SUMMARY
Differentiation proceeds along a continuum of increasingly fate-restricted intermediates, referred to as canalization1–4. Canalization is essential for stabilizing cell fate, but the mechanisms underlying robust canalization are unclear. Here we show that deletion of the BRG1/BRM-associated factor (BAF) chromatin remodeling complex ATPase gene Brm (encoding Brahma) results in a radical identity switch during directed cardiogenesis of mouse embryonic stem cells (ESCs). Despite establishment of well-differentiated precardiac mesoderm, Brm-null cells subsequently shifted identities, predominantly becoming neural precursors, violating germ layer assignment. Trajectory inference showed sudden acquisition of non-mesodermal identity in Brm-null cells, consistent with a new transition state inducing a fate switch referred to as a saddle-node bifurcation3,4. Mechanistically, loss of Brm prevented de novo accessibility of cardiac enhancers while increasing expression of the neurogenic factor POU3F1 and preventing expression of the neural suppressor REST. Brm mutant identity switch was overcome by increasing BMP4 levels during mesoderm induction, repressing Pou3f1 and re-establishing a cardiogenic chromatin landscape. Our results reveal BRM as a compensable safeguard for fidelity of mesoderm chromatin states, and support a model in which developmental canalization is not a rigid irreversible path, but a highly plastic trajectory that must be safeguarded, with implications in development and disease.
Main Text
Our previous studies indicated the prevalence of BRM in the cardiomyocyte-enriched chromatin remodeling complex BAF1705. BRM has been reported to be dispensable for mouse development6, but it is implicated in human developmental syndromes7,8, mouse skeletal muscle function9, and several cancers10,11, and can partly compensate for loss of BRG112–16. To determine the role of BRM in cardiac differentiation, we deleted Brm in ESCs and directed cardiac differentiation (Fig. 1a, Extended data Fig. 1a). Brm−/− cells failed to generate beating cTnT+ cardiomyocytes (Extended Data movies. 1-3) and as measured by immunofluorescence (Fig. 1b) and flow cytometry (Fig. 1c). This was confirmed in three independent Brm−/− lines, while two heterozygous lines from the same set of clones differentiated well. RNA-seq during the differentiation time course showed that at D4, mesoderm gene expression was unaffected in Brm−/− cells, while D6 (cardiac precursor, CP) and D10 (cardiomyocyte, CM) gene expression was significantly altered (FDR<0.05, ±2-fold change) (Extended data Fig. 1b, c). At D6, several important cardiac TFs were not induced in Brm−/− cells (Isl1, Hand2, Nkx2-5, Mef2c, Tbx20), whereas osteoblast- and neural-associated TFs were upregulated (Tcf15, Sox2). At D10, when WT cells had reached the beating cardiomyocyte stage, Brm−/− cells completely failed to activate cardiac genes and instead expressed genes associated with neural (Ascl1, Pax6, Neurod1, Neurog1, Olig2 and Sox2) or other (e.g. erythrocyte Gata1, Tal1) cell types (Fig. 1d).
BRM safeguards CP differentiation to CM and represses neural and other gene programs
The drastic gene expression changes in Brm−/− cells upon cardiac differentiation suggested either 1) the formation of new non-cardiac populations or 2) that a relatively homogeneous population activated normally mutually-exclusive expression modules. To differentiate between these scenarios, we performed single cell RNA-seq on 21,991 WT and Brm−/− cells at D10 of differentiation using the 10X Genomics dropseq platform. Uniform Manifold Approximation and Projections (UMAP)17 plots showed that Brm−/− cells at D10 were radically distinct from their WT counterparts (Fig. 1e), lacked expression of cardiac genes, and clustered in multiple sub-populations. These included cells with signatures of neural stem cells (Sox2, Sox9, Neurod1, and Ascl1), neural progenitors or immature neurons (Dcx, Otx2, Gap43, Tubb2b, Tubb3), glial (Gfap, Olig2) and Schwan cells (Gap43), and retinal neuronal precursors (Rax, Lhx2, Lmo1) (Figs. 1f, g, Extended Data Fig. 1d). We also identified a cluster of cells expressing markers of osteoblast development (Postn, Bgn, Col1a1, Fbln2, Twist2), indicating that some Brm−/− cells adopt non-cardiac mesodermal fates. Immunofluorescence showed TUBB3+ staining in D10 Brm−/− cells, displaying neuron like outgrowths (Fig. 1h). Notably, no other mesodermal or ectodermal derivatives, nor endodermal cell types, were observed, indicating that BRM deletion induced a specific fate switch. Loss of BRM did not seem to affect directed neuronal precursor differentiation (Extended Data Fig. 1e).
BRM controls the mesoderm to cardiac precursor transition
To elucidate the events underlying the anomalous differentiation in absence of BRM, we examined the time courses of differentiation of WT and Brm−/− cells by single cell RNA-seq (Fig. 2a). Consistent with bulk RNA-seq, D4 (mesoderm) Brm−/− cells were statistically similar to WT cells and both occupied the same UMAP space (Fig. 2b-d). Brm−/− cells segregated slightly based on modest (less than 2-fold) changes in expression of a few genes (Mesp1, Lhx1, Fn1, Rps28) (Supplementary Sheet 1). In sharp contrast, at D6 Brm−/− clustered separately from WT cells (Fig. 2b). Most D6 WT cells expressed well-defined CP markers (Smarcd3, Mef2c, Hand2; Fig. 2d, Extended Data Fig. 2a), whereas Brm−/− cells mostly expressed genes involved in neural lineages (Gbx2, Sox2, Irx3, Crabp1, Crabp2 and Prtg; Fig. 2d, Extended Data Fig. 2a). The few D6 WT and Brm−/− cells that clustered together expressed markers of hematopoietic lineages (Gata1, Klf1, Hbb1; Supplementary Sheet 2), suggesting a low level of BRM-independent hematopoietic differentiation. As expected, D10 WT and Brm−/− cells clustered largely in different UMAP space (Fig. 2b-d). Our time course therefore indicates a crucial early role of BRM immediately following cardiac mesoderm formation. Partition Based Graph Abstraction18 revealed genotype-dependent connectivity (Fig. 2e-g). WT and Brm−/− cells at D4 predominantly connected with their respective D6 and D10 genotype-specific clusters. A small percentage of WT and Brm−/− D6 cells connected to clusters 5 and 12 forming hematopoietic and endothelial clusters (Fig. 2e-g). To further assess the differentiation paths, we built differentiation trajectories in the form of a branching tree using URD19. Pou5f1+ −WT and Brm−/− clusters were selected as the root, while D10 clusters were defined as tips. Only WT cells progressed in stepwise pseudotime to CPs and their derivatives. In contrast, Brm−/− cells directly transitioned from mesoderm to non-cardiac neural lineages after D4 (Fig. 2h, i, Extended Data Fig. 2b, c).
Quantitative gene expression analysis at D4 revealed minimal expression of pluripotency (Nanog and Sox2) and paraxial mesoderm markers (Tbx6, Msgn1), while primitive streak marker (T) and mesoderm precursor (Pou5f1 or Oct4) showed minimal changes between genotypes (Extended Data Fig. 2a), confirming proper cardiogenic mesoderm differentiation of Brm−/− cells. The absence of Tbx6+/Sox2+ or T+/Sox2+ cells precluded the possibility that Brm−/− cells represent neuromesodermal precursors, especially given that our differentiation lacks retinoic acid20,21. At D4, we did not observe transcriptional changes in neuroectodermal markers22 between WT and Brm−/− cells. Indeed, the continuous expression of POU5F1 and absence of SOX2 indicated that D4 cells are derived from a mesendoderm rather than an neuroectoderm lineage23,24, suggesting Brm−/− cells acquire neural lineage after D4 (Fig. 2c, Extended Data Fig. 2a).
BRG1, a paralog of BRM, has important roles in cardiogenesis5,25,26. To compare the role of Brg1 to that of Brm in cardiac differentiation, we induced genetic Brg1 deletion25,27 at D2 and analyzed the effects at D4 and D10 by single cell RNA-seq (Extended Data Fig. 3a). Brg1 loss did not affect D4 transcription broadly (Extended Data Fig. 3b, c). However, at D10, WT and Brg1 KO cells clustered separately (Extended Data Fig. 3b). Marker analysis revealed that Brg1-deficient cells formed very few cardiac myocytes and instead formed endothelial cells, fibroblasts, neural progenitors, and developmentally-arrested progenitors (Extended Data Fig. 3b-e). Unlike Brm−/− cells, we did not observe TUBB3 staining in Brg1 KO cells (Extended Data Fig. 3f).
These results show that under directed differentiation conditions, Brm−/− cells form specified cardiac mesodermal cells, however then undergo a radical fate change towards a constellation of non-cardiac, largely neuronal cell types. Loss of Brg1 also had a similar but less severe phenotype, indicating essential but distinct roles of BAF complex ATPases in regulating lineages during differentiation28,29.
BRM modulates dynamic chromatin accessibility
We next used ATAC-seq30 to examine BRM’s role in modulating chromatin accessibility during cardiac differentiation (Fig. 3a). Although gene expression was minimally affected at D4, we found significant changes (FDR<0.05, fold change >2) at 3320 chromatin regions, 98.3% of which showed reduced accessibility upon Brm loss. These sites were enriched for genes involved in cardiac and other developmental pathways (Fig. 3b). At D6 and D10 8814 and 5391 regions were significantly changed (FDR<0.05, fold change >2), respectively, between genotypes (Supplementary Sheet 3). Consistent with the RNA-seq data, at D6 the differentially closed chromatin in Brm−/− cells was near regulatory elements of cardiac development genes, including TFs such as Gata4, Tbx5, Nkx2-5, Myocd, Hand2, Mef2c, and cardiac functional genes such as Ttn, Myl2, Myh6, Myh7, Actc1. In contrast, D6 Brm−/− cells had newly open chromatin near genes involved in non-cardiac differentiation processes, including some neural genes (e.g. Pax6) (Fig. 3c), consistent with initiation of neural gene expression. By D10, a clear and strong association was found with accessible chromatin near neural differentiation-related genes (Fig. 3d). Comparison of temporal accessibility patterns for select genes revealed that BRM maintains accessibility near cardiac genes throughout differentiation beginning at D4 (Fig. 3e, Extended Data Fig. 4a). In contrast, loss of BRM induced or maintained accessibility near neural genes at or after D6 (Fig. 3e, Extended Data Fig. 4b).
Altered chromatin accessibility was highly correlated with gene expression changes at D6 and D10 (Figs. 3f and 3g). Comparing BRM-mediated chromatin accessibility to active enhancers identified by H3K27ac marks31,32 showed that BRM promoted accessible regions at CP and CM enhancers (Figs. 3h and 3i). Conversely, BRM-mediated closed chromatin associated significantly with neural progenitor enhancers in CMs32. Motifs enriched in ATAC-seq peaks that were significantly depleted in Brm−/− included those for cardiac-related transcription factors (GATA4, MEF2C, HAND2) while peaks newly opened in absence of BRM were enriched for motifs for neuronal TFs (SOX2, OCT6/8 (POU3F1/3), OTX2, LHX2/3, RFX; Extended Data Fig. 4c). Thus, BRM promotes open chromatin at cardiogenic genes, while subsequently establishing or maintaining the inaccessible state of non-cardiac (including neural) enhancers.
Timing of BRM function
To more precisely pinpoint the timing of BRM function, we created an auxin-inducible degron ES cell line33 to rapidly deplete BRM (Extended Fig. 4d). Continuous auxin application largely recapitulated the impaired cardiac differentiation of Brm−/− cells, with a low level of remaining TNNT2+ cells likely due to incomplete BRM depletion in a proportion of cells (Extended Data Fig. 4e)33. Depletion of BRM prior to D4 impaired differentiation, whereas subsequent depletion did not greatly affect global cardiac differentiation (Extended Data Fig. 4f). BRM loss did not affect ESC pluripotency or self-renewal (Extended Data Fig. 5a). These results indicate a critical role for BRM after exit from pluripotency, and during cardiac mesoderm formation, but before mesoderm to CP transition.
Epigenetic regulation of chromatin by BRM
BAF complex subunits been implicated in modulation of chromatin in lineage-specific enhancers25,34. BRM specifically facilitates acetylation of H3K27 residues at H3K27me3 enriched Polycomb targets35,36. To understand if BRM modulates enhancer landscape via histone modifications, we profiled the effect of BRM loss on H3K27ac and H3K27me marks. At D4 of cardiac differentiation, very few regions were differentially enriched with H3K27me3. At later stages, regions near cardiovascular genes gained H3K27me3 marks upon BRM loss, while PcG-repressed genes involved in early embryo development lost their H3K27me3 marks (Extended Data Fig. 6a).
Conversely, in D6 and D10 Brm−/− cells, H3K27ac was reduced near genes associated with cardiac muscle development and contraction (Fig. 4a, clusters c and f, Extended Data Fig. 5c,d, 6c,d), and increased near genes involved in cell fate specification, neurogenesis and regulation of neuron differentiation (Fig. 4a, clusters a, g and I, Extended Data Fig. 5c, d, 6c,d), concordant with ATACseq data. Sites reduced in Brm−/− cells were enriched for cardiac TFs, while sites that gained H3K27ac marks were enriched for neural TF motifs (Extended Data Fig. 5f, g). At D4, sites with reduced H3K27ac in Brm−/− cells were enriched near cardiovascular development genes (Fig. 4a, cluster b, Extended Data Fig 5b, e), and had motifs for cardiac TFs. Despite the absence of accessibility changes we observed sites that gained H3K27ac in D4 Brm−/− cells, including genes involved in stem cell maintenance and neurogenesis (Fig. 4a, cluster h), which were enriched for POU or OCT motifs (Fig. 4b).
The enrichment of POU motifs suggested a potential involvement of these TFs in neural induction in Brm−/− cells. Bulk RNAseq with lower statistical cutoff (raw p-value <0.05) showed a modest increase in Pou3f1 (Oct6) mRNA in Brm−/− D4 cells; other POU factors were not detected. POU3F1 promotes neural fate by activating neural lineage genes and inhibiting BMP4-dependent transcription37. We found that POU3F1 protein was expressed at D2, and thereafter reduced during WT differentiation (Fig. 4c). Despite the low level of mRNA induction, POU3F1 protein was robustly increased at D4 and D6 in Brm−/− cells (Fig. 4c), suggesting prolonged POU3F1 may initiate the neurogenic gene expression program. Indeed, knockdown of Pou3f1 in Brm−/− cells resulted in fewer TUBB3+ neuronal progenitor cells and absence of filamentous extensions (Fig. 4d).
To delineate BRM occupancy, we performed chromatin immunoprecipitation followed by sequencing (ChIP-seq) using anti-FLAG antibody on a BRM-3xFLAG tagged strain at D4, D6 and D10 (Extended Data Fig. 7a). BRM bound to 110 regions at D4, 521 regions at D6 and 1188 regions at D10 (Extended Data Fig. 7b-d and Supplementary Sheet 4), consistent with its increasing expression pattern during cardiac differentiation. BRM-bound regions were enriched for genes involved in transcriptional and post-transcriptional regulation of gene expression at D4 and D6, and for regulation of muscle development at D10 (Extended Data Fig. 7e). Motif analysis revealed enrichment of REST motifs at all stages, along with cardiac TFs at D10 (Fig. 4e). Moreover, BRM deletion abrogated REST expression at D10 (Fig. 4f). REST knockdown from D4 to D7 resulted in ectopic expression of TUBB3+ cells at D10 (Fig. 4g), suggesting BRM controls expression of REST to specifically represses neural lineage genes during cardiac differentiation.
Signal-dependent rescue of anomalous differentiation
The BRM degron experiments suggested that BRM is most critical during cardiac mesoderm induction (Extended Data Fig. 4f). At this stage, BRM regulates neuronal lineage inducing factor POU3F1, which can counteract BMP signaling37. BMP4 concentration at this stage is finely regulated to ensure proper cardiac differentiation 38,39. In our system, supraphysiological exogenous BMP4 concentration inhibited cardiac differentiation of WT ESCs but rescued cardiac differentiation of Brm−/− ESCs (Fig 5a, Extended Data Fig. 8a). High BMP4 also repressed prolonged POU3F1 expression in Brm−/− cells and normalized expression of BAF60c and REST (Fig. 5b). Loss of BRG1, however was not similarly compensable by increasing BMP4 concentrations (Extended Data Fig 8b).
BMP4-mediated BRM rescue restored accessibility of cardiac enhancers (Fig. 5c-e, Extended Data Fig. 8c-e, see boxed regions and 8f). Conversely, neural progenitor enhancers were inaccessible in Brm−/− cells with high BMP4 (Fig. 5e, Extended Data Fig. 8g).
Single cell RNAseq analyses revealed BMP4-dependent changes in gene expression. At D4, WT and Brm−/− cells differentiated with normal BMP4 concentrations (3.2 ng/μl) clustered together in the same UMAP space, but separated from the high BMP4 (12.8 ng/μl) samples (Fig. 5f). Later in differentiation, high BMP4 WT cells clustered separately from normal BMP4-treated cells, forming endothelial and hematopoietic progenitors, and eventually fibroblasts and blood cells (Fig. 5f,g). As expected, Brm−/− cells at normal BMP4 formed mostly neural progenitor clusters (Fig. 5h) separate from WT cells. In contrast, Brm−/− cells at high BMP4 clustered with WT cells treated with normal BMP4 at D6 and D10, and expressed CP and CM markers (Figs. 5f, i, Extended Data Fig. 9a and Supplementary Sheet 5).
URD trajectory analysis with Pou5f1+ expressing clusters as root and D10 clusters as tips placed WT and Brm−/− cells at normal and high BMP4 concentrations at the trajectory root. At normal BMP4, WT cells followed a step-wise trajectory forming cardiomyocytes, while Brm−/− cells followed a different path (Fig. 5j), however Brm−/− cells at high BMP4 formed cardiac progenitors and cardiomyocytes by following an almost identical trajectory to WT cells at normal BMP4 (Fig. 5k, Extended Data Fig. 9b). Thus, simply modulating BMP4 signaling compensated for the absence of BRM and completely restored the normal path of cardiac differentiation.
How BMP4 might change differentiation path of Brm−/− cells is not clear. We confirmed that loss of BRM simply did not change BMP4 availability to the cells (Extended Data Fig. 9c). BMP-dependent gene regulatory networks are highly robust, and self-regulate in part by modulating transcriptional noise40,41. To evaluate if gene expression noise is regulated by BRM or BMP4, we calculated cell-to-cell gene expression variability42 and single cell entropy43 to predict differentiation potential. At normal BMP4 both WT and Brm−/− cells had similar dispersion and entropy metrics, both of which increased in presence of high BMP4 at D4 (Extended Data Fig 9c and 9d). This suggests increased intrinsic gene expression noise could participate in BMP4-dependent modulation of the Brm−/− transcriptional state.
In vivo requirement for BRM
Brm−/− mice are viable, although the nature of the original allele has been questioned6,44. Our independent mouse line with an 8bp deletion at Exon 2 resulting a premature stop codon and loss of BRM protein produced pups at Mendelian ratios (Extended Data Figs. 10a-c) confirming viability of Brm knockout. Viability may be due to compensatory overexpression of BRG1 in Brm knockout tissues (Extended Data Fig. 10d), which did not occur in our directed differentiation system (Extended Data Fig. 10e).
Discussion
Along the “landscape” of cell fate decisions, epigenetic regulators are key determinants of transition states. This is apparent in cancer, where new attractor states are formed that result in anomalous differentiation or dedifferentiation. In normal development however, only scant examples exist of natural transdifferentiation21,45,46, pointing to stability and robustness of canalization in vertebrate differentiation. Here we show that this stability during cardiac differentiation requires safeguarding by BRM.
Artificially forced reprogramming overcomes cell states47, and certain chromatin remodeling factors including BRM are important safeguards against reprogramming48–51. Conversely, other BAF complex subunits (e.g. BRG1, BAF60c) enhance reprogramming52–54. Reprogramming of fibroblasts to neurons involves transient competition between myogenic and neural gene expression programs, evidence that genome plasticity can transcend germ layer specification55. It is likely that in the absence of BRM, deregulation of neurogenic TFs, e.g. POU3F1, activates a cascade of neural gene expression in the context of broadly deregulated enhancer accessibility (Fig. 5l). That we observe this “self-reprogramming” in a directed differentiation context, but not in the complete organism indicates that the cues provided in vitro are strictly narrow parameters, while in vivo they are likely highly buffered. Indeed, increased BRG1 in Brm−/− mice and rescue of Brm−/− cells by elevated BMP signaling, indicates that loss of BRM is compensable.
Our findings indicate that BRM maintains developmental canalization of committed mesodermal precursors by providing an epigenomic state that favors a limited range of transition states, and that in its absence an unstable state induces a transition akin to a saddle node bifurcation (Fig. 5m). We highlight the fragility of the differentiation path, challenging the concept of highly robust developmental canalization, with important implications for understanding the stability of gene regulation in differentiation, and for deregulated gene expression in disease.
Methods
Cell culture and in-vitro differentiations
Mouse embryonic stem cells (ESCs) were cultured in media containing fetal bovine serum (FBS) and leukemia inhibitory factor (LIF) without feeder mouse embryonic fibroblast cells with daily media change at 37°C, 7% CO2 and 85% relative humidity. CMs were differentiated as described previously31,38. Briefly, mouse ESCs were cultured in presence of ascorbic acid (50μg/ml) in suspension cultures without LIF and Serum for 2 days to form embryoid bodies (EBs). EBs were dissociated and treated for 2 days with VEGF, Activin A and BMP4 to induce cardiac mesoderm which were subsequently dissociated and cultured as monolayer in presence of FGF-basic (FGF2) and FGF10 for 6 days to form beating cardiac myocytes. Brg1 was deleted in presence of 200 nM 4-hydroxytamoxifen for 48 h with control cells treated similarly with tetrahydrofuran5,25,56. Neural stem cell differentiations were carried out in presence FGF2 and epidermal growth factor (EGF), with growth factor removal forming neuronal progenitor cells as described previously57.
Cell line and mouse line generation
BRM was targeted using CRISPR/Cas9 with sgRNA targeting exon 2 following the described protocol58. sgRNA were cloned to a BbsI-digested pX330 vector (Addgene Cat #42230) by annealing the following primers: 5′ caccg GTCCACTGTGGATCCATGAA 3′ and 5′ aaac TTCATGGATCCACAGTGGAC c 3′ (bold indicates the BbsI digestion site). For construction of BRM-3xFLAG tag line, we followed a similar strategy to insert 3xFLAG tag sequence between the stop and penultimate codon using the following primers to clone sgRNA to the BbsI site of pX330 vector : 5’ caccg CTGATAACGAGTGACCATCC 3’ and 5’ aaac GGATGGTCACTCGTTATCAG C 3’. The following sequence was inserted to the upstream homology sequence for the insertion of 3x-FLAG tag: 5’ ggaggcggtggagcc GAC TAC AAG GAC CAC GAC GGC GAC TAC AAG GAC CAC GAC ATC GAC TAC AAG GAC GAC GAC GAC AAG TGA 3’. BRM targeting vectors were constructed by cloning 450 to 500 bp of DNA upstream and downstream of midpoint of sgRNA target site into KpnI-XhoI and BamH1-NotI sites of pFPF (a derivative of Addgene plasmid #22678 in which neomycin is replaced with puromycin cassette). BRM-AID strain was constructed following a previously-described strategy33. Briefly, pEN244-CTCF-AID_71-114_-eGFP-FRT-Blast-FRT plasmid (addgene Cat#92140) was digested with BamH1 and Sal1 and replaced the 3’ and 5’ homology of Ctcf with that of Brm respectively. The following primers were used to clone an sgRNA to pX330 vector: 5’ CAC CCT GAT AAC GAG TGA CCA TCC 3’ and 5’ GAC TAT TGC TCA CTG GTA GGC AAA 3’. 2.5 μg of each of the sgRNA plasmid, plus 20 μg of Brm targeting constructs were used for transfection. Single clones were selected, grown, PCR genotyped and DNA sequenced.
For constructing a Brm mouse strain, we used CRISPR/Cas9 with the exact same exon 2 sgRNAs as in the cell line cloned to a BbsI-digested pX330 vector by annealing oligos: 5’ caccg GTCCACTGTGGATCCATGAA 3’ & 5’ aaac TTCATGGATCCACAGTGGAC c 3’. In-vitro transcribed RNA and CAS9 protein complex and were injected to the embryos and transferred to 0.5 dpc pseudo-pregnant female mice. We obtained a mouse line with 8bp deletion resulting in a premature stop codon, confirmed by genotyping PCR sequencing and loss of BRM protein by western blot
siRNA mediated knockdown
RNA knockdown were carried out using Lipofectamine-RNAiMax reagent (ThermoFisher, 13778150) and pre-designed siRNA against POU3F1 (Sigma, SASI_Mm02_00319981) and REST(Sigma, SASI_Mm01_00196017) mRNAs. Control siRNA were used as negative controls (Sigma, SIC001-10NMOL). Briefly, cells were split, washed and suspended in suspension culture plates (for D0 differentiation) or monolayer (D4 differentiation). siRNAs (3μl of 10μM conc.) and RNAiMax (7 μl) were mixed separately with 75 μl Optimem (Thermofisher, 31985062). Knockdown was initiated by mixing both siRNA and RNAiMAX suspensions together, incubated for 5 mins at RT. The entire 160 μl of silencing mix were added dropwise to 1ml culture or scaled accordingly.
Nuclear extracts and Western blot
Nuclear extracts were prepared using protocols described previously59. Western blotting was performed using standard techniques with PVDF membranes. Primary antibodies used were anti-BRG1 (Abcam, ab110641, 1:1000), anti-FLAG (Sigma, F1804, 1:1000), anti-BAF170 (Bethyl, 1:1000, A301-39A), anti-BAF60c (Cell Signaling Technology, 62265, 1:1000), anti-REST (EMD-Millipore, 07-579, 1:1000), anti-POU3F1 (Abcam, ab126746, 1:1000), or anti-TBP (Abcam, ab51841, 1:2000), Vinculin (Sigma-Aldrich V9131, 1:1000), phospho-Smad (CST 9511, 1:1000) and, Smad1 (CST 9743, 1:1000). Secondary antibodies used were donkey anti-rabbit IRDye 800cw (Licor, 926-32213, 1: 10,000), donkey anti-mouse IRDye 800cw (Licor, 925-32212, 1: 10,000) and donkey anti-goat IRDye 680cw (Licor, 925-68074-1:10,000), HRP-linked-anti-mouse (Cell Signaling Technology, 7076, 1:10000) or HRP-linked-anti-rabbit (Cell Signaling Technology, 7074, 1:10000).
Immunofluoresence
Cells in monolayer were fixed for 30mins in 4% para-formaldehyde, permealized in 0.1% Triton X and 5% goat serum in PBS for 1 hr and incubated with primary antibody (anti-FLAG (Sigma, F1804, 1:300), anti-OCT4 (R&D, MAB1759, 1:300), anti-SOX2 (Abcam, ab97959, 1:300), anti-NANOG (Abcam, ab80892, 1:300), anti-cardiac Troponin T (Thermo Scientific, MS-295-P, 1:100), or TUBB3 (BioLegend, 8012 1:5000) overnight. They were then washed thrice with 0.1% triton X in PBS, incubated with secondary antibody (Goat anti-mouse Alexa 594 (Invitrogen, A11005, 1:1000), Goat anti-rabbit Alexa594 (Invitrogen, 110037, 1:1000) or Donkey-anti-goat AlexaFluor594, 1:1000) for 1hr at RT. Wells were washed thrice and stained with DAPI (1:1000 dilution) for 1-2 min followed by a PBS wash. Images were taken in Keyence confocal microscope at 10x or Zeiss Spinning Disk microscope at 63x (for Extended Data Fig. 5a) magnification.
Flow cytometry
At D10 of differentiation, WT and BRM KO cells were dissociated using TrypLE and fixed with 4% methanol-free formaldehyde. Cells were washed with PBS and permeabilized using FACS buffer (0.5% w/v saponin, 4% Fetal Bovine Serum in PBS). For evaluation of differentiation efficiency, cells were stained with a mouse monoclonal antibody for cardiac isoform Ab-1 Troponin at 1:100 dilution (ThermoFisher Scientific #MS-295-P) or the isotype control antibody (ThermoFisher Scientific #14-4714-82) for 1 hour at room temperature. After washing with FACS buffer, cells were stained with goat anti-mouse IgG Alexa 594 secondary antibody at 1:200 dilution (ThermoFisher Scientific #A-11005) for 1 hour at room temperature. Cells were then washed with FACS buffer, stained with DAPI for 2 minutes, rinsed, and filtered with a 40-micron mesh. At least 10,000 cells were analyzed using the BD FACS AriaII and results were processed using FlowJo (BD Bioscience).
Bulk RNA-seq
Total RNA was isolated from biologically triplicate samples using miRNeasy micro kit with on-column DNase I digestion (Qiagen). RNA-seq libraries were prepared using the Ovation RNA-seq system v2 kit (NuGEN). Libraries from the SPIA amplified cDNA were made using the Ultralow DR library kit (NuGEN). RNA-seq libraries were analyzed using Bioanalyzer, quantified using KAPA QPCR and paired-end 100 bp reads were sequenced using a HiSeq 2500 instrument (Illumina). RNA reads were aligned with TopHat260, counts per gene calculated using feature Counts61 and edgeR62 was used for the analysis of differential expression. K-means clustering and pheatmap functions in R were used to cluster and generate heatmaps. GO enrichment analysis were performed using GO Elite63.
Single cell RNA-seq
Single-cell libraries were prepared using Single Cell 3’ Library Kit v2 (10x Genomics) according to the manufacturer’s protocol. Briefly, about 10, 000 cells were suspended in 0.04% ultrapure BSA–PBS (McLab, #UBSA-500) for GEM generation. GEMs were reverse transcribed, and single stranded DNA were isolated and cleaned. Then cDNA was amplified twice, fragmented, end-repaired, A-tailed and index adaptor ligated, with Ampure cleanup (Beckman Coulter) after each step. Libraries were PCR amplified and cleaned with Ampure beads before shallow sequencing in a NextSeq 500. Read depth normalized libraries were re-sequenced in a NovaSeq sequencer (Illumina).
Sequencing reads were aligned using CellRanger 2.0.2 or 3.0 to the mm9 mouse reference genome. cellranger aggr was used to generate an aggregated read normalized data matrix of samples. The filtered gene matrix was subsequently used to create a Seurat object for QC and tSNE or UMAP visualizations as described in https://satijalab.org/seurat/tutorial64.
Seurat analysis
Seurat package v2.3.4 was used to analyze single cell RNA sequencing data. Cells were filtered to remove dead cells and doublets. After log-normalization, sources of unwanted variation, including differences in the number of UMI, number of genes, percentage of mitochondrial reads and differences between G2M and S phase scores were regressed using the ScaleData function. Clustering was performed using the top 30 principal components and visualized using Uniform Manifold Approximation and Projection (UMAP)17.
Differential gene expression tests were run using the FindMarkers function with min.pct set to 0.1 and logfc.threshold set to 0.25. Selected differentially expressed genes with an adjusted p-value less than 0.05 from the Wilcoxon Rank Sum test were then displayed using the Dotplot function.
Cell trajectories and pseudotime analysis
Single Cell Analysis in Python (Scanpy), version 1.4.5, was used for finding highly variable genes (HVGs), computing dimensionality reduction, regressing unwanted sources of variation, and building developmental trajectory. Two thousand HVGs were selected within each differentiation time separately and merged, to capture differentiation-specific genes64 Variations were regressed from HVGs that encode for ribosomal and mitochondrial proteins. HVGs were then scaled to unit variance and zero mean. Next, the regressed two thousand HVGs were decomposed to fifty principal components using the SciPy, version 1.41, ARPACK Singular Value Decomposition (SVD) solver. A k-nearest neighbor graph was then constructed from a local neighborhood size of ten, using thirty principal components (PCs), the euclidean distance metric, and the connectivity estimation of the manifold set to Unified Manifold Approximation Projection (UMAP) The louvain-graph based clustering algorithm was then run at a resolution of 1.0 on the k-nearest neighbor graph65. A developmental trajectory was resolved by assessing the connectivities of the louvain clusters from the k-nearest neighbor graph, using partition-based graph abstraction (PAGA)18. Finally, the UMAP embedding was recomputed using the PAGA initialization to visualize the developmental trajectory at single cell resolution.
Pseudotime analysis was performed using the URD package19 (version 1.0.2). A single expression matrix with data from three timepoints and WT and Brm−/− in low and high BMP4 conditions was processed in Seurat v2.3.4, as described above. The object was then down-sampled to retain 5000 cells per sample in the low BMP4 dataset or 3000 cells per sample in the combined low and high BMP4 dataset. The down-sampled object was converted to an URD object using the seuratToURD function. Cell-to-cell transition probabilities were calculated by setting the number of nearest neighbors (knn) to the square root of total cells in the object. POU5F1+ clusters from day 4 were set as ‘root’ and all day 10 clusters were set as ‘tip’ cells. An URD tree was constructed by simulating biased random walks from each tip cluster to root.
Signaling Entropy Analysis
Gene-barcode matrices from single-cell RNA-sequencing of day 4 and 6 differentiation samples were first filtered and normalized using the Seurat package implemented in R. The “LogNormalize” method with a default scaling factor of 10,000 was applied for normalization. Differentiation potency was next estimated for each cell within the datasets using the SCENT algorithm implemented in R, which integrates a cell’s transcriptomic profile with existing protein-protein interaction (PPI) maps to quantify signaling entropy43. Higher entropy is an indication of greater developmental potency. A human PPI map compiled from Pathway Commons was used as input for an adjacency matrix (https://github.com/aet21/SCENT). Mouse Ensembl IDs were converted into their human homologues using the AnnotationTools Bioconductor package. The resulting set of genes were then integrated with the human PPI network. The entropy value for each cell was normalized to the largest eigenvalue (maximum possible entropy) of the adjacency matrix. Distributions of normalized entropy values for each sample were then plotted for comparison.
Differential Variability Testing with BASiCS
To assess changes in gene expression variability while accounting for artefactual technical noise and the confounding relationship between variance and mean, single-cell RNA-seq datasets were analyzed via the BASiCS framework as implemented in R42. This approach produces gene-specific estimates of residual over-dispersion: a metric describing how greatly a gene’s variability departs from what is expected given its mean expression. Quality control and filtering of gene-barcode matrices was performed using the BASiCS_Filter function with default parameters. Posterior estimates of mean and residual over-dispersion for each gene were computed using a Markov chain Monte Carlo (MCMC) simulation with 40,000 iterations, log-normal prior and regression analysis.
ATACseq
Assay for transposase-accessible chromatin using sequencing (ATAC-seq) was performed as described30 in two to four biological replicates. Briefly, 50,000 cells (>95% viability) were lysed, washed and tagmented for 45 mins and 3 h for CP and CM cells, respectively. DNA was purified and amplified using universal Ad1 and barcoded reverse primers30. Libraries were purified, quantified and analyzed on a bioanalyzer and sequenced on a NEB NextSeq 550 sequencer using Illumina NextSeq 500/550 High Output v2 kit (150 cycles). Sequencing image files were de-multiplexed and fastq files generated. Reads (paired end 75 bp) were trimmed and aligned to mouse genome mm9 assembly using Bowtie 266 with a minimum mapping quality score of 30.
Statistically enriched bins with a P-value threshold set to 1×106 were used to call peaks67. UCSC genome browser and IGV were used to view the browser tracks. Deeptools package in Galaxy68 (usegalaxy.org) was used to pool multiple replicates to generate 1x genome coverage (average of multiple samples) browser tracks. GREAT69 was used to generate gene lists near ATACseq sites within 100Kb.
ChIPseq
Chromatin immunoprecipitations of histone modifications (H3K27ac and H3K27me3) were performed as described70 with modifications. Briefly, cells were crosslinked with 1% formaldehyde, and quenched with 0.125 M glycine. Frozen pellets (1 ×107) were thawed, washed, dounced and digested with MNase. Chromatin was sonicated at output 4 for 30s twice with a 1 min pause between cycles then centrifuged at 10,000 g for 10 min at 4°C and stored at −80°C. Chromatin was diluted to fivefold, pre-cleared for 2 h followed by immunoprecipitation with primary antibodies for 12-16 hours at 4°C (H3K27ac, Active motif 39133; H3K27me3, CST 9733s). 5% of samples were used as input DNA. Antibody-bound protein-DNA complexes were immunoprecipitated using 25 μl of M-280 goat anti-rabbit IgG or anti-mouse IgG dyna beads for 2 h, washed a total of ten times with buffers [twice with IP wash buffer 1 containing 50 mM Tris.Cl (pH 7.4), 150 mM NaCl, 1% NP-40, 0.25% sodium deoxycholate and 1 mM EDTA), five times with IP wash buffer 2 containing 100 mM Tris.Cl (pH 9.0), 500 mM LiCl, 1% NP-40 and 1% sodium deoxcholate, and then thrice with IP wash buffer 2 along with 150 mM NaCl for increasing stringency and eluted with 200 μl of elution buffer [10 mM TrisCl (pH 7.5), 1 mM EDTA and 1%SDS) at 65°C for 30 mins. Samples were reverse crosslinked, digested with proteinase K and RNAse A, and purified using AMPure XP beads (Beckman Coulter). To prepare libraries for sequencing, DNA was end repaired, A-tailed, adapter ligated (Illumina TrueSeq) and PCR amplified for 14 cycles. PCR-amplified libraries were size selected (200 – 500 bps)and ampure purified. The concentration and size of eluted libraries was measured (Qubit and Bioanalyzer) before single-end 75bp sequencing using a NEBNextSeq sequencer.
Chromatin IP with anti-FLAG antibodies (Sigma, F1806) to probe for BRM binding sites were performed similarly except following modifications. 1) Cells were double crosslinked with 2 mM disuccinimidyl glutarate (DSG) and 1% formaldehyde. 2) MNase digestion conditions were adjusted to have optimal chromatin digestion yielding fragments sizes of 400 to 1Kb. 3) Chromatin binding to antibody and initial two washes contained either 0.05% (low SDS) or 0.2% (high SDS) conditions. 4) Bound protein was competitively eluted with 0.1mg/ml FLAG peptides (ELIM biopharma) and remaining material at 65°C. We observed better ChIP signal over noise at high SDS samples eluted with the FLAG peptides.
Reads (single end 75 bp) were processed as in ATACseq analysis. The HOMER71 motif enrichment package was used to enrich DNA motifs in both ATACseq and ChIP-binding sites. HOMER calculates the q-value of known motifs to statistically confirm to Benjamini-Hochberg multiple hypothesis testing corrections.
ATAC-seq and ChIP-seq analysis
The raw sequence data in fastq files were aligned to the mouse genome build mm9 using bowtie2 aligner 66. Open chromatin regions and regions marked by H3K27Ac for each sample were called using the narrowPeak output of the MACS2 peak caller 72. Regions marked by H3K27me3 were called using the BCP 72 peak caller. A consensus set of peaks across replicates (across samples for each of the ATAC-seq and the histone modification ChIPs) is defined using the -everything followed by -merge options of the bedops program 73. A peak is included in the consensus set of peaks (for the ATAC-seq data or the particular histone modification ChIP-seq data) if it includes a peak called by the relevant peak caller for at least one of the associated replicates. The number of reads mapping to each of the consensus regions for each of replicates using the subread featureCounts program 61. This creates a matrix of raw counts - the number of rows equals the number of consensus regions and the number of columns equals the number of samples. Regions that don’t have at least 5 reads in at least 2 of the samples are filtered out. The raw counts matrix is then normalized using edgeR bioconductor 62,73 R package. For each data set, a linear model is fit for the mean normalized signal in each of the filtered consensus region. This model allows for the main effects of genotype (BRM KO versus Wild type), differentiation time (D0, D4, D6 and D10), conditions (normalBMP4 vs high BMP4) and the interaction between these two variables. The significance of the regions associated with genotype, condition and/or differentiation time is estimated by testing the combined null hypothesis that the main effects of genotype, differentiation time and the interaction effect of between these two variables are all equal to zero. The heatmap of significantly associated regions (FDR < 0.1) is done using the pheatmap package in R.
Data availability
Bulk and single cell RNAseq, ATACseq, and ChIPseq datasets have been deposited in GEO under accession number GSE150186.
Code availability
Codes used to analyze single cell data on Seurat, generate heat maps, UMAPs and URD figures will be available upon request.
Author Contribution
Project design and direction: B.G.B. and S.K.H. ES cell engineering, in vitro differentiation, gene expression, scRNAseq, ATACseq, ChIPseq and data analysis: S.K.H. Additional scRNA-seq analysis: A.P.B, K.R., under direction of B.G.B and I.S.K. Cell culture: A.M.B., K.S. Data analysis: R.V.D. under direction of L.S.W. Manuscript writing: S.K.H and B.G.B with contribution from all authors.
Competing interests
B.G.B. is a co-founder of Tenaya Therapeutics. The work presented here is not related to the interests of Tenaya Therapeutics
Acknowledgements
We thank Natasha Carli, Y. Hao, M. Bernardi, and J. McGuire (Gladstone Genomics Core) for RNA-Seq and 10X Genomics library preparation, the UCSF Center for Applied Technologies for sequencing, Elphege Nora for help with Brm-AID strain construction, Vasumathi Kameswaran for help with BRM ChIP-seq, Reuben Thomas for ATAC-seq and ChIP-seq data analysis, J. Zhang (Gladstone Transgenic Core) for knockout mouse production, and K. Claiborn for editorial assistance. This work was supported by grants from the NIH/NHLBI (P01HL089707, Bench to Bassinet Program UM1HL098179, and R01HL114948), to B.G.B; and postdoctoral fellowships from the American Heart Association (13POST17290043), Tobacco Related Disease Research Program (22FT-0079) and NIH training grant (2T32-HL007731 26) to S.K.H. I.S.K. was supported by funds from the Society for Pediatric Anesthesia, Hellman Family Fund, UCSF REAC Award and the UCSF Department of Anesthesia and Perioperative Care. This work was also supported by an NIH/NCRR grant (C06 RR018928) to the J. David Gladstone Institutes and by The Younger Family Fund (B.G.B.).