Kemkemer et al. 2002). Single-cell transcriptomic approaches, such as those used in this study, can shed light on the specific molecular differences between cells that result in the heterogeneity of cellular responses.

Our study also has important implications for improving the efficiency of cellular assays. For example, reprogramming assays can be made more efficient if we understand why some cells are successfully reprogrammed while other cells result in “dead-end” states (Graf and Enver 2009; Biddy et al. 2018; Francesconi et al. 2019). Fluctuations of TFs in cells prior to reprogramming might underlie the heterogenous outcomes in cellular reprogramming protocols. For example, in hematopoietic progenitor cells, levels of a stem cell marker determine which lineage a cell differentiates towards (Chang et al. 2008). Isolating cells with different levels of such proteins may improve reprogramming efficiency. Identification of TFs, like Prrx1 in this study, that dampen noise of a system could also aid in the design of synthetic circuits where noise is a significant obstacle (Murphy et al. 2010).

We have inferred the Hedgehog response trajectories using methods that identify gene expression similarities between single cells. A complementary measurement of trajectories would involve using transcribing molecular barcodes to tag the cells prior to scRNA-seq (Kong et al. 2020). This approach could reveal whether cells can transition back and forth between the fast and slow responder states (Hormoz et al. 2016; Stumpf et al. 2017; Larsson et al. 2021). In addition, using fluorescently tagged versions of Prrx1 might help separate the fast responder population from the slow responders and facilitate biochemical assays on these subsets of cells. We speculate that other TFs also contribute to the fast-responding state, but that we are underpowered to detect them. The current expense of scRNAseq experiments limits us from performing the many replicate experiments that would be required to identify still smaller differences between fast- and slow-responding cells. Finally, we have used a widely accepted cell-culture model of Hedgehog signaling, and how our findings apply to in vivo Hedgehog signaling remains to be tested.

One future direction of this work will be to determine whether the fast-responding state is specific for Hedgehog signaling or if this state makes cells differentially responsive to other signaling pathways and perturbations. We speculate that similar variability in the activities of different TFs may underlie other phenotypic differences among genetically identical cells and experimental approaches similar to ours can be used to uncover this variability.

Data availability

The raw single-cell and bulk RNA-sequencing data from this publication are available from GEO under the accession numbers GSE203134 and GSE206154. Analysis notebooks used for the analysis of single-cell data are available for download at Zenodo: https://doi.org/10.5281/zenodo.6981764

Supplemental material available at GENETICS online.

Author contributions

A.R and B.A.C. conceptualized and designed the project. A.R designed and performed all experiments and analyses. A.R and B.A.C. wrote the manuscript.

Funding

This work by supported by grants R01 GM140711 and R01 GM092910 from National Institute of General Medical Sciences (NIGMS) to Barak A Cohen.