Technology
Lineage motifs as developmental modules for
control of cell type proportions
Graphical abstract
Highlights
d
Lineage motifs are overrepresented patterns of cell fates on
lineage trees
d
Lineage motifs could reflect committed progenitors or
extrinsic interactions
d
Lineage motifs are identified in existing retina and early
embryo lineage datasets
d
Lineage motifs could facilitate adaptive variation in cell type
proportions
Authors
Martin Tran, Amjad Askary,
Michael B. Elowitz
Correspondence
amjada@g.ucla.edu (A.A.),
melowitz@caltech.edu (M.B.E.)
In brief
Biological tissues require fine-tuned cell
type proportions for optimal function, but
how this process is regulated remains
poorly understood. Tran et al. suggest
that lineage motifs reflect modular
developmental programs that could
constrain variation in cell type
proportions.
Tran et al., 2024, Developmental Cell
59
, 812–826
March 25, 2024
ª
2024 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.devcel.2024.01.017
ll
Technology
Lineage motifs as developmental modules
for control of cell type proportions
Martin Tran,
1
Amjad Askary,
2
,
*
and Michael B. Elowitz
1
,
3
,
4
,
*
1
Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, CA 91125, USA
2
Department of Molecular, Cell and Developmental Biology, University of California, Los Angeles, Los Angeles, CA 90095, USA
3
Howard Hughes Medical Institute, Chevy Chase, MD 20815, USA
4
Lead contact
*Correspondence:
amjada@g.ucla.edu
(A.A.),
melowitz@caltech.edu
(M.B.E.)
https://doi.org/10.1016/j.devcel.2024.01.017
SUMMARY
In multicellular organisms, cell types must be produced and maintained in appropriate proportions. One way
this is achieved is through committed progenitor cells or extrinsic interactions that produce specific patterns
of descendant cell types on lineage trees. However, cell fate commitment is probabilistic in most contexts,
making it difficult to infer these dynamics and understand how they establish overall cell type proportions.
Here, we introduce Lineage Motif Analysis (LMA), a method that recursively identifies statistically overrepre-
sented patterns of cell fates on lineage trees as potential signatures of committed progenitor states or
extrinsic interactions. Applying LMA to published datasets reveals spatial and temporal organization of
cell fate commitment in zebrafish and rat retina and early mouse embryonic development. Comparative anal-
ysis of vertebrate species suggests that lineage motifs facilitate adaptive evolutionary variation of retinal cell
type proportions. LMA thus provides insight into complex developmental processes by decomposing them
into simpler underlying modules.
INTRODUCTION
Most tissues comprise multiple specialized cell types that
appear in appropriate proportions to support proper tissue-level
functions. In many cases, cell type proportions vary spatially
within the tissue. For example, the center of the primate retina
is cone-dense, allowing for high visual acuity, while the periph-
ery is rod-dense, enabling greater sensitivity in low light condi-
tions.
1
Cell type proportions also vary between species. For
instance, the ratio of rod and cone photoreceptors varies de-
pending on the visual needs associated with the lifestyle of
each species.
2
Tissue development thus faces the fundamental
challenges of (1) generating cell types in correct proportions,
and (2) facilitating spatial and evolutionary changes in those
proportions.
3
,
4
One prevalent mechanism for specifying cell type propor-
tions occurs through regulating cell fate differentiation. As pro-
genitor cells undergo successive rounds of cell division, they
progressively become restricted in their fate potential, eventu-
ally committing to terminal cell fates. This process can be
described in terms of a collection of cell states and the rates
at which cells in each state transition to other states, i.e., a
cell state transition map
5
(
Figures 1
A and 1B). In some cases,
like the nematode
C. elegans
, cell state transitions are deter-
ministic, producing a stereotyped lineage tree in all individuals.
6
However, in most other organisms, one cannot infer a quantita-
tive cell state transition map from any single lineage tree due to
variability. For example, in the mammalian retina, individual
progenitor cells can give rise to a wide distribution of cell
numbers and types with no apparent fixed ratios between
different types. This observation prompted investigators to
initially suggest a stochastic view of cell fate determination.
7
,
8
However, other studies of terminally dividing progenitors with
particular expression patterns provided evidence for consistent
cell-intrinsic biases in cell fate decisions.
9–15
These biases also
appear in earlier, non-terminal divisions.
11
,
16–22
Cell state tran-
sition dynamics can also integrate extrinsic signals, develop-
mental time, and stochastic ‘‘noise’’ with internal progenitor
states.
23
,
24
Thus, even in well-studied systems such as the
retina, it remains a major challenge to quantitatively elucidate
cell state transition maps.
Different cell state transition maps can generate distinct dis-
tributions of cell fates on lineage trees. One simple transition
map comprises a multipotent progenitor that can directly and
probabilistically differentiate into multiple terminal fates (
Fig-
ure 1
A). A system employing such a direct, memoryless transi-
tion map would not exhibit fate correlations between related
cells. Alternatively, a more complex transition map could
involve the probabilistic generation of various types of
committed progenitors, each predetermined to give rise to an
invariant set of descendant cell types (
Figure 1
B). In this
case, each type of progenitor would produce a characteristic
distribution of descendant cell fates, introducing fate correla-
tions on lineage trees. These fate correlations represent lineage
812
Developmental Cell
59
, 812–826, March 25, 2024
ª
2024 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (
http://creativecommons.org/licenses/by/4.0/
).
ll
OPEN ACCESS
motifs that reflect otherwise hidden progenitor states. Further-
more, based on what motifs are used to specify cell types in
developing tissues, this could in turn limit variation in overall
cell type proportions (
Figure 1
C).
Recently, new methods have begun to allow for lineage tree
reconstruction at scale. Long-term
in toto
live imaging allows
direct tracking of dividing progenitor cells.
25
Additionally, a
new generation of engineered lineage reconstruction systems
has emerged.
26–32
These advances provoke the question of
how fully resolved lineage trees with endpoint cell fates can be
used to infer cell state transition maps.
To address this challenge, we introduce Lineage Motif Anal-
ysis (LMA), a computational approach for inferring statistically
overrepresented patterns of cell fates on lineage trees. LMA
is based on the principle of motif detection, which has been
used to identify the building blocks of complex regulatory
networks,
33
DNA sequences,
34
,
35
and other biological fea-
tures,
36
,
37
but has not to our knowledge been applied to under-
stand cell fate differentiation. As a ‘‘bottom-up,’’ data-driven
approach, LMA does not require specific assumptions about
underlying molecular mechanisms and can be applied to
diverse systems for which sufficient cell lineage information is
available. Biologically, motifs could be generated by progeni-
tors intrinsically programmed to autonomously give rise to spe-
cific patterns of descendant cell fates. They could also reflect
more complex cell state transition maps involving extrinsic
cues and cell-cell signaling that generate correlated cell fate
patterns on lineage trees.
Here, we first define LMA and de
monstrate how accurately
it performs using simulated datasets. We then identify lineage
A
Stochastic, memoryless
cell state transition map
20
40
40
Fate 1
Fate 2
Multipotent progenitor
B
40
100
100
40
Partially stochastic
cell state transition map
Motif A
Motif B
Multipotent progenitor
Committed
progenitor
20
C
Generating cell types through motifs can limit variation in cel
l type proportions across tissues
Cell type
proportions
100%
100%
33%
67%
50%
50%
42
%
58%
67%
33%
58%
42%
Inaccessible
Inaccessible
Accessible
Moti
f
proportion
100%
100%
75%
25%
75%
25
%
50
%
50%
Tissue #1
#2
#3
#4
#5
2
0%
80%
80%
20%
Figure 1. Cell type proportions can be controlled using partially stochastic cell state transition maps that specify defined groups of cell types
as motifs
(A) A completely stochastic cell state transition map where a multipotent progenitor can self-renew or give rise to different fates in a memoryless ma
nner. Lineage
trees (only triplets shown) generated under this transition map would not exhibit fate correlations between related cells.
(B) A partially stochastic cell state transition map where a multipotent progenitor can self-renew or give rise to different types of committed proge
nitors. The
committed progenitors differentiate, and each gives rise to a defined set of cell types (motif A or B). Lineage trees generated under this transition ma
p would
exhibit fate correlations between related cells, representative of the committed progenitors present within the transition map.
(C) In tissues that specify cell types solely by modulating the frequency of motif A and B, variation in cell type proportions is capped such that a cell t
ype can be at
most twice as abundant as the other type.
ll
OPEN ACCESS
Technology
Developmental Cell
59
, 812–826, March 25, 2024
813
motifs in published zebrafish and rat retina development data-
sets, as well as a dataset of early mouse embryonic develop-
ment. These results reveal spatial and temporal differences in
cell fate determination across d
ifferent progenitors. Further,
the appearance of shared retinal motifs across different spe-
cies suggests that motifs may be evolutionarily conserved fea-
tures of development. Computationally, we explore how
various dataset characteristics affect motif identification.
We demonstrate how the use of lineage motifs facilitates
adaptive variation in retinal cell type composition and show
that this theory is consistent with known variation in vertebrate
retinal cell type proportions. Together, these results support
LMA as a broadly useful tool to understand cell fate
differentiation.
DESIGN
A previous study analyzed sister cell fate correlations by
comparing the frequency of two-cell clones with that pre-
dicted by random association of cell types given their
observed proportions.
24
Another study analyzed triplet fate
correlations by comparing the frequency of triplet patterns
with that observed in simulated lineage trees using a stochas-
tic model where each starting progenitor can self-renew or
differentiate into all possible cell types within the dataset un-
der a set of probabilities.
38
These studies provide evidence
for fate correlations between related cells. However, a frame-
work that can be recursively applied to any lineage tree data-
set to systematically identify lineage motifs of varying size re-
mains lacking.
We first simulated a dataset of lineage trees with two termi-
nal cell fates (
Figure 2
A;
STAR Methods
). We then applied
LMA to analyze the tree dataset, starting by enumerating all
possible doublet and triplet cell fate patterns (with varying
fate composition and order of fate differentiation) and counting
the number of times each occurred within the observed trees
(
Figure 2
B). Then, we compared these counts with those ex-
pected in a ‘‘null’’ model without fate correlations. This can
be done by randomly shuffling the cell fates at the leaves of
the lineage trees to generate resampled trees, followed
by counting the number of times each pattern occurs
across the resampled trees. We then repeat the resampling
process many times. Since the arrangement of cell types
in the resampled trees are randomized, the average of
counts obtained within the null model represents the
expected count if there is no relationship between lineage
and cell fate. To identify larger motifs that span more than
one cell division, the resampling was done in a manner that
preserves the frequency of sub-patterns within each pattern
(
STAR Methods
).
For each pattern, we computed a
Z
score to quantify the de-
gree of over-representation, as well as a false discovery rate
(FDR)-adjusted p value
39
,
40
to measure significance (
STAR
Methods
). In the identified lineage motifs, higher over-represen-
tation can be interpreted as stronger intrinsic commitment of a
given progenitor toward generating a particular fate pattern.
Alternatively, it could represent the strength of extrinsic interac-
tion that generates a particular fate pattern. Finally, anti-motifs,
defined as patterns that are underrepresented in the observed
trees, were identified using the same approach.
LMA is distinct from a related approach termed Kin Correlation
Analysis (KCA). KCA infers cell state transition dynamics
from lineage trees and endpoint cell state datasets but is mainly
applicable to systems governed by Markovian dynamics,
in which sister cell transitions are independent of one
another.
41
,
42
To demonstrate that LMA can recover lineage motifs that
reflect progenitor states in cell state transition maps, we simu-
lated lineage tree datasets using either a competence progres-
sion model (
Figure 3
A) or a binary fate model (
Figure S1
A). We
used differentiation probabilities that generate roughly equal
cell type proportions in the overall dataset (
Figures 3
B and
S1
B). Applying LMA to both datasets, we found that the resulting
motifs reflected the structure of the generative model and
captured multiple levels of progenitor commitment over time.
For example, in trees generated using a competence progres-
sion model (
Figures S2
A–S2C), where cell fates A through F
are generated progressively over time, only symmetric doublet
patterns, such as (F,F), were statistically overrepresented within
all possible doublet patterns (
Figure 3
C).
We next analyzed triplet patterns, in which a single progeni-
tor divides to produce a terminal cell, X, and a second progen-
itor cell that divides once more to produce a doublet of terminal
cells, Y and Z, producing a triplet denoted as (X,(Y,Z)). Only
triplet patterns including two sequential levels of progenitor
commitment, such as (E,(F,F)), were significantly overrepre-
sented (
Figure 3
D).
LMA can be scaled up to analyze larger asymmetric patterns.
Given a reasonable number of trees (500 total), the motifs suc-
cessfully captured up to 5 levels of the competence progression
model. Similar to the triplet results, the significant higher-order
motifs exclusively involved sequentially generated cell fates. As
motif size grows larger, the size of the dataset required for detec-
tion also increases (
Figure 3
E). Together, these results confirm
that LMA can be used to recursively identify lineage motifs in
large patterns.
We also analyzed trees generated using a binary fate model in
which progenitors make binary choices which restrict their fate
potential over time (
Figures S2
D–S2G). The doublet and quartet
motifs reflect the structure of the generative model as expected
(
Figures S1
C and S1D). However, no octet patterns were signif-
icantly over- or underrepresented (
Figures S1
E and S1F). Taken
together, these results indicate that LMA is capable of recur-
sively identifying lineage motifs of multiple sizes in different
models of development and is especially powerful when applied
to the competence progression dynamics, likely due to the lower
number of possible patterns per level of progenitor commitment.
Having demonstrated that LMA can recover motifs in lineage
trees generated using an intrinsic program, we next sought to
demonstrate that LMA could do so in trees generated using
an extrinsic program. More specifically, we considered a
simplified model of the classic developmental mechanism of
lateral inhibition, in which cells of one fate inhibit similar fates
in their neighbors
43–45
(
Figures S3
A–S3C). Our model assumes
a two-dimensional grid of progenitors, which self-renew or
differentiate into two cell fates, A or B, each of which inhibits
differentiation of its neighbors into its own fate. The inhibitory
ll
OPEN ACCESS
Technology
814
Developmental Cell
59
, 812–826, March 25, 2024
effects of multiple neighbors are assumed to combine addi-
tively (
Figure S3
B). As expected, applying LMA to a null dataset
generated without lateral inhibition revealed no significantly
overrepresented doublet patterns (
Figure S3
D). By contrast,
symmetric sister doublets (A,A) and (B,B) were underrepre-
sented in the lateral inhibition model, whereas the asymmetric
sister doublet (A,B) was overrepresented. These results show
that extrinsic developmental programs can generate signatures
of fate correlation on lineage trees, which can be reliably de-
tected using LMA.
To enable the identification of lineage motifs across diverse
developmental contexts, we created a Python package, termed
‘‘linmo.’’ The package is available on a GitHub repository
(
https://github.com/labowitz/linmo
), which includes supporting
documentation and tutorials for processing the following lineage
tree datasets analyzed here.
Lineage tree dataset
Fate 1
Fate 2
Lineage Motif Analysis
A
B
Normalized
count
0.5
0.4
0.5
0.7
z-score = 4.44
adjusted p-value = 0.002
Possible committed progenitor
or extrinsic interaction
Enumerate all possible
cell fate patterns
1
ABC D
E
E
EE
EE
E
E
FGHI
Resample many times
and count occurences
in resampled trees
3
Compare observed count
vs. distribution of null
counts across resamples
4
Identify over-represented
pattern as lineage motif
5
Count occurences
in observed trees
for each pattern
2
0
.
3
0
.2
0.
4
0
.5
0
.
6
0.7
Occurences
Expected count
Count across
resamples
Observed count
Pattern E
Shuffle cell fate labels
while preserving
subpattern frequency
Resample 10,000 times
Figure 2. Lineage Motif Analysis identifies fate correlations in lineage trees by statistical resampling
(A) Lineage trees with two cell types were simulated (
STAR Methods
).
(B) The LMA workflow consists of the following steps. First, all possible cell fate patterns are enumerated. Second, the occurrence of each pattern wit
hin the
observed lineage trees is counted (triplet pattern ‘‘E’’ is shown here as an example). Third, the cell fate labels at the leaves of the trees are randoml
y shuffled to
obtain a resampled set of trees with no fate correlations. This process is then repeated across many resamples. To identify the higher-order motifs th
at span
multiple cell divisions, the shuffling process is done in a manner that preserves sub-pattern frequency (
STAR Methods
). The occurrence of each cell fate pattern is
then counted for each resample. Fourth, the count in the observed lineage trees is compared with the distribution of counts across resamples, whose av
erage is
approximately equal to the expected count if there were no fate correlations. Finally, overrepresented patterns are classified as lineage motifs, wh
ich represent
possible committed progenitors or extrinsic interactions.
ll
OPEN ACCESS
Technology
Developmental Cell
59
, 812–826, March 25, 2024
815
Competence
progression model
20
48.5
31.5
a
A
20
46.5
33.5
b
B
20
44.5
35.5
c
C
20
42.5
37.5
d
D
20
40.5
39.5
e
E
20
80
f
F
Pattern
10
0
10
1
10
2
z-score
F
F
E
F
F
D
E
F
F
C
D
E
F
F
B
C
D
E
F
F
A
B
C
D
E
F
F
Motif significance in simulations
Number of trees
50
500
5000
50000
16.16%
A
17.59% B
18.07% C
16.97% D
14.83% E
16.38% F
Cell type proportions
A
B
C
E
D
z-score = 3
Doublet combinations (top 12 by |z-score|)
0
20
40
z-score
**
**
**
**
**
**
**
**
**
**
**
**
F
F
E
E
D
D
C
C
B
B
A
A
B
F
A
C
B
D
C
F
C
D
B
C
Deviation from resamples
Observed count
Null z-score across 100 resamples
Average null z-score
Triplet combinations (top 12 by |z-score|)
−10
−5
0
5
10
15
20
z-score
***
***
***
***
***
***
***
***
***
***
***
***
E
F
F
D
E
E
C
D
D
B
C
C
A
B
B
A
D
D
A
C
C
B
D
D
D
F
F
A
F
F
C
F
F
B
F
F
Deviation from resamples
Observed count
Null z-score across 100 resamples
A
verage null z-score
Figure 3. Lineage motifs reflect sequential progenitor states in a competence progression model
(A) Lineage trees were simulated using a competence progression model.
(B) Cell type proportions in 500 simulated lineage trees.
(C) Deviation score for doublet patterns. Null
Z
scores were calculated by comparing a random resample dataset with the rest of the resample datasets. 10
datasets containing 500 simulated trees each were used, with the standard deviation across the datasets plotted as error bars (** and *** represent ad
justed
p value < 0.005 and < 0.0005, respectively).
(D) Deviation score for triplet patterns.
(E) Deviation score for select patterns that reflect sequential differentiation of cell types using datasets of varying size. Shading indicates 95% c
onfidence interval
across 10 datasets for each point.
See also
Figures S1
–
S3
.
ll
OPEN ACCESS
Technology
816
Developmental Cell
59
, 812–826, March 25, 2024
RESULTS
LMA reveals spatial organization of zebrafish retina
development
Retina development provides a well-studied example of cell
fate diversification. It involves generation of a conserved set
of terminal cell fates across diverse vertebrate species. At
the same time, it also exhibits substantial spatial and inter-
species variation in cell type proportions,
1
making it an ideal
target tissue for LMA. Therefore, we examined a zebrafish
retina development dataset spanning 32 to 72 h post fertiliza-
tion (hpf),
46
during which progenitors terminally differentiate to
form major neuronal and glial cell types, including ganglion
(G), amacrine (A), bipolar (B), photoreceptor (P), horizontal
(H), and M
€
uller glia (M) (
Figure 4
A). He et al. used time-lapse
confocal microscopy in reporter zebrafish lines to track every
cell division event for 60 retinal progenitors spanning the
nasal-temporal axis. Their data supported previous work
showing that a wave of differentiation starts in the nasal region
and gradually progresses to the temporal region.
47
,
48
Clonal
cell type composition was generally observed to be variable,
with weak fate correlations between related cells. A key
exception, however, was the frequent appearance of symmet-
ric terminal pairs of photoreceptor, bipolar, and horizon-
tal cells.
We sought to identify lineage motifs and characterize how
their frequency varies across spatial regions in the zebrafish
retina. Therefore, we partitioned lineage trees based on the
progenitor spatial location and applied LMA, beginning with
doublet patterns, representing the terminal cell division. We
found that the (H,H), (B,B), and (P,P) doublet patterns had
significantly higher observed counts in the lineage trees,
compared with the distribution of counts across resamples
and expected count, in a similar manner across the three
spatial regions (
Figures 4
B–4D). Therefore, these doublet pat-
terns are statistically overrepresented in the dataset and repre-
sent lineage motifs (
Figure 4
E). The exception was a lack of
(H,H) and (B,B) doublets in the nasal region, likely because
those cell types were only present at very low levels in this re-
gion (
Figures 4
D and 4E). These results were consistent with
key findings from He et al., while extending the analysis to
assess regional variation.
LMA also found motifs not previously identified in the He et al.
study and revealed how their frequency varies across space. For
example, even though amacrine and bipolar cells appear at
similar frequencies across all three retinal regions, the (A,B)
doublet was specifically overrepresented in the nasal region
(
Figures 4
D and 4E). Also, doublets comprising one P cell and
all other cell types were generally underrepresented across all
regions, constituting anti-motifs. We also searched for higher-or-
der motifs that involve multiple cell divisions but found that no
patterns were significantly over- or underrepresented, possibly
due to the limited size of the dataset (
Figure S4
). Overall, the
observed motif profile suggests that amacrine and bipolar cells
frequently share a common progenitor at the terminal cell divi-
sion, specifically in the nasal region of the zebrafish retina,
whereas photoreceptor and non-photoreceptor cells do not
share a common progenitor at the terminal cell division in all
regions.
Shared retinal lineage motifs across species suggest
conservation of cell fate determination
Are retinal lineage motifs conserved between different species?
To address this question, we analyzed a dataset of post-natal rat
retinal progenitor cells grown
in vitro
at clonal density, consisting
of 129 lineage trees with at least 3 cells.
38
During post-natal
development, rat retinal progenitor cells gave rise to mostly
rod cells (R), some bipolar and amacrine cells (respectively,
B and A), and few M
€
uller glia (M) (
Figure 5
A). In this work, the au-
thors showed that a stochastic model based on independent fate
decisions could explain the observed frequencies of most triplet
patterns. However, some triplets may be generated by fate-
committed progenitors that give rise to sets of correlated
cell fates.
Applying LMA to this rat retina dataset confirmed some of
these conclusions, such as over-representation of (B,(A,B)) trip-
lets (
Figures 5
C and 5E). However, it also revealed additional fea-
tures of rat retinal development. For example, using LMA, we
found that (A,B), (B,M), and (A,A) doublets were overrepre-
sented, whereas (B,R) doublets were underrepresented
(
Figures 5
B and 5D). Correcting for sub-pattern frequencies in
the triplet analysis revealed that the apparent over-representa-
tion of the (R,(A,A)) triplet in the previous study
38
could be entirely
explained by the (A,A) doublet motif frequency. This highlights
the importance of the recursive nature of LMA.
Because this dataset excluded two-cell lineages, this could
potentially introduce biases in three-cell motifs. Therefore, we
analyzed cell type proportions in triplets and compared this
with those across all other cells (
Table S1
). We found that there
are no obvious differences in cell type proportions between
triplet and non-triplet populations, suggesting that the lack of
two-cell lineages in the dataset does not substantially bias the
triplet motifs detected here.
We next compared the motif profile between zebrafish and rat
retina. Because the time period analyzed in these datasets is
different and involves the generation of different cell types, we
limited this analysis specifically to cell types that are shared be-
tween the analyzed datasets (i.e., amacrine, bipolar, and M
€
uller
glia). Notably, the (A,B) and (A,A) motifs are observed in both
species, suggesting that the committed progenitors that these
motifs possibly represent are at least partially evolutionarily
conserved. In contrast, the (B,B) motif appears specifically in
the zebrafish retina, whereas the (B,M) motif appears specifically
in the rat retina. Overall, these data suggest that cell fate alloca-
tion in retina across species can occur in a biased and evolution-
arily conserved manner, in which amacrine and bipolar cells
share a common progenitor at the terminal cell division. At the
same time, other aspects of cell fate differentiation may be
more species-specific. For example, bipolar and M
€
uller glia
tend to share a common progenitor in rat, but not zebrafish,
retina at the terminal cell division. More generally, these results
provide a case example for how LMA can be used to assess
the evolution of cell fate differentiation.
Computational simulations reveal how various dataset
characteristics affect motif identification
To what degree the limited size of available lineage tree datasets,
coupled with sampling variation, affects the accuracy of motif
identification is unclear. Therefore, we simulated lineage tree
ll
OPEN ACCESS
Technology
Developmental Cell
59
, 812–826, March 25, 2024
817
A
Outer nuclear laye
r
Inner nuclear layer
Ganglion cell layer
Bipolar (B)
Horizontal (H)
Amacrine (A)
Ganglion (G)
Bip
p
Bi
p
(
n
e
n
(
(
n
e
n
Ga
G
Ga
G
n
n
n
an
an
n
)
)
)
ontal (H)
z
zontal (H)
(H)
ntal
on
z
)
(
Ho
m
(
Ho
m
zo
r
zo
cr
(H
A
(H
A
e (A)
()
(A)
)
e
e
e
Photoreceptor (P)
Müller glia
(M)
Pho
Pho
ept
ept
or (
or (
)
)
ore
ore
Zebrafish
72 hours
post-fertilization
Retina
Side view
Temporal
Middle
Nasal
Front view
RL
B
E
C
D
Doublet patterns
(top 13 by |z-score| across all regions)
−4
−2
0
2
4
6
z-score
H
H
B
B
A
B
P
P
A
A
M
P
B
G
A
G
B
H
H
P
A
P
G
P
B
P
Deviation from resamples
No deviation
Temporal region
Middle region
Nasal region
Not determined
Doublet patterns
(top 13 by |z-score| across all regions)
0
5
10
15
20
25
Counts
**
**
**
*
*
*
*
H
H
B
B
A
B
P
P
A
A
M
P
B
G
A
G
B
H
H
P
A
P
G
P
B
P
Temporal region doublet frequency
Doublet patterns
(top 13 by |z-score| across all regions)
0
5
10
15
20
25
Counts
**
**
*
H
H
B
B
A
B
P
P
A
A
M
P
B
G
A
G
B
H
H
P
A
P
G
P
B
P
Nasal region doublet frequency
Doublet patterns
(top 13 by |z-score| across all regions)
0
5
10
15
Counts
*
*
**
H
H
B
B
A
B
P
P
A
A
M
P
B
G
A
G
B
H
H
P
A
P
G
P
B
P
Middle region doublet frequency
Observed count
Count across resamples
Expected count
Observed count
Count across resamples
Expected count
Observed count
Count across resamples
Expected count
Figure 4. Doublet lineage motifs in zebrafish retina development show spatial organization of fate commitment
(A) Schematic of cell type organization in the zebrafish retina.
(B) Counts for doublet patterns in the observed zebrafish retina trees from He et al.
46
in the temporal region and across 10,000 resamples (* and ** represent
adjusted p value < 0.05 and < 0.005, respectively). The expected count was calculated analytically (
STAR Methods
).
(C) Counts for doublet patterns in the middle region of zebrafish retina and across 10,000 resamples.
(D) Counts for doublet patterns in the nasal region of zebrafish retina and across 10,000 resamples.
(E) Deviation score for doublet patterns in the temporal, middle, and nasal region. Doublet patterns with an observed and expected count of 0 were omit
ted from
the analysis.
See also
Figure S4
.
ll
OPEN ACCESS
Technology
818
Developmental Cell
59
, 812–826, March 25, 2024