Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Dissecting the regulatory logic of specification and differentiation during vertebrate
embryogenesis
Jialin Liu1,2,*, Sebastian M. Castillo-Hair3,+, Lucia Y. Du1,2,+, Yiqun Wang1,4, Adam N. Carte1,5, Mariona
Colomer-Rosell1,2, Christopher Yin3, Georg Seelig3,6, and Alexander F. Schier1,2,*
1Biozentrum, University of Basel, Basel, 4056, Switzerland
2Allen Discovery Center for Cell Lineage Tracing, University of Washington, Seattle, WA, 98195, USA
3Department of Electrical & Computer Engineering, University of Washington, Seattle, WA, 98195, USA
4Center for Marine Biotechnology and Biomedicine, Scripps Institution of Oceanography, UCSD, La
Jolla, CA, 92037, USA
5Department of Genetics, Blavatnik Institute, Harvard Medical School, Boston, 02115, USA
6Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA,
98195, USA
*Correspondence: jialin.liu@unibas.ch, alex.schier@unibas.ch
+Equal contribution: these authors contributed equally and are listed alphabetically
Abstract
The interplay between transcription factors and chromatin accessibility regulates cell type
diversification during vertebrate embryogenesis. To systematically decipher the gene regulatory logic
guiding this process, we generated a single-cell multi-omics atlas of RNA expression and chromatin
accessibility during early zebrafish embryogenesis. We developed a deep learning model to predict
chromatin accessibility based on DNA sequence and found that a small number of transcription factors
underlie cell-type-specific chromatin landscapes. While Nanog is well-established in promoting
pluripotency, we discovered a new function in priming the enhancer accessibility of mesendodermal
genes. In addition to the classical stepwise mode of differentiation, we describe instant differentiation,
where pluripotent cells skip intermediate fate transitions and terminally differentiate. Reconstruction of
gene regulatory interactions reveals that this process is driven by a shallow network in which maternally
deposited regulators activate a small set of transcription factors that co-regulate hundreds of
differentiation genes. Notably, misexpression of these transcription factors in pluripotent cells is
sufficient to ectopically activate their targets. This study provides a rich resource for analyzing
embryonic gene regulation and reveals the regulatory logic of instant differentiation.
Main
Cell differentiation can generally be conceptualized as a stepwise process. It starts with
specification to a developmental fate and progresses through sequential changes in gene expression
and morphology to ultimately result in the acquisition of specialized structures and functions (Liberali &
Schier, 2024). For example, insulin-producing beta cells derive from a stepwise specialization through
pluripotent, endodermal, pancreatic and endocrine states (Murtaugh, 2007). However, certain cell types
rapidly differentiate during early embryogenesis without undergoing multi-step transitions These include
the trophectoderm in mammals (Lim et al., 2020), and the liver-like yolk syncytial layer (YSL) and the
skin-like enveloping layer (EVL) in fish (Kimmel et al., 1990, 1995). This phenomenon of 'instant
differentiation' is characterized by its speed and simplicity compared to stepwise differentiation. For
example, at the onset of gastrulation, the EVL has already differentiated, displaying distinct morphology
(Kimmel et al., 1990), cytoskeleton and adhesion structures (Zalik et al., 1999), with hundreds of genes
differentially expressed compared to embryonic cells (Satija et al., 2015). Although this phenomenon
has long been described, the gene regulatory logic of instant differentiation has remained elusive.
Recent advances to systematically study gene regulatory networks (GRNs) provide the
opportunity to address this question (Badia-i-Mompel et al., 2023; Fleck et al., 2023; Janssens et al.,
2022; Kamimoto et al., 2023; Saunders et al., 2023). For example, single-cell RNA sequencing (scRNA-
seq) has led to the systematic reconstruction of the gene expression trajectories during cell type
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
specification and differentiation (Briggs et al., 2018; Farrell et al., 2018; Qiu et al., 2022; Wagner et al.,
2018). More recently, these efforts have been complemented with single-cell measurements of
chromatin accessibility (Argelaguet et al., 2019; Cao et al., 2018; Ma et al., 2020). Deep learning models
trained on these data can predict cell type accessibility from sequence and can be interrogated to reveal
cis-regulatory elements (Ameen et al., 2022; Eraslan et al., 2019; Janssens et al., 2022; Zhou &
Troyanskaya, 2015). Furthermore, the capacity to correlate changes in the epigenome with alterations
in the transcriptome facilitates the construction GRNs (Badia-i-Mompel et al., 2023; Bravo González-
Blas et al., 2023; Fleck et al., 2023). In this study, we created a single-cell multi-omics (scMultiome)
atlas that jointly tracks RNA expression and chromatin accessibility throughout early zebrafish
embryogenesis. We utilized this resource to develop a deep learning model and construct GRNs to
dissect the gene regulatory logic of instant differentiation.
Single-cell multi-omics identifies gene expression and chromatin accessibility dynamics during
embryogenesis
Single-cell gene expression (Farrell et al., 2018; Wagner et al., 2018) or chromatin accessibility
(Sun et al., 2024) have been profiled separately during zebrafish embryogenesis. However, we lack
one-to-one cell correspondence and even stage correspondence between the two modalities, which
would provide a more detailed and integrated understanding of cellular states and regulatory
mechanisms. To systematically dissect the gene regulatory logic that governs embryogenesis, we
simultaneously measured gene expression and chromatin accessibility within individual nuclei during
early zebrafish embryogenesis (Figure 1 A, 10x Genomics). Our comprehensive dataset spans nine
developmental stages, from pluripotency to early organogenesis (high stage at 3.3 hours post-
fertilization (hpf) to 6-somite stage at 12 hpf). In total, 40,992 high-quality single nuclei were captured,
with a median of 2,082 expressed genes and 16,925 ATAC fragments detected per nucleus (Table S1),
an improvement to previous standalone single-cell RNA sequencing (scRNA-seq) (Farrell et al., 2018)
and single-nucleus ATAC sequencing (snATAC-seq) (Sun et al., 2024) datasets. For quality control we
generated standalone snATAC-seq data using the same 10x Genomics technology for 21,050 nuclei at
the onset of gastrulation (50% epiboly) and 6-somite stage. The chromatin accessibility between the
scMultiome and standalone snATAC-seq was highly concordant (Figure S1). Using the snRNA-seq
modality of the scMultiome, we identified 95 cell states based on previously described cell state markers
(Farrell et al., 2018; Qiu et al., 2022; Wagner et al., 2018). These cell states were also independently
found in the snATAC-seq modality-based UMAP space, revealing that both modalities can capture
cellular diversity (Figure 1A and Figure S2). These results indicate that the single-cell multi-omics data
is of high quality.
To characterize the chromatin accessibility profiles of each cell state, we used a cluster-specific
and replicate-aware peak calling approach (Granja et al., 2021). We identified 444,530 peaks,
representing putative cis-regulatory elements (CREs). Distal CREs (dCREs), including putative
enhancers, were defined as peaks >500 bp from an annotated transcription start site (TSS). Peaks
within <500 bp of a TSS were annotated as promoters. To link cell states and types across different
stages into trajectories, we used gene expression similarity between adjacent stages(Briggs et al., 2018;
Qiu et al., 2022) (Figure 1B, Table S2). Briefly, we connected each cell in a given stage to its most likely
ancestors in the preceding stage based on Euclidean distances in a low-dimensional space (see
methods). The reconstructed trajectories recapitulated the developmental paths described in previous
studies (Farrell et al., 2018; Qiu et al., 2022; Wagner et al., 2018). The earliest branching events include
the segregation between embryonic and extra-embryonic cell types; i.e. the enveloping layer (EVL) and
the yolk syncytial layer (YSL). Early differentiation of EVL and YSL provides epidermal and lipid-
metabolizing functions, respectively, which are essential for embryo development. From 50% epiboly
(onset of gastrulation), we observed the expansion of mesendodermal cell types, while ectodermal cell
types expand from the bud stage (end of gastrulation), culminating in the presence of two dozen cell
types at 6-somite stage, including heart, adaxial muscle, forebrain, epidermis, hatching gland, and
notochord. In summary, the scMultiome dataset provides an extensive resource to explore chromatin
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
accessibility and gene expression during early zebrafish embryogenesis. In the following sections we
highlight how the scMultiome dataset can provide new biological insights.
Figure 1: Zebrafish single-cell multi-omics atlas reveals gene expression and chromatin
accessibility dynamics
A. Upper panel: single-cell multi-omics data were collected from zebrafish embryos at nine
developmental stages (represented by colored dots) using 10x Genomics technology. Middle
panel: Uniform Manifold Approximation and Projection (UMAP) visualization of single cells, with
each cell colored according to its developmental stage. The UMAP coordinates are based on
either RNA or ATAC-seq data. Lower panel: UMAP visualization of single cells at 6-somite stage,
colored by clusters. The clusters are defined by gene expression, and the UMAP coordinates are
derived from either RNA or ATAC-seq data.
B. Inferred relationships between cell states during early zebrafish embryogenesis. Each row
corresponds to a specific cell state, while each column represents a developmental stage. The
nodes are color-coded according to different germ layers. All edges with weights above 0.2 are
displayed. olf.+adeno., olfactory + adenohypophyseal; NPB, neural plate border; Pron.,
pronephros; PSM, presomitic mesoderm; EVL, enveloping layer; YSL, yolk syncytial layer.
B
D
High
50%
Epiboly
75%
Epiboly 6-Somite
Dome
CREs acquired from High stage
dCREs acquired from Oblong to 6-Somite
Maintained
Disappear
E
High
Oblong
Dome to 50% Epiboly
Shield
75% Epiboly to Bud
0.0 0.2 0.8
Promoters
Genes active during
75% Epiboly to Bud
0.40.61.0
dCREs
0.0 0.2 0.8
Promoters
Genes active during
Dome to 50% Epiboly
Proportion of accessible CREs
0.40.61.0
dCREs
C
3.3 hpf 5.3 hpf 12 hpf
ZGA Gastrulation Somitogenesis
10x scMultiome
RNA ATAC
6-Somite
RNA ATAC
UMAP1
UMAP2
8 hpf
High 50% Epiboly 75% Epiboly 6-Somite
UMAP1
UMAP2
0 2000 4000 6000 8000
EVL
Notochord
Hatching gland
Tailbud
Endoderm
Adaxial cells
PSM
Somite
Heart field
Cephalic mesoderm
Pron.duct+Blood.island
Optic primordium
Telencephalon
Dorsal diencephalon
Ventral diencephalon
Midbrain
Hindbrain
Spinal cord
Differentiating neuron
Neural crest
Epidermis
Posterior NPB
Placode(otic)
Placode(olf.+adeno)
Placode(lens)
Oblong
Dome
30% Epiboly
50% Epiboly
Shield
75% Epiboly
Bud
6−somite
Number of maintained dCRE
Edge weight
Neural ectoderm
Non-neural ectoderm
Mesoderm
Endoderm
EVL
Others
YSL
A
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
C. Analysis of distal cis-regulatory elements (dCREs) acquired at high stage (left panel) or any stage
from oblong to 6-somite (right panel), examining which elements are maintained at or disappear
by the 6-somite stage. Note that most dCREs at high stage are maintained in most cell types
throughout early embryogenesis, whereas dCREs acquired later disappear or become cell-type
-specific.
D. The number of acquired dCREs at each stage that are maintained by the 6-somite stage. The
figure is divided into three groups: EVL, mesendoderm, and ectoderm, shown from left to right.
Note that the timing of CRE appearance generally coincides with the timing of cell type
specification or differentiation.
E. The proportion of promoters or dCREs that are accessible at each stage. Note that promoters
are often accessible before the expression of the corresponding gene.
Chromatin accessibility dynamics reveal cell-type- and stage-specific cis-regulatory element
usage
To analyze the temporal dynamics of putative distal cis-regulatory element (dCREs) usage, we
asked how many dCREs acquired at a given stage were still present at 6-somite stage (Figure 1C and
1D). We found that more than 50% of the dCREs acquired at the high stage were maintained at the 6-
somite stage. These dCREs were ubiquitously accessible in all cell types (Figure 1C) and associated
with housekeeping genes (e.g. genes involved in metabolic processes and RNA biosynthetic processes)
(Table S3). In contrast to these ubiquitous dCREs, dCREs acquired and maintained after the high stage
were mostly cell-type-specific (Figure 1C). The timing at which these cell-type-specific dCREs are
acquired coincided with the specification and differentiation of the associated cell types (Figure 1D and
Figure S3). Three waves were apparent: EVL and YSL (Figure S3) dCREs at the 6-somite originate
from the oblong to the 50% epiboly stages (from blastula to the onset of gastrulation); mesoderm-
associated dCREs appear at 30% epiboly (before the onset gastrulation); dCREs in ectoderm-derived
cell types emerge from the bud (end of gastrulation) to the 6-somite stage. These results reveal a
hierarchical appearance of dCREs, in which ubiquitous dCREs are acquired early, whereas cell-type-
specific dCREs appear during cellular specification and differentiation.
In addition to studying the relationship of dCRE usage with cell type diversification, the
simultaneously measured two modalities allow us to compare the temporal emergence of CREs with
the timing of gene activation. We identified cell-type-specific genes that become active from dome to
50% epiboly (onset of gastrulation). We inferred their associated dCREs (association score>0.1 and p-
value<0.05) based on the correlation analysis between chromatin accessibility and gene expression
(see methods). We observed that the majority of dCREs become accessible at the time when gene
expression starts (Figure 1E). In contrast, promoter peaks generally preceded gene expression. For
example, for genes active during 75% epiboly to bud stage (mid- to end of gastrulation), the majority of
promoters are already accessible at the high stage (Figure 1E). These results suggest that promoters
are generally primed for future gene expression (Pálfy et al., 2020; Reddington et al., 2020), whereas
the accessibility of most dCREs coincides with gene activation.
Nanog primes chromatin accessibility of putative mesendodermal distal cis-regulatory
elements
While most putative dCREs become accessible when gene expression starts, some dCREs are
accessible beforehand (Figure 1E), reminiscent of the concept of enhancer priming (Spitz & Furlong,
2012). In this process, transcription factors (TFs) bind to an enhancer earlier in embryogenesis,
rendering it accessible for binding by subsequent TFs that activate gene expression during later stages.
Enhancer priming has been suggested to enable rapid and sustained transcriptional responses (Falo-
Sanjuan et al., 2019). Multi-omics profiling during mouse embryogenesis indicated that
neuroectodermal but not mesendodermal enhancers are primed by chromatin accessibility in the
epiblast (Argelaguet et al., 2019). To investigate if this property is conserved in zebrafish, we analyzed
if and when dCREs associated with mesendodermal and neuroectodermal genes are primed. In
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
contrast to mouse embryogenesis, we found that 13% of dCREs associated with mesendodermal
genes active between dome and 50% epiboly are already accessible at the high stage (Figure 2A and
2B) when most cells are pluripotent. This percentage is significantly higher than that of non-associated
dCREs (6%). In contrast to the mesendoderm, we did not find a higher proportion of high-stage
accessible dCREs associated with neuroectodermal genes (active from 75% epiboly to the bud stage,
the end of gastrulation) (Figure 2C). Instead, at 50% epiboly (onset of gastrulation), the proportion of
accessible neuroectodermal associated dCREs is significantly higher than that of non-associated
dCREs (Figure 2C and D). These results suggest that dCRE priming occurs in both mesendoderm and
neuroectoderm and that the timing of priming correlates with the temporal specification of these tissue
types.
Figure 2: Nanog primes putative mesendodermal distal cis-regulatory elements
A. Left panel: representative chromatin accessibility tracks for all CREs near TSS of the noto gene
across the dorsal mesoderm trajectory. The aggregated pseudo-bulk data for chromatin
accessibility in each cell state is presented. Right panel: violin plot shows the expression of noto
across the dorsal mesoderm trajectory. Bottom panel: the numbers above chromosome 13
indicates the genomic coordinates.
B. Proportion of mesendodermal putative distal cis-regulatory elements (dCREs) primed at the high
stage. For each mesendodermal gene, we inferred its associated dCREs from all dCREs located
within 50 kb upstream or downstream of the transcription start site (TSS). dCREs were
considered associated if they had an association score >0.1 and a p-value <0.05. "High stage
priming" indicates that these dCREs were already accessible at the high stage.
C. The left bar graph shows the proportion of neuroectodermal dCREs that were primed at the high
stage. The right bar graph illustrates the proportion of neuroectodermal dCREs that were primed
at the 50% epiboly stage. Similar to mesendodermal genes, for each neuroectodermal gene we
inferred its associated dCREs. "High stage priming" means these dCREs were accessible at the
high stage, while "50% Epiboly priming" means they were accessible at the 50% epiboly stage.
D. Similar to A, but for telencephalon trajectory.
E. Chromatin accessibility of primed versus non-primed dCREs across various rescue conditions.
Here, the dCREs only refer to the associated ones. The lower and upper intervals indicated by
the dashed lines (“whiskers”) represent 1.5 times the interquartile range (IQR), and the box
shows the lower and upper intervals of IQR together with the median.
AB
DEF
Mesoendodermal dCRE
High stage priming
0.0 0.1
Proportion of primed dCRE
0.30.5
Associated
dCREs
Non−associated
dCREs
p=2e−08
0.20.4
Neuroectodermal dCRE
High stage priming
0.0 0.1
Proportion of primed dCRE
0.4 0.5
p=0.6
0.3
0.2
Neuroectodermal dCRE
50% Epiboly priming
0.0 0.1
Proportion of primed dCRE
0.4 0.5
p=2.1e−07
0.3
0.2
C
High
Oblong
Dome
noto
14960000 14970000 14980000 14990000 15000000
Chromsome 13
0.15
0.20
0.25
Association
score
Expression
Accessibility
Putative CREs
0123
50% Epiboly
30% Epiboly
Dorsal posterior trajectory
Accessibility
01234
six7
7400000 7405000 7410000 7415000 7420000
Chromsome 7
0.10
0.15
0.20
Association
score
High
Oblong
Dome
50% Epiboly
Shield
75% Epiboly
Bud
30% Epiboly
Telencephalon trajectory
Expression
Putative CREs
High stage primed dCRE
50% Epiboly primed dCRE
Promoter
Promoter
1
4
Chromatin accessibility
Primed dCREs Non−primed dCREs
MZnps
Wild type
Nanog rescue
pou5f3 rescue
Sox19a rescue
Nanog+Pou5f3 rescue
Nanog+Sox19a rescue
Nanog+Pou5f3+Sox19a rescue
Pou5f3+Sox19a rescue
Primed
dCREs
p=4.2e−13
0.5
0.0 0.1 0.2 0.3 0.4
Associated
dCREs
Non−associated
dCREs
Associated
dCREs
Non−associated
dCREs
023
Proportion of dCREs with Nanog binding
Non−primed
dCREse
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
F. The proportion of primed versus non-primed dCREs that overlap with Nanog peaks at the high
stage. Here, the dCRES only refer to the associated ones.
The p-values for B, C and F were determined using pairwise Fisher's exact tests.
To investigate which TFs prime mesendodermal dCREs at high stage, we focused on three
maternally deposited pioneer TFs, Nanog, Pou5f3 and Sox19b. We compared the chromatin
accessibility of primed mesendodermal dCREs between triple mutant and various rescue conditions,
using previously published data (Miao et al., 2022). First, non-primed dCREs showed much lower
accessibility than primed dCRE (Figure 2E). Second, we discovered that chromatin accessibility
increases whenever the rescue includes Nanog, with the greatest increase observed in the triple rescue.
In contrast, the rescue conditions with just Pou5f3 and Sox19b contributed little to chromatin
accessibility. These observations suggest that Nanog is the primary factor that primes mesendodermal
dCREs. To further test this idea, we examined the overlap between primed mesendodermal dCREs
and Nanogbinding sites using CUT&RUN data from the high stage (X. Wang et al., 2022). We found
that approximately 30% of primed mesendodermal dCREs overlap with NanogCUT&RUN peaks, a
proportion much higher than the overlap observed with non-primed mesendodermal dCREs (Figure 2F).
Overall, these results indicate that the dCREs of numerous mesendodermal genes are primed primarily
by the pioneer TF Nanog.
Deep learning identifies a small number of transcription factors that predict cell type- specific
chromatin landscapes
Deep-learning models, such as convolutional neural networks (CNNs), have emerged as
powerful tools for predicting genomic "activity profiles" based on DNA sequences (Eraslan et al., 2019).
Interpretative methods (Novakovsky et al., 2022), can extract sequence motifs responsible for the
predicted "activity profiles" and help decode the cis-regulatory logic of gene expression. To
systematically investigate the sequence motifs of chromatin accessibility, we developed DeepDanio, a
deep learning model (Figure 3A, see methods). DeepDanio was trained to predict chromatin
accessibility in all 95 cell states across all developmental stages given a 500bp DNA sequence. To
improve accuracy and generalization, we designed DeepDanio as an ensemble of three deep CNNs
with residual connections (Figure 3A and Table S7), where each CNN was trained on CREs from a
different subset of chromosomes (~80% of all CREs), with the remaining CREs used for early stopping
to prevent overfitting (2 chromosomes, 10% of all CREs) and for performance evaluation (2
chromosomes, 10% CREs). On sequences held out from training, DeepDanio showed good
performance in predicting accessibility (Figure 3B). Overall, 92.4% of test CREs were predicted with a
statistically significant correlation coefficient (Figure 3C, permutation test, p-value <0.01). Additionally,
across all CREs within each specific cell state, DeepDanio attained a high correlation between observed
and predicted chromatin accessibility (Figure S4A).
To extract sequence motifs that influence cell state accessibility predictions from DeepDanio,
we used deep learning interpretation methods. Starting from the top 10,000 most specific putative CREs
in each cell state, we used DeepExplainer (Lundberg & Lee, 2017) to calculate the contribution of each
nucleotide to cell state predictions. We then used TF-MoDISco (Shrikumar et al., 2018) to identify, align,
and cluster regions with high contribution scores into de novo motifs for each cell state. We identified
an average of 17 motifs per cell state (see data access). Scanning the top 10,000 specifically accessible
CREs for each cell state using TF-MoDISco-identified motifs provides cell-type-specific, genome-wide
transcription factor binding site (TFBS) predictions. For example, nucleotides important for predicting
accessibility in mesendoderm map to the motif for Tbxta (Figure 3D). For YSL, significant nucleotides
for predicting accessibility emerge as a motif for Mxtx2 (Figure 3D). These results agree with previous
findings that Tbxta and Mxtx2 are key players in the development of the mesoderm and YSL,
respectively (Schier & Talbot, 2005; Schulte-Merker et al., 1994; Xu et al., 2012). We further validated
the predicted TFBSs for the motifs that have available whole embryo ChIP-seq data (Nelson et al.,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
2017a; Xu et al., 2012), confirming that ChIP-seq signals are enriched at the predicted TFBS locations
(Figure 3D and Figure S4B).
The above results indicate that DeepDanio can accurately predict chromatin accessibility and
identify contributing motifs. To explore how many motifs are required to define cell-type-specific
chromatin accessibility, we calculated the distribution of the predicted TFBSs for each motif across the
top 10,000 specifically accessible CREs for each cell state. For many cell states, over 80% of TFBSs
come from fewer than 6 motifs (Figure 3E-F). For example, among the YSL-specific CREs at 50%
epiboly (onset of gastrulation), the Gata6 motif corresponds to 40% of the TFBSs, the Mxtx2 motif to
another 30%, and the Hnf4 motif to another 10% (Figure 3E and Figure S4C). Many of these motifs are
associated with TFs known to be fate regulators of the respective cell types. For example, Grhl3 is
critical for EVL (Miles et al., 2017), Myod1 for adaxial cells (Weinberg et al., 1996), and Tp63 for the
epidermis (H. Lee & Kimelman, 2002) (Figure 3E and Figure S4C). These results indicate that small
sets of TFs play crucial roles in shaping the accessible chromatin landscape in each cell type (Wei et
al., 2018).
Figure 3: DeepDanio predicts chromatin accessibility based on DNA sequence and identifies
sequence motifs of chromatin accessibility
A. Architecture of DeepDanio. In each model, a 500bp sequence is processed via a series of
residual blocks containing convolutional layers and skip connections. The model output is the
predicted log10(CPM) (accessibility) of each of the 95 cell states, obtained as the average of
three independently trained models.
B. Measured vs. predicted accessibility for all cell states on a selected CRE held out from training,
along with their Pearson correlation coefficient (r) and p-value.
C. Distribution of Pearson correlation values across all CREs held out from training (blue). CREs in
the evaluation set of all three individual models and their predictions were pooled together. To
A
−0.5 0.0 0.5 1.0 1.5
0.0 0.5 1.0
dCRE
(chr7: 65810735−65811235)
True Accessibility
Predicted Accessibility
r=0.92
p=1.8e−40
B
Pearson's r
Frequency
−0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0
0 1000 3000 5000
mean=0.02 mean=0.47
C
D
2.5
5.0
7.5
10.0
12.5
-2000 -1000 0 1000
tbxta
2000
ChIP−Seq Signal
Distance from TFBS (bp)
0
5
10
15
20
-2000 -1000 01000
mxtx2
2000
Distance from TFBS (bp)
ChIP−Seq Signal
F
E
Cumulative percentage of predicted TFBSs
0.5 0.6 0.7 0.8 0.9 1.0
YSL
(50% Epiboly)
gata6
mxtx2
hnf4
0.4 0.9 1.0
EVL
(50% Epiboly)
0.5 0.6 0.7 0.8
grhl3
grhl3-variant
klf17
cebpb
0.75 0.8 0.85 0.9 0.95 1.0
Adaxial cells
(6-somite)
myod1
snai1a
hoxa9b
1.0
Telencephalon
(6-somite)
0.4 0.6 0.8
sox3
zic1
patz1
otx2
0.5 0.6 0.7 0.8 0.9 1.0
Epidermis
(6-somite)
tfap2a
tp53
tead3
0.3 0.4 0.9 1.0
Hatching gland
(6-somite)
0.5 0.6 0.7 0.8
foxa
gsc
klf17
foxa-varaint
Frequency
2 4 6 8 10
0 5 10 15 20
Mesoderm peak1
Mesoderm peak2
Nucleotide contribution
tbxta
YSL peak1
YSL peak2
mxtx2
Nucleotide contribution
Number of motifs accounting for more than
80% of predicted TFBSs
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
assess the statistical significance of the observed correlation, a null distribution was generated
by randomly shuffling the cell type labels of the predictions (red, permutation test). The p-value
was inferred by comparing the observed correlation to this null distribution.
D. Left panel, illustration of nucleotide importance scores for mesoderm and YSL CREs. The higher
the letter, the greater its importance in predicting chromatin accessibility. The predicted TFBSs
are highlighted in grey, and the motifs identified from ChIP-seq are displayed in the top right
corner. Right panel: The ChIP-seq signal in either 2kb upstream or downstram of TFBS. The gray
area indicates 95% confidence interval.
E. Cumulative percentage of TFBSs from all motifs identified by TF-MoDISco within the top 10,000
cell-state-specific CREs. Note that over 80% of TFBSs come from the first three or four motifs.
F. Distribution of the number of distinct motifs required to account for 80% of predicted TF binding
sites for all 95 cell states. Note that for most of the 95 states, over 80% of TFBSs come from
fewer than 6 motifs.
GRN reconstruction identifies critical regulators of cell identity
Since our data contain both gene expression and chromatin accessibility information from the
same nuclei, it provides a way to incorporate CREs to infer enhancer-driven gene regulatory networks
(eGRNs). We utilized the SCENIC+ pipeline (Bravo González-Blas et al., 2023) to reconstruct eGRNs
during early zebrafish embryogenesis. Overall, SCENIC+ utilizes a three-step workflow (Figure 4A):
First, we identified each TF's potential CREs based on motif scanning using motifs collected in DANIO-
CODE (Baranasic et al., 2022). Second, we linked these CREs to their target genes based on the
correlation between gene expression and CRE accessibility. Third, we connected TFs to their target
genes through CREs (see methods). This process resulted in a set of 100 enhancer-driven regulons
(eRegulons), with each eRegulon consisting of a TF, its enhancers, and the target genes of these
enhancers (Table S4). The 100 eRegulons contain 4,741 genes and 26,419 CREs. The size of each
eRegulon varies from 10–976 genes, with a median size of 70 genes (Figure 4B).
For each eRegulon, we evaluated its activities across 95 cell states (Table S5) using target
gene expression and defined its cell state specificity score (Table S6) based on Jensen-Shannon
divergence (see methods). The identified cell-state-specific eRegulons are consistent with current
knowledge. For instance, the eRegulons specific to 6-somite adaxial cells include Myod1 (Weinberg et
al., 1996) and Mespaa (Sawada et al., 2000), and the eRegulons specific to 6-somite telencephalon
include Foxg1 (Zhao et al., 2009)and Nr2f2 (Chowdhury et al., 2024). At 50% epiboly (onset of
gastrulation), the EVL-specific eRegulons include Klf17 (Liu et al., 2016; Miles et al., 2017) and Grhl3
(Miles et al., 2017), while the YSL-specific eRegulons include Hnf4a and Gata6 (Xu et al., 2012) (Figure
4C).
The UMAP plot provides additional support that the activities of these eRegulons are highly
specific to their corresponding cell types (Figure 4C). Clustering cell types based on eRegulon activity
shows that developmentally related cell types share similar eRegulon activity (Figure 4D). For instance,
neural ectoderm cell types tend to cluster together, reflecting their shared eRegulon activity. Similarly,
paraxial mesoderm cell types also exhibit clustering based on their shared regulatory activity. Taken
together, our comprehensive network analysis connects accessible chromatin regions, putative
enhancers, TFs and their binding motifs, and target genes.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure 4: SCENIC+ infers enhancer-driven gene regulatory networks
A. SCENIC+ workflow. TF., transcription factor; CRE., cis-regulatory element; TG., target gene. A
eRegulon consists of a TF, all the putative CREs regulated by this TF, and all the target genes
regulated by this TF.
B. Histogram of the number genes regulated by each TF.
C. Upper panel: eRegulons ranked by specificity score. UMAP visualization showing eRegulon
activity at the single-cell level, with cells colored according to their eRegulon activity levels.
D. Heatmap displaying the mean eRegulon activity for all eRegulons across different cell types at
the 6-somite stage. For each cell type, mean eRegulon activity was calculated by averaging the
activity across all cells within the type. Hierarchical clustering was then performed based on these
mean eRegulon activities to group cell types with similar activity patterns.
Combining GRNs and deep learning uncovers a shallow network driving instant differentiation
With the GRN, eRegulons and DeepDanio in hand, we wished to dissect the regulatory logic of
the instant differentiation of EVL and YSL. We focused on the top 10 specific eRegulons in each stage
of the EVL trajectory and YSL trajectory (Figure 5A). We discovered that the eRegulons associated with
the EVL and YSL were remarkably large, including 12 of the 15 eRegulons with more than 360 target
genes (Figure S5). For example, the two largest eRegulons, each including nearly 1000 genes, are
Grhl3 in EVL and Onecut1 in YSL. We found that many of these eRegulons become active shortly after
zygotic genome activation and reach peak activity at 50% epiboly (onset of gastrulation), in line with the
observation that EVL and YSL have differentiated by this stage.
Specificity score
AC
eRegulon activity
Telencephalon
Optic primordium
Ventral diencephalon
Hindbrain
Dorsal diencephalon
Midbrain
Neural crest
Spinal cord
Differentiating neuron
Somite
Adaxial cells
PSM
Tailbud
Placode(olf.+adeno)
Placode(otic)
Cephalic mesoderm
Endoderm
Heart field
Pron.duct+Blood.island
Endothelial progenitors
Neural floor plate
Notochord
Placode(lens)
Posterior NPB
Epidermis
Ionocyte.progenitors
Hatching gland
EVL
YSL
hnf1ba
atf3
hnf4a
hnf4b
onecut1
gata4
bhlhe41
gata6
nr2f6b
gata5
rxrgb
sox9a
foxp2
klf13
pitx2
foxo1a
bhlha15
rreb1b
foxa
foxa2
mespab
mespaa
myod1
zbtb18
fli1a
erfl3
meis3
wt1a
pknox2
rxraa
tcf12
tbx1
tcf3b
tbx16
tbx16l
snai1a
tbxtb
tbxta
hoxc9a
hoxa9a
tcf7
ybx1
e2f3
e2f7
etv5b
ctcf
etv5a
nfya
erf
tcf7l1a
tcf7l2
tcf7l1b
zeb1a
mycn
zic1
foxg1a
otx2a
otx1
nr2f2
otx2b
her6
gsc
gli1
rfx4
rfx3
gli3
gli2b
sox2
sox19b
sox19a
sox3
egr2b
egr2a
en2a
glis1b
zic3
zic2b
her9
zic2a
dmbx1a
rfx2
e2f5
foxd3
ets2
tead1b
tfap2c
hic1
rreb1a
elf1
tfap2a
klf6a
gata2a
hey1
tead3b
tead3a
grhl3
cebpb
TCF4
klf17
prox1a
B
D
eRegulons
0400 600 800
Number of target genes
200
0.17 0.18 0.19 0.20 0.21
Adaxial cells (6-Somite)
myod1
zbtb18
pknox2
hoxa9a
mespaa
tcf12
0.17 0.18 0.19 0.20 0.21 0.22 0.23
Telencephalon (6-Somite)
foxg1a
gli2b
gli3
zic1
nr2f2
egr2b
0.18 0.24 0.26
EVL (50% Epiboly)
0.20 0.22
klf17
cebpb
tcf4
tead3a
grhl3
0.170 0.175 0.180 0.185 0.190
YSL (50% Epiboly)
hnf4a
gata6
hic1
hnf1ba
onecut1
UMAP1
UMAP2
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Plotting the targets of these eRegulons reveals that the majority of differentiation genes are
regulated by only a few TFs (Figure 5B). Importantly, these TFs share many target genes, indicating
that most differentiation genes are regulated combinatorially. By overlapping the TFs identified by
SCENIC+ with the TFs identified by TF-MoDISCo and DeepDanio (Figure 5C and Figure S6), we
defined high-confidence candidate TFs involved in the differentiation of EVL (Grhl3, Klf17, Cebpb,
Tead3a, Tead3b, and Tfap2a) and YSL (Onecut1, Gata6, Gata4, Hnf4a, and Hnf4b). Notably, among
these TFs, we also found auto-regulatory and cross-regulatory interactions. For example, Grhl3 can
regulate itself and also engage in cross-regulation with Klf17, Cebpb, Tead3a and Tead3b (Figure 5D).
Overall, the inferred GRN suggests that instant differentiation is regulated by a handful of TFs that
function together to activate a large number of differentiation genes.
To explore this potential combinatorial regulation further, we conducted in-silico ablation(Yin et
al., 2024) analysis of the motifs from DeepDanio. In this analysis, motifs were replaced with dinucleotide
shuffles to preserve sequence composition while disrupting the motif. The resulting decrease in
accessibility was quantified as an 'ablation score' (see methods), with higher scores indicating a greater
impact on chromatin accessibility. Our findings reveal that while all motifs influence chromatin
accessibility, Grhl3 has a much higher impact than the others, aligning with previous research that
categorizes it as a pioneer TF(Jacobs et al., 2018)(Figure 5E). To measure the interaction score for
pairs of motifs co-occurring on the same CRE, we calculated the sum of two individual motif ablation
scores minus the score from ablating both motifs simultaneously. A score above zero indicates
cooperativity between co-binding TFs, contributing to chromatin accessibility. For example, we
discovered that the majority of motif pairs between Grhl3 and Klf17 exhibit scores greater than zero
(Figure 5E), supporting the existence of cooperative regulation between the two TFs. Additionally, we
observed that this regulatory cooperation is influenced by the distance between motifs, with the
interaction score decreasing as the distance increases (Figure 5E). Other motif pairs show similar
patterns as Grhl3 and Klf17 (Figure S7). These simulations suggest that there is extensive distance-
dependent cooperativity between co-binding TFs.
To identify the putative regulators that might activate these EVL and YSL differentiation TFs,
we first identified DeepDanio motifs exhibiting high contribution scores at the onset of differentiation
(Figure 5C). We then narrowed our focus to motifs whose corresponding TFs are expressed in the EVL
or the YSL. Following these criteria, we identified two primary candidates: Irf6 for EVL and Mxtx2 for
YSL (Figure 5F). This finding aligns with the important roles of these factors in the EVL and YSL and
suggests that Irf6 and Mxtx2 activate the differentiation TFs to initiate EVL and YSL differentiation (Liu
et al., 2016; Sabel et al., 2009; Xu et al., 2012). Further inspection of putative Mxtx2 TFBSs showed
that these motifs are not only found in differentiation TF genes but also in their target genes. Moreover,
Mxtx2 TFBSs frequently co-localize with the TFBSs of the differentiation TFs (gata6, hnf4, and onecut1;
Figure S8). This prediction suggests that Mxtx2 functions in a feedforward loop in which it activates
differentiation TFs and then together with differentiation TFs activates their target genes (Figure 5F).
Notably, the expression of mxtx2 is driven by Nanog (Xu et al., 2012), a maternally deposited TF, as is
Irf6 (Sabel et al., 2009). These observations indicate that the GRN from pluripotency to the terminal
differentiation of EVL and YSL is very shallow (Figure 5F).
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure 5: Gene regulatory logic underlying the instant differentiation of EVL and YSL
A. eRegulon activity dynamics for the EVL and YSL trajectories. The top 10 specific eRegulons in
each cell state of the EVL (from oblong to 6-somite) and YSL (from dome to 6-somite) were
selected, combined, and filtered to create a unique set of eRegulons. For each eRegulon, the
mean activity across all cells within each cell state was then displayed.
B. Overlap of target genes for all eRegulons displayed in panel A, showing that many targets are
shared among a small set of TFs colored in red.
C. TFBS contribution score dynamics for the EVL and YSL trajectories. The top 5 motifs identified
by DeepDanio in each cell state of the EVL (from oblong to 6-somite) and YSL (from dome to 6-
somite) were selected, combined, and filtered to create a unique set of motifs. For each motif,
the mean contribution score across all its TFBSs was then displayed.
D. Gene regulation among differentiation TFs, highlighting numerous auto-regulatory and cross-
regulatory interactions.
E. Motif ablation analysis. Left Panel: Distribution of ablation scores for each motif. Right Panel:
Correlation between distances of two TFBSs and their interaction scores. The data show a trend
where longer distances correspond to lower interaction scores. The Spearman’s correlation
coefficient (Rho) and the P-value are displayed in the top-left corner of the plot.
F. Topology of the gene regulatory network from maternally deposited TFs to differentiation genes.
The initial input is maternally deposited TFs. Irf6 and Mxtx2 were not recovered using SCENIC+
because they are not included in the list of motifs in DANIO-CODE.
A B
atf3
hic1
nfya
ets2
klf13
hey1
TCF4
cebpb
elf1
grhl3
klf17
klf6a
tead3a
tead3b
tfap2a
nr2f6b
prox1a
High
Oblong
Dome
30% Epiboly
50% Epiboly
Shield
75% Epiboly
Bud
6-Somote
eRegulon
activity
0
0.5
1
atf3
e2f7
prox1a
hic1
snai1a
hnf1ba
gata6
hnf4a
onecut1
rreb1a
rxrgb
hnf4b
foxo1a
gata4
bhlhe41
nr2f6b
High
Oblong
Dome
30% Epiboly
50% Epiboly
Shield
75% Epiboly
Bud
6-Somote
eRegulon
activity
0
0.5
1
High
Oblong
Dome
30% Epiboly
50% Epiboly
Shield
75% Epiboly
Bud
6 Somote
Contribution
score
0
0.5
1
sp1
irf6
tbx16
pou5f3
klf17
grhl3
cebpb_variant
cebpb
grhl3_variant
tfap2a
tead3a
high
Oblong
Dome
30% Epiboly
50% Epiboly
Shield
75% Epiboly
Bud
6 Somote
Contribution
score
0
0.5
1
irf2a
mgaa
znf740a
patz1
mxtx2
foxa3
mxtx2_variant
gata6
gata6_variant
hnf4a
onecut1
hnf1ba
eRegulon targets
onecut1
hnf4b
hnf4a
gata6
gata4
hic1
rxrgb
rreb1a
nr2f6b
e2f7
bhlhe41
hnf1ba
snai1a
atf3
foxo1a
prox1a
grhl3
TCF4
tead3a
klf17
cebpb
tfap2a
hey1
elf1
prox1a
klf6a
klf13
hic1
atf3
ets2
nr2f6b
tead3b
C
D
0 1 3 4 5
0.0 0.5 1.0 1.5
EVL (50% Epiboly)
2
Density
grhl3
cebpb
klf17
tfap2a
tead3a
Ablation score
01 345
0.0 0.2 0.4 0.6 0.8 1.0 1.2
YSL (50% Epiboly)
2
Density
gata6
onecut1
hnf4
Ablation score
−1.0
−0.5
0.0
0.5
1.0
0 300
100 200
Predicted interaction score
250
500
750
# pairs
grhl3 and klf17 interaction
Rho= −0.284 / p=8.7e−39
Motif pair distance (bp)
−1.0
−0.5
0.0
0.5
1.0
0 50 100 150 200 250
Motif pair distance (bp)
Predicted interaction score
200
400
600
800
# pairs
gata6 and onecut1 interaction
Rho= −0.238 / p=1.6e−26
EF
YSL
EVL EVL YSL
EVL YSL
EVL YSL
EVL
YSL
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Mis-expression of differentiation transcription factors inactivates hundreds of target genes
The SCENIC+ and DeepDanio analyses predict that a small set of differentiation TFs (Grhl3,
Klf17, Cebpb, Tead3a, Tead3b, and Tfap2a) activates a large number of EVL differentiation genes
(Figure 5F). To test this prediction experimentally, we mis-expressed these TFs and performed scRNA-
seq at 50% epiboly (onset of gastrulation, when the EVL has formed, Table S8). Strikingly, the
differentiation TFs ectopically activated hundreds of EVL differentiation genes. Compared to the
mCherry control (Figure 6A), we observed the emergence of a new gene expression cluster that
expresses EVL genes. For example, cldne is expressed in both the EVL and the “EVL-like” clusters
upon mis-expression of differentiation TF (Figure 6A). We found that almost all SCENIC+ predicted
EVL-specific target genes are highly expressed in the EVL-like cells (Figure 6B). In addition, EVL-like
cells and deep cells (mesendodermal and ectodermal cells) had distinct gene expression profiles,
indicating that EVL-like cells are not in a hybrid state but transcriptionally resemble EVL cells (Figure
6B). Overall, these results demonstrate that the differentiation TFs are sufficient to activate EVL
differentiation genes.
We performed the same experiment for the YSL and mis-expressed Mxtx2 together with the
differentiation TFs (Onecut1, Gata6, Gata4, Hnf4a, and Hnf4b). TF mis-expression ectopically activated
almost all SCENIC+ predicted YSL differentiation genes (Figure 6C). To further test the feedforward
role of Mxtx2 predicted by DeepDanio, we mis-expressed the differentiation TFs without Mxtx2 and
found that the majority of YSL-specific genes were not activated in deep cells, supporting a feedforward
loop involving Mxtx2 (Figure S9).
In summary, the mis-expression of differentiation TFs identified by SCENIC+ and DeepDanio
ectopically activates hundreds of EVL and YSL differentiation genes. The transformed cells do not
express genes specific to other cell types and express the large majority of EVL- or YSL-specific genes.
These results indicate that a shallow network of differentiation TFs underlies the process of instant
differentiation.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure 6: Mis-Expression of differentiation transcription factors ectopically activates EVL
and YSL differentiation genes
A. UMAP projection of single-cell transcriptomes, with cells colored by cell type (upper panel) and
by marker gene expression (lower panel). cldne and cobll1b are used as markers for EVL and
YSL, respectively. In addition to the native EVL, we identified an “EVL-like” cluster expressing
EVL markers. It should be noted that since we performed single-cell RNA-seq rather than single-
nucleus RNA-seq, we couldn't capture YSL cells as they are multinucleated and exist in a
syncytium However, a “YSL-like” cluster was identified, expressing YSL marker genes.
B. Heatmap showing gene expression of cell type-specific genes across different cell types under
two conditions: EVL mix (mis-expression of all differentiation TFs) and EVL mCherry (mis-
expression of mCherry). Notably, most EVL-specific genes are activated in the “EVL-like” cluster,
although their expression levels are lower than in native EVL. The 77 “EVL-like” specific genes
refer to those highly expressed in the “EVL-like” cluster (compared to native EVL from the
mCherry control, fold change (log2) > 1, p-value < 0.01). These genes are typically weakly
expressed housekeeping genes during normal development and are primarily involved in RNA
biosynthetic processes.
C. Heatmap showing gene expression of cell type-specific genes across different cell types under
the YSL mix condition (mis-expression of all differentiation TFs plus Mxtx2). Most YSL-specific
genes are activated in the “YSL-like” cluster.
mCherry EVL mix
Apoptosis-like
Mesoderm
EVL
EVL-like Ectodetm
EVL
Mesoderm
Ectodetm
Apoptosis-like
UMAP1
UMAP2
cldne cldne
YSL mix
Ectodetm YSL-like
EVL
Mesoderm
cobll1b
012345
AC
2 1 0 -1 -2
YSL-like Mesoderm Ectoderm
EVL
Ectoderm specific genes
YSL specific genes
Mesoderm specific genes
EVL specific genes
EVL-like specific genes
Ectoderm specific genes
Mesoderm specific genes
Apoptosis-like specific genes
EVL specific genes
EVL-like
(mix)
EVL
(mix)
EVL
(mCherry)
Mesoderm
(mCherry)
Mesoderm
(mix)
Ectoderm
(mCherry)
Ectoderm
(mix)
Apoptosis-like
(mCherry)
Apoptosis-like
(mix)
B
2 1 0 -1 -2
log2(exp)
scaled(exp)
scaled(exp)
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Discussion
In this study, we generated and explored a high-quality single-cell multi-omics atlas for early
zebrafish embryogenesis (Figure 1 and Table S1). The atlas provides a rich resource for future studies
of zebrafish development, including integration with spatial transcriptomics (Wan et al., 2024) and in-
toto live imaging (McDole et al., 2018). The reconstructed GRN and eRegulons will help dissect the
regulatory logic of cell type specification and differentiation. In addition, the DeepDanio deep learning
model developed here can be applied to dissect cell-type- or stage-specific enhancer architecture and
design synthetic enhancers with specific temporal and spatial activities (de Almeida et al., 2024;
Taskiran et al., 2024).
Our study also presents a generalizable framework to dissect the regulatory logic of embryonic
specification and differentiation through integrating single-cell multi-omics, deep learning, GRN
reconstruction and in vivo assays. We provide three main insights:
First, we defined the emergence of accessible chromatin regions and their relationships with
gene expression. We identified three distinct temporal waves of dCRE expansion: EVL and YSL >
mesendoderm > neuroectoderm (Figure 1D). This pattern closely follows the emergence of these cell
types but contrasts with the early accessibility of promoters (Figure 1E), which mark genes for later
expression (Pálfy et al., 2020; Reddington et al., 2020). Comparing the temporal emergence of dCRE
with the timing of gene activation revealed that dCRE priming (chromatin is accessible before gene
transcription) is common: mesendodermal dCREs are primed in pluripotent cells at the high stage, while
neuroectodermal dCREs are primed at the 50% epiboly stage (onset of gastrulation) (Figure 2B and
2C). This early priming of mesendodermal dCREs might facilitate mesendoderm specification during
gastrulation by enabling a rapid response to inductive signals such as Nodal (Schier & Talbot, 2005).
Notably, we discovered that hundreds of mesendodermal dCREs were primed by Nanog (Figure 2E
and 2F). This pioneer TF is known to activate several hundred genes to promote pluripotency (Boyer et
al., 2005; M. T. Lee et al., 2013). Our results reveal that Nanog has a dual role: it promotes pluripotency
while also prepares the embryo for subsequent mesoderm induction. These results extend and
generalize studies of individual genes found to be primed in flies (Falo-Sanjuan et al., 2019), frogs
(Charney et al., 2017), and human and mouse embryonic stem cells (Kim et al., 2018; Liber et al., 2010;
A. Wang et al., 2015). Our results contradict an earlier report that suggested that mesendodermal
dCREs are not primed in the mouse epiblast (Argelaguet et al., 2019), but align with a recent preprint
that does suggest priming (Sendra et al., 2024).
Second, the predictions from DeepDanio suggest that a small number of TFs play a critical role
in shaping the accessible chromatin landscape (Figure 3). This observation might seem surprising
considering the broad array of TFs typically expressed in cells, but it complements a recent study (Wei
et al., 2018) that developed a massively parallel protein activity assay to measure the DNA-binding
activity of all TFs in cell or tissue extracts through electrophoretic mobility shift assays (EMSAs). Only
a small set of TFs displayed strong DNA-binding activity, and the sequence features underlying the
DNA-binding sites of these TFs can accurately predict the cell type-specific accessible chromatin
landscape. Collectively, these results support the idea that a limited number of TFs establish the overall
chromatin landscape and thus facilitate the binding of TFs with weaker chromatin-modulating
capabilities.
Third, we discovered a shallow gene regulatory network underlying instant differentiation: an
upstream TF activates a set of differentiation TFs, which in turn activate hundreds of effector genes that
define cell-type-specific structures and functions. In the case of the EVL, the maternally deposited TF
Irf6 serves as the upstream TF that directly activate multiple differentiation TFs to initiate differentiation
(Figure 5F). In the case of the YSL, Nanog is the upstream TF that activates Mxtx2, which then activates
differentiation TFs. Together, Mxtx2 and these differentiation TFs form a feedforward loop to initiate
differentiation. Since the expression of mxtx2 is transient (Figure S10), its feedforward role may involve
cooperation with the differentiation TFs to open the chromatin of enhancers in the YSL-specific genes
(Figure S11). The differentiation TFs then activate and maintain the expression of differentiation genes.
This shallow network model suggests that instant differentiation occurs because differentiation
TFs are quickly activated by maternally deposited TFs, thereby skipping multiple cell fate transitions
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
typically found during the stepwise differentiation of cell (Davidson, 2010; Liberali & Schier, 2024;
Murtaugh, 2007; Y. Wang et al., 2023). Indeed, TF mis-expression analysis showed that the identified
EVL and YSL differentiation TFs are sufficient to ectopically activate nearly all effector genes at 50%
epiboly (onset of gastrulation, when the EVL and YSL have formed). These findings support the
proposed shallow network model and extend previous studies (De La Garza et al., 2013; Liu et al., 2016;
Sabel et al., 2009; Xu et al., 2012) that examined the effects of some differentiation TFs on a few marker
genes.
While direct activation of effector genes by maternal TFs could theoretically speed up
differentiation, the EVL/YSL regulatory design offers advantages. Activating multiple differentiation TFs
enhances the specificity of effector gene expression through cooperative regulation, as indicated by in-
silico motif ablation analysis (Figure 5E). Additionally, we observed auto-regulatory and cross-
regulatory interactions among these TFs, similar to those in embryonic stem cell networks that help
maintain stemness (Boyer et al., 2005) (Figure 5D). Given the transient expression of maternal factors,
these interactions might sustain and amplify effector gene expression, ensuring sufficient production of
the gene products necessary for differentiation.
The observation that cell types like the EVL and YSL can undergo instant differentiation
suggests that stepwise differentiation programs might not be necessary for differentiation TFs to activate
their effector genes. Future studies will address the necessity of stepwise differentiation: Can stepwise
differentiation be transformed into instant differentiation?
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Methods
Culturing and collecting embryos
Embryos from wild-type (Tupfel longfin/AB) crosses were collected 20 minutes after fertilization. They
were dechorionated by incubation in 1 mg/ml pronase (Sigma-Aldrich) for 5 minutes until chorions
began to blister, submerged in ~200ml of zebrafish “blue water” embryo medium (5 mM NaCl, 0.17 mM
KCl, 0.33 mM CaCl2, 0.33 mM MgSO4, 0.1% methylene blue, dissolved in fish system water) in a glass
beaker, and then the blue water was decanted and vigorously replaced 3 times. It was critical to prevent
embryos from contacting air or plastic during this process or at any point thereafter. Embryos were then
cultured at 28°C in blue water in plastic Petri dishes that had previously been coated with 2% agarose
(dissolved in blue water).
Cell dissociation and nuclei isolation
Single cell suspensions were obtained using an embryo fractionation protocol. Approximately 200
dechorionated embryos for each experimental condition were transferred into a 2 ml LoBind Eppendorf
tube (EP-022431048) containing 1.5 ml of chilled deyolk buffer (55 mM NaCl, 1.8 mM KCl, 1.25 mM
NaHCO3). The embryos were gently pipetted up and down 3-4 times using a p1000 tip and then placed
in a thermal shaker at 1200 rpm at 4°C for 20 seconds. The tubes were centrifuged at 4°C at 250g for
4 minutes. After centrifugation, the supernatant was discarded, and 1 ml of deyolking wash buffer (10
mM Tris pH 8.5, 110 mM NaCl, 3.5 mM KCl, 2.7 mM CaCl2) was added to each tube to resuspend the
pelleted embryos. Resuspension was performed by pipetting six times with a p1000 tip. The tubes were
then centrifuged again, the supernatant discarded, and the cell pellets resuspended in 1 ml DMEM
(Gibco, 11594426) with 0.1% BSA. This centrifugation and resuspension step with DMEM (0.1% BSA)
was repeated once more. After another round of centrifugation, 150 μl of chilled 0.5X Lysis Buffer (10
mM Tris pH 7.5, 10 mM NaCl, 3 mM MgCl2, 1% BSA, 0.1% Tween, 1 mM DTT, 1 U/μl RNaseIn
(Promega), 0.1% NP40, 0.01% Digitonin) was added to the pellet. The mixture was pipetted 5 times
with a p200 tip and incubated on ice for 5 minutes. Next, 1.5 ml of chilled Wash Buffer (identical to NE
buffer but without NP40 and digitonin) was added to the lysed cells. The cells were pipetted five times
and then centrifuged at 450g for 5 minutes at 4°C. The supernatant was carefully removed without
disturbing the nuclei pellet. The pellet was resuspended in at least 300 μl of chilled Diluted Nuclei Buffer
(10x Genomics). The suspension was then passed through a Flowmi Cell Strainer (40 μm). Finally, the
nuclei concentration and morphology were checked using a hemocytometer.
10x Multiome library preparation and sequencing
Nuclei were diluted to ensure a maximum of 8,000 were used for the 10x Multiome library preparation.
For the high stage, oblong, dome, and 30% epiboly stages, we have one technical replicate each. For
the 50% epiboly and 6-somite stages, we have two technical replicates each. For the shield, 75%
epiboly, and bud stages, we have two biological replicates, and for each biological replicate, we have
two technical replicates. Libraries were prepared from the single nuclei suspensions using the 10x
Chromium Next GEM Single Cell Multiome ATAC + Gene Expression kit, following the standard 10x
protocol. The libraries were sequenced on a NovaSeq platform (Illumina) using the recommended read
lengths. This sequencing yielded an average of 348 million RNA-seq reads and 536 million ATAC reads
per sample. We recovered an average of 2,500 cells per sample prior to quality control.
scRNA data and scATAC data processing
Raw sequencing files were processed with CellRanger arc 2.0.0 using default parameters. Reads were
mapped to the Ensembl GRCz11 reference genome (Ensembl Release 100)(Yates et al., 2020). Low-
quality cells were filtered out based on several quality control metrics: the number of expressed genes,
the proportion of mitochondrial reads for RNA; the number of fragments, and the TSS enrichment score
for ATAC. For RNA: For the high and oblong stages, since they mark the onset of transcription, the
number of expressed genes is lower than in later stages. Therefore, we filtered out cells with fewer than
200 (low quality cells) or more than 3000 (to exclude potential doublets) expressed genes. For later
stages, we filtered out cells with fewer than 500 or more than 5000 expressed genes. We also excluded
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
cells with more than 10% of reads mapping to the mitochondrial genome. For ATAC: We filtered out
cells with a TSS enrichment score lower than 4. We also excluded cells with fewer than 1000 unique
nuclear fragments.
Cell clustering and annotation
The scRNA-seq part of the scMultiome data was processed and analyzed using Seurat version
4.1.0(Hao et al., 2021). We processed each developmental stage separately and clustered the cell
types accordingly. Marker genes for each cell cluster were identified using the Wilcoxon test, comparing
the expression levels within the cluster to those in the rest of the cells. To annotate the cell types
represented by each cluster, we checked the expression profiles of the significant marker genes with
high fold changes against the ZFIN database and our previously constructed developmental trajectories.
Peak calling
The scATAC-seq part of the scMultiome data was processed and analyzed using ArchR version
1.0.1(Granja et al., 2021). Using the cell types identified from the scRNA-seq data as groups, pseudo-
bulk replicates were generated for each cell type. Peak calling was then performed using MACS2
version 2.2.7.1 through ArchR with the following parameters: “--shift -75 --extsize 150 --nomodel --call-
summits --nolambda --keep-dup all -q 0.01”. We did not call peaks for cell types with fewer than 80 cells.
These excluded cell types were: YSL(dome), YSL(30% epiboly), YSL(6-somite), forerunner cells(75%
epiboly), ionocyte progenitors(6-somite), endothelial progenitors(6-somite), neural floorplate(6-somite).
Calling peaks in a cell type-aware manner resulted in multiple peak sets that needed to be consolidated
into a single peak annotation. Using ArchR’s iterative overlap merging procedure, we obtained a
consensus peak set consisting of 444,653 peaks, each 500 bp wide. Distal CREs (dCREs), including
putative enhancers, were defined as peaks >500 bp from an annotated transcription start site (TSS).
Peaks within <500 bp of TSS were annotated as promoters.
10x standalone scATAC-seq library preparation and sequencing
To compare scATAC-seq data from scMultiome with standalone scATAC-seq, we also generated
standalone scATAC-seq data for two stages: 50% epiboly and the 6-somite stage. For the 50% epiboly
stage, we have two technical replicates, and for the 6-somite stage, we have three technical replicates.
Libraries were prepared from the single nuclei suspensions using the Chromium Next GEM Single Cell
ATAC Reagent Kits v1.1, following the standard 10x protocol. The libraries were sequenced on a
NovaSeq platform (Illumina) using the recommended read lengths. This sequencing yielded an average
of 368 million ATAC reads per sample. We recovered an average of 4’010 cells per sample prior to
quality control.
Cellular trajectories reconstruction
To connect each cell cluster observed at a given stage with its pseudoancestor in the preceding stage,
we applied a k-NN (k-nearest neighbors) method as described in earlier studies. First, we merged all
cells from the given stage and the preceding stage using Seurat (version 4.1.0) and projected them into
a common UMAP embedding space. We then calculated the Euclidean distances between individual
cells from the given stage and the preceding stage within this UMAP space. For each cell cluster at the
given stage, we identified their five nearest neighbors from the preceding stage and calculated the
proportion of these neighbors derived from each cell cluster in the preceding stage. This process was
repeated 500 times with 80% subsampling from the same embedding. The median proportions of
neighbors were then used as weights for edges between a cell cluster and its potential antecedents.
Only edge weights greater than 0.2 were retained (Table S2) for constructing the resulting acyclic
directed graph, which is shown in Figure 1B.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
CRE usage during embryogenesis
We applied the AccessiblePeaks function from the Signac package to identify accessible CREs in
different cell types. A CRE is considered "accessible" in a cell type if it is accessible in at least 5% of
the cells of that cell type(Kelley et al., 2016; Sarropoulos et al., 2021; Stuart et al., 2021). For each
trajectory analysis (Figure 1C and D), at each stage, we examined how many CREs became accessible
and whether their accessibility was maintained up to the 6-somite stage.
The timing differences between CRE chromatin opening and target gene activation
1. Identify cell type specific genes
For cell types at the 50% epiboly stage (EVL, YSL, dorsal anterior, dorsal posterior, ventrolateral
mesendoderm, and margin tail), we identified the genes specifically expressed in each cell type using
the Seurat FindMarkers function (fold change (log2) > 1, p-value < 0.01). We then filtered out genes
that were already expressed at the oblong stage. A gene is considered expressed in a cell type if it is
expressed in at least 10% of the cells in that cell type(Bravo González-Blas et al., 2023; Karlsson et al.,
2021; Kotliar et al., 2019; Meng et al., 2024). Similarly, for cell types at the bud stage (forebrain, midbrain,
hindbrain, epidermis, hatching gland, notochord, tailbud, lateral plate mesoderm, presomitic mesoderm,
adaxial cells, and endoderm), we identified the genes specifically expressed in each cell type using the
Seurat FindMarkers function (fold change (log2) > 1, p-value < 0.01). We then filtered out genes that
were already expressed at the shield stage.
2. Infer the putative distal CREs (dCREs) of cell type specific genes
For each cell type at the 50% epiboly stage, we extracted the cells along the trajectory from the high
stage to the 50% epiboly stage. For each cell type at the bud stage, we extracted the cells along the
trajectory from the 50% epiboly stage to the bud stage. We then used the LinkPeaks function from
Signac to infer the putative dCREs of each gene. The LinkPeaks function computes the Pearson
correlation coefficient (r) between gene expression and chromatin accessibility for each dCREs located
within 50 kb upstream or downstream of the transcription start site (TSS). This analysis includes all cells
extracted from the trajectory. For each dCREs, LinkPeaks also computes a background set of expected
correlation coefficients by randomly sampling 200 CREs located on a different chromosome from the
gene, matched for GC content, accessibility, and sequence length to estimate a p-value. An associated
dCRE of a gene is defined as having a Pearson correlation coefficient > 0.1 and a p-value < 0.05.
3. Calculate the proportion of accessible CREs in each stage
For the cell type-specific genes in each cell type, we calculated the proportion of their promoters or
dCREs that are accessible at each stage of the trajectory as shown in Figure 1E.
Enhancer priming analysis
1. Infer the associated distal CREs (dCREs) for mesendodermal and neuroectodermal genes
For mesendodermal cell types at the 50% epiboly stage (dorsal anterior, dorsal posterior, ventrolateral
mesendoderm, and margin tail), we identified genes specifically expressed in each cell type using the
Seurat FindMarkers function (fold change (log2) > 1, p-value < 0.01). We then filtered out genes already
expressed at the oblong stage. A gene is considered expressed in a cell type if it is expressed in at
least 10% of the cells in that cell type(Bravo González-Blas et al., 2023; Karlsson et al., 2021; Kotliar
et al., 2019; Meng et al., 2024). Similarly, for neuroectodermal cells at the bud stage (forebrain, midbrain,
hindbrain), we identified genes specifically expressed in each cell type using the Seurat FindMarkers
function (fold change (log2) > 1, p-value < 0.01). Genes already expressed at the shield stage were
filtered out. For each gene, we identified dCREs located within 50 kb upstream or downstream of the
transcription start site (TSS). A dCRE was considered associated with a gene if the Pearson correlation
coefficient (association score) between gene expression and dCRE chromatin accessibility was > 0.1,
with a p-value < 0.05. For more details, refer to “Infer the putative distal CREs (dCREs) of cell type
specific genes”.
2. Compare the proportion of primed dCREs between associated and non-associated dCREs,
as shown in Figure 2B and C
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
A primed dCRE at a given stage is defined as a dCRE that is accessible at that stage, but its target
gene is not yet expressed. We applied the AccessiblePeaks function from the Signac package to identify
accessible CREs in a given cell type. Check “CRE usage during embryogenesis” for details. A gene is
considered expressed in a cell type if it is expressed in at least 10% of the cells in that cell type(Bravo
González-Blas et al., 2023; Karlsson et al., 2021; Kotliar et al., 2019; Meng et al., 2024).
3. Chromatin accessibility analysis
The chromatin accessibility data (bedGraph file) comes from Miao et al., (2022). They performed bulk
ATAC-seq at the sphere stage (4 hpf) in nanog, pou5f3, and sox19b triple mutants (MZnps) condition,
wild type condition, as well as in various rescue conditions by injecting mRNAs into MZnps embryos.
For each primed and non-primed dCREs, we calculated the chromatin accessibility in different
conditions using the bedtools “map” command(Quinlan & Hall, 2010), as shown in Figure 2E.
4. Nanog binding site analysis
The Nanog binding peaks, identified using CUT&RUN, come from Wang et al., (2022). For each dCRE,
if it overlapped with a Nanog binding peak by at least 1 bp, we defined this dCRE as binding with Nanog.
Then, we calculated the proportion of primed and non-primed dCREs with Nanog binding, as shown in
Figure 2F.
Deep learning analysis
1. Data preprocessing
We started from a 444,653 x 95 matrix containing aggregated Counts Per Million (CPM) for each
scATAC peak and cell state. Log10(CPM) values were then calculated with a per cell-state pseudocount
obtained from the corresponding minimum non-zero CPM value. Finally, quantile normalization was
applied across cell states. 500nt-long peak sequences were extracted from the GRCz11 version of the
Danio Rerio genome. We additionally extracted 200,000 500nt-long fragments from random non-peak
genome locations (negative samples).
2. Training splits
We distributed all 25 Danio Rerio chromosomes into 10 subsets with a roughly similar number of
scATAC peaks (mean: 44.5 thousand peaks per subset) using the prtpy python package
(https://github.com/coin-or/prtpy). For each individual model in the DeepDanio ensemble, eight subsets
were used directly for model fitting (training), one was used for early stopping (validation), and one was
held out from training for performance evaluation (test). Each model was assigned a different test set.
The list of chromosomes used in each model can be found in Table S7.
3. Training
DeepDanio is an ensemble of three residual neural networks as shown in Figure 3A. The input is a one
hot-encoded 500bp sequence. Outputs are predicted log10CPM for each of the 95 cell states. Model
training was performed with python 3.10 and tensorflow 2.10 using AWS EC2 g5.2xlarge instances.
The loss was the sum of a standard MSE component and a per-sequence Pearson r component that
incentivizes learning the relative order of predicted log10 CPMs across cell states. More specifically,
we adapted the AI-TAC Pearson r loss implementation(Maslova et al., 2020)where we maximize the
cosine similarity between the per-sequence mean-normalized prediction and the mean-normalized
measurement. Before training, positive (i.e. the preprocessed scATAC matrix) and negative data were
split into training/validation/test sets according to chromosomes as described above. Training data was
further augmented with the reverse complement of every sequence. At the beginning of each epoch,
we randomly selected enough negative samples to reach a 5:1 ratio of positive-to-negative training
peaks. Negative peaks were assigned the lowest log10CPM value on each cell state. We used the
Adam optimizer with default settings, a learning rate of 2e-4, and early stopping on the validation loss
with a patience of two. We included negative samples for validation loss evaluation but no reverse
complements. Once training finished, we performed another training round starting from the optimized
weights, which resulted in a slight additional performance improvement. Three models were
independently trained on different chromosome splits as described above. For performance evaluation
(i.e. Figure 3B and C and Figure S4A) we aggregate predictions of each model over its own test set.
For all other analyses in this manuscript, we used the mean of all three models.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
4. De novo motif discovery
We extracted motifs that contribute to accessibility in each cell state using the contribution scores of the
10,000 most cell state-specific peaks. We first filtered out peaks where the square prediction error
averaged across cell states was 0.125 or greater, retaining 377,312 peaks (84.86%). Then, for each
cell state, we obtained the most specific peaks by sorting by their predicted specificity score, defined
as the predicted log10CPM in the target cell state minus the average predictions across all other states,
and retained the top 10,000. Next, we calculated nucleotide contribution scores of each set of cell state-
specific peaks with respect to the model output corresponding to that cell state. We used a custom
version of DeepExplainer/DeepSHAP
(https://github.com/castillohair/shap/tree/castillohair/genomics_mod) modified to work with tensorflow 2
and generate hypothetical contribution scores needed for TFModisco. As background, we used 10
random dinucleotide shufflings per sequence. Next, we ran TFModisco (Shrikumar et al., 2018),
specifically the modisco-lite implementation (https://github.com/jmschrei/tfmodisco-lite), on the
contributions of each set of cell state-specific peaks, with 25,000 as the maximum number of seqlets
per metacluster. This resulted in an average of 17 motifs per cell state (Table S3), each with a
corresponding PWM and contribution weight matrix (CWM). Finally, to generate the plots in Figures 3D-
F, we realigned motifs to their corresponding cell state-specific peaks using the motif scan method in
Avsec et al (Avsec et al., 2021)which considers nucleotide contributions in addition to motif PWM
similarity.
5. Validation of deep learning-identified transcription factor binding sites (TFBS) using ChIP-
Seq data
The ChIP-seq data for nanog performed at the high stage were obtained from Xu et al.(Xu et al., 2012).
In this study, C-terminally myc-tagged zebrafish nanog was overexpressed, and myc antibodies were
used for immunoprecipitation; The ChIP-seq data for pou5f3 performed at the sphere stage were
obtained from Miao et al. ((Miao et al., 2022). Similarly, C-terminally myc-tagged zebrafish pou5f3 was
overexpressed, and myc antibodies were used for immunoprecipitation. The ChIP-seq data for mxtx2
performed at the dome stage were obtained from Xu et al. (Xu et al., 2012). Here again, C-terminally
myc-tagged zebrafish mxtx2 was overexpressed, and myc antibodies were used for
immunoprecipitation. The ChIP-seq data for tbxta performed at the 75% epiboly stage were obtained
from Nelson et al. (Nelson et al., 2017). In this case, anti-tbxta antibodies were used for
immunoprecipitation. The ChIP-seq data for cdx4 performed at the bud stage were obtained from Paik
et al. (Paik et al., 2013). C-terminally myc-tagged zebrafish cdx4 was overexpressed, and myc
antibodies were used for immunoprecipitation.
For nanog and pou5f3, the deep learning-identified TFBS were derived from the high stage. For mxtx2,
the TFBS were identified in 50% epiboly YSL cells. For tbxta, the TFBS were found in 50% ventrolateral
mesoendoderm cells. For cdx4, the TFBS were identified in the lateral plate mesoderm, presomitic
mesoderm, and tailbud cells at the bud stage, as cdx4 is expressed in all three cell types. For each
TFBS identified through deep learning, the region was extended by 2 kb on either side. These extended
regions were then divided into 20 bp bins, and the ChIP-seq signal (bedGraph file) for each bin was
calculated using the bedtools "map" command(Quinlan & Hall, 2010), as shown in Figure 3D and Figure
S4B.
6. Motif in-silico ablation analysis
For each TF involved in EVL differentiation (grhl3, cepbp, klf17, tfap2a, and tead3a) and YSL
differentiation (mxtx2, gata6, and hnf4), we first performed single TFBS ablation analysis, as shown in
Figure 5E left panel and Figure S11A. If the TFBSs of one TF colocalized with the TFBSs of other TFs
within the same dCREs, we also conducted double ablation analysis for each pair of co-binding TFBSs.
For ablation analysis, the identified TFBSs were replaced with dinucleotide shuffles to preserve
sequence composition while disrupting the motif. Chromatin accessibility of both the original and ablated
sequences was calculated using DeepDanio. Each ablation was repeated 10 times, and the mean
chromatin accessibility value was computed for the ablated sequences. The ablation score was
calculated using the formula:
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
AS(Ablation+Score) = Accessibility+of+original+sequence −Mean+accessibility+of+ablated+sequences
Accessibility+of+original+sequence
To study pairwise motif interactions, as shown in Figure 5E (right panel), Figure S7, and Figures S11B,
C, D, we calculated ablation scores for single motif ablations (AS(A) for motif A and AS(B) for motif B)
and for double ablations (AS(AB)). We then compared the sum of the individual ablation scores with
the double ablation score to determine the type of interaction:
Additive Effect: AS(A)+ AS (B) = AS (AB)
Redundant Effect: AS (A)+ AS (B) < AS (AB)
Cooperative Effect: AS (A)+ AS (B) > AS (AB)
GRN analysis:
We followed the tutorial from SCENIC+: https://scenicplus.readthedocs.io/en/latest/.
SCENIC+ is a three-step workflow consisting of: Identifying candidate enhancers; Identifying enriched
TF binding motifs on candidate enhancers; Linking TFs to candidate enhancers and target genes, as
shown in Figure 4A.
1. Identifying candidate enhancers
SCENIC+ uses both differentially accessible regions (DARs) and topics (sets of co-accessible regions)
across cell types or states as enhancer candidates. In this work, for each of 95 cell states, we identified
DARs using ArchR (FDR<0.01, Log2FC>0.5), which resulted in a total of 122,554 DARs. In addition,
the topic modeling was performed with pycisTopic using Latent Dirichlet Allocation (LDA) with the
collapsed Gibbs sampler to iteratively optimize the probability of a region belonging to a topic. A model
of 200 topics was selected based on the stabilization of metrics described in the references and log-
likelihood. Region–topic probabilities were binarized using either the Otsu method or by selecting the
top-3,000 regions per topic. These DARs and region–topic associations served as the starting point for
further analyses to identify enhancer–gene links and eGRNs.
2. identifying enriched TF-binding motifs on candidate enhancers (cistrome)
To discover potential transcription factor binding sites (TFBSs) in candidate enhancers, we conducted
motif enrichment analysis. We started by collecting 590 motifs from DANIO-CODE(Baranasic et al.,
2022), corresponding to 912 transcription factors (TFs), since some TFs share the same motif. We used
motifMatch(Schep et al., 2017) to create a motif match score for each motif in each candidate enhancer
(region). We then performed motif enrichment analysis using pycisTarget. Motif enrichment was
conducted with both the cisTarget and differential enrichment of motifs (DEM) algorithms on cell-type-
based DARs, the top 3,000 regions per topic, and topics binarized using the Otsu method. cisTarget is
a ranking-and-recovery-based algorithm, where enrichment is calculated as a normalized area under
the curve (AUC) at the top 0.5% ranking, resulting in a normalized enrichment score (NES). Motifs with
an NES greater than 3.0 were retained. To identify the target regions for each motif (motif-based
cistrome), regions at the top of the ranking (leading edge) were retained, defined by an automated
thresholding method that keeps regions below the rank at maximum enrichment. For the DEM algorithm,
a Wilcoxon rank-sum test was performed between a foreground and a background set of regions using
score distributions for each motif or cluster of motifs. Motifs with an adjusted p-value less than 0.05
(Bonferroni) and log fold change greater than 0.5 were kept. Regions containing the motif (motif-based
cistrome) were obtained by selecting regions with a cis-regulatory module score greater than 3 for each
enriched motif.
3. linking TFs to candidate enhancers and target genes (eRegulon)
3.1 Calculating region-to-gene and TF-to-gene importance scores
We used the default parameters from SCENIC+ to quantify region-to-gene importance scores. In
SCENIC+, the importance score for each region was calculated using gradient-boosting machine
regression to predict target gene expression based on region accessibility. All regions within a gene’s
search space, defined as a minimum of 1 kb and a maximum of 150 kb upstream or downstream of the
gene's start or end, or the promoter of the nearest upstream or downstream gene, were considered.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
The promoter of a gene was defined as the transcription start site ±10 bp. Since the importance score
does not indicate directionality, we calculated the Spearman rank correlation between accessibility and
expression, using the correlation coefficient to distinguish positive interactions (>0.03) from negative
interactions (<−0.03). Similarly, for each TF and gene pair, we calculated the importance score of TF
expression to predict target gene expression using gradient-boosting machine regression, and used
Pearson correlation to separate positive (>0.03) from negative (<−0.03) interactions.
3.2 Binarizing region-to-gene importance scores
We used the default parameters from SCENIC+ to binarize region-to-gene importance scores. In
SCENIC+, region-to-gene importance scores were binarized using multiple methods: taking the 85th,
90th, and 95th quantiles of the importance scores, selecting the top 5, 10, and 15 regions per gene
based on these scores, and applying a custom implementation of the BASC88 method on the region-
to-gene importance scores.
3.3 eRegulon creation
We used the default parameters from SCENIC+ to create eRegulons based on the cistrome, the region-
to-gene, and the TF-to-gene relationships. For each TF, TF–region–gene triplets were identified by first
selecting all regions enriched for motifs associated with the TF, based on cistrome data. Next, we used
binarized region-to-gene links to assign genes to these regions. To eliminate false positives, we
conducted gene set enrichment analysis (GSEA), ranking genes by their TF-to-gene importance score
and calculating the enrichment of the gene set within the TF–region–gene triplet using the
gsea_compute function from GSEApy. Genes at the top of the ranking (the leading edge) were retained
as target genes of the eRegulon. eRegulons with fewer than ten predicted target genes were discarded.
4. eRegulon activity
All consensus peaks and genes were ranked based on their chromatin accessibility and raw gene
expression counts per cell, respectively. Enrichment for eRegulon target regions and target genes was
defined as the Area Under the Curve (AUC) at the top 5% of the ranking. This enrichment score is
defined as eRegulon activity, as shown in Figure 4C, D and Table S5. There are two types of eRegulon
activity: one based on target regions and the other based on target genes.
5. eRegulon filter
eRegulons were filtered based on the correlation coefficient between the AUC scores (eRegulon activity)
of the target regions and the target genes. eRegulons with a correlation coefficient greater than 0.4
were considered high quality. For downstream analysis, we further refined the selection to include only
activator eRegulons, defined as those with both region-to-gene and TF-to-gene correlations greater
than 0. This filtering process resulted in 100 eRegulons (as shown in Figure 4B), with a median of 70
genes and 93 regions per eRegulon.
6. eRegulon specificity score
For each eRegulon, we calculated an eRegulon specificity score (RSS, as shown in Figure 4C and
Table S6) in each of the 95 cell states using Jensen-Shannon divergence, which measures the similarity
between two probability distributions. This calculation utilized the target genes based on eRegulon
activity as input. We determined the Jensen-Shannon divergence by comparing each vector of binary
eRegulon activity overlaps with the assignment of cells to specific cell types.
Mis-expression
1. Cloning and vector construction
For EVL TFs, open reading frames (ORFs) were PCR-amplified with Platinum SuperFi II DNA
Polymerase (Invitrogen) using primers with overhanging restriction sites (BstI for forward primers, SpeI
or AgeI for reverse primers). PCR products were purified using MinElute PCR Purification Kit (QIAGEN)
and digested using BstI and SpeI or BstI and AgeI (New England Biolabs) at 37°C for 1 hour. The
digested amplicons were then purified using MinElute PCR Purification Kit (QIAGEN). An in vitro
transcription vector containing a beta-globin 5′ UTR sequence with an optimized translation initiation
site was used for optimal mRNA stability (https://doi.org/10.1101/2023.11.23.568470). This vector was
digested using BstI and AgeI or BstI and SpeI (New England Biolabs, NEB) at 37°C for 1 hour,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
dephosphorylated using Antarctic Phosphatase (New England Biolabs #M0289) at 37°C for 30 minutes,
followed by purification using MinElute PCR Purification Kit (QIAGEN). The ORF amplicon was
directionally ligated to the vector using T4 DNA ligase (New England Biolabs #M0202S) and
transformed into One Shot™ TOP10 Chemically Competent E. coli cells (Invitrogen) using the heat
shock method. Plasmid DNA was isolated from these cultures using QIAprep Spin Miniprep Kit
(QIAGEN). The resulting constructs were validated by Sanger sequencing. For YSL TFs, synthetic gene
fragments with flanking BstI and SpeI/AgeI restriction sites were ordered from Twist Bioscience for
cloning.
2. mRNA synthesis and purification
Validated constructs were used as templates for PCR amplification with Platinum SuperFi II DNA
Polymerase (Invitrogen) using an SP6 forward primer (5′ CACGCATCTGGAATAAGGAAGTGC 3′) and
a 3′ UTR-specific reverse primer with a 36 nt-long poly(T) overhang (5′
TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTCCTGTGAGTCCCATGGGTTTAAG 3′) as
described in (https://doi.org/10.1101/2023.11.23.568470). PCR products were purified and transcribed
using mMESSAGE mMACHINE™ SP6 Transcription Kit (Invitrogen). The resulting mRNA was purified
using RNA Clean & Concentrator™-5 (Zymo Research).
3. mRNA misexpression and embryo dissociation
Embryos were obtained from wild-type (Tupfel longfin/AB) and microinjected with 15 pg of each TF
mRNA (total of ~100 pg) at the one-cell stage. This was determined to be the optimal dose with minimal
embryo lethality at 50% epiboly. Injected embryos were incubated at 28.5°C in embryo medium. At
approximately 30% epiboly, ~50 embryos per experimental condition were dechorionated using 1 mg/ml
Pronase (Protease type XIV from Streptomyces griseus, Millipore Sigma). At 50% epiboly,
dechorionated embryos were added to a protease solution (10 mg/ml BI protease [Sigma, P5380], 125
U/ml DNaseI, 2.5 mM EDTA in DPBS) in 2 ml LoBind Eppendorf tube (EP-022431048) and incubated
on ice for 4 minutes followed by the addition of a stop solution stop solution (30% FBS, 0.875 mM CaCl2
in DPBS). The supernatant was removed and replaced with chilled Ringer’s solution (140 mM NaCl, 2
mM KCl, 1.5 mM K2HPO4, 1 mM MgSO4, 2 mM MgCl2, 10 mM HEPES, 10 mM D+ glucose) followed
by gentle pipetting (5-10 times) to dissociate the embryos. The cells were pelleted at 4°C with 250g for
4 minutes to discard the supernatant. This dissociation process was repeated two more times with
chilled Ringer’s solution followed by two more times with chilled DMEM (Gibco, 11594426) with 0.1%
BSA followed by resuspension in a final volume of 500 μl. The cells were filtered through a 70 μm cell
strainer (Flowmi Cell Strainer, BAH136800070) into 15 mL polypropylene Falcon tubes (Corning
352196) pre-coated with 1% BSA. Cell viability and concentration were evaluated using AO-PI viability
dye (0.002% acridine orange and 0.02% propidium iodide) on a hemocytometer.
4. Single-cell RNA sequencing using Parse Bioscience
Single-cell suspensions were immediately fixed using Evercode™ WT v2 (Parse Biosciences)
according to the manufacturer's protocol. Approximately 4000-5000 cells per condition were used for
library construction using either Evercode™ WT Mini v2 or Evercode™ WT v2 (Parse Biosciences).
Two sub-libraries were generated using Evercode™ WT Mini v2 and eight sub-libraries using
Evercode™ WT v2.
All sub-libraries were sequenced on a NovaSeq platform (Illumina).
5. Data analysis
The resulting FASTQ files were demultiplexed using the split-pipe pipeline (v1.2.1) from Parse
Biosciences and aligned to Ensembl GRCz11 reference genome (Ensembl Release 100). The data
was processed and analyzed using Seurat version 4.1.0(Hao et al., 2021). Cells with fewer than 1,000
expressed genes or more than 6,000 expressed genes were filtered out. On average, we obtained
3,466 cells per sample, with each cell expressing an average of 3,070 genes (Table S8). Each sample
was processed separately, with cell types clustered and annotated accordingly.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Data access
The processed data are saved in:
https://drive.google.com/drive/folders/1NTRCoIOviDVmsVhDiaPXpbdDqs40D7rz?usp=drive_link
The folder includes the following files:
• Seurat object of the nine-stage integrated single-cell multi-omic datasets
• Peak file
• Single-cell metadata file
• eRegulon file
• eRegulon specificity file
• DeepDanio models and example code for predicting chromatin accessibility based on DNA
sequences
• DeepDanio identified motifs and TFBSs for each of the 95 cell states
Acknowledgements
We thank the Schier lab for input and discussions. We thank Alba Aparicio Fernandez, Rita Gonzalez
Dominguez and Diana Medeiros Gomes for support with fish husbandry. We thank Fabien Cubizolles
for support with experimental work. High-throughput sequencing was performed at the Genomics
Facility Basel. Part of the computations were performed at the sciCORE Center for scientific computing
at the University of Basel. This work was funded by ERC Advanced grant 834788 and the Allen
Discovery Center for Cell Lineage Tracing to A.F.S., NIH Award R33CA255893 and NSF Award
2021552 to G. S.
Contributions
J.L., and A.F.S. conceived and designed the study. J.L., Y.W., and A.N.C. collected single-cell multi-
omics data. J.L. performed single-cell multi-omics related analyses. S.M.C-H. designed and trained
deep learning model. S.M.C-H., J.L., and C.Y. performed deep learning related analyses. J.L.
reconstructed GRN and performed related analyses. L.Y.D. led the mis-expression experiments with
help from J.L., and M.C-R. J.L. analyzed the mis-expression data. J.L., A.F.S., L.Y.D., S.M.C-H., M.C-
R. C.Y. and G.S. interpreted the results. J.L. wrote the original draft, J.L., A.F.S., G.S. L.Y.D., and
S.M.C-H. finalized the paper with input from all the other authors. All authors read and approved the
final manuscript.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
References
Ameen, M., Sundaram, L., Shen, M., Banerjee, A., Kundu, S., Nair, S., Shcherbina, A., Gu, M.,
Wilson, K. D., Varadarajan, A., Vadgama, N., Balsubramani, A., Wu, J. C., Engreitz, J. M., Farh,
K., Karakikes, I., Wang, K. C., Quertermous, T., Greenleaf, W. J., & Kundaje, A. (2022).
Integrative single-cell analysis of cardiogenesis identifies developmental trajectories and non-
coding mutations in congenital heart disease. Cell, 185(26), 4937-4953.e23.
Argelaguet, R., Clark, S. J., Mohammed, H., Stapel, L. C., Krueger, C., Kapourani, C. A., Imaz-
Rosshandler, I., Lohoff, T., Xiang, Y., Hanna, C. W., Smallwood, S., Ibarra-Soria, X., Buettner,
F., Sanguinetti, G., Xie, W., Krueger, F., Göttgens, B., Rugg-Gunn, P. J., Kelsey, G., … Reik, W.
(2019). Multi-omics profiling of mouse gastrulation at single-cell resolution. Nature, 576(7787),
487–491.
Avsec, Ž., Weilert, M., Shrikumar, A., Krueger, S., Alexandari, A., Dalal, K., Fropf, R., McAnany, C.,
Gagneur, J., Kundaje, A., & Zeitlinger, J. (2021). Base-resolution models of transcription-factor
binding reveal soft motif syntax. Nature Genetics, 53(3), 354–366.
Badia-i-Mompel, P., Wessels, L., Müller-Dott, S., Trimbour, R., Ramirez Flores, R. O., Argelaguet, R.,
& Saez-Rodriguez, J. (2023). Gene regulatory network inference in the era of single-cell multi-
omics. Nature Reviews. Genetics, 24(11), 739–754.
Baranasic, D., Hörtenhuber, M., Balwierz, P. J., Zehnder, T., Mukarram, A. K., Nepal, C., Várnai, C.,
Hadzhiev, Y., Jimenez-Gonzalez, A., Li, N., Wragg, J., D’Orazio, F. M., Relic, D., Pachkov, M.,
Díaz, N., Hernández-Rodríguez, B., Chen, Z., Stoiber, M., Dong, M., … Müller, F. (2022).
Multiomic atlas with functional stratification and developmental dynamics of zebrafish cis-
regulatory elements. Nature Genetics, 54(7), 1037–1050.
Boyer, L. A., Tong, I. L., Cole, M. F., Johnstone, S. E., Levine, S. S., Zucker, J. P., Guenther, M. G.,
Kumar, R. M., Murray, H. L., Jenner, R. G., Gifford, D. K., Melton, D. A., Jaenisch, R., & Young,
R. A. (2005). Core Transcriptional Regulatory Circuitry in Human Embryonic Stem Cells. Cell,
122(6), 947.
Bravo González-Blas, C., De Winter, S., Hulselmans, G., Hecker, N., Matetovici, I., Christiaens, V.,
Poovathingal, S., Wouters, J., Aibar, S., & Aerts, S. (2023). SCENIC+: single-cell multiomic
inference of enhancers and gene regulatory networks. Nature Methods, 20(9), 1355–1367.
Briggs, J. A., Weinreb, C., Wagner, D. E., Megason, S., Peshkin, L., Kirschner, M. W., & Klein, A. M.
(2018). The dynamics of gene expression in vertebrate embryogenesis at single-cell resolution.
Science (New York, N.Y.), 360(6392).
Cao, J., Cusanovich, D. A., Ramani, V., Aghamirzaie, D., Pliner, H. A., Hill, A. J., Daza, R. M.,
McFaline-Figueroa, J. L., Packer, J. S., Christiansen, L., Steemers, F. J., Adey, A. C., Trapnell,
C., & Shendure, J. (2018). Joint profiling of chromatin accessibility and gene expression in
thousands of single cells. Science, 361(6409), 1380–1385.
Charney, R. M., Forouzmand, E., Cho, J. S., Cheung, J., Paraiso, K. D., Yasuoka, Y., Takahashi, S.,
Taira, M., Blitz, I. L., Xie, X., & Cho, K. W. Y. (2017). Foxh1 Occupies cis-Regulatory Modules
Prior to Dynamic Transcription Factor Interactions Controlling the Mesendoderm Gene Program.
Developmental Cell, 40(6), 595-607.e4.
Chowdhury, G., Umeda, K., Ohyanagi, T., Nasu, K., & Yamasu, K. (2024). Involvement of nr2f genes
in brain regionalization and eye development during early zebrafish development. Development,
Growth & Differentiation, 66(2), 145–160.
Davidson, E. H. (2010). Emerging properties of animal gene regulatory networks. Nature, 468(7326),
911.
de Almeida, B. P., Schaub, C., Pagani, M., Secchia, S., Furlong, E. E. M., & Stark, A. (2024).
Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo. Nature,
626(7997), 207–211.
De La Garza, G., Schleiffarth, J. R., Dunnwald, M., Mankad, A., Weirather, J. L., Bonde, G., Butcher,
S., Mansour, T. A., Kousa, Y. A., Fukazawa, C. F., Houston, D. W., Manak, J. R., Schutte, B. C.,
Wagner, D. S., & Cornell, R. A. (2013). Interferon Regulatory Factor 6 promotes differentiation of
the periderm by activating expression of Grainyhead-like 3. The Journal of Investigative
Dermatology, 133(1), 68.
Eraslan, G., Avsec, Ž., Gagneur, J., & Theis, F. J. (2019). Deep learning: new computational
modelling techniques for genomics. Nature Reviews. Genetics, 20(7), 389–403.
Falo-Sanjuan, J., Lammers, N. C., Garcia, H. G., & Bray, S. J. (2019). Enhancer Priming Enables Fast
and Sustained Transcriptional Responses to Notch Signaling. Developmental Cell, 50(4), 411-
425.e8.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Farrell, J. A., Wang, Y., Riesenfeld, S. J., Shekhar, K., Regev, A., & Schier, A. F. (2018). Single-cell
reconstruction of developmental trajectories during zebrafish embryogenesis. Science (New
York, N.Y.), 360(6392).
Fleck, J. S., Jansen, S. M. J., Wollny, D., Zenk, F., Seimiya, M., Jain, A., Okamoto, R., Santel, M., He,
Z., Camp, J. G., & Treutlein, B. (2023). Inferring and perturbing cell fate regulomes in human
brain organoids. Nature, 621(7978), 365–372.
Granja, J. M., Corces, M. R., Pierce, S. E., Bagdatli, S. T., Choudhry, H., Chang, H. Y., & Greenleaf,
W. J. (2021). ArchR is a scalable software package for integrative single-cell chromatin
accessibility analysis. Nature Genetics 2021 53:3, 53(3), 403–411.
Hao, Y., Hao, S., Andersen-Nissen, E., Mauck, W. M., Zheng, S., Butler, A., Lee, M. J., Wilk, A. J.,
Darby, C., Zager, M., Hoffman, P., Stoeckius, M., Papalexi, E., Mimitou, E. P., Jain, J.,
Srivastava, A., Stuart, T., Fleming, L. M., Yeung, B., … Satija, R. (2021). Integrated analysis of
multimodal single-cell data. Cell, 184(13), 3573-3587.e29.
Jacobs, J., Atkins, M., Davie, K., Imrichova, H., Romanelli, L., Christiaens, V., Hulselmans, G., Potier,
D., Wouters, J., Taskiran, I. I., Paciello, G., González-Blas, C. B., Koldere, D., Aibar, S., Halder,
G., & Aerts, S. (2018). The transcription factor Grainy head primes epithelial enhancers for
spatiotemporal activation by displacing nucleosomes. Nature Genetics, 50(7), 1011–1020.
Janssens, J., Aibar, S., Taskiran, I. I., Ismail, J. N., Gomez, A. E., Aughey, G., Spanier, K. I., De Rop,
F. V., González-Blas, C. B., Dionne, M., Grimes, K., Quan, X. J., Papasokrati, D., Hulselmans,
G., Makhzami, S., De Waegeneer, M., Christiaens, V., Southall, T., & Aerts, S. (2022). Decoding
gene regulation in the fly brain. Nature, 601(7894), 630–636.
Kamimoto, K., Stringa, B., Hoffmann, C. M., Jindal, K., Solnica-Krezel, L., & Morris, S. A. (2023).
Dissecting cell identity via network inference and in silico gene perturbation. Nature, 614(7949),
742–751.
Karlsson, M., Zhang, C., Méar, L., Zhong, W., Digre, A., Katona, B., Sjöstedt, E., Butler, L., Odeberg,
J., Dusart, P., Edfors, F., Oksvold, P., von Feilitzen, K., Zwahlen, M., Arif, M., Altay, O., Li, X.,
Ozcan, M., Mardonoglu, A., … Lindskog, C. (2021). A single–cell type transcriptomics map of
human tissues. Science Advances, 7(31).
Kelley, D. R., Snoek, J., & Rinn, J. L. (2016). Basset: learning the regulatory code of the accessible
genome with deep convolutional neural networks. Genome Research, 26(7), 990–999.
Kim, H. S., Tan, Y., Ma, W., Merkurjev, D., Destici, E., Ma, Q., Suter, T., Ohgi, K., Friedman, M.,
Skowronska-Krawczyk, D., & Rosenfeld, M. G. (2018). Pluripotency factors functionally premark
cell-type-restricted enhancers in ES cells. Nature, 556(7702), 510–514.
Kimmel, C. B., Ballard, W. W., Kimmel, S. R., Ullmann, B., & Schilling, T. F. (1995). Stages of
embryonic development of the zebrafish. Developmental Dynamics : An Official Publication of
the American Association of Anatomists, 203(3), 253–310.
Kimmel, C. B., Warga, R. M., & Schilling, T. F. (1990). Origin and organization of the zebrafish fate
map. Development (Cambridge, England), 108(4), 581–594.
Kotliar, D., Veres, A., Nagy, M. A., Tabrizi, S., Hodis, E., Melton, D. A., & Sabeti, P. C. (2019).
Identifying gene expression programs of cell-type identity and cellular activity with single-cell
RNA-Seq. ELife, 8.
Lee, H., & Kimelman, D. (2002). A dominant-negative form of p63 is required for epidermal
proliferation in zebrafish. Developmental Cell, 2(5), 607–616.
Lee, M. T., Bonneau, A. R., Takacs, C. M., Bazzini, A. A., Divito, K. R., Fleming, E. S., & Giraldez, A.
J. (2013). Nanog, Pou5f1 and SoxB1 activate zygotic gene expression during the maternal-to-
zygotic transition. Nature, 503(7476), 360–364.
Liber, D., Domaschenz, R., Holmqvist, P. H., Mazzarella, L., Georgiou, A., Leleu, M., Fisher, A. G.,
Labosky, P. A., & Dillon, N. (2010). Epigenetic priming of a pre-B cell-specific enhancer through
binding of Sox2 and Foxd3 at the ESC stage. Cell Stem Cell, 7(1), 114–126.
Liberali, P., & Schier, A. F. (2024). The evolution of developmental biology through conceptual and
technological revolutions. Cell, 187(14), 3461–3495.
Lim, H. Y. G., Alvarez, Y. D., Gasnier, M., Wang, Y., Tetlak, P., Bissiere, S., Wang, H., Biro, M., &
Plachta, N. (2020). Keratins are asymmetrically inherited fate determinants in the mammalian
embryo. Nature, 585(7825), 404–409.
Liu, H., Leslie, E. J., Jia, Z., Smith, T., Eshete, M., Butali, A., Dunnwald, M., Murray, J., & Cornell, R.
A. (2016). Irf6 directly regulates Klf17 in zebrafish periderm and Klf4 in murine oral epithelium,
and dominant-negative KLF4 variants are present in patients with cleft lip and palate. Human
Molecular Genetics, 25(4), 766.
Lundberg, S. M., & Lee, S. I. (2017). A Unified Approach to Interpreting Model Predictions. Advances
in Neural Information Processing Systems, 2017-December, 4766–4775.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Ma, S., Zhang, B., LaFave, L. M., Earl, A. S., Chiang, Z., Hu, Y., Ding, J., Brack, A., Kartha, V. K.,
Tay, T., Law, T., Lareau, C., Hsu, Y. C., Regev, A., & Buenrostro, J. D. (2020). Chromatin
Potential Identified by Shared Single-Cell Profiling of RNA and Chromatin. Cell, 183(4), 1103-
1116.e20.
Maslova, A., Ramirez, R. N., Ma, K., Schmutz, H., Wang, C., Fox, C., Ng, B., Benoist, C., & Mostafavi,
S. (2020). Deep learning of immune cell differentiation. Proceedings of the National Academy of
Sciences of the United States of America, 117(41), 25655–25666.
McDole, K., Guignard, L., Amat, F., Berger, A., Malandain, G., Royer, L. A., Turaga, S. C., Branson,
K., & Keller, P. J. (2018). In Toto Imaging and Reconstruction of Post-Implantation Mouse
Development at the Single-Cell Level. Cell, 175(3), 859-876.e33.
Meng, X., Zheng, Y., Zhang, L., Liu, P., Liu, Z., & He, Y. (2024). Single-Cell Analyses Reveal the
Metabolic Heterogeneity and Plasticity of the Tumor Microenvironment during Head and Neck
Squamous Cell Carcinoma Progression. Cancer Research, 84(15).
Miao, L., Tang, Y., Bonneau, A. R., Chan, S. H., Kojima, M. L., Pownall, M. E., Vejnar, C. E., Gao, F.,
Krishnaswamy, S., Hendry, C. E., & Giraldez, A. J. (2022). The landscape of pioneer factor
activity reveals the mechanisms of chromatin reprogramming and genome activation. Molecular
Cell, 82(5), 986-1002.e9.
Miles, L. B., Darido, C., Kaslin, J., Heath, J. K., Jane, S. M., & Dworkin, S. (2017). Mis-expression of
grainyhead-like transcription factors in zebrafish leads to defects in enveloping layer (EVL)
integrity, cellular morphogenesis and axial extension. Scientific Reports 2017 7:1, 7(1), 1–14.
Murtaugh, L. C. (2007). Pancreas and beta-cell development: from the actual to the possible.
Development (Cambridge, England), 134(3), 427–438.
Nelson, A. C., Cutty, S. J., Gasiunas, S. N., Deplae, I., Stemple, D. L., & Wardle, F. C. (2017a). In
Vivo Regulation of the Zebrafish Endoderm Progenitor Niche by T-Box Transcription Factors.
Cell Reports, 19(13), 2782–2795.
Nelson, A. C., Cutty, S. J., Gasiunas, S. N., Deplae, I., Stemple, D. L., & Wardle, F. C. (2017b). In
Vivo Regulation of the Zebrafish Endoderm Progenitor Niche by T-Box Transcription Factors.
Cell Reports, 19(13), 2782–2795.
Novakovsky, G., Dexter, N., Libbrecht, M. W., Wasserman, W. W., & Mostafavi, S. (2022). Obtaining
genetics insights from deep learning via explainable artificial intelligence. Nature Reviews
Genetics 2022 24:2, 24(2), 125–137.
Paik, E. J., Mahony, S., White, R. M., Price, E. N., Dibiase, A., Dorjsuren, B., Mosimann, C.,
Davidson, A. J., Gifford, D., & Zon, L. I. (2013). A Cdx4-Sall4 regulatory module controls the
transition from mesoderm formation to embryonic hematopoiesis. Stem Cell Reports, 1(5), 425–
436.
Pálfy, M., Schulze, G., Valen, E., & Vastenhouw, N. L. (2020). Chromatin accessibility established by
Pou5f3, Sox19b and Nanog primes genes for activity during zebrafish genome activation. PLoS
Genetics, 16(1).
Qiu, C., Cao, J., Martin, B. K., Li, T., Welsh, I. C., Srivatsan, S., Huang, X., Calderon, D., Noble, W.
S., Disteche, C. M., Murray, S. A., Spielmann, M., Moens, C. B., Trapnell, C., & Shendure, J.
(2022). Systematic reconstruction of cellular trajectories across mouse embryogenesis. Nature
Genetics, 54(3), 328–341.
Quinlan, A. R., & Hall, I. M. (2010). BEDTools: a flexible suite of utilities for comparing genomic
features. Bioinformatics (Oxford, England), 26(6), 841–842.
Reddington, J. P., Garfield, D. A., Sigalova, O. M., Karabacak Calviello, A., Marco-Ferreres, R.,
Girardot, C., Viales, R. R., Degner, J. F., Ohler, U., & Furlong, E. E. M. (2020). Lineage-
Resolved Enhancer and Promoter Usage during a Time Course of Embryogenesis.
Developmental Cell, 55(5), 648-664.e9.
Sabel, J. L., d’Alençon, C., O’Brien, E. K., Otterloo, E. Van, Lutz, K., Cuykendall, T. N., Schutte, B. C.,
Houston, D. W., & Cornell, R. A. (2009). Maternal Interferon Regulatory Factor 6 is required for
the differentiation of primary superficial epithelia in Danio and Xenopus embryos. Developmental
Biology, 325(1), 249–262.
Sarropoulos, I., Sepp, M., Frömel, R., Leiss, K., Trost, N., Leushkin, E., Okonechnikov, K., Joshi, P.,
Giere, P., Kutscher, L. M., Cardoso-Moreira, M., Pfister, S. M., & Kaessmann, H. (2021).
Developmental and evolutionary dynamics of cis-regulatory elements in mouse cerebellar cells.
Science, 373(6558).
Satija, R., Farrell, J. A., Gennert, D., Schier, A. F., & Regev, A. (2015). Spatial reconstruction of
single-cell gene expression data. Nature Biotechnology 2015 33:5, 33(5), 495–502.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Saunders, L. M., Srivatsan, S. R., Duran, M., Dorrity, M. W., Ewing, B., Linbo, T. H., Shendure, J.,
Raible, D. W., Moens, C. B., Kimelman, D., & Trapnell, C. (2023). Embryo-scale reverse
genetics at single-cell resolution. Nature, 623(7988), 782–791
Sawada, A., Fritz, A., Jiang, Y. J., Yamamoto, A., Yamasu, K., Kuroiwa, A., Saga, Y., & Takeda, H.
(2000). Zebrafish Mesp family genes, mesp-a and mesp-b are segmentally expressed in the
presomitic mesoderm, and Mesp-b confers the anterior identity to the developing somites.
Development, 127(8), 1691–1702.
Schep, A. N., Wu, B., Buenrostro, J. D., & Greenleaf, W. J. (2017). chromVAR: Inferring transcription
factor-associated accessibility from single-cell epigenomic data. Nature Methods, 14(10), 975.
Schier, A. F., & Talbot, W. S. (2005). Molecular genetics of axis formation in zebrafish. Annual Review
of Genetics, 39(Volume 39, 2005), 561–613.
Schulte-Merker, S., Van Eeden, F. J. M., Halpern, M. E., Kimmel, C. B., & Nüsslein-Volhard, C.
(1994). no tail (ntl) is the zebrafish homologue of the mouse T (Brachyury) gene. Development,
120(4), 1009–1015.
Sendra, M., McDole, K., Jimenez-Carretero, D., Hourcade, J. de D., Temiño, S., Raiola, M., Guignard,
L., Keller, P. J., Sánchez-Cabo, F., Domínguez, J. N., & Torres, M. (2024). Epigenetic priming of
embryonic lineages in the mammalian epiblast. BioRxiv, 2024.01.11.575188.
Shrikumar, A., Tian, K., Avsec, Ž., Shcherbina, A., Banerjee, A., Sharmin, M., Nair, S., & Kundaje, A.
(2018). Technical Note on Transcription Factor Motif Discovery from Importance Scores (TF-
MoDISco) version 0.5.6.5.
Spitz, F., & Furlong, E. E. M. (2012). Transcription factors: from enhancer binding to developmental
control. Nature Reviews. Genetics, 13(9), 613–626.
Stuart, T., Srivastava, A., Madad, S., Lareau, C. A., & Satija, R. (2021). Single-cell chromatin state
analysis with Signac. Nature Methods 2021 18:11, 18(11), 1333–1341.
Sun, K., Liu, X., Xu, R., Liu, C., Meng, A., & Lan, X. (2024). Mapping the chromatin accessibility
landscape of zebrafish embryogenesis at single-cell resolution by SPATAC-seq. Nature Cell
Biology, 26(7), 1187–1199.
Taskiran, I. I., Spanier, K. I., Dickmänken, H., Kempynck, N., Pančíková, A., Ekşi, E. C., Hulselmans,
G., Ismail, J. N., Theunis, K., Vandepoel, R., Christiaens, V., Mauduit, D., & Aerts, S. (2024).
Cell-type-directed design of synthetic enhancers. Nature, 626(7997), 212–220.
Wagner, D. E., Weinreb, C., Collins, Z. M., Briggs, J. A., Megason, S. G., & Klein, A. M. (2018).
Single-cell mapping of gene expression landscapes and lineage in the zebrafish embryo.
Science (New York, N.Y.), 360(6392), 981–987.
Wan, Y., Jakob El Kholtei, Ignatius Jenie, Mariona Colomer-Rosell, Jialin Liu, Joaquin Navajas
Acedo, Lucia Y. Du, Mireia Codina-Tobias, Mengfan Wang, Ahilya Sawh, Edward Lin, Serena
Chuang, Susan E. Mango, Guoqiang Yu, Bogdan Bintu, & Schier, A. F. (2024). Whole-embryo
Spatial Transcriptomics at Subcellular Resolution from Gastrulation to Organogenesis. BioRxiv.
Wang, A., Yue, F., Li, Y., Xie, R., Harper, T., Patel, N. A., Muth, K., Palmer, J., Qiu, Y., Wang, J.,
Lam, D. K., Raum, J. C., Stoffers, D. A., Ren, B., & Sander, M. (2015). Epigenetic priming of
enhancers predicts developmental competence of hESC-derived endodermal lineage
intermediates. Cell Stem Cell, 16(4), 386–399.
Wang, X., Wang, W., Wang, Y., Chen, J., Liu, G., & Zhang, Y. (2022). Antibody-free profiling of
transcription factor occupancy during early embryogenesis by FitCUT&RUN. Genome Research,
32(2), 378–388.
Wang, Y., Liu, J., Du, L. Y., Wyss, J. L., Farrell, J. A., & Schier, A. F. (2023). Gene module
reconstruction elucidates cellular differentiation processes and the regulatory logic of specialized
secretion. BioRxiv, 2023.12.29.573643.
Wei, B., Jolma, A., Sahu, B., Orre, L. M., Zhong, F., Zhu, F., Kivioja, T., Sur, I., Lehtiö, J., Taipale, M.,
& Taipale, J. (2018). A protein activity assay to measure global transcription factor activity
reveals determinants of chromatin accessibility. Nature Biotechnology 2018 36:6, 36(6), 521–
529.
Weinberg, E. S., Allende, M. L., Kelly, C. S., Abdelhamid, A., Murakami, T., Andermann, P., Doerre,
O. G., Grunwald, D. J., & Riggleman, B. (1996). Developmental regulation of zebrafish MyoD in
wild-type, no tail and spadetail embryos. Development, 122(1), 271–280.
Xu, C., Fan, Z. P., Müller, P., Fogley, R., DiBiase, A., Trompouki, E., Unternaehrer, J., Xiong, F.,
Torregroza, I., Evans, T., Megason, S. G., Daley, G. Q., Schier, A. F., Young, R. A., & Zon, L. I.
(2012). Nanog-like regulates endoderm formation through the Mxtx2-Nodal pathway.
Developmental Cell, 22(3), 625–638.
Yates, A. D., Achuthan, P., Akanni, W., Allen, J., Allen, J., Alvarez-Jarreta, J., Amode, M. R., Armean,
I. M., Azov, A. G., Bennett, R., Bhai, J., Billis, K., Boddu, S., Marugán, J. C., Cummins, C.,
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Davidson, C., Dodiya, K., Fatima, R., Gall, A., … Flicek, P. (2020). Ensembl 2020. Nucleic Acids
Research, 48(D1), D682–D688.
Yin, C., Hair, S. C., Byeon, G. W., Bromley, P., Meuleman, W., & Seelig, G. (2024). Iterative deep
learning-design of human enhancers exploits condensed sequence grammar to achieve cell
type-specificity. BioRxiv, 2024.06.14.599076.
Zalik, S. E., Lewandowski, E., Kam, Z., & Geiger, B. (1999). Cell adhesion and the actin cytoskeleton
of the enveloping layer in the zebrafish embryo during epiboly. Biochemistry and Cell Biology =
Biochimie et Biologie Cellulaire, 77(6), 527–542.
Zhao, X. F., Suh, C. S., Prat, C. R., Ellingsen, S., & Fjose, A. (2009). Distinct expression of two foxg1
paralogues in zebrafish. Gene Expression Patterns, 9(5), 266–272.
Zhou, J., & Troyanskaya, O. G. (2015). Predicting effects of noncoding variants with deep learning-
based sequence model. Nature Methods, 12(10), 931–934.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S1: Comparison of scMultiome ATAC and standalone scATAC
A and B. UM AP v isual ization of s ingle cell s, with e ach cel l col ored ac cording to its resp ective conditio n. S1,
S2, and S3 represent different technical replicates.
C and D. Count correlation (spearman’s rank correlation) between scMultiome ATAC and standalone
scATAC. The genome was divided into 500 bp bins, and the fragment counts were calculated for each bin
in each cell. These counts were then aggregated across single cells to create a pseudobulk-count. The
correlation was performed on these pseudobulk-counts.
0 500 1000 1500 2000 2500
0 500 1000 1500 2000
6 somite
scATAC counts
scMultiome counts
Rho=0.934
0 500 1000 2000 3000
0 500 1000 1500 2000
50% Epiboly
scATAC counts
scMultiome counts
Rho=0.928
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
UMAP Dimension 1
UMAP Dimension 2
color 1−scATAC_S1 2−scATAC_S2 3−scMultiome_S1 4−scMultiome_S2
UMAP Dimension 1
UMAP Dimension 2
color 1−scATAC_6Som_S1 2−scATAC_6Som_S2 3−scATAC_6Som_S3 4−scMultiome_6Som_S1 5−scMultiome_6Som_S2
A B
CD
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
deep cells(oblong)
evl(oblong)
−4
0
4
0 5
UMAP_1
UMAP_2
Oblong
−4
−2
0
2
−6−3 0 3 6
umap.atac_1
umap.atac_2
Oblong
ysl(dome)
apoptosis like(dome)
evl(dome)
ectoderm
non−dorsal margin
−20
−10
0
10
−10 −5 0 5
UMAP_1
UMAP_2
dorsal margin
Dome
−5.0
−2.5
0.0
2.5
5.0
−5.0 −2.5 0.0 2.5
umap.atac_1
umap.atac_2
Dome
ectoderm(30epi)
EVL
YSL
non-dorsal margin
apoptosis like(30epi)
−10
−5
0
5
10
−5 0 5 10 15
UMAP_1
UMAP_2
30% Epiboly
dorsal margin
−3
0
3
6
−10 −5 0 5
umap.atac_1
umap.atac_2
30% Epiboly
dorsal ectoderm
ventral ectoderm
v
YSL
apoptosis like
EVL
−5
5
10
15
−10 0 10
UMAP_1
UMAP_2
0
50% Epiboly
dorsal posterior
dorsal anterior
margin tail
ventrolateral mesoendoderm
evl(50epi)
−6
−3
0
3
6
−5 0 5 10
umap.atac_1
umap.atac_2
50% Epiboly
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S2: UMAP visualization of cell state diversification from the oblong to bud stages
The clusters/cell states are defined by gene expression, and the UMAP coordinates derived from either
RNA-seq data (left) or ATAC-seq data (right).
ysl(shield)
evl(shield)
dorsal ectoderm
endoderm(shield)
ventral ectoderm
ventrolateral mesoderm
−10
0
−10 0 10
UMAP_1
UMAP_2
dorsal anterior
dorsal posterior
margin tail
Shield
−5
0
5
10
−15 −10 −5 0 5
umap.atac_1
umap.atac_2
Shield
YSL
epidermis
endoderm
EVL
neural plate anterior
lateral plate mesoderm
prechordal plate
neural plate posterior
margin tail
notochord
forer
adaxial cells
−10
0
10
−10 0 10
UMAP_1
UMAP_2
psm
forerunner cells
75% Epiboly
−10
−5
0
5
10
0 10
umap.atac_1
umap.atac_2
75% Epiboly
ysl
psm)
anterior neural plate border
hatching gland
adaxial cells
dorsal
diencephalon
telencephalon
epidermis
evl(bud)
posterior neural plate border
cephalic mesoderm
notochord
endoderm
lateral plate mesoderm
tailbud
−10
0
10
20
−10 0 10
UMAP_1
UMAP_2
optic primordium
ventral
diencephalon
hindbrain
midbrain
spinal cord
Bud
ud)
−10
0
10
−15 −10 −5 0 5
umap.atac_1
umap.atac_2
Bud
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S3: The number of acquired dCREs at each stage that are maintained by the 6-somite stage for the
YSL trajectory. Due to the low number of YSL cells at the dome stage (only 22 cells), this stage was
excluded from the analysis.
Figure S4: Deep learning quality control
A. For each cell type, we calculated the Pearson or Spearman correlation coefficient between the predicted
and actual chromatin accessibility. The boxplot shows these correlations, with the lower and upper whiskers
representing 1.5 times the interquartile range (IQR). The box itself displays the IQR, with the median
indicated.
0.0 0.2 0.4 0.6 0.8
Pearson's r
Spearmna's rho
A
10 20 30 40 50
Cdx4
Disatnce from TFBS (bp)
ChIP−Seq signal
−2000
−1000
0
1000
2000
0.1 0.2 0.3 0.4 0.5 0.6
Nanog
Disatnce from TFBS (bp)
ChIP−Seq signal
−2000
−1000
0
1000
2000
0.10 0.15 0.20 0.25 0.30 0.35
Pou5f3
Disatnce from TFBS (bp)
ChIP−Seq signal
−2000
−1000
0
1000
2000
YSL 50%epi
gata6
hnf4
mxtx2
EVL 50%epi
grhl3
klf17
cebpb
Grhl3_variant
Adaxial cells(6somite)
myod1
tbx15
hoxa9b
patz1
telencephalon (6somite)
sox3
zic1
otx2b
epidermis(6somite)
tfap2a
tead3
Hatching gland (6somite)
foxa
tp63
gsc
klf17
Foxa_variant
Motif DanioDeep Motif DanioCode
Motif DanioDeep Motif DanioCode
Motif DanioDeep Motif DanioCode
Motif DanioDeep Motif DanioCode
Motif DanioDeep Motif DanioCode
Motif DanioDeep Motif DanioCode
B
C
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
B. ChIP-seq signal in either 2kb upstream or downstram of TFBS.
C. De novo motifs identified by DeepDanio were annotated using motifs from the DANIO-CODE
database.
Figure S5: Histogram showing the number of genes regulated by each TF
Blue represents EVL eRegulons, while green represents YSL eRegulons. The top 10 specific eRegulons in
each cell state of the EVL (from oblong to 6-somite) and YSL (from dome to 6-somite) were selected,
combined, and filtered to create unique sets of EVL and YSL eRegulons.
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S6: De novo motifs identified by DeepDanio were annotated using motifs from the DANIO-
CODE database
A. De novo motifs identified by DeepDanio for the EVL trajectory. For each cell state along the EVL
trajectory, the top 5 motifs were selected. Since some motifs were shared between stages, all top 5
motifs from each cell state were combined and filtered to obtain a unique set of motifs. This figure
displays the resulting unique set of motifs for the EVL trajectory.
B. De novo motifs identified by DeepDanio for the YSL trajectory. For each cell state along the YSL
trajectory, the top 5 motifs were selected. Since some motifs were shared between stages, all top 5
motifs from each cell state were combined and filtered to obtain a unique set of motifs. This figure
displays the resulting unique set of motifs for the YSL trajectory.
C. Expression of irf2a and irf6. The same motif was assigned to different transcription factors based on
expression levels. The second motif in EVL was annotated as irf6 due to irf6 being expressed in EVL,
while the first motif in YSL was annotated as irf2a because irf2a was expressed in YSL.
EVL trajectory motifs
irf6
sp1
tbx16
pou5f3
klf17
grhl3
cebpb_variant
cebpb
grhl3_variant
tfap2a
tead3a
Motif DanioDeep Motif DanioCode
YSL trajectory motifs
irf2a
mgaa
znf740a
patz1
mxtx2
foxa3
mxtx2_variant
gata6
gata6_variant
hnf4a
onecut1
hnf1ba
Motif DanioDeep Motif DanioCode
EVL trajectory
YSL trajectory
ABC
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
0.0
0.4
0.8
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
50
100
n_neighbors
Rho= −0.317 / p= 1.723411e−09
cebpb and tead3a interaction (50% Epiboly EVL)
0.0
0.5
1.0
1.5
2.0
2.5
0 100 200
Motif pair distance (bp)
Predicted cooperativity
50
100
150
n_neighbors
Rho= −0.441 / p= 2.376967e−28
cebpb and tfap2a interaction (50% Epiboly EVL)
−0.5
0.0
0.5
1.0
1.5
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
100
200
300
n_neighbors
Rho= −0.305 / p= 2.313172e−25
grhl3 and cebpb interaction (50% Epiboly EVL)
0
1
2
0 100 200
Motif pair distance (bp)
Predicted cooperativity
50
100
150
n_neighbors
Rho= −0.343 / p= 1.102409e−16
grhl3 and tead3a interaction (50% Epiboly EVL)
−1
0
1
0 50 100 150 200
Motif pair distance (bp)
Predicted cooperativity
100
200
n_neighbors
Rho= −0.174 / p= 2.774054e−07
grhl3 and tfap2a interaction (50% Epiboly EVL)
−1.0
−0.5
0.0
0.5
1.0
0 100 200 300
Motif pair distance (bp)
Predicted cooperativity
250
500
750
n_neighbors
Rho= −0.261 / p= 1.203079e−28
klf17 and cebpb interaction (50% Epiboly EVL)
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S7: Correlation between distances of two TFBSs and interaction scores
The Spearman’s correlation coefficient (Rho) and the P-value are displayed in the top-left corner of the plot.
−0.5
0.0
0.5
1.0
0 100 200
Motif pair distance (bp)
Predicted cooperativity
50
100
150
200
n_neighbors
Rho= −0.199 / p= 7.708311e−07
klf17 and tead3a interaction (50% Epiboly EVL)
−0.5
0.0
0.5
1.0
0 100 200 300
Motif pair distance (bp)
Predicted cooperativity
100
200
n_neighbors
Rho= −0.135 / p= 1.603117e−04
klf17 and tfap2a interaction (50% Epiboly EVL)
0
1
2
100 200
Motif pair distance (bp)
Predicted cooperativity
25
50
75
n_neighbors
Rho= −0.108 / p= 6.606988e−02
tfap2a and tead3a interaction (50% Epiboly EVL)
−1.0
−0.5
0.0
0.5
1.0
0 100 200 300
Motif pair distance (bp)
Predicted cooperativity
250
500
750
1000
n_neighbors
Rho= −0.374 / p= 1.252148e−86
gata6 and hnf4 interaction (50% Epiboly YSL)
−1.0
−0.5
0.0
0.5
1.0
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
100
200
300
400
n_neighbors
Rho= −0.314 / p= 6.574145e−31
onecut1 and hnf4 interaction (50% Epiboly YSL)
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S8: Venn Diagrams of CREs containing gata6, hnf4, onecut1, and mxtx2 Motifs
Figure S9: YSL misexpression without mxtx2
A. UMAP projection of single-cell transcriptomes, with cells colored by cell type.
9
64
18
1328
589
689
19
77
406
974
449 2306
79
9
1933
gata6.peaks mxtx2.peaks
onecut1.peaks hnf4.peaks
9
64
18
1328
589
689
19
77
406
974
449 2306
79
9
1933
gata6.peaks mxtx2.peaks
onecut1.peaks hnf4.peaks
AB
C
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
B. UMAP projection of single-cell transcriptomes, with cells colored by marker gene expression. ctsh,
slmapb, and plk3 are YSL marker genes, tbxta is a marker for mesoderm, sox3 is a marker for ectoderm,
and krt4 is a marker for EVL.
C. Heatmap showing the expression of YSL-specific genes that are activated by the misexpression of all
differentiation TFs, excluding Mxtx2.
Figure S10: mxtx2 expression across YSL trajectory
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
Figure S11: ablation analysis including mxtx2 motif
A: Distribution of ablation scores for each TF
B, C, D: Correlation between distances of two TFBSs and interaction scores. The Spearman’s correlation
coefficient (Rho) and the P-value are displayed in the top-left corner of the plot.
0 1 2 3 4 5
0.0 0.2 0.4 0.6 0.8 1.0 1.2
YSL (50% Epiboly)
Ablation score
Density
mxtx2
gata6
onecut1
hnf4
−0.5
0.0
0.5
1.0
1.5
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
100
200
300
n_neighbors
Rho= −0.379 / p= 4.110166e−51
gata6 and mxtx2 interaction (50% Epiboly YSL)
−1.0
−0.5
0.0
0.5
1.0
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
n_neighbors
100
200
300
400
Rho= −0.449 / p= 4.204944e−63
mxtx2 and hnf4 interaction (50% Epiboly YSL)
−1.0
−0.5
0.0
0.5
1.0
0 50 100 150 200 250
Motif pair distance (bp)
Predicted cooperativity
n_neighbors
100
200
300
Rho= −0.316 / p= 2.753228e−23
mxtx2 and onecut1 interaction (50% Epiboly YSL)
A
D
C
B
.CC-BY-NC-ND 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 27, 2024. ; https://doi.org/10.1101/2024.08.27.609971doi: bioRxiv preprint
