Epithelial-Mesenchymal Micro-Niches Govern Stem Cell Lineage Choices

Summary

Adult tissue stem cells (SCs) reside in niches, which through intercellular contacts and signaling, influence SC behavior. Once activated, SCs typically give rise to short-lived transit-amplifying cells (TACs), which then progress to differentiate into their lineages. Here, using single cell RNA-sequencing, we unearth unexpected heterogeneity among SCs and TACs of hair follicles. We trace the roots of this heterogeneity to micro-niches along epithelial-mesenchymal interfaces, where progenitors display molecular signatures reflective of spatially distinct local signals and intercellular interactions. Using lineage-tracing, temporal single cell analyses and chromatin landscaping, we show that SC plasticity becomes restricted in a sequentially and spatially choreographed program, culminating in seven spatially arranged uni-lineage progenitors within TACs of mature follicles. By compartmentalizing SCs into micro-niches, tissues gain precise control over morphogenesis and regeneration: Some progenitors specify lineages immediately; others retain potency, preserving self-renewing features established early while progressively restricting lineages as they experience dynamic changes in microenvironment.

INTRODUCTION

Adult stem cells (SCs) maintain tissue homeostasis and repair wounds. They reside in specialized niches where neighboring cells provide the inputs that keep these cells in an undifferentiated state (Losick et al., 2011; Scadden, 2014; Schofield, 1978). When SCs mobilize to make tissue, they generate shorter-lived progeny, sometimes called transit-amplifying cells (TACs), which then progress to differentiate into the lineages that fuel tissue growth. At some point, cells with proliferative capacity must choose the particular lineage to embark upon. The underlying mechanisms involved in making these fate choices can become daunting, especially when considering that lineages need to be balanced and coordinated to guard against disease states. This becomes particularly critical for progenitors of epithelial-based tissues, which in response to injury must rapidly restore the lineages that exclude microbes and retain body fluids.

While the ability of progenitors to choose a defined lineage has universal importance in tissue biology, it remains largely unanswered when and where SCs and/or their progeny become specified. The traditional view has been that lineage decisions are made downstream, and that SCs and TACs are molecularly homogeneous (Barker et al., 2009; Bryder et al., 2006; Clayton et al., 2007; Lien et al., 2011; Orkin and Zon, 2008). Recent studies using elegant lineage marking of hematopoietic stem cells (HSCs), and then tracing their behavior upon transplantation, challenge this view and suggest that at least for HSCs, heterogeneity may exist within the bone marrow (Yu et al., 2016). Single cell RNA-sequencing supports this notion, unmasking molecularly distinct subpopulations of HSCs within the bone marrow. Although transplantation assays did not reveal functional distinctions, signs of diversity at transcriptional and epigenetic levels provide compelling evidence in support of heterogeneity at the SC level (Paul et al., 2015; Wilson et al., 2015; Zhou et al., 2016). How this early specification of HSCs into subtypes might occur and whether it is important for hematopoiesis, remain unknown.

A general difficulty in pinpointing when and how spatial molecular heterogeneity arises in the governance of homeostasis and tissue regeneration is that SCs and their proliferative progeny rely on inputs from their native niche microenvironment. Thus, although transplantations have been the gold standard for measuring self-renewal and tissue regenerative properties, these methods interrogate SC behavior outside their native context, where they often display broader lineage options than those imposed by their niche microenvironment (Blanpain and Fuchs, 2014; Paul et al., 2015). Laser ablation studies further highlight the importance of niche microenvironment, illustrating that when one cell is removed from its niche, another cell nearby can replace it, even if it comes from a more committed progenitor population (Rompolas et al., 2013; Tetteh et al., 2016; Tian et al., 2011). To evaluate whether SCs and/or their shorter-lived progeny have heterogeneity, where it comes from, and whether it has a functional purpose for the tissue or is merely stochastic, it is essential to study them in their native niche.

The HF is well-suited for tackling these problems, given that the lineages afforded to its SCs are spatially and temporally well-defined. In contrast to the hematopoietic system, consecutive SC progeny can be monitored along a lineage, so that choices made along the way can be analyzed at a molecular level and in context of microenvironment. Additionally, HFs undergo natural bouts of regeneration as they cycle through sequential phases of active hair growth (anagen), destruction (catagen) and then rest (telogen). In a physiologically unperturbed state, HFSCs transition from bouts of quiescence to tissue-regeneration, where hair growth can reach 20μm per hour.

Deceptively simple as an epidermal appendage, a HF’s cycling portion consists of at least 7 morphologically distinct lineages, which are spatially organized and coordinately regulated (Figures 1A and 1B). The SCs that fuel the process are located in a niche, called the bulge. Quiescent ‘primed’ SCs, located at the bulge base (hair germ, HG) become activated to proliferate at the telogen→anagen transition when two positive SC stimulators, BMP-inhibitors and WNTs, reach threshold levels in the niche that overcome BMP signals emanating from inner bulge cells (Greco et al., 2009; Hsu et al., 2011). These changes in niche signals arise through cross-talk between HG-SCs and specialized stimulatory mesenchymal cells, called dermal papilla (DP) (Woo et al., 2012). As the HG expands to form a hair bulb of short-lived, TACs, the DP becomes distanced from the bulge by the downward growing ‘outer root sheath’ (ORS). When the HF reaches maturity, TACs divide in equilibrium with fueling the upward differentiation of the remaining concentric lineages of the hair shaft (HS) and its channel, the inner root sheath (IRS) (Figure 1B).

Figure 1. The Eight Concentric Layers of the Mature HF.

(A) Schematic depicting resting (Telogen) and regenerating (Anagen) phases of hair follicles. Bu-HFSC, HF stem cells residing in the bulge niche; HG, hair germ, a niche housing the ‘primed’ HFSCs that will be activated at the start of the new hair cycle; DP; dermal papilla, specialized mesenchymal cells required to fuel the hair cycle. (B) (left), Schematic depicting the complex organization of the mature HF. TACs, transit amplifying cells, which in full anagen balance proliferation and differentiation to produce the hair shaft (HS) and its channel, the inner root sheath (IRS). ORS, outer root sheath; Cp, companion layer; He, Henle’s layer; Hu, Huxley’s layer; Ci, IRS cuticle; Ch, HS cuticle; Co, cortex; Me, medulla. (right), Ultrastructure of sagittal section just above the mature hair bulb. Scale bar = 10μm. (C) Lineage heterogeneity revealed by immunofluorescence for: GATA3 (Hu, Ci); K71 (He); K6 (Me, Cp); K32 (Ch); K31 (Co). Scale bars=50μm.

The bulk transcriptomes of SCs of bulge and HG are largely indistinguishable (Greco et al., 2009), and laser ablation studies suggest that they are interchangeable (Rompolas et al., 2013). While TACs exhibit a bulk molecular profile distinct from that of their SC parents, they are still proliferative and undifferentiated (Lien et al., 2011). Thus, it has remained a mystery why and how HFSCs and their progeny execute temporally and spatially well-defined movements in tissue regeneration (Hsu et al., 2011; Rompolas et al., 2013). When along the SC lineage does the specification into discrete HF lineages occur? What drives diversification and when? Do SCs go through sequentially restricted intermediates? How do TACs generate so many lineages? The answers to these questions are at the crux of our understanding of how multipotent SCs and their progenitors become progressively restricted in their fate choices and whether SC and/or TAC heterogeneity play a physiological role in tissue regeneration.

In the current study, we tackle these questions. Taking advantage of the synchronized bouts of tissue regeneration in the murine hair coat, we employ temporal single-cell RNA-sequencing on purified populations of HF cells and assign molecular signatures as its progenitors progress through their journey from quiescence to active tissue regeneration. We show that in contrast to prevailing notions, molecular heterogeneity exists within the HG, TAC and DP. Moreover, by lineage tracing and mapping the heterogeneity, we show that the heterogeneity reflects lineage priming, and is spatially and temporally blueprinted within micro-niches along an epithelial-mesenchymal ridge. This lineage priming begins in the SCs at a stage when they are quiescent. As the tissue grows and expands, so too does the epithelial-DP interface and its localized cross-talk, maintaining and elaborating these micro-niches to guide the complex lineages of the beautifully patterned hair follicle. This organizational plan also allows for the temporal restriction of SC plasticity during lineage progression, as our single cell data show.

RESULTS

‘Transit Amplifying Cells’ Consist of at Least Seven Spatially Arranged Progenitors

Ultrastructural and immunofluorescence microscopy reveals eight distinct concentric layers of the mature HF (Figure 1). Lineage tracing experiments with ubiquitously expressed, tamoxifen-induced CreERT2 suggest that lineages emerge from spatially organized cells within the hair bulb (Legué and Nicolas, 2005; Legué et al., 2010). Two fundamentally distinct models could explain this behavior: 1) TACs are homogeneous, adopting spatially defined fates as they divide and bud from the basement membrane niche; or 2) TACs are molecularly heterogeneous, launching distinct terminal differentiation programs according to a spatially pre-determined transcriptional program.

To distinguish between these two possibilities, we analyzed TACs from mature (Anagen VI) follicles at the single cell level. We devised a purification scheme based upon fluorescence activated cell sorting (FACS) of the P-cadherin- and integrin-rich TACs associated with the basement membrane, and then further fractionated them according to Lhx2-GFP (Figures 2A and S1). We then used Smart-seq2 (Picelli et al., 2013) to profile whole transcriptomes of Ana-VI TACs at the single cell level. External RNA spike-ins were used to measure technical noise (Brennecke et al., 2013). Overall, the batch to batch variation of pooled TAC sequences was comparable to that of conventional RNA-seq of bulk populations, indicating the fidelity of our single cell data (Figure 2B).

Figure 2. Single Cell RNA-seq Analysis Reveals Molecular Heterogeneity of TACs.

(A) Experimental strategy to purify the TACs at the epithelial-mesenchymal border. (left), Expression patterns of P-cadherin, Integrin β1 and transgenic Lhx2-GFP in the hair bulb. (right), Purification scheme. (B) Scatter plots showing correlation between (left) two batches of single cell RNA-seq and (right) two biological replicates of RNA-seq on bulk FACS-sorted populations. (C) (left), Unbiased clustering of transcriptomes of individual basal TACs from Ana-VI HFs. Each cell is represented as a dot, colored by a clustering algorithm and plotted on the tSNE graph. (right), Heatmap showing transcriptome similarities between single cells measured by Pearson’s correlation coefficient matrix. Clusters are color-coded along the axes. (D) (top) tSNE plot showing the cell cycle status of each cluster. (bottom), tSNE plot showing Tchh expression. Although a marker of differentiating medulla and IRS cells, trichohyalin (Tchh) was only expressed in basal medulla progenitors (arrows). The dotted line denotes epidermal-dermal border. Note that Cluster 1 (C1) is non-proliferative and expresses many medulla genes. (E) Immunofluorescence showing EdU⁺ S-phase cells, Integrin α6 and P-cadherin in the mature hair bulb. (F) tSNE plots showing 2^nd-level subclustering of C2 and C3 into six (left) and two (right) subclusters, respectively. Scale bars = 50μm. See also Figures S1 and S2; Tables S1 and S2.

Reflecting the high quality of our sequenced libraries, an average of nearly 10⁶ reads and ~8,000 genes per TAC were detected, with <5% of reads mapping to the mitochondrial genome (Figure S2A–2C; Table S1). Only 17 of 384 individual Ana-VI TACs displayed < 4000 genes and were excluded from further analyses. Dimensional reduction by principle component (PC) analysis, t-distributed stochastic neighbor embedding (tSNE), and unsupervised hierarchical clustering revealed 5 distinct subsets of expression patterns (Figures 2C and S2D–S2F). Importantly, cell transcriptomes within a single cluster were highly similar, while comparative analyses against other clusters revealed ~500 mRNAs differentially expressed by ≥ 2X (Figure 2C, Table S2; FDR < 0.1). Two clusters (C4 and C5) were small and displayed a signature characteristic of non-epithelial cells (for representative examples of their signature transcripts, see Figure S2G). Hence, we did not study these clusters further.

Cells in cluster C1 were non-dividing, as revealed by ‘performed machine learning-based cell cycle allocation analysis’ (Scialdone et al., 2015). This cluster showed signs of an early medulla signature (Figure 2D). By contrast, S-phase and G2/M were enriched in the clusters C2 and C3, and based upon S-phase labeling with 5-Ethynyl-2′-deoxyuridine (EdU), these cells were TACs (Figures 2D and 2E). 2^nd-level clustering analysis revealed that the largest proliferative cluster (C2) contained 6 distinct sub-clusters, while the smaller (C3) contained at least two distinct sub-clusters (Figure 2F).

Each subpopulation displayed a ‘molecular signature’ of ~100 genes whose expression changed by ≥2X relative to the other subtypes (FDR < 0.3) (Table S3). Within each signature were identity genes diagnostic for a specific lineage (Figure 3A). Similar subsets of distinct molecular signatures were identified using the ‘K-means’ clustering algorithm (Figure S2H).

Figure 3. Assigning the Identities of Molecularly Distinct TAC Progenitors.

(A) Heatmap showing relative expression levels of signature genes for each cluster. Cells were ordered by clusters. (B) (left), Gene ontology analyses of subclusters displaying enriched IRS and HS genes (>2X). (right), tSNE plots show that Cluster C2 is composed of HS (C2-a,b,c) and IRS (C2-d,e,f) subclusters, which were identified according to transcriptome. Genes in red are regulated by TAC ‘super-enhancers’, reflective of key lineage identity genes (Adam et al., 2015). (C) (left), tSNE plot showing high Gata3 expression in C2-d and C2-e IRS populations. (right), Immunofluorescence confirms the identity of these sub-clusters. (D) (left), tSNE plot showing high Prdm1 expression in C2-a, C2-b, C2-c and C2-f. (right), Immunofluorescence confirms the identity of these sub-clusters. (E) tSNE plot showing two sub-clusters of C3, which represent companion layer (CP, Krt75⁺) and lower proximal cup (LPC, Lgr5⁺). (F) Table summarizing certain genes diagnostic for assigning identities to proliferative basal TAC subclusters. In all, each subcluster displayed a signature of ~100 genes up by ≥2X relative to other subclusters. Scale bars = 50μm. See Table S3 for full signature lists.

By coupling these transcriptome landscapes with immunolabeling for well-established lineage markers, we validated the identities of each sub-cluster (Figures 3B–3D). The 6 C2 sub-clusters showed GO-terms and specific transcript profiles of the three lineages of the IRS (C2d-C2f) and HS (C2a-C2c) (Table S3). Known fate markers were differentially expressed across signatures, enabling assignments for the basal TAC progenitors of the Henle, Huxley or cuticle of IRS, and the medulla, cortex and cuticle of HS, respectively (representative examples in Figures 3B–3D). Analogously, C3a expressed Krt75, a marker of the companion (Cp) layer, while C3b expressed Lgr5, active in the ‘lower proximal cup’ (LPC), a term for the cells at the very base of the HF bulb (Figure 3E). The LPC transcriptome bore similarities to HFSCs and ORS, albeit with reduced overall marker expression. Although their function remains unknown, previous lineage tracing and/or pulse-chase experiments suggest that LPC cells are short-lived (Sequeira and Nicolas, 2012) and may be precursors to Cp progenitors (Alonso et al., 2005).

Figure 3F highlights key distinctive features of each signature. Taken together, our data suggest that the proliferative subset of FACS-purified, integrin-rich TACs on the basement membrane represent an already distinct cohort of 7 different progenitors and the LPC.

Molecularly Distinct TAC Progenitors Are Spatially Arranged and Divide Perpendicularly Relatively to the Basement Membrane

Their distinctive molecular signatures suggested that lineage-specific TAC progenitors may be spatially organized along the basement membrane. We examined this possibility by conducting lineage-tracing, making sure to mark these progenitors after follicle regeneration had begun but prior to HF maturation. We therefore devised a strategy based upon a tamoxifen inducible CreER driven by active TGFβ signaling, known to be triggered transiently during late telogen to Ana-II (Oshimori and Fuchs, 2012)1.

Using our in utero lentiviral delivery system (Beronja et al., 2010), we selectively and stably transduced skin epithelium of Rosa26-Lox-Stop-Lox-Confetti mice, and then at Ana-I, we administered tamoxifen. As judged by pSMAD2 immunolabeling, TGFβ signaling was active in emerging nascent TACs (Figure 4A). At Ana-VI, columnar patterns of lineage-traced cells were seen, similar to those described previously with Rosa26-LacZ mice and a ubiquitously expressed promoter for Cre (Legué and Nicolas, 2005). Importantly, however, with Confetti and a spatially and temporally defined Cre driver, we identified single lineages marked with different fluors within a HF (Figures 4A and S3A). Together, these data provided compelling evidence that the epithelial-DP interface within the hair bulb is composed of micro-niches that confer unique molecular signatures and lineage specificity in a spatially organized fashion.

Figure 4. TACs Consist of Spatially Arranged Uni-Lineage Progenitors That Divide Asymmetrically Relative to the Basement Membrane.

(A) (left), Lineage tracing strategy with R26-Confetti reporter mouse and lentivirus expressing CreER driven by pSMAD2-binding elements (SBE, TGFβ-responsive element), active only at late Telo-AnaII. (middle), In Ana-II, pSMAD2 labels progenitors adjacent to the DP. (right), Lineage-tracing marks cellular columns of uni-lineage cells. Scale bars = 50μm. (B) Ultrastructure of mature hair bulb. TACs, orange; pre-medulla cells, purple; mitotic cells, green. (C) Basal TACs divide perpendicularly to the basement membrane. Late anaphase daughters are marked by SURVIVIN (example shown), allowing quantification of division angles relative to the underlying basement membrane. Scale bars = 10μm. (D) Schematic depicting basal TACs as unipotent progenitors. See also Figure S3.

To understand how these progenitors generate their respective clonal lineages, we first examined the hair bulb by ultrastructure (Figure 4B). Focusing on the basal epithelial cells juxtaposed to the DP, the upper cells (purple) harbored melanin granules, early features of the medulla lineage and consistent with the transcriptome and non-dividing status of the C1 sub-cluster. By contrast, the lower cells (orange) frequently displayed mitotic figures (green), in agreement with the proliferative status of EdU⁺ IRS, HS and Cp TAC progenitors. Tracing a single columnar trail of cells curving upward, mitoses were both on and off the basal lamina that separated the thin strand of DP from the epithelium (shown). With the exception of the ORS however, mitoses were confined to the hair bulb.

Probing deeper, we determined the mitotic division planes of progenitors by using immunofluorescence for integrins to mark the basal cell surface, and SURVIVIN to mark the septum of late anaphase daughter cells. Quantifications revealed that these TACs divided nearly exclusively asymmetrically to produce one basal progenitor and one suprabasal cell (Figure 4C). Moreover, since sequential suprabasal divisions followed the plane of a single layer, the results implied that even when still proliferative, suprabasal daughters were primed to differentiate.

Our findings predicted that basal TACs should be less differentiated than their suprabasal progeny. To test this, we FACS-purified and transcriptionally profiled the two populations. As shown in Figure S3B, suprabasal progeny were enriched for hair differentiation markers, whereas basal progenitors preferentially expressed signaling factors of the WNT, BMP, FGF and NOTCH pathways, suggestive of their ability to influence the differentiation programs fated in their suprabasal progeny. Together with our single cell, lineage tracing and spindle orientation analyses, these data establish that TACs are spatially organized and already primed to divide asymmetrically and generate the concentric differentiating layers of the hair and its channel (Figure 4D).

Tracing the Origins of Heterogeneity: Evidence for Multipotent Progenitors

Since TGFβ-signaling in developing TACs waned after Ana-II, the TACs that were Confetti-marked by activating Cre expression with a pSMAD2 binding element (SBE) must have existed well-before HF maturation. This became particularly intriguing when we noticed that some cells marked by Confetti gave rise to multiple lineages (Figures 5A and S4). Since adjacent proliferative basal cells along the basement membrane displayed distinct molecular signatures, the presence of multiple closely juxtaposed progenitors marked by a single Confetti fluor and displaying similarly marked upward descendants suggested that some progenitors at earlier stages of the regenerative phase must be multipotent. Given the concentric organization of lineages, immunofluorescence with one lineage marker was often sufficient to allow assignments of lineages derived from a single fluor. By these criteria, HS, IRS and even IRS/HS lineages were often marked by a single fluor.

Figure 5. Lineage Restriction Occurs in a Temporally Sequential Manner.

(A) Lineage-tracing (as in Figure 4) also gave rise to adjacent lineages marked by same fluor, suggestive of multi-lineage progenitors born at an earlier stage in the regenerative phase. (B) Single cell analysis of Ana-II TACs reveals multi-lineage progenitors with features of both IRS and HS, as well as unipotent progenitors of the lower proximal cup (LPC) and companion layer (CP). (C) Heatmap shows relative expression levels of signature genes for each cluster. Cells are ordered by clusters. (D) (top), tSNE plots show co-expression of IRS lineage marker (Nrp2) and HS lineage marker (Hoxc13) in multipotent progenitors. (bottom), Multipotent progenitors are enriched for key IRS and HS lineage progression genes. Genes in red are enriched in Ana-VI IRS TACs, genes in blue are enriched in Ana-VI HS TACs and genes in black are enriched in both lineages. (E) Dual expression of key IRS (GATA3) and HS (HOXC13) lineage TFs in a subset of Ana-II TACs. Note that by Ana-III, these TFs have become restricted to their different lineages. (F) (top), tSNE plots showing high Krt79 and Gata6 expression in companion (CP) progenitor population. (bottom), Some representative genes enriched in CP cluster (>2-fold). (right), GATA6 immunostaining illustrates the emergence of CP unipotent progenitors (arrows) at Ana-II. (I) Venn Diagrams show similarity overlaps between Ana-II and Ana-VI progenitors. Scale bars = 50μm. See also Figures S4 and S5 and Table S4.

Seeking the roots of these putative multipotent progenitors, we devised a strategy to purify Ana-II TACs corresponding to the TGFβ-signaling basal progenitors (Figures S5A–S5D). We then performed RNA-seq analyses on 206 single cells from this population, achieving similar efficacy to what we had obtained with Ana-VI TACs (Figures S5E–S5G). Principal component analysis revealed the emergence of a small cohort of Ana-II cells that co-expressed transcripts critical for HS and IRS lineage specification and differentiation (Figures 5B–5D and S5H; Table S4). Notably, however, even with 2^nd-level clustering analysis, progenitors displaying restricted features of only one of these two tissue structures were not detected (data not shown). Together with our lineage tracings, these data indicated that nascent IRS/HS-TACs possess multi-lineage potential.

Immunofluorescence of differentiation-specific markers corroborated the existence of multi-lineage progenitors, as revealed by co-expression of their master regulators, GATA3 (IRS) and HOXC13 (HS) (Figure 5E). By monitoring the resolution of these and other factors from our single cell analyses, we confirmed that IRS-specific lineages only emerged at Ana-IIIa while HS-specific lineages emerged at Ana-IIIb (Figures 5E and S5I). These findings reveal a temporal hierarchy, with multipotent progenitors existing at Ana-II and then bifurcating into discrete lineage progenitors in Ana-III.

Single cell analyses revealed two basal progenitor populations at Ana-II. Similar to the LPC of mature HFs, one cohort displayed characteristics of ORS, but at lower levels. The other cohort expressed Cp markers, Krt79 and Krt6a, as well as hitherto undescribed transcription factor (TF) genes Gata6, Grhl3, Nfatc2 and Oct6 (Figures 5F and S5J; Table S4). GATA6 immunostaining assigned these cells as the outermost Ana-II TACs (Figure 5F). Interestingly, however, inner Ana-II suprabasal cells flanking the ORS were GATA6⁺ K79⁺, establishing the Cp layer as the first TAC progenitor to emerge in the regenerative phase of the hair cycle. Its appearance was only superceded by the ORS, which generates the multipotent HFSCs for the next hair cycle (Hsu et al., 2011).

Finally, we compared all three basal populations of Ana-II nascent TACs to the 8 different basal populations of mature Ana-VI TACs. As shown in Figure 5G, Ana-II and Ana-VI LPC progenitors showed significant overlap as did Cp progenitors. By contrast, Ana-II multipotent progenitors shared only minor similarities to unilineage IRS and HS progenitors, reflective of their uncommitted state. Thus, lineage restriction is a dynamic process that occurs in a step-wise fashion from the outside→in during the tissue regenerative phase.

Lineage Diversity Begins During Quiescence

We next probed for possible diversity within the bulge and HG stem cell populations. In the resting (telogen) phase, both bulge and HG are morphologically indistinguishable and quiescent. At anagen entry, HG cells closest to the DP become proliferative (Greco et al., 2009; Hsu et al., 2011). We observed that like their TAC progeny, these cells divided asymmetrically relative to the basement membrane (Figures 6A and S6A). Employing Lgr5-eGFP and surface markers, we FACS-purified and performed RNA-seq on 39 telogen bulge cells, 71 telogen HG cells, and 74 Ana-I HG cells (Figures S6B–S6D). Clustering analysis of 174 of these cells revealed distinct patterns (Figure 6B).

Figure 6. Lineage Diversity Begins During Quiescence.

(A) (left), Ultrastructure of HF at telogen and Ana-I. Bulge-HFSCs in green; hair germ (HG) in red; DP in blue. Note perpendicular anaphase mitosis in yellow. Scale bars = 25μm. (B) tSNE plot showing that unbiased clustering of telogen HFSCs, telogen HG cells and Ana-I HG cells identifies 6 distinct progenitor states. Each cell is represented as a dot and colored by a clustering algorithm. (C) Heatmap shows relative expression levels of signature genes for each cluster. Cells are ordered by clusters. (D) (left) Table shows a list of representative signature genes for each cluster. (right) Marker-based identification (tSNE plots) of HFSC and HG clusters. Shown are expression patterns of 6 representative signature transcripts. (E) The switch between super-enhancer regulated HFSC and TAC identity genes occurs precociously in Telo-HG3, revealing a strong regional bias within primed basal SCs of the telogen HG. (F) Chromatin accessibility (profiled by ATAC-seq) of the genes encoding the 6 signature transcripts from (D) and assayed in HG versus bulge SCs in mid-telogen. Note that the four representative signature genes of the bulge HFSC cluster have reduced chromatin accessibility in HG, while the two representative signature genes of HG clusters have lower accessibility in bulge. See also Figure S6 and Tables S5.

Bulge HFSCs (Telo-HFSC) were homogenous and distinguished by Cd34+Krt24+ (Figures 6B–6D), consistent with a recent study on unfractionated telogen skin cells (Joost et al., 2016). By contrast, telogen-phase HG cells constituted three distinct populations, which by EdU and cell cycle analyses were quiescent like the bulge (Figure S6E). While HG cells still displayed features of HFSCs, they varied in expression patterns. The most divergent were HFSC-determinants whose genes are regulated by “super-enhancers,” and thus signify their importance in stemness (Adam et al., 2015) (Figure 6E). Conversely, a small cohort of key super-enhancer regulated TAC lineage determinants showed the opposite gradient of activity. In total, ~25% of transcripts upregulated by ≥2X in Telo-HG3 versus Telo-HG1 cells were encoded by genes which were expressed by TACs within mature HFs (Table S5).

Upon anagen entry, two additional populations of the expanding HG emerged (Figures 6B–6D). Compared to telogen phase HG populations, these Ana-I HG cells robustly expressed cell cycle-related genes, reflecting their proliferative status at this time (Figures 6A–6B and S6E–S6F) (Greco et al., 2009). Overall, comparative analyses of telogen and Ana-I subclusters revealed molecular signatures with ≥ 100 transcripts changed by ≥ 2X (FDR<0.3) relative to any of the other populations (Figures 6C and 6D, Table S5).

The distinctions in telogen-phase transcriptomes of HG subpopulations were particularly insightful. Such early indications of lineage priming had not been unearthed by bulk RNA-sequencing (Greco et al., 2009). If our early transcriptome heterogeneities are genuine, then they should be reflected at the level of chromatin dynamics. To test this, we purified telogen-phase HG and bulge populations and subjected them to ATAC-seq (Buenrostro et al., 2013), which reveals open chromatin domains of their genomes. As illustrated by the representative comparisons, the chromatin patterns mirrored transcriptome signatures (Figure 6F). These data further confirmed the heterogeneity within quiescent primed (HG) SCs. Notably, this heterogeneity preceded SHH, one of the earliest functionally relevant markers known for nascent TACs (Figure S6G) (Hsu et al., 2014).

BMP Inhibition and WNT Signaling Are At the Roots of Diversity Within the Quiescent Hair Germ

To probe the possible downstream functional importance of the molecular heterogeneity within the HG, we first examined the status of BMP inhibition and WNT activation in these cells, as these antagonistic signals are known to be critical for activating the hair cycle (Greco et al., 2009). Both Telo-HG3 and AnaI-HG2 cells downregulated negative WNT regulators, e.g. Sfrp1 and Dkk3, featured by HFSCs. They also upregulated Wnts (3, 4, 6, 10b) and Ctnnb1, encoding WNT effector β-catenin (Figures S6F and S6H), and suppressed positive regulators of BMP-signaling, e.g. Nfatc1 (Figures 6D–6F). Importantly, the WNT-inhibitory/BMP-high pathways that normally maintain SC quiescence were progressively and dynamically changing in unison across individual HG cells at the telogen→anagen transition.

Immunofluorescence lent spatial context to the emerging heterogeneity in these signaling states. Thus, when coupled with EdU labeling to monitor proliferation, it was clear that even during the quiescent phase, BMP signaling, as evidenced by phosphorylated SMAD1/5/9, was waning in the HG cells juxataposed to the DP (Figure 7A). Associating telogen transcriptomes with these patterns, BMP target genes, e.g. Id2 and Sbsn, were expressed robustly in bulge SCs and HG1 cells, but progressively declined in HG2 and HG3 cells (Figure 7A).

Figure 7. Differential BMP and WNT Signaling Establishes Compartmentalization of Epithelial-Mesenchymal Micro-niches.

(A) (left), pSMAD1/5/9 immunofluorescence reveals heterogeneity within the telogen HG, reflected in corresponding repression of BMP target genes Id2 and Sbsn. (middle) pSMAD1/5/9 in Ana-I and Ana-III HF. (right), Genes regulated by BMP (Genander et al., 2014) correlate with SMAD1 activity in primed SC (HG) populations. (B) Gene set enrichment analysis showing similarities between Bmpr1a-null HFSCs and HG3 showing low BMP signaling. (C) (left), Immunofluorescence and single cell analyses reveals heterogeneity among quiescent HG cells in their suppression of SOX9/Sox9 concomitant with induction of LEF1/Lef1 induction. Note the appearance of LEF1⁺ DP cells opposite the Lef1^Hi Telo-HG3 cells (arrow), and the broadening of this zone with onset of regenerative phase. (D) Ana-VI DP was purified as LEF1-RFP⁺Integrinα9⁺K14-H2BGFP⁻ cells. (E) Heterogeneity in basal TACs is mirrored by heterogeneity in DP cells. (left), tSNE plot showing that unbiased clustering of FACS-purified Ana-VI DP cells identifies 4 subpopulations. Each cell is represented as a dot and colored by a clustering algorithm. (right) Heatmap shows relative expression levels of signature genes for each cluster. Cells are ordered by clusters. (F) In situ hybridization and immunofluorescence to spatially map Ana-VI DP subpopulations. (G) Patterns of WNT and BMP signaling and expression of downstream target genes in Ana-VI DP clusters. (H) Schematic depicting spatial organization of epithelial-mesenchymal cross-talk between opposing TAC and DP micro-niches in Ana-VI HF. Scale bars = 25μm. See also Figure S7 and Table S6.

Next, we asked how these spatial changes in signaling pathways impact the transcriptome as a whole. In comparing our single cell analyses to previously identified pSMAD1-bound target genes influenced by ≥ 2X in WT versus Bmpr1a null bulge SCs (Genander et al., 2014), we found that positively influenced BMP target genes were progressively down-regulated from HG1→HG3, while negatively influenced BMP targets showed an upward, albeit more modest, concomitant trend (Figure 7A). Thus, BMP responsiveness is one of the primary driving forces that generates spatial heterogeneity of telogen HG cells. Moreover, based upon gene set enrichment analyses (Subramanian et al., 2005), the HG3 signature, associated with reduced BMP-signaling, reflected a gradual process of the transitioning to SC activation and fate specification (Figure 7B). This was particularly evident by the enrichment of genes related to translational regulation, the mTOR pathway and TCA cycle, specifically induced in HG3 (Figure S6I).

By contrast, nuclear LEF1, a faithful proxy for WNT-signaling in HF lineages (Ouspenskaia et al., 2016), was seen in telogen HG cells closest to DP, with expansion upon anagen entry (Figures 7C). Shortly thereafter, nuclear LEF1 appeared in the DP neighboring the LEF1⁺ HG cells. As the transition to anagen progressed, both epithelial and DP LEF1 signals subsequently spread outward (shown). Conversely, SOX9, a master regulator of the SC and ORS fates, was downregulated in a pattern spreading inward (Figure 7C).

Building Upon Stem Cell Niche Heterogeneity During Tissue Regeneration

Together, our findings suggested that the spatial arrangement within the niche of primed, but quiescent HFSCs establishes the blueprint for the orchestrated plan in temporal and spatial lineage establishment during the regenerative phase. If indeed, the progressive choice of lineages and restriction of lineage plasticity arises by exploiting ever-changing downstream epithelial-DP dynamics, then this should be reflected in the DP. Indeed, when we purified the DP of mature HFs and performed single cell RNA-seq analyses, we identified four distinct DP subpopulations (Figures 7D, 7E and S7; Table S6). Interestingly, many of the known DP markers (Greco et al., 2009; Hsu et al., 2014; Rendl et al., 2008; Woo et al., 2012) were differentially distributed within the subpopulations.

We used immunofluorescence and in situ hybridization to spatially map the clusters along the highly elongated strand of mature DP cells that flanked the unilineage TAC progenitors (Figure 7F–7G and S7G). Together with their distinctive molecular signatures, signs of potential epithelial-DP crosstalk between opposing TAC and DP micro-niches began to emerge. For example, Wnt10b, encoding the major WNT ligand in the hair bulb, localized asymmetrically to two IRS progenitors, while nuclear LEF1, a proxy for WNT signaling (Ouspenskaia et al., 2016), was seen both in IRS lineages and mid-DP. Notably, R-spondins, potent stimulators of WNT-signaling, were produced by these mid-DP cells, providing an explanation for the symmetric patterns of WNT-signaling despite asymmetric ligand distribution. A number of DP genes known to be affected by loss of β-catenin, including Fgf7, Fgf10 and Hoxa9 (Enshell-Seijffers et al., 2010), also showed strong expression in these mid-DP clusters.

Digging deeper, we unearthed a gradient in BMP-signaling from the upper to lower DP cells. Strikingly, BMP-signaling effector pSMAD1 and potent BMP inhibitor NOGGIN, were at opposite poles of the DP (Figures 7F–7G and S7G). Previously reported BMP targets, including Sbsn, Alpl and Bmp6, were highly expressed in these upper DP clusters where BMP-signaling was high. These data not only enabled spatial assignments of the DP clusters, but also revealed new insights into how different patterns of BMP and WNT signaling become established in TAC lineages.

It was particularly intriguing to place the WNT and BMP signaling patterns we found in mature DP clusters in the context of patterns that arose temporally in the DP during HF regeneration. In telogen DP, where BMP inhibitors are known to be high (Greco et al., 2009), both WNT and BMP signaling were absent (Figures 7A–7C). During early anagen, DP WNT-signaling became prevalent, but clearly preceded DP BMP-signaling. These temporal DP patterns during the regenerative phase recapitulated the bottom→top DP heterogeneity in the mature follicle (Figures 7F–7H). Moreover, the temporal establishment of the TAC progenitors matched these temporal DP states. Strikingly, the first TAC lineage progenitors to emerge resided in micro-niches opposite the bottom DP cluster, while the last TAC progenitors to emerge resided opposite to the top DP cluster.

Together, these findings provide compelling evidence that heterogeneity established in the primed quiescent SCs and determined by proximity to the DP triggers downstream crosstalk at the epithelial-DP interface. Our data further suggest that by compartmentalizing progeny into distinct micro-niches at this interface, where WNT-BMP crosstalk occurs, the niches can be dynamically elaborated and stem cell lineage plasticity can become increasingly restricted during the regenerative phase.

DISCUSSION

Stem Cells Generate Progenitors in Response to Spatially and Temporally Orchestrated Patterns

Studies over the past decade have placed increasing emphasis on the importance of the niche in dictating the behavior of tissue SCs (Morrison and Spradling, 2008; Scadden, 2014). Much less is known about the cues that instruct activated SCs along particular lineage pathways. Even for epithelial tissues where spatial and temporal dissection of lineage pathways is possible and has been tracked by pulse-chase experiments and live imaging (Hsu et al., 2011; Rompolas et al., 2013), it has remained largely a mystery as to how primed SCs give rise to distinct progenitors.

Single cell analyses enabled us to gain inroads into this important problem through a high-resolution view of lineage diversification. Starting with the established proliferative SC progeny of the mature HF, we learned that in contrast to prevailing notions, these cells are not homogeneous. Rather, at least 7 distinct TAC populations reside at the epithelial-DP interface, where they divide asymmetrically to generate the differentiating layers of the HS and IRS and companion layer. During the peak of hair regeneration, this single basal layer of specified lineage progenitors provides the elaborate tissue heterogeneity seen in this complex mini-organ.

Our TGFβ-driven CreER reporter allowed us to trace the roots of progenitor diversity. Interestingly, formation of unipotent progenitors occurs spatially and sequentially during the early stages of follicle regeneration. The HF’s synchronized and orchestrated program of tissue regeneration permitted temporal transcriptome landscaping of single cells in a way not possible with other SC systems. This revealed that bulge SCs are homogeneous, but heterogeneity already exists within the niche of primed SCs and continues to become more elaborate in downstream proliferative progenitors.

Importantly, the timing and complexity of progenitor heterogeneity unfolded concomitantly with lineage diversification. The heterogeneity arose in a highly choreographed process of temporal and spatial steps. This enabled us to identify clear lineage relations between long-term SCs, primed SCs, committed multi-lineage progenitors and unipotent progenitors. It also revealed links to the dynamically changing climates that arise as the follicle regenerates and the progenitor’s neighbors and signaling inputs elaborate.

Heterogeneity with a Purpose: The Existence of Micro-Niches

The spatially and temporally orchestrated plan of progenitor heterogeneity implies the importance of a progenitor’s microenvironment in these lineage choices, and in this regard is similar to the intimate relation between SCs and their niche. Our studies revealed that whether in the SCs or progenitors, the heterogeneity was rooted at the epithelial-DP interface and was found in both compartments. The first signs of heterogeneity were detected in telogen, where SCs juxtaposed to DP cells displayed nuclear LEF1, reflective of WNT-signaling and inhibition of BMP-signaling. By full anagen, heterogeneity in LEF1 was still found, this time not only in the unilineage TACs but also in the DP. In this case, however, the heterogeneity within DP and TACs was elaborate, with signs of WNT-signaling sandwiched between BMP inhibition at the bottom and BMP signaling at the top. Notably, we unearthed striking parallels between a) the spatial arrangement of TAC progenitors and WNT/BMP signaling in DP clusters in mature HFs and b) the concomitant timing at which TAC lineages were specified and DP signaling patterns emerged during the regenerative phase. These findings suggest that during tissue regeneration, both stem cell progeny and DP continue to generate diversity as they encounter ever-changing environmental inputs.

When coupled with the heterogeneity involved, the newfound dependency upon the epithelial-mesenchymal interface adds a new layer of complexity in stem cell biology and lineage determination, namely the existence of ‘micro-niches.’ Our data support this concept even within a quiescent population of stem cells. Indeed, at every step, we found that heterogeneity among the cells within a stem or progenitor was not random or stochastic, but rather existed according to a detailed tissue blueprint. This phenomenon allows stem cells and their downstream progenitors to exploit the full regalia of their environment, from external signaling to spatial and intercellular context, irrespective of whether they rest in quiescence or generate tissue. Moreover, given the recent finding that SC progeny can have a profound impact on their parents (Hsu et al., 2014), it is notable that the tissue and its progenitors dynamically remodel during HF reformation. Although many details of the heterogeneous epithelial-DP cross-talk await future investigation, our findings provide new insights into why the failure to either respond to a niche signal or execute a specific lineage causes an excess in the progenitor population for the adjacent lineage(s) (Genander et al., 2014; Kaufman et al., 2003).

Fine Tuning Stem Cell Behavior Through Micro-Niches: Implications for Tissue Regeneration

Our results also highlight why heterogeneity needs to arise early in the regenerative process, even in the quiescent SC population: each step in organ development or regeneration must be spatially and temporally choreographed to set the stage for the subsequent step. Through polarized epithelial-mesenchymal interactions, not only are some SCs primed to proliferate first, as reported previously (Greco et al., 2009), but as we show here, they are also primed in a rigid spatial context to launch this protocol for tissue regeneration such that each lineage becomes specified at the right time and place. By sub-compartmentalization, a niche can dole out its SCs to perform distinct lineage tasks. This feature has broad implications for epithelial-based organs, where spatial context and intercellular interactions are dynamic during embryogenesis and wound-repair, as well as in natural bouts of adult tissue morphogenesis, which occurs for instance in HFs and mammary glands.

The existence of micro-niches also allows for coordinating multi-lineage tissue growth both spatially and temporally. Our single cell analyses show that the lineages afforded to the SCs of the regenerating hair follicle become progressively defined from the outside→in, and in a spatially and temporally orchestrated pattern. By building in this way, morphogenesis of complex epithelia such as the HF, mammary gland, sweat gland, and lung, can first establish their surface/lining barrier, and then exploit newly emerging environmental signals and intercellular interactions that become more elaborate as the tissue forms.

The stem cell road map that we’ve unfolded here for HFs bears some intriguing similarities to the well-established lineage hierarchy of the hematopoietic system, where HSCs go through stages of multipotency and finally unipotency (Kondo, 2010; McKinney-Freeman et al., 2012; Morrison and Weissman, 1994). While it has long been surmised that circulating HSCs initiate these developmental programs, recent studies with single cell RNA sequencing and barcoding techniques have identified lineage restricted SCs in the bone marrow (Paul et al., 2015; Perié et al., 2015). Although it remains unclear where and how these HSC fate decisions and lineage restrictions are made, it is tempting to speculate that there are distinct micro-niches within bone marrow that either impact fate choice or serve as a local residence for distinct SC populations that were specified at an earlier developmental stage.

In closing, our combination of in vivo single-cell RNA-seq, lineage tracing and chromatin analysis uniquely poised us to exploit the spatial and temporally well-defined landscape of the cycling HF to tease apart the regulatory circuitries that govern progressive stages of lineage specification that restrict SC lineage plasticity. This afforded us unprecedented high-resolution precision mapping of the repertoire of TFs and signal transduction pathways that are key for each particular hair lineage. In so doing, we unearthed new insights into how SCs undergo fate specification in response to external cues and how mechanistically this leads to lineage diversification at the chromatin and transcriptional levels. By layering new signaling platforms on previously established ones, the progenitors of a SC lineage can respond to their new environment while preserving features of their behavior established at an earlier stage. In this way, SC progenitors become increasingly set in their destined behavior, as do their mesenchymal neighbors. Finally, although our focus here was on the HF, the paradigms we’ve unearthed suggest that just as the hematopoietic system has helped to shape our view of SC biology, so too will the unique advantages of the HF help to fuel our knowledge of other tissue stem cells.

STAR Methods

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Elaine Fuchs (fuchslb@rockefeller.edu).

EXPERIMENTAL MODELS AND SUBJECT DETAILS

Mice and Procedures

Lhx2-GFP (CD1 background), Lgr5-eGFP-IRES-CreER (C57BL6/J background), Rosa26-tdTomato (C57BL6/J background), Rosa26-Confetti (C57BL6/J background), K14-H2BGFP (CD1 background), LEF1-RFP (CD1 background), CD1 and C57BL6/J strains were used and cross-bred as needed in this study. Mice were maintained in the Association for Assessment and Accreditation of Laboratory Animal Care-accredited animal facility (AAALAC) of The Rockefeller University (RU), and procedures were performed with Institutional Animal Care and Use Committee (IACUC)-approved protocols. For LEF1-RFP transgenic mice generated by a 6.7kb fragment of the human LEF1 promoter, previous studies showed that RFP expression does not faithfully recapitulate the endogenous expression pattern of murine LEF1 protein, but it is active in all DP cells (Rendl et al., 2008). Male mice were used to minimize hair cycle variation and randomly assigned to experimental groups.

In utero lentivirus injections were performed as previously described (Beronja et al., 2010). For lineage tracing of basal TACs, Rosa26-Confetti mice were crossed to outbred CD1 mice, and E9.5 embryos were used for in utero injection of lentivirus harboring a TGFβ-sensitive SMAD binding element (SBE)-CreER. After birth, transduced mice were tested by genotyping using a set of Cre specific primers, and positive mice were injected with 20μg/g of tamoxifen intraperitoneally at indicated days and sacrificed at P32 (Ana-VI).

Basal TACs were isolated from Lhx2-GFP mice whose HFs were in Ana-VI. To isolate nascent TACs at Ana-II, Lgr5-eGFP-IRES-CreER mice were crossed with Rosa26-tdTomato mice to create Lgr5-eGFP-IRES-CreER;Rosa26-tdTomato offspring. 50μg/g of tamoxifen was administered to the mice at P21 and P22 (telogen and Ana-I) to label bulge and hair germ (HG) stem cells (SCs), and mice were then sacrificed at Ana-II. To isolate bulge and HG SCs at telogen and Ana-I, Lgr5-eGFP-IRES-CreER mice were used. To isolate dermal papillae cells at Ana-VI, K14-H2BGFP;LEF1-RFP mice were used. For labeling of proliferating cells, EdU (25μg/g) was intraperitoneally injected 4hr before lethal administration of CO₂.

Hair Cycle Timing

The six subdivisions of the hair cycle were determined based on previous studies (Hsu et al., 2014 and references therein). Since hair cycles vary among strains, sexes and individuals, stages instead of exact mouse ages were evaluated and carefully monitored for each experiment. For FACS isolation of bulge and HG SCs at telogen and nascent TACs at Ana-II, the mouse backskin was further divided into three regions (upper, middle and lower) and a small piece of skin for each region was embedded in OCT. Before FACS sorting, hair cycle stages were determined by histological analysis and EdU staining, and the regions showing the desired substage were chosen for FACS-purification.

METHOD DETAILS

Fluorescence Activated Cell Sorting and Analysis

The backskin at indicated hair cycle stages was placed dermis side down in 0.25% collagenase (Sigma) in HBSS (Gibco) for 35 min (Ana-VI) or 20 min followed by additional 10 min in 0.25% trypsin (Gibco) (Telogen, Ana-I and Ana-II). After digestion, single cell suspensions were obtained by scraping the dermal side gently using a blunt scalpel. After washing with FACS buffer (5% FBS in PBS), the cell suspensions were filtered through a 45-μm strainer. Cells were pelleted and washed once with FACS buffer. Cell suspensions were incubated with the appropriate antibodies in FACS buffer for 10 min on ice and washed with FACS buffer.

The following antibodies were used for staining: eFluor660-CD34 (1:100), PerCP-Cy5.5-Sca1 (1:1,000), Alexa647-CD29 (1:500), PECy7-CD29 (1:500), PerCP-Cy5.5-CD49f (1:500), P-Cadherin (1:100), Alexa555-anti-goat secondary antibody (1:500), Itga9 (1:100), Alexa647-anti-goat secondary antibody, Biotin-CD31 (1:100), Biotin-CD45 (1:500), Biotin-CD117 (1:100) and Biotin-CD140a (1:100). For un-conjugated or biotin-conjugated primary antibodies, after washing with FACS buffer, cells were incubated with the following secondary antibodies: Alexa647-anti-goat IgG secondary antibody, BV421-Streptavidin (1:300) and PerCPCy5.5-Streptavidin (1:300). DAPI (0.2 μg/ml) was used to exclude dead cells. Single-cells were sorted using a BD FACSAria II (BD Biosciences) into 96 well PCR plates (Bio-Rad) containing 2μl of lysis buffer (0.2% Triton X-100, 2U/μl RNaseOUT (Thermo Scientific), 0.25μM oligo-dT₃₀VN primer and 1:8,000,000 diluted ERCC spike-in RNAs (Ambion)). Doublets and dead cells were excluded based on forward scatter, side scatter and DAPI fluorescence. Sorting was done using the 100 μm nozzle setting with the sort mode set to ‘single cell’. After sorting, plates were briefly centrifuged, snap-frozen in liquid nitrogen and stored at −80°C. For bulk RNA-seq, 30,000 cells were directly sorted into 500μl TRIzol LS (Invitrogen) and stored at −80°C. FACS analyses were performed by FACSDiva software (BD Biosciences).

Single cell cDNA synthesis and sequencing library generation

Single cell RNA-seq libraries were prepared using the Smart-seq2 protocol with a few modifications (Picelli et al., 2014). Briefly, plates were thawed on ice and incubated for 3 min at 72°C to lyse cells. Revers transcription (4U/μl Maxima H-transcriptase), template switching reaction and PCR pre-amplification (15 cycles) were performed according to the protocol. PCR product was cleaned up using 0.8X AMPure XP beads (Beckman Coulter) and the quality and quantity of cDNA libraries were measured by Agilent 2200 TapeStation and Qubit Florometer (Thermal Fisher Scientific). To exclude empty wells and low quality libraries, the cDNA libraries were assessed by qPCR with a primer pair of GAPDH, a housekeeping gene (Table S8). Samples showing high Ct value (> 35 cycles) or no PCR amplification were discarded. 50–100pg of each cDNA library was used for generating Illumina sequencing library using a Nextera XT DNA library preparation kits (Illumina). After the final PCR amplification, samples were pooled and cleaned by 0.9X AMPure XP beads. The pooled sequencing libraries were sequenced on an Illumina Nextseq 500 instrument using a 75 bp paired-end-reads setting.

Bulk RNA-seq and Quantitative RT-PCR

Total RNA was purified using the Direct-zol RNA MicroPrep kit (Zymo Research). Briefly, after adding 500μl of 100% ethanol to samples, the lysate was loaded to RNA binding column. DNase I treatment was done on the column for 15 min at room temperature. After several washing steps, the RNA was eluted by DNase/RNase-free water. Quality of RNA samples was determined using an Agilent 2100 Bioanalyzer, and all samples for sequencing had RNA integrity (RIN) numbers >9. cDNA library construction using the Illumina TrueSeq mRNA sample preparation kit was performed by the Weill Cornell Medical College Genomic Core facility (New York, NY), and cDNA libraries were sequenced on an Illumina HiSeq 2000 instrument. For real-time qRT-PCR, equivalent amounts of RNA were reverse-transcribed by Maxima reverse transcriptase (Thermo Fisher Scientific). cDNAs were normalized to equal amounts using primers against Gapdh or Ppib2. cDNAs were mixed with indicated gene-specific primers listed in Table S8 and SYBR green PCR Master Mix (Sigma), and qRT-PCR was performed on an Applied Biosystems 7900HT Fast Real-Time PCR system.

ATAC-seq

ATAC-seq libraries were made from freshly FACS-sorted SCs. Library preparation and analyses were performed as described (Buenrostro et al., 2013). Briefly, FACS-sorted cells were pelleted and incubated with cold lysis buffer (10 mM Tris-HCl, pH 7.4, 10 mM NaCl, 3 mM MgCl2, 0.1% IGEPAL CA-630). After removing lysis buffer by centrifugation, samples were immediately subjected to a transposition reaction at 37°C for 30 min with 10 μL transposase enzyme (Illumina Nextera DNA Preparation Kit). Transposed DNA was purified using QIAGEN MiniElute PCR purification kit and PCR amplified with 10–15 cycles. Library concentration and quality was assayed by D1000 Tape Station (Agilent) prior to sequencing. The samples were sequenced on Illumina HiSeq 2000 using a 50bp paired-end-reads setting. Extensive data analyses will be published elsewhere (Y Ge and E Fuchs). For the purposes of the current study, ATAC data were only used to substantiate the transcriptome differences that were found in 6 key regulatory genes (Figure 6F).

Immunofluorescence and Microscopy

For immunofluorescence analysis, mouse backskins were dissected and either embedded directly in OCT (Tissue Tek), or fixed with 2% paraformaldehyde (PFA) in PBS for 1 hours at 4°C to preserve the cytoplasmic fluorescent signal and then incubated with 30% sucrose for overnight followed by embedding in OCT. Frozen tissue blocks were sectioned at 12 μm on a cryostat (Leica), and mounted on SuperFrost Plus slides (Fisher). Non-prefixed tissue sections were incubated with 2% PFA for 5 minutes and rinsed 5 times with PBS. For SHH antibody staining, antigen retrieval was performed with 1X HIER buffer (10mM Tris, 1mM EDTA. pH 9.0) at 60°C for 1 hour. The tissue sections were blocked for 1 hour at room temperature in blocking solution (5% normal donkey serum, 0.5% bovine serum albumin, 2.5% fish gelatin and 0.3% Triton X-100 in PBS). Sections were incubated with primary antibodies diluted in blocking solution at 4°C over night. For mouse primary antibodies, prior to tissue blocking, endogenous mouse IgG was blocked by M.O.M blocking for 1 hour at room temperature. Sections were then washed three times with PBS and incubated with secondary antibodies in blocking solution at room temperature for 1 hour. Finally, sections were washed three times with PBS and mounted with ProLong Gold Antifade Mountant (Thermo Fisher Scientific).

The following antibodies and dilutions were used: GFP (1:1000), RFP (1:2000), SOX9 (1:1000), LEF1 (1:200), CUX1 (1:200), GATA3 (1:200), HOXC13 (1:100), GATA6 (1:100), PRDM1 (1:100), pSMAD1/5/9 (1:100), pSMAD2 (1:1000), HES1 (1:500), SEMA5A (1:100), NOGGIN (1:50), K6 (1:1000), K24 (1:2000), K31 (1:1000), K32 (1:1000), K40 (1:1000), K71 (1:1000), K79 (1:200), TCHH (1:1000), P-CADHERIN (1:200), SURVIVIN (1:1000), CD104/β4-INTEGRIN (1:100), CD29/β1-INTEGRIN (1:100) and CD49f/α6-INTEGRIN (1:100). Alexa Fluor-488, 546, or 647-conjugated secondary antibodies (Life Technologies) were used. For pSMAD2 immunostaining, we used TSA Plus Cyanine 3 System (PerkinElmer) to amplify signals. Nuclei were stained using 4′6′-diamidino-2-phenylindole (DAPI). EdU click-it reaction was performed according to manufacturer’s directions. Images were acquired with an Axio Observer.Z1 epifluorescence microscope equipped with a Hamamatsu ORCA-ER camera (Hamamatsu Photonics), and with an ApoTome.2 (Carl Zeiss) slider that reduces the light scatter in the fluorescent samples, using 20X objective, controlled by Zen software (Carl Zeiss). Images were processed using ImageJ and Adobe Photoshop CS5.

Lentiviral Expression Construct and High Titer Lentivirus Production

To generate pLKO-SBE-CreER construct, we subcloned Smad-binding elemets (SBE) (Oshimori et al., 2015) and CreER into pLKO-MCS vector. Production of VSV-G pseudotyped lentivirus was performed by calcium phosphate transfection of 293FT cells (Invitrogen) with pLKO-SBE-CreER plasmid and helper plasmids pMD2.G and pPAX2 (Addgene plasmids 12259 and 12260). Viral supernatant was collected 46 h after transfection and filtered through a 0.45-μm filter. For in utero lentiviral transduction, viral supernatant was concentrated by ultracentrifugation. Final viral particles were resuspended in viral resuspension buffer (20 mM Tris pH 8.0, 250 mM NaCl, 10 mM MgCl2, 5% sorbitol) and 1μl of viral suspension was injected in utero into E9.5 embryos.

QUANTIFICATION AND STATISTICAL ANALYSIS

Single-Cell RNA-Seq Analysis

1) Read alignment and gene quantification

Sequencing reads from single cell RNA-seq libraries were aligned to the mouse reference genome (Version mm10 from UCSC) combined with sequences for ERCC spike-ins as artificial chromosomes using Bowtie2 (version 2.2.9) (Langmead and Salzberg, 2012) with default parameters for paired-end reads. The expression values of each gene were quantified as transcript per million (TPM) using RSEM (v1.2.30) (Li and Dewey, 2011). TPMs were transformed to log₂(TPM+1). For downstream analyses, cells with <4000 genes detected and genes expressed in <5 cells were removed.

2) Identification of highly variable genes, cell clustering and tSNE visualization

All analyses and visualization of data were conducted in a Python environment built on the Numpy, SciPy, matplotlib, scikit-learn package and pandas libraries (Pedregosa et al., 2011). To distinguish true biological variability from technical noise in the single-cell experiments, we used the statistical model for identifying highly variable genes compared to ERCC spike-ins as described by (Brennecke et al., 2013). Briefly, genes with higher level of variation (above the technical variation) and false discovery rate (FDR) value less than 0.1 were considered as highly variable. Table S7 summarizes the minimal biological dispersion parameters and the number of highly variable genes that were used in the study.

To identify cell clusters, principle component analysis (PCA) was performed on the list of highly variable genes, and the number of significant PCs for clustering analysis was determined by North’s rule of thumb, which means the difference between neighboring eigenvalues must be greater than their associated error (Elbow plot). See Table S7 for the number of significant PCs used for clustering analysis in each data set. The significant PCs were applied to unsupervised hierarchical clustering using Euclidean distance and ward’s method.

For 2nd-level sub-clustering, we applied the analyses of identification of highly variable genes, PCA and unsupervised hierarchical clustering to a selected population. To further validate the 2nd-level sub-clusters, we performed k-means clustering method, which is an independent algorithm of unsupervised clustering. Briefly, to apply k-means clustering method on the C2 dataset, we first determined the optimal number of clusters according to the Bayesian Information Criterion (BIC) for expectation-maximization (Fraley and Raftery, 2011). Mclust (R package) was used to generate multivariate Gaussian mixture models (Fraley and Raftery, 2011). K-means clustering with predetermined cluster numbers resulted in very similar sub-clustering results to the hierarchical clustering analyses.

To present high dimensional data in two-dimensional space, we performed t-distributed stochastic neighbor embedding (tSNE) (Maaten and Hinton, 2008) using the results of PCA with significant PCs as input. The expression level of individual genes was represented by colors indicated in the tSNE plots.

3) Differential expression test and analyses

To identify differentially expressed genes between clusters, normalized raw counts were applied to DEseq package (Anders and Huber, 2010) in R software, and then genes with fold change >2 and FDR < 0.1 or 0.3 were considered to be differentially expressed. Corrected P values (FDR) were calculated using the Benjamini and Hochberg method. Signature genes of each cluster were defined as genes that are >2x differentially expressed, and unique for each subcluster/lineage compared to all other clusters.

Gene Ontology (GO) analyses of differentially expressed genes between HS and IRS lineages and between telogen HG and Ana-I HG were performed using DAVID v6.721,22 (Huang et al., 2009). Gene Set Enrichment Analysis (GSEA) was conducted using Pre-ranked model of GSEA software (Subramanian et al., 2005). Pre-defined gene sets were extracted from our previous study (Genander et al., 2014). Also, we queried the KEGG pathway and REACTOME category curated in the Molecular Signatures Database (MSigDB). Genes were ranked by “Signal2Noise” metric. Pathways enriched with FDR < 0.25 were considered to be significant.

Bulk RNA-seq Data Processing

Sequencing reads from bulk RNA-seq were aligned to the mouse reference genome (Version mm10 from UCSC) using Bowtie2 (version 2.2.9) (Langmead and Salzberg, 2012) with default parameters. The expression values of each gene were quantified as transcript per million (TPM) using RSEM (v1.2.30) (Li and Dewey, 2011). Differential gene expression analyses were performed on normalized raw counts using DEseq package (Anders and Huber, 2010) in R software. Genes with fold change >2 and FDR < 0.1 were considered to be differentially expressed. A list of selected genes was presented by a heatmap with z-scores normalized expression value.

ATAC-Seq Alignment and Visualization

50-bp paired-end reads were aligned to mm10 using bowtie with the parameters –X 2000 and –m 1. Duplicates were removed using Picard. ATAC-seq signal tracks were presented by Integrative Genomics Viewer (IGV) software (Robinson et al., 2011).

Statistics

Statistical and graphical data analyses were performed using Microsoft Excel and Prism 6 (Graphpad) software. All experiments shown were replicated at least twice, and representative data are shown. For all measurements, three biological replicates and two or more technical replicates were used. To determine the significance between two groups, comparisons were made using unpaired two-tailed Student’s t-test.

DATA AND SOFTWARE AVAILABILITY

The accession numbers for the sequencing data reported in this paper are NCBI GEO: GSE90848 (single cell RNAseq), GSE90847 (bulk RNAseq) and GSE96782 (ATACseq).

Supplementary Material

Figure S1. Figure S1. Purification of Hair Follicle Cells from Mouse Skin. Related to Figure 2.

(A) FACS-purification strategy of ORS, basal TACs and suprabasal TACs. Epidermal and upper HF progenitors (Sca1+) and bulge HFSCs (CD34+) were removed along with non-epithelial cells (antibodies shown). (B) q-PCR verified the FACS strategy and measured enrichment of cell-type specific marker genes relative to ORS. Mean and standard deviation are shown (n = 3). P-values from t-test: *P<0.05.

Table S3

Table S3. List of signature genes of C2 and C3 sub-clusters of basal TACs population and HS and IRS signature genes at Ana-VI. Related to Figure 3.

Table S4

Table S4. List of signature genes of clusters of nascent TACs at Ana-II. Related to Figure 5.

Table S5

Table S5. List of signature genes of clusters of bulge and HG SCs at telogen and Ana-I and genes that are differentially expressed between Telo-HG1 and Telo-HG3. Related to Figures 6 and 7.

Table S6

Table S6. List of signature genes of clusters of DP at Ana-VI. Related to Figure 7.

Table S7

Table S7. Parameters for single cell analysis used in this study. Related to STAR methods.

Table S8

Table S8. Primer used for genotyping and qRT-PCR. Related to STAR methods.

Figure S2. Figure S2. Quality Control of Single-Cell RNA-seq Analysis of Basal TACs. Related to Figure 2.

(A) Box plot showing the number of total RNA-seq reads mapped to annotated genes. (B) Box plot showing the percentage of RNA-seq reads mapped to the mitochondrial genome. (C) Box plot showing the number of detected genes (TPM >1). (D) Graph showing the percentage of variance explained by each principle component (PC). The significant PCs determined by North’s rule of thumb (North et al., 1982) are to the left of the red-dotted vertical line. (E) Dot plot showing technical noise fit and inference of highly variable genes using External RNA Controls Consortium (ERCC) spike-in RNAs. Each dot represents each gene plotted with average normalized read count (x-axis) and squared coefficient of variation (y-axis). Magenta dots represent highly variable genes (< 10% false discovery rate). Blue dots correspond to ERCC spike-in data points. The solid red line represents the technical noise fit, and the dashed magenta line marks the expected position of genes with 120% biological squared coefficient of variation (CV²). (F) t-Distributed Stochastic Neighbor Embedding (tSNE) analysis of 2 batches of single-cell RNA-seq libraries shows minimal batch to batch variation. (G) Differential gene expression analyses of clusters 4 and 5 revealed that they have distinct gene expression pattern (1170 and 460 mRNAs differentially expressed by ≥ 2X compared to other clusters) including non-epithelial genes such as Pax7 and Pitx2 (skeletal muscle) and Myocd and Mef2c (smooth muscle), respectively. They were hence omitted from all subsequent analyses. (H) 2nd level sub-clustering of C2 by ‘K-means’ clustering algorithm. (left) Estimating the optimum number of clusters based on the Bayesian information criterion (BIC) (Kumar et al., 2014). (right) tSNE plot showing the results of unsupervised k-means clustering. Note that the clustering results are similar to the results of unsupervised hierarchical clustering (Figure 2F).

Figure S3. Figure S3. Lineage Tracing of Uni-lineage Cells and Transcriptomic Comparison of ORS, basal TACs and their Suprabasal Progeny. Related to Figure 4.

(A) Lineage-tracing marks cellular columns of uni-lineage. Scale bars = 50μm. (B) Heatmap showing relative expression levels of selected genes by bulk RNA-seq of FACS-purified cells. Note that basal TACs and their suprabasal progeny do not express ORS marker genes (Sox9, Lhx2 and Foxe1), but express HF lineage differentiation genes (Lef1, Shh, Wnt10b and Sox21). Suprabasal TACs were enriched for differentiation markers of the IRS and HS (mature keratins), whereas basal TACs preferentially expressed signaling factors of the WNT, BMP, FGF and NOTCH pathways, suggestive of their ability to influence the differentiation programs fated in their suprabasal progeny.

Figure S4. Figure S4. Lineage Tracing of Multi-Lineage Cells. Related to Figure 5.

SBE (TGFβ-responsive element)-CreER driven Confetti-lineage-tracing experiments show co-existence of adjacent lineages marked by same fluor, indicative of multi-lineage progenitors in early stage of anagen (Ana-II). GATA3 (IRS) and K6 (medulla and companion layer) were used for lineage markers. Scale bars = 50μm.

Figure S5. Figure S5. Purification Strategy and Analysis of Bulge and Hair Germ SCs and Nascent Committed TACs. Related to Figure 5.

(A) Genetic labeling strategy of bulge SCs and HG using Lgr5-eGFPiresCreER;R26-tdTomato mice. Lgr5-eGFP labels bulge and HG SCs in telogen, Lgr5 is silenced in the nascent TACs that forms in Ana-II. CreER, also driven by the Lgr5 promoter, can be induced to permanently activate the R26-tdTomato locus. Tamoxifen administered in telogen and Ana-I results in bulge and HG SCs labeled by both eGFP and tdTOMATO. By Aan-II, stem cell progeny are only labeled by tdTOMATO. (B) Epifluorescence analysis from labeling strategy in (A). Scale bars = 25μm. (C) FACS-purification strategy of bulge and HG SCs in telogen and Ana-I and TACs in Ana-II. (D) q-PCR to verify the FACS sorting strategy and measure enrichment of Shh a marker of nascent TACs at Ana-II. (E) Box plot showing the total number of unique RNA-seq reads mapped to annotated genes. (F) Box plot showing the percentage of RNA-seq reads mapped to the mitochondrial genome. (G) Box plot showing the number of detected genes (TPM >1). (H) Heatmap showing progressive temporal induction of hair differentiation markers and concomitant suppression of HFSC markers (252 single cells from Ana-I HGs and Ana-II TACs were analyzed and ranked by PC1 scores). (I) Immunofluorescence pinpoints the appearance of specified GATA3+ IRS progenitors in Ana-IIIa (left) and keratin K40+ differentiating HS cells in Ana-IIIb (right). Scale bars = 50μm. (J) Immunostaining for K79 illustrates the emergence of CP at Ana-II. Scale bars = 50μm.

Figure S6. Figure S6. Differences between Telogen and Ana-I Primed SC Populations. Related to Figure 6.

(A) Immunofluorescence shows proliferative transition between telogen and Ana-I in primed HGs (arrow). Scale bars = 25μm. (B) Box plot showing the total number of unique RNA-seq reads mapped to annotated genes. (C) Box plot showing the percentage of RNA-seq reads mapped to the mitochondrial genome. (D) Box plot showing the number of detected genes (TPM >1). (E) tSNE plot showing the cell cycle status of telogen/Ana-I bulge and HG stem cell clusters. Cell cycle status is predicted by Cyclone (Scialdone et. al., 2015). (F) (left), Gene ontology analysis showing differences between the telogen and Ana-I HG. (right), tSNE plots showing expression of Mcm5 (DNA replication) and Dkk3 (WNT inhibitor). (G) tSNE plot showing the expression level of Shh. Note that Shh is not expressed in telogen and Ana-I HGs. (H) tSNE plots showing that the expressions of Ctnnb1, Wnt10b, Wnt6, Wnt4 and Wnt3 were enriched in Telo-HG3 and AnaI-HG2. (I) Gene Set Enrichment Analysis of genes enriched in Telo-HG3 relative to Telo-HG1. Reactome and KEGG pathway genes were tested.

Figure S7. Figure S7. Purification Strategy and Analysis of DP at Ana-VI. Related to Figure 7.

(A) Immunofluorescence showing the expression pattern of K14-H2BGFP, LEF1-RFP and Integrin α9 at Ana-VI. Note that LEF1-RFP is expressed by DP and upper dermal fibroblasts. After collagenase digestion, only the lower part of dermis was collected, indicative of the faithful separation of LEF1-RFP+ DP from upper dermal fibroblasts. Scale bars = 50μm. (B) FACS-purification strategy of Ana-VI DP. Epithelial cells (K14-H2BGFP+), melanocytes (CD117+), endothelial cells (CD31+) and hematopoietic cells (CD45+) were removed. LEF1-RFP and Integrin α9 were used as positive markers (C) q-PCR verified the FACS strategy and measured enrichment of cell-type specific marker genes. Mean and standard deviation are shown (n = 3). P-values from t-test: *p<0.01. (D) Box plot showing the total number of unique RNA-seq reads mapped to annotated genes. (E) Box plot showing the percentage of RNA-seq reads mapped to the mitochondrial genome. (F) Box plot showing the number of detected genes (TPM >1). (G) tSNE plots showing the differential expression pattern of Sostdc1, Lef1, Fgf10, Sema5a, Nog, Hes1 in the Ana-VI DP.

Table S1

Table S1. Single cell RNAseq quality control metrics. Related to Figures 2, 5, 6 and 7.

Table S2

Table S2. List of signature genes of clusters of basal TACs at Ana-VI. Related to Figure 2.