To date, the transcriptome of attachment cells has not been characterized thoroughly. We therefore analyzed the transcriptome of embryonic day (E) 14.5 attachment cells from the prominent deltoid tuberosity and greater tuberosity of the humerus. With the goal to isolate these cells specifically, we generated a compound mouse by crossing three mouse lines, namely Col2a1-Cre, R26R-tdTomato, and Scx-GFP (see Materials and methods) (Blitz et al., 2013; Sugimoto et al., 2013). Thus, the fluorescent reporter tdTomato labeled Col2a1-expressing chondrocytes, whereas GFP fluorescently labeled Scx-expressing tenocytes. Unexpectedly, the two reporters failed to label the attachment cells that were located in between these two populations. This failure might be due to a missing regulatory element in one of the constructs that were used to produce each transgenic reporter. Nevertheless, the borders between tendon and attachment cells and between cartilage and attachment cells were clearly demarcated (Figure 1A, Figure 1—figure supplement 1A,B). We therefore used laser capture microdissection (LCM) to subdivide the attachment site into three cellular compartments, namely attachment cells, adjacent tenocytes, and adjacent chondrocytes. As controls, samples were also taken from two more compartments, remote tenocytes and remote chondrocytes. Initial analysis of the different transcriptomes using principal components analysis (PCA) showed that the transcriptomes of tenocytes and chondrocytes were clearly separated, whereas attachment cells were located between the two cell types, recapitulating their anatomical positions. This suggests that the attachment cell transcriptome is shared with both chondrocytes and tenocytes (Figure 1B, PC1 31.66%).

To further support our initial observation that the transcriptome of the attachment cells is a mixture of chondrocyte and tenocyte transcriptomes, we clustered the statistically significant differentially expressed genes between all samples into five clusters, using CLICK (Figure 1—figure supplement 2 and see Materials and methods). Out of 865 identified genes, 735 genes were found in two clusters. The first cluster contained mainly known tenogenic genes and the second contained chondrogenic genes. From these two clusters, 374 genes, 320 of them tenogenic and 54 chondrogenic, were also found to be expressed by attachment cells. They included major regulators and marker genes of the two tissues, such as Sox9, Col2a1, and Acan for chondrocytes and Col1a1, Col1a2, Scx, and Col5a1 for tenocytes (Figure 1C).

The third cluster comprised 54 genes that were found to be upregulated in cartilage adjacent to attachment cells alone (Figure 1—figure supplement 2). Interestingly, our analysis identified 24 and 23 genes that were found to be down- or upregulated in attachment cells, shown by the fourth and fifth clusters, respectively (Figure 1—figure supplement 2). The genes that were found to be uniquely upregulated in attachment cells included transcription factors, such as the Krüppel-like factors (KLFs), Lmo1, and Gli1, which could act as regulators of the genetic program of attachment cells. In addition, this set included differentiation markers such as Thy1, regulators of bone for example Acp5 and Alpl, protein kinases such as Mapk12 and Mast2, and signaling molecules such as Nod, Traip, Aplnr and others (Figure 1—figure supplement 3A,B). GO analysis of these genes yielded terms relating to regulation of cytokine and IL-12 production, as well as response to laminar fluid shear stress (Figure 1—figure supplement 3C). These results clearly show that the transcriptome of the attachment cells includes a mixture of tenocyte and chondrocyte genes, many of which are involved in ECM organization and developmental processes, in addition to a unique subgroup of genes that are upregulated in these cells. To determine the relative proximity of attachment cells to the other two cell types, we performed hierarchical clustering on the 865 differentially expressed genes. Results showed that these cells are closer to tenocytes (Supplementary file 1).

To validate our transcriptome analysis, we performed single- and double-fluorescent in situ hybridization (FISH) using marker genes for tenocytes and chondrocytes that were selected from the transcriptomic results, namely Igfbp5, biglycan (Bgn), Col5a1, and Col1a1 for tenocytes and Col11a1 and Wwp2 for chondrocytes. As seen in Figure 1D and Figure 1—figure supplement 4, in agreement with the transcriptome analysis, the selected markers were co-expressed by the attachment cells.

Finding that the bulk transcriptome of attachment cells represents a mixture of tenocyte and chondrocyte genes led us to examine if this phenotype also exists at a single-cell resolution by performing scRNA-seq. To isolate E13.5 Sox9⁺/Scx⁺ attachment progenitors, we generated a compound mouse line harboring Sox9-CreER, tdTomato and Scx-GFP transgenes (Figure 1—figure supplement 1C–E). Then, FACS-sorted attachment progenitors from the proximal side of the forelimb underwent single-cell 10x Genomics RNA sequencing (Figure 2A,B and see Materials and methods). Sequenced progenitors were then filtered, normalized, and clustered (Seurat 3.1.5).

Results showed five transcriptionally distinct subpopulations (Figure 2A). Examination of the top-10 differentially expressed genes across clusters revealed that they were subdivided to three main categories. Clusters 1 and 3 were highly rich with chondrogenic genes (Sox9, Sox5, Col2a1, Wwp2), whereas clusters 2 and 4 were highly rich with tenogenic genes (Scx, Hic1, Ptn, Igfbp5). By contrast, the identity of cluster 0 was less specified, as it was highly rich with genes such as Col1a1, Alpl, Spp1, and Runx2 in addition to Bicc1, Maf, and Mef2c (Figure 2—figure supplement 1).

To study the bi-fated nature of attachment cells, we analyzed the relative expression of tenogenic (Scx, Col1a2, biglycan) and chondrogenic (Sox9, Col2a1, Wwp2) marker genes in each subpopulation. As seen in Figure 2B, all five clusters contained both chondrogenic and tenogenic marker genes. However, while tenogenic genes were relatively high in all five clusters, the expression levels of chondrogenic markers were more variable, being the lowest in cluster 2. Overall, the scRNA-seq results suggest that the Sox9⁺/Scx⁺ attachment progenitors are bi-fated. The analysis also revealed a molecular sub-classification of this progenitor population.

Finally, to confirm the co-expression of tenogenic and chondrogenic genes at the attachment site at a single-cell resolution, we performed single-molecule FISH for Wwp2 (chondrogenic marker) and Bgn (tenogenic marker). As seen and quantified in Figure 2C,D, Wwp2 and Bgn were indeed co-expressed in the attachment cells.

Overall, these results indicate that attachment cells are bi-fated, expressing in parallel both chondrogenic and tenogenic genes (referred to in the following as mixed transcriptome or mixed expression profile).

To gain a mechanistic understanding of how attachment cells activate a combination of two transcriptomes, we compared chromatin accessibility in these cells with open chromatin signatures defining chondrocytes and tenocytes by conducting an assay for transposase-accessible chromatin with high-throughput sequencing (ATAC-seq) (Buenrostro et al., 2015). This method allows to profile open chromatin regions, some of which may act as enhancers. To isolate by FACS E13.5 humeral chondrocytes, tenocytes and attachment cells, we generated a compound mouse line harboring Sox9-CreER, tdTomato, and Scx-GFP transgenes (Figure 1—figure supplement 1C–E). However, because the number of isolated chondrocytes was insufficient, chondrocytes were FACS-sorted from E13.5 Col2a1-CreER^T-tdTomato-Scx-GFP mouse. These three cell populations were then subjected to ATAC-seq (see Materials and methods).

Initial PCA analysis of accessible chromatin profiles for each FACS-sorted cell population once again revealed that tenocytes and chondrocytes were clearly separated, while attachment cells resided between these two cell types (Figure 3—figure supplement 1A, Figure 1B). Next, we compared global chromatin accessibility among the three cell types by calculating the level of overlap among the ATAC-seq peaks (Figure 3A). While the majority of the peaks were shared by all three cell types, attachment cells had a significantly lower number of unique peaks (p<1e-4 relative to both cell types, chi-square with Yates correction), and a significantly higher overlap with the other two cell types (p<2.2e-1).

Analysis of the ATAC-seq signal revealed that 13,017 peaks were located near transcription start sites (TSSs), whereas 31,856 peaks were in intergenic or intron regions. Most of the peaks that were located near TSSs were accessible in all three cell types (87%), and only 13% were accessible in one or two cell types (Figure 3—figure supplement 1C,D).

Next, we studied the ATAC-seq signal of peaks associated with the genes that were differentially expressed at E14.5, using HOMER default parameters. We found that 819 peaks were located near transcription start sites (TSSs), whereas 2340 peaks were in intergenic or intron regions. Most of the peaks that were located near TSSs (708, 86%) were accessible in all three cell types, and only 111 (13%) were accessible in one or two cell types (Figure 3B,C). This low level of differential accessibility is inconsistent with the possibility that promoter accessibility is the main mechanism regulating the bi-fated attachment cells. Interestingly, a significantly higher fraction of intergenic peaks were specific to one or two cell types (1767, 74.7%, p=0, chi-square test, Figure 3B,C).

Out of these 1767 intergenic peaks, 920 peaks were accessible in attachment cell; 672 were shared with either tenocytes or chondrocytes; and 248 were accessible only in attachment cells (Supplementary file 3). These results suggest that the 920 intergenic elements that were accessible in attachment cells may act as enhancers that drive the transcriptome of these cells. Moreover, since most of these intergenic elements were shared between attachment cells and either chondrocytes or tenocytes, they may serve as part of the mechanism that drives the mixed transcriptome of attachment cells.

To identify such dual cell-type-specific enhancers likely regulating attachment cell differentiation, we next screened for shared enhancers of 15 bona fide markers of tenocytes or chondrocytes that were found to be expressed in E14.5 attachment cells (Figure 1C). To improve the prediction of these enhancers, we selected our ATAC-seq peaks based on their proximity to genes with verified expression in attachment cells and another cell type (Figure 1—figure supplement 4, Figure 1D) and computationally intersected them with ENCODE datasets of histone modification marks associated with enhancers and promoters in mouse limbs at E13.5 (H3K27Ac, H3K9ac, H3K4me3, H3K4me1 ChIP-Seq), and other datasets (Andrey et al., 2017; Guo et al., 2017; Figure 4 and Table 1), revealing the degree of evolutionary conservation of each core sequence (Casper et al., 2018). For example, as shown in Figure 3D, we identified a region at −42 kb from the TSS of Mgp, a bona fide chondrogenic marker (Barone et al., 1991), which was accessible in chondrocytes and attachment cells, whereas in tenocytes this site was closed. Another example is Sox9, a bona fide chondrogenic marker. At +303 kb from Sox9, we identified a region that was accessible in attachment cells and chondrocytes, but not in tenocytes. The same pattern was observed for a region at −330 kb from the TSS of a third bona fide chondrogenic marker, namely Col11a1 (Li et al., 1995). The opposite pattern was observed at −17 kb from the TSS of Col1a2, a bona fide tenogenic marker, where we identified a region that was accessible in attachment cells and tenocytes, but not in chondrocytes. Similar results were obtained for additional chondrogenic markers, such as Sox6, and for tenogenic markers Tnc and Col1a1 (data not shown). Importantly, we found that the chromatin accessibility patterns of these putative enhancers were in agreement with the transcriptomic and ISH results, as shown, for example, by Sox9 and Col5a1 (Figure 1C,D, Figure 1—figure supplement 4). This suggests that the mechanism for the activation of a mixed transcriptome in attachment cells is based on sharing regulatory elements with chondrocytes or tenocytes.

The identification of multiple predicted enhancer regions near genes expressed by attachment cells and tenocytes or chondrocytes suggests that attachment cells are regulated predominantly by enhancers with shared activities. To test this hypothesis in vivo, we took advantage of a recently developed, site-directed transgenic mouse enhancer reporter system (Kvon et al., 2020). Using this system, we studied the activity of eight elements, which were selected because they were associated with bona fide marker genes for tenocytes or chondrocytes, and were found to be expressed in the attachment cells. Moreover, they were predicted to drive transcription in attachment cells and one of the flanking tissues (Figure 3D). The activity of these representative elements was examined at E14.5, a stage at which chondrocytes, tenocytes, and attachment cells have already been established.

Five elements were found to be active in the mouse forelimbs, as well as in other anatomical areas (Figure 4). The Col1a1-associated element drove lacZ expression at the teres major insertion into the scapula, in agreement with the ATAC-seq results, which predicted its activity in tenocytes and attachment cells (n = 2/9). This result suggests that Col1a1 element is active in tenocytes and attachment cells. Klf2 and Sox9 elements were predicted to be active in chondrocytes and attachment cells (Figure 3D). The Klf2 element indeed drove reporter activity in hypertrophic chondrocytes and perichondrium at the humerus and forelimb digits, in addition to the skull and mandible (Figure 4, n = 3/3). The Sox9 element showed activity solely in hypertrophic chondrocytes of the humerus (n = 4/7). These results suggest a chondrocyte-specific function of these two enhancers. Mgp element was also predicted to be active in chondrocytes and attachment cells (Figure 3D). Its activity was seen in forelimb and hindlimb, specifically in hip, metacarpal joints and digits as well as in the posterior distal side of the femur, a site where ligaments (e.g. the cruciate ligaments) are inserted into the femur at the knee area and at ligament insertion into to the hip (e.g. iliofemoral ligament; Figure 4, n = 3/3), verifying its activity in chondrocytes and attachment cells. Lastly, Col11a1 element activity was predicted in chondrocytes and attachment cells (Figure 3D). Its activity verified the bioinformatic analysis, showing LacZ staining in the greater tuberosity insertion, as well as in the posterior side of the skull, the nasal bone area and the anterior distal side of the femur (Figure 4, n = 11/11).

These results suggest that Mgp and Col11a1 elements are active in chondrocytes and attachment cells, whereas Col1a1 element is active in tenocytes and attachment cells, as the chromatin analysis predicts. Overall, these results provide a proof of concept for the ability of the accessible intergenic elements we have identified to act as enhancers that drive expression in both attachment cells and chondrocytes or tenocytes. This supports our hypothesis that shared enhancers activate a mixed transcriptome in attachment cells.

Our finding of enhancers that can drive the transcription of the mixed transcriptome of the attachment cells raised the question of the identity of the transcription factors (TFs) that can potentially bind to these elements. To identify such factors, we used Genomatix to analyze accessible elements that were associated with differentially expressed genes for over-representation of transcription-factor-binding sites (TFBS), selecting the top 50 TFBS families, and then mining our transcriptomic data for the expression of these TFBS families (see Materials and methods). Among the differentially expressed genes at E14.5 we identified NFIs (Nfia), GLIs (Gli1), KLFs (Klf2 and Klf4), ZBTBs (Zbtb48) and RUNXs (Runx3; Table 2, Figure 5A), whose expression was upregulated in attachment cells. Further support for these results was provided by HOMER motif analysis (Heinz et al., 2010), which showed significant over-representation of KLFs and RUNXs TFBSs. We therefore sought to explore the possible role of KLFs as regulators of attachment cells.

Focusing on Klf2, which was found to be differentially expressed by the bulk transcriptome analysis of the attachment site (Figure 1—figure supplement 2, Figure 1—figure supplement 3), we first validated its expression in the forming attachment site by ISH (Figure 5B). Next, we analyzed the enhancers that were shown by the bioinformatic analysis to be active in attachment cells and either tenocytes or chondrocytes (Figure 4). Three of these enhancers had Klf2-binding sites in their sequence (Table 1), further supporting a potential role for KLFs during attachment site development.

Previous studies demonstrated that Klf2 and Klf4 are functionally redundant, as KLF4 has ~90% sequence similarity to KLF2 in its zinc finger DNA-binding domain, suggesting that these factors could have common target sequences (Chiplunkar et al., 2013). In our scRNA-seq analysis, the expression of both genes was found in Sox9+/Scx+ attachment progenitors (Figure 5C). We therefore proceeded to study attachment cell development upon blocking the expression of both Klf2 and Klf4 in limb mesenchyme, using Prx1-Cre as a deleter and focusing on E15.5 and E18.5, a period during which the attachment site of the deltoid tuberosity undergoes differentiation and consequently grows in size. Transverse histological sections through the deltoid tuberosity of E15.5 control mice showed that the attachment cells were packed together and surrounded by ECM (Figure 5Da’). In contrast, in the putative attachment site of Prx1-Klf2-Klf4 double conditional knockout (dcKO) embryos, the cells were sparse with reduced ECM (Figure 5Db’). By E18.5, this difference was more pronounced, as attachment cells failed to differentiate (Figure 5Dc’,d’). To gain a molecular understanding, we studied the expression of several genes that were previously shown to be expressed at these stages in the attachment site (Felsenthal et al., 2018; Figure 1D). Indeed, we found that the expression of Col1a1, Gli1, Bsp, Bgn, and Col5a1 was reduced in the dcKO attachment site, relative to the control (Figure 5De-p), supporting a role for KLF2/4 in attachment cells differentiation.

Finally, to further validate the involvement of KLFs in activation of gene expression in the attachment site, we searched for KLF2/4 binding sites in ATAC-seq peaks associated with the 374 genes that were shown to be expressed by attachment cells (Figure 1C). Interestingly, we found that many of these genes had KLF2/4 binding sites in their regulatory regions (72% of the 374 attachment genes relative to 53% in the whole genome, p<1e-4, chi-square). We then searched for KLF2/4 binding sites in ATAC-seq peaks that were associated with genes whose expression was reduced in the dcKO attachment site (Figure 5De-p). For Gli1, we found KLF2/4 TFBSs in peaks that reside −2.1 and −1.5 kb from its TSS (Table 3). For Col5a1, we found multiple binding sites for KLF2 or KLF4. Together, these results indicate that KLF2/4 play an essential role in regulating attachment cell gene expression.

Combined with the bioinformatic analysis of chromatin and transcriptomic data, these results suggest that KLF2/4 are major regulators of tendon-to-bone attachment, playing a central role in attachment cell differentiation.

The generation of floxed Klf2 (Lee et al., 2006), floxed Klf4 (Katz et al., 2002), Prx1-Cre (Logan et al., 2002), Sox9-CreER (Soeda et al., 2010), Col2-CreER^T (Nakamura et al., 2006), Col2a1-Cre (Ovchinnikov et al., 2000), R26R-tdTomato (Madisen et al., 2010), and Scx-GFP (Pryce et al., 2007) mice have been described previously.

To create Col2a1-CreER, R26R-tdTomato, Sox9-CreER, R26R-tdTomato, and Col2a1-Cre, R26R-tdTomato reporter mice, all on the background of Scx-GFP, floxed R26R-tdTomato mice were mated with Col2a1-CreER, Sox9-CreER or Col2a1-Cre mice, respectively. These strains were mated on a mixed background of C57BL/6 and B6.129 (ICR) mice and used for LCM and FACS experiments. To create Prx1-Klf2-Klf4 mutant mice, floxed Klf2-Klf4 mice were mated with Prx1-Klf2-Klf4 mice. As a control, we used embryos that lack Cre alleles. E14.5 wild-type C57BL/6 mice were used for ISH experiments as well.

For FACS experiments, Sox9-CreER or Col2a1-CreER mice were crossed with Rosa26-tdTomato reporter mice, all on the background of Scx-GFP. Induction of Cre recombinase was performed at various pregnancy stages by administration of 0.03 mg/gr tamoxifen/body weight in corn oil by oral gavage (stock concentration was 5 mg/ml). In all timed pregnancies, plug date was defined as E0.5. For harvesting of embryos, timed-pregnant females were sacrificed by cervical dislocation. Tail genomic DNA was used for genotyping by PCR.

E14.5 Col2a1-tdTomato-Scx-GFP mouse forelimbs were dissected, shortly fixated in 4% PFA, washed with PBS and cryo-embedded (as described by Bhattacherjee et al., 2004 Bhattacherjee et al., 2004). Next, samples were cryo-sectioned and mounted on LCM slides (PET, Zeiss), washed with RNase-free water and EtOH (Arcturus dehydration component kit) according to an altered protocol of Pazin et al., 2012. LCM (PALM MicroBeam C system, Zeiss) was calibrated for refined tissue cutting. Isolated cells were collected to LCM caps (Adhesive Cap 500 clear, Zeiss) and RNA was purified using RNeasy FFPE Kit (Qiagen). The resulting RNA was the input for the bulk MARS-seq protocol (RNA sequencing).

Flow cytometry analysis and sorting were performed at the Weizmann Institute of Science Flow Cytometry Core Facility on a BD FACS AriaIII instrument (BD Immunocytometry Systems) equipped with 488, 407, 561, and 633 nm lasers, using a 70 μm nozzle, controlled by BD FACS Diva software v8.0.1 (BD Biosciences). Further analysis was performed using FlowJo software v10.2 (Tree Star). For collection of cells, Sox9-CreER^T2-tdTomato;ScxGFP or Col2a1-CreER-tdTomato;ScxGFP mice were crossed with Rosa26-tdTomato;ScxGFP reporter mice. Embryos were harvested at E13.5 following tamoxifen administration at E12.0, as described above. Forelimbs were dissected and suspended in cold PBS. To extract cells from tissues, PBS was replaced with 1 ml heated 0.05% trypsin and collagenase type V (dissolved in DMEM, Sigma) and incubated for 15 min at 37°C, gently agitated every 5 min. Tissues were then dissociated by vigorous pipetting using 1 ml tips. Next, 4 ml of DMEM supplemented with 10% FBS and 1% Pen-Strep was added and cell suspensions were filtered with 40 μm filter net. Finally, tubes were centrifuged at 1000 rpm for 7 min, supernatant was removed and cells were resuspended in 0.5–1 ml of cold PBS and used immediately for FACS. Single-stained GFP and tdTomato control cells were used for configuration and determining gate boundaries. Live cells were gated by size and granularity using FSC-A versus SSC-A and according to DAPI staining (1 μg/ml). FSC-W versus FSC-A was used to further distinguish single cells. In addition, unstained, GFP-stained only and tdTomato-stained only cells were mixed in various combinations to verify that the analysis excluded false-positive doublets. GFP was detected by excitation at 488 nm and collection of emission using 502 longpass (LP) and 530/30 bandpass (BP) filters. tdTomato was detected by excitation at 561 nm and collection of emission using a 582/15 BP filter. DAPI was detected by excitation at 407 nm and collection of emission using a 450/40 BP filter.

Total RNA was purified from LCM-isolated samples of E14.5 mouse forelimbs using RNeasy FFPE Kit (Qiagen). Reverse transcription was performed with High Capacity Reverse Transcription Kit (Applied Biosystems) according to the manufacturer's protocol. Analysis of Col2a1 and Scx was performed to monitor RNA quality during LCM calibrations, whereas RNA quantity was monitored by analysis of β–actin. RT-PCR was performed using Fast SYBR Green master mix (Applied Biosystems) on the StepOnePlus machine (Applied Biosystems). Values were calculated using the StepOne software. Data were normalized to 18S rRNA or β-actin in all cases.

Section ISH were performed as described previously (Riddle et al., 1993). Single-and double-fluorescent ISH on paraffin sections were performed using DIG- and/or FITC-labeled probes (listed in Table 4; Shwartz and Zelzer, 2014). After hybridization, slides were washed, quenched, and blocked. Probes were detected by incubation with anti-DIG-POD (Roche; 1:300) and anti-FITC-POD (Roche, 1:200), followed by Cy2-tyramide- and Cy3-tyramide-labeled fluorescent dyes according to the instructions of the TSA Plus Fluorescent Systems Kit (Perkin Elmer).

Harvested E14.5 forelimbs were fixed with cold 4% formaldehyde (FA) in PBS and incubated first in 4% FA/PBS for 3 hr, then in 30% sucrose in 4% FA/PBS overnight at 4°C with constant agitation. Fixed tissues were embedded in OCT and sectioned at a thickness of 10 µm. The preparation of the probe library, hybridization procedure, and imaging conditions were previously described (Itzkovitz et al., 2012; Lyubimova et al., 2013; Raj et al., 2008). In brief, probe libraries were designed against biglycan (Bgn) and Wwp2 mRNA sequences using the Stellaris FISH Probe Designer (Biosearch Technologies, Inc, Petaluma, CA) coupled to Quasar 670 and CAL Fluor Red 610, respectively. Libraries consisted of 17–96 probes each of length 20 bps, complementary to the coding sequence of each gene (Supplementary file 2). Nuclei were stained with DAPI. To detect cell borders, Alexa Fluor 488 conjugated phalloidin (Thermo Fisher, A12379) was added to the GLOX buffer, which was wash for 15 min. Slides were mounted using ProLong Gold (Molecular Probes, P36934).

For smFISH image acquisition, we used a Nikon-Ti-E inverted fluorescence microscope equipped with a Photometrics Pixis 1024 CCD camera to image 10-µm-thick cryosections. For image analysis, we used ImageM, a custom MATLAB program (Lyubimova et al., 2013), which was used to compute single-cell mRNA concentrations by segmenting each cell manually according to the cell borders and the nucleus. The size of the nucleus was detected automatically by the program according to the DAPI signal. For each cell, the concentration of cytoplasmic mRNA of each gene was calculated by measuring the number of dots per volume. Images were visualized and processed using ImageJ 1.51 hr (Schindelin et al., 2012) and Adobe Illustrator CC2018.

For this analysis, we adapted the MARS-seq protocol (Jaitin et al., 2014; Keren-Shaul et al., 2019) to generate RNA-seq libraries for expression profiling of the purified RNA from E14.5 LCM-isolated samples. Briefly, RNA from each sample was barcoded during reverse transcription and pooled. Following Agencourt Ampure XP beads cleanup (Beckman Coulter), the pooled samples underwent second strand synthesis and were linearly amplified by T7 in vitro transcription. The resulting RNA was fragmented and converted into a sequencing-ready library by tagging the samples with Illumina sequences during ligation, RT, and PCR. Libraries were quantified by Qubit and TapeStation as well as by qPCR for actb gene as previously described (Jaitin et al., 2014; Keren-Shaul et al., 2019). Sequencing was done on a Hiseq 2500 SR50 cycles kit (Illumina).

The data were analyzed using the Pipeline Pilot-designed pipeline for transSeq (by INCPM, https://incpmpm.atlassian.net/wiki/spaces/PUB/pages/36405284/tranSeq+on+Pipeline-Pilot). Briefly, the analysis included adapter trimming, mapping to the mm9 genome, collapsing of reads with the same unique molecular identifiers (UMI) of 4 bases (R2) and counting of the number of reads per gene with HTseq-count (Anders et al., 2015), using the most 3’ 1000 bp of each RefSeq transcript. DESeq2 (version 1.4.5, Love et al., 2014) was used for normalization and differential expression analysis with betaPrior set to true, cooksCutoff=FALSE, independentFiltering=FALSE. Benjamini-Hochberg method was used to adjust the raw p-values for multiple testing. Genes with adjusted p-value≤0.05 and fold change ≥ 2 between every two conditions were considered as differential. PCA analysis was done using log-transformed normalized data (DESeq2 function rlogTransformation with parameter blind=TRUE) with a modified plotPCA function of DESeq2. Clustering of the log-normalized read count of differentially expressed genes was done using CLICK algorithm (Expander package version 7.1, Ulitsky et al., 2010), followed by visualization by R (R Development Core Team, 2013). Further analysis was performed using GSEA (Broad institute) and Gorilla (Eden et al., 2009; Subramanian et al., 2005). The crude data of the work has been deposited on NCBI GEO (GSE144306).

Cells were counted and diluted to a final concentration in PBS supplemented with 0.04% BSA. Cellular suspension was loaded onto Next GEM Chip G targeting 4000 cells and then ran on a Chromium Controller instrument to generate GEM emulsion (10x Genomics). Single-cell 3’ RNA-seq libraries were generated according to the manufacturer’s protocol (10x Genomics Chromium Single Cell 3’ Reagent Kit User Guide v3/v3.1 Chemistry).

Single-cell 3’ RNA-seq libraries were quantified using NEBNext Library Quant Kit for Illumina (NEB) and high sensitivity D1000 TapeStation (Agilent). Libraries were pooled according to targeted cell number, aiming for ~50,000,000 reads per cell. Pooled libraries were sequenced on a NovaSeq 6000 instrument using an SP 100 cycles reagent kit (Illumina) (R1 28 bases, R2 82 bases, and I1 eight bases), specifically 290M fragments of the relevant library were sequenced.

Demultiplexing and alignment were performed using cellranger (10x Genomics version 3.0.2) bioinformatics pipeline using mm10 genome, followed by detecting swapped barcodes between libraries, since the NovaSeq run contained a mix of three libraries. For this, we used R (3.6.3) and the package DropletUtiles (Lun et al., 2019) (1.6.1) with the function swappedDrops (parameter min.frac = 0.9) and the molecule_info.h5 input. 1.5% of the molecules were detected as swapped. Detecting empty droplets was done with the function emptyDrops (parameter lower = 300), the number of cells detected was 2213.

For additional analysis, the R package Seurat (3.1.5) was used (Butler et al., 2018). The analysis included filtering genes (must be expressed in at least three cells), filtering cells with over 20% expression of mitochondria genes and high and low 3% percentiles of total number of genes and UMIs; this resulted in 1218 cells. In addition, one of the clusters had significantly lower RNA counts as compare to other collected cells, which might have indicated that it contained mainly low-quality cells. Indeed, examination of the 10 most highly expressed genes using heatmap demonstrated this; therefore, this cluster was removed from further analysis. A total of 1076 cells were clustered using 2000 variable genes and 15 principal components (PCs) (resolution = 0.4). UMAP plot was made with the Seurat functions RunUMAP and DimPlot and violin plots were made with the VpnPlot function. The crude data of the work have been deposited on NCBI GEO (GSE160090).

ATAC-seq data were trimmed from their adaptors and filtered from low quality reads using Cutadapt followed by alignment to the mm10 genome (GRCm38.p5) using Bowtie2 (version 2.3.4.1) (Langdon, 2015). PCR-duplicate reads were removed with Picard ‘MarkDuplicates’ (http://broadinstitute.github.io/picard/). Mitochondrial reads were removed from the alignment, and the data were further filtered to contain only reads with a unique mapping with SAMtools (-F 4 f 0 × 2). Read pairs with inner distance of up to 120 bp were selected as representing the accessible chromatin region. MACS2 (version 2.1.1.20160309) (Zhang et al., 2008) was applied for peak calling using the setting: callpeak -f BAMPE--nomodel. Peaks from all samples were combined and merged with BEDTools (Quinlan, 2014), followed by extension to a minimum length of 500 bp. For every tissue, a set of reproducible peaks was obtained by voting, which means that a normalized read count ≥ 30 was detected in at least 50% of the replicates. Peaks that were not reproducible in any tissue were removed. Peaks that reside in the ENCODE ‘Blacklist’ regions, that is regions that were previously found by ENCODE (ENCODE Project Consortium, 2012) to produce artificial signal (http://mitra.stanford.edu/kundaje/akundaje/release/blacklists/), were also eliminated. Peak quantification was done with BedTools Quinlan, 2014 following by DESeq2 (Love et al., 2014) normalization. Peaks with an averaged normalized read count ≥ 30 in at least one of the studied tissues were selected for the downstream analyses. The crude data of the work have been deposited on NCBI GEO (GSE144306).

Annotation of ATAC-seq peaks was performed using HOMER (Heinz et al., 2010) and GREAT (McLean et al., 2010). When a peak was associated by GREAT to multiple genes, the two closest genes were selected for further analysis. ATAC-seq peaks that were at a distance of up to −2 kb down or +0.5 kb up from a TSS of their annotated gene (HOMER) were considered as promoter peaks; otherwise, peaks were considered as distal. To rank distal ATAC-seq peaks as putative cis-regulatory elements, we calculated the overlap between the peaks and relevant histone modification datasets (ChIP-seq) performed by the ENCODE project (Yue et al., 2014) on E13.5 C57BL/6 mouse embryo limb. The overlap was calculated using BEDTools intersect (Quinlan, 2014). The following datasets were used: ENCSR905FFU (H3K27ac) and ENCSR426EZM (H3K4me1) as markers of enhancers, ENCSR416OYH (H3K4me3) as a marker of promoters and ENCSR022DED (H3K9me3). Overlap with the phastConsElements60wayPlacental track downloaded from the UCSC site (Casper et al., 2018) was calculated to account for evolutionary conservation. Enrichment analysis of over-representation of TFBSs in the ATAC-seq peaks was performed with the RegionMiner tool of Genomatix and HOMER.

Candidate enhancers were PCR-amplified and cloned upstream of a Shh-promoter-LacZ-reporter cassette. We used a mouse enhancer-reporter assay that relies on site-specific integration of a transgene into the mouse genome (Kvon et al., 2020). In this assay, the reporter cassette is flanked by homology arms targeting the safe harbor locus (Tasic et al., 2011). Cas9 protein and a sgRNA targeting H11 were co-injected into the pronucleus of FVB single cell-stage mouse embryos (E0.5) together with the reporter vector (Kvon et al., 2020). Embryos were sampled and stained at E14.5. Embryos were excluded from further analysis if they did not carry the reporter transgene. All mouse works were reviewed and approved by the Lawrence Berkeley National Laboratory Animal Welfare and Research Committee.