Introduction

To generate organs in a developing embryo, cells progressively become more specialized as they differentiate. Cell differentiation needs to be tightly regulated to produce the correct cell types at the right developmental time. Impairment of the temporal sequence of differentiation can have detrimental consequences for organismal development, including incorrect organ size or cellular arrangements (Manto & Jissendi, 2012; Tonchev et al, 2016). It is thus important to unveil the factors that ensure the production of the right cell types with the correct timing. This is particularly true for the formation of the central nervous system (CNS), where timely emergence of the different neurons is an important step to later ensure correct neuronal connectivity and the formation of functioning neuronal networks (Macagno, 1978; Tomassy et al, 2010; Tonchev et al, 2016). Unsurprisingly, changes in the timing of neuronal specification and differentiation can impair brain formation. This in turn can lead to severe cognitive deficits including mental retardation and impairments in motor coordination (Guarnieri et al, 2018). Nevertheless, particularly in vertebrates, the factors involved in different neuronal fate decisions during development are not yet fully revealed. This is different for the Drosophila nervous system, where the temporal regulation of fate decisions has been more thoroughly explored, for example, in the CNS and the optic lobe (Brody & Odenwald, 2000; Isshiki et al, 2001; Grosskortenhaus et al, 2006; Chia et al, 2008). In these areas, a defined sequence of progenitor divisions leads to the formation of the different neurons in a consecutive manner. This sequence arises during development as multipotent progenitors progressively acquire and lose the competence to form different neuronal types through the sequential expression of defined transcription factors (Doe & Technau, 1993; Campos‐Ortega, 1995; Brody & Odenwald, 2000; Cleary & Doe, 2006; Homem & Knoblich, 2012; Li et al, 2013).

In the vertebrate CNS, it is currently less understood how single progenitor competence changes during development to give rise to different neuronal cell types at the right time and in the right proportions. So far, some progress has been made to understand neuronal birth orders in the neocortex, spinal cord, olfactory bulb and retina (Holt et al, 1988; Livesey & Cepko, 2001; Butt et al, 2005; Batista‐Brito et al, 2008; Kao & Lee, 2010; Kohwi & Doe, 2013) but these studies have been mostly performed at the clonal or the population level. This eventually led to different interpretations on whether the outcome of progenitor divisions is pre‐determined by fixed competence windows during development, resulting in stereotypic and fully predictable lineages (Livesey & Cepko, 2001; Wong & Rapaport, 2009; Cepko, 2014), or whether it is variable and influenced by stochastic processes acting on fate decision mechanisms, resulting in not fully predictable lineages (Gomes et al, 2011; He et al, 2012; Boije et al, 2015; Llorca et al, 2019).

To better understand the variability and stereotypicity of fate decisions in the vertebrate CNS, a quantitative appreciation of the predictability of the fate outcome of progenitor divisions and its variability over development is needed. To date, this has been challenging as a plethora of different neuronal and progenitor cell types inhabit the different vertebrate brain areas, of which many parts are not easily accessible for experimental manipulation (Paridaen & Huttner, 2014). An attractive system to circumvent these issues is the developing retina, the part of the CNS responsible for light collection and transmission. It is populated by only five major neuronal types that are clearly distinguishable by their final position, morphology and mode of migration (Hoon et al, 2014; Amini et al, 2018). In different systems including zebrafish and Xenopus, it has been shown that these neurons arise from a homogeneous pool of multipotent progenitor cells (Holt et al, 1988; Wetts & Fraser, 1988; Fadool, 2001). However, so far, the sequence of divisions and fate decisions that leads to the emergence of the correct neuronal types at the right time remained elusive. To shed light on this question, several studies used the imaging potential of the zebrafish embryo to follow the exact sequence and outcome of progenitor divisions over time and the lineage relationships between cell types in a quantitative manner (Poggi et al, 2005; He et al, 2012; Nerli et al, 2020; Engerer et al, 2021). Indeed, it has been shown that multipotent progenitors divide asymmetrically to produce the first neurogenic Atoh7‐expressing progenitors (Nerli et al, 2020). Furthermore, many transcription factors that specify different neuronal fates are well described (Table 1) and, by interfering with their expression, it is possible to manipulate neuronal emergence (Almeida et al, 2014). However, even in the zebrafish, it is not yet clear whether deterministic lineage patterns exist among the general stochasticity in fate decisions (He et al, 2012; Boije et al, 2015). A fate decision x is defined as deterministic when it occurs with probability P(x) = 1, which is only rarely the case, for example, in the Caenorhabditis elegans embryonic lineages (Sulston et al, 1983). If the P(x) < 1 for each possible fate outcome, then stochastic elements are involved (Zechner et al, 2020). As in the mathematical sense, the terms probabilistic and stochastic are synonyms, we will mainly use the term probabilistic in this work. For further definitions, we refer to Zechner et al (2020).

Table 1 Transcription factors specifying zebrafish retinal neuronal fates.
Full size table

To understand which lineage patterns lead to neuronal formation, we here explore the probabilities and the constraints of different fate decisions of zebrafish retinal progenitors by investigating the emergence of a specific neurogenic lineage. Interestingly, we find that neurogenic Atoh7‐expressing progenitors give rise to a photoreceptor precursor with a probability of 1 and a sister cell of different fate. While photoreceptor production is deterministic, the fate of the sister cell is acquired with different, time‐dependent probabilities. If the deterministic photoreceptor fate decision is impaired, neurogenic progenitors do not generate a different known neuronal fate, resulting in lineage and tissue impairments. In contrast, interference with the probabilistic branch of the lineage resulted in changes in the probabilities of generating different fates and maintenance of structural tissue integrity. Stochastic modelling of these changed probabilities revealed a simple gene regulatory network (GRN) that faithfully predicts the proportions and timing of fate decisions during wild‐type retinal neurogenesis.

Results

Neurogenic Atoh7+ progenitors produce one photoreceptor precursor and a sister cell of variable and different fate

To investigate the possible fate decisions of neurogenic progenitor in the zebrafish retina, we analysed the neurogenic lineage expressing the pro‐neural transcription factor Atoh7 (atonal bHLH transcription factor 7, also called Ath5). Atoh7‐positive (Atoh7+) progenitors arise from asymmetric divisions of multipotent progenitors that also produce an Atoh7‐negative (Atoh7−) progenitor (Nerli et al, 2020; Fig 1A). Previous work on few Atoh7+ progenitor divisions showed that these divisions give rise to a retinal ganglion cell (RGC) and an apical sister cell (Poggi et al, 2005; Fig 1A). However, as most neurons in the retina express Atoh7 (Vitorino et al, 2009; Jusuf et al, 2011; preprint: Rocha‐Martins et al, 2021), we aimed to develop a more exhaustive picture of division outcome distribution. To gain a quantitative understanding of which neurons arise from Atoh7+ progenitors and in what proportions at different developmental stages, we analysed fate distributions using Atoh7‐driven reporter construct to mosaically label these progenitors (atoh7:GFP‐CAAX: of note, while we are aware that the original name for this construct is ath5:GFP‐CAAX, we use the other nomenclature for consistency; Zolessi et al, 2006; Icha et al, 2016a). We followed the Atoh7+ progenitor division modes and fate outcomes using long‐term light sheet imaging (Icha et al, 2016b) performed between the onset of neurogenesis at 28 h post‐fertilization (hpf) and 60 hpf, the time when most progenitors have entered neurogenesis (Hu & Easter, 1999; Schmitt & Dowling, 1999; Martinez‐Morales et al, 2005). Sister cell fate could be unambiguously assigned due to the distinct migration modes of the different emerging neurons (Appendix Figs S1 and S2), their morphology during and after migration and their final position within the tissue (Fig 1B, see Materials and Methods for details). We also performed live imaging of the Spectrum of Fates line (Almeida et al, 2014) in which different cell types can be identified with a unique combination of fate markers: Atoh7 only for RGCs, Atoh7 and Ptf1a for amacrine cells and horizontal cells (here referred to as inhibitory neurons, INs), Atoh7 and Crx for PRs and Crx only for bipolar cells.

Figure 1
figure 1

Atoh7+ progenitors divide asymmetrically to produce a PRpr and a sister cell

A. Schematic of the Atoh7+ lineage and open questions in fate decisions.

B. Cell fate assignment strategy based on (left) mode of migration and (right) morphology and final neuronal position in the laminated retina. Plot on the left indicates typical movement trajectories for each cell type.

C. Montage of Atoh7+ progenitor division generating an RGC (magenta dot) and a PRpr (cyan dot). Dashed line indicates the apical side and arrowhead points to RGC axon. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm.

D. Montage of Atoh7+ progenitor division generating an HC (orange dot) and a PRpr (cyan dot). Dashed line indicates the apical side. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm.

E. Montage of Atoh7+ progenitor division generating an AC (yellow dot) and a PRpr (cyan dot). Dashed line indicates the apical side and arrowhead points to basal dendrites. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm.

F. Photoconversion experiment. A 405 laser was used to photoconvert isolated cells labelled with Crx:H2BDendra at 42 hpf. Twenty‐four hours after photoconversion, photoconverted PRs were assessed. (Top) Schematic of the experiment and (bottom) representative images of the different steps of the experiment. Crx:H2BDendra (green) and photoconverted Crx:H2BDendra (magenta). Scale bar 10 μm, white line indicates apical side of the retina.

G. Schematics of the Atoh7+ lineage showing possible outcomes of Atoh7+ divisions and PRpr divisions.

H. Distribution of fates for the sister cell of the PRpr acquired during early and late neurogenic windows. Amacrine cells and horizontal cells are pooled into the Inhibitory neurons (IN) category. N = 13 embryos, n = 96 Atoh7+ divisions. Mean and 95% CI are indicated.

I. Event plot of all divisions analysed in (H).

Source data are available online for this figure.

Full size image

Analysis of 96 divisions in two neurogenic windows, between 28 and 42 hpf (from here on referred to as “early,” N = 9 embryos) and between 36 and 60 hpf (from here on referred to as “late,” N = 4 embryos), in mosaically injected embryos revealed that Atoh7+ progenitors divide asymmetrically and reproducibly produce one cell with columnar morphology and a sister cell which was an RGC or an amacrine cell (AC) or a horizontal cell (HC; of note, when by referring to “horizontal cell,” this can be a “horizontal cell” as well as a “horizontal cell precursor” (Godinho et al, 2007; Amini et al, 2019) as not all imaging experiments ran long enough to catch the final division leading to two HC). We never observed a division giving rise to bipolar cells (BCs, Figs 1C–E and EV1A–F, Movie EV1). This finding was confirmed with live imaging of the SoFA line (n = 21 divisions, 2 embryos), as Atoh7+ divisions always produced an Atoh7+, Crx+ cell (PR) and a sister cell that was either Atoh7+ only (RGC; Fig EV1B) or Atoh7+, Ptf1a+ (IN; Fig EV1D and F).

Apically positioned cells with columnar morphology were previously suggested to be immature photoreceptors, so‐called photoreceptor cell precursors (PRprs; Poggi et al, 2005; Icha et al, 2016a; preprint: Rocha‐Martins et al, 2021). These cells were characterized by the expression of crx (cone‐rod homeobox; Suzuki et al, 2013; Weber et al, 2014) and presumably divide symmetrically to produce two PRs (Suzuki et al, 2013; Weber et al, 2014). However, whether all PRs result from such committed precursor divisions remained unclear. We found that the majority of Crx+ cells at the apical side incorporated EdU at 48 hpf, indicating that most PRprs are cycling at this developmental stage (Fig EV1G). To test whether indeed all PRprs undergo an additional division, we performed photoconversion experiments using a Crx:H2B‐Dendra construct at 42 hpf as shown in Fig 1F. Twenty‐four hours after photoconversion, 18/19 PRprs (N = 16 embryos) divided (Fig 1F), while 1/19 did not divide. In 16 of these 18 PRpr divisions, two PRs were produced (Fig 1F and G), while in 2/18 cases, three or four PRs were produced. However, in these cases, we cannot exclude that initially two PRprs were photoconverted. These results indicate that most PRs arise from immature committed precursors (PRprs) that divide once to produce two PRs (Fig 1G).

Atoh7+ progenitor divisions undergo a deterministic and a probabilistic fate decision

Our data show that each Atoh7+ progenitor division gives rise to one PRpr throughout the neurogenic window (Fig 1G). This indicates that the PRpr fate decision is deterministic, as it occurs at each Atoh7+ progenitor division with a probability of 1 (Zechner et al, 2020). The sister cell of the PRpr acquired a different, variable fate. It either became an RGC (Poggi et al, 2005; Zolessi et al, 2006; Icha et al, 2016a; Figs 1C, and EV1A and B, Movie EV1) or an inhibitory Neuron (IN), that is, a horizontal cell (Godinho et al, 2007; Weber et al, 2014; Amini et al, 2019; HC; Figs 1D, and EV1C and D, Movie EV1) or an amacrine cell (Chow et al, 2015; Icha et al, 2016a; Engerer et al, 2021; AC; Figs 1E, and EV1E and F, Movie EV1), but never a second PRpr or a BC. This means that this branch of the lineage entails stochastic elements as fate decisions occur with a probability smaller than 1 (see also Zechner et al, 2020).

We should note that previous studies also reported different clonal compositions namely RGC‐AC (Jusuf et al, 2012) and AC‐BC (Wang et al, 2020; Engerer et al, 2021) divisions. While we cannot exclude that RGC‐AC divisions occur, we did not find such division outcome in the 96 divisions of 13 embryos analysed. As we here concentrate on Atoh7+ progenitors, the AC‐BC divisions would not occur if a BC arises from an Atoh7‐negative lineage (Vitorino et al, 2009, and this work, Fig EV4).

To understand whether and how the fate distribution of the sister cell of the PRpr changed over time, we analysed neuronal fate proportions during the early and late neurogenic windows. In the early window, 80.8% of divisions (confidence interval CI = [53.5%, 100%]) produced an RGC, while 19.2% (CI = [0.0%, 46.5%]) produced an inhibitory neuron that was an AC or HC (Figs EV1H and 1H). As neurogenesis progressed, the proportion of RGCs decreased (42.9%, CI = [36.0%, 53.7%]) while the proportion of inhibitory neurons increased (57.1%, CI = [46.3%, 64.0%]; Fig 1H), with 35.5% of ACs and 21.5% of HCs (Fig EV1H). This showed that the sister cell fate decision is probabilistic as different fates are acquired with different, time‐dependent probabilities. This was also reflected by the live imaging experiments with the SoFa line (n = 21 divisions), from which 62% of divisions generated RGC‐PRpr, 10% generated AC‐PRpr and 28% generated HC‐PRpr.

While it has been suggested that retinal progenitor competence changes during development, whether this change occurs gradually or more abruptly was still debated. To answer this question, we plotted the time of each progenitor division from all acquired time‐lapse experiments (N = 9 embryos for early and N = 4 embryos for late window) and its fate outcome for the PRpr sister cell (from here on referred to as “event plots”). This analysis showed that Atoh7+ progenitors gradually change competence during development (Fig 1I), switching from a prevalent production of RGCs at early stages of neurogenesis to a progressive decrease in their production and an increased generation of INs. This is consistent with the previously suggested overlapping birth order of retinal neurons (Turner & Cepko, 1987; Cepko et al, 1996; Livesey & Cepko, 2001; Cepko, 2014).

We conclude that deterministic and probabilistic fate decisions co‐exist within the same lineage. Atoh7+ progenitor divisions always generate one PRpr throughout the neurogenic window and a sister cell of variable fate. The sister cell can become an RGC or an inhibitory neuron. We never observed a BC or a second PRpr. Furthermore, we find that the probabilities of producing RGCs or INs gradually change over development.

Perturbing the deterministic fate decision affects overall tissue development and generates non‐canonical cell fates

The fact that one PRpr was always produced during the entire neurogenic window made us ask how the outcome of Atoh7+ division would change upon perturbation of this branch of the lineage. To prevent the development of PRprs and consequently PRs, we evaluated target genes resulting from a transcriptomic analysis of 42 hpf retinae, a stage at which photoreceptors are still in their committed precursor state, before terminal division (Fig EV1G). These cells already express genes known to be involved in PR development and maturation as crx, prdm1a and otx2b (Nishida et al, 2003; Shen & Raymond, 2004; Wang & Harris, 2005; Brzezinski et al, 2010; Emerson et al, 2013; Ghinia Tegla et al, 2020; Fig EV2A and B, reference GSE194158 in NCBI). Differential gene expression analysis revealed that only Prdm1a and Crx were significantly enriched in the PRpr population when comparing PRpr to their sister RGC (Fig EV2C). While crx is linked to PR differentiation and its knockdown leads to PR degeneration (Furukawa et al, 1999; Shen & Raymond, 2004), prdm1a is involved in PR fate specification (Brzezinski et al, 2010; Goodson et al, 2020). A mouse knockout of prdm1a led to a reduction in the number of PRs (Brzezinski et al, 2013), without severe consequences on tissue development. Further, our transcriptomics analysis showed that Prdm1a expression is specific to PRprs (Fig EV2D). Taking these findings into account, we decided to use a previously established Prdm1a morpholino knockdown approach to interfere specifically with the emergence of PRprs (Lee & Roy, 2006; Liu et al, 2012; see Materials and Methods). While control embryos at 48 hpf feature a layer of Atoh7+ Crx+ photoreceptors at apical positions, this layer is missing in most of the Prdm1a morphants (Fig 2A and A′). At 72 hpf, a layer started to form at the outer nuclear layer (ONL) location, but this layer was mostly occupied by Crx‐negative and Zpr1‐negative cells, confirming a significant reduction in the production of red and green cone PRs (Fig 2B and B′, N = 10 embryos). In extreme cases, we found a total depletion of this cell layer (3/10 embryos).

Figure 2
figure 2

Perturbing the deterministic PRpr fate decision affects lineage and overall tissue development

A. Two examples of retinas at 48 hpf in (left) control and (right) Prdm1a morphant (MO). Atoh7+ cells (magenta), inhibitory neurons (yellow) and photoreceptors (cyan) are labelled. Scale bar 50 μm. (A′) Close‐up of Crx (cyan), signal (upper panel) and DAPI (grey, lower panel) from (A) for controls (left) and Prdm1a morphant (right). Scale bar 20 μm.

B. Staining for the photoreceptor cell marker zpr‐1 at 72 hpf in (left) control and (right) Prdm1a knockdown. Atoh7+ cells (magenta), inhibitory neurons (yellow), photoreceptors (cyan) and zpr‐1 (grey). Scale bar 50 μm. Arrowheads indicate zpr‐1 staining. (B′) Close‐up of Atoh7 (magenta) and Crx (cyan) signal (upper panel), together with Zpr‐1 (grey, lower panel) from (B), for controls (left) and Prdm1a morphant (right). Scale bar 20 μm.

C. Schematics of layer thickness and retinal diameter measurements in the central part of the retina.

D. Measurements of retinal diameter in control (black dots) and Prdm1a knockdown (empty dots) at 24, 36, 48 and 72 hpf. N and P‐values are found in Table 2. **** for P < 0.0001, Two‐way ANOVA with Bonferroni correction. Mean and single values are indicated.

E. Layer thickness analysis in control and Prdm1a knockdown embryos at 72 hpf. N = 4 embryos (control) and 6 embryos (Prdm1a morphant). Total thickness comparison: P = 0.0055; GCL comparison, P < 0.0001; INL comparison, ns; ONL comparison, P = 0.0369. Two‐way ANOVA with Bonferroni correction. Mean and SD are indicated, as well as single values.

F. Montage of Atoh7+ progenitor division upon Prdm1a knockdown, generating an RGC (magenta dot) and a non‐canonical sister cell (violet dot). Dashed line indicates the apical side and arrows indicate the dynamic basal process of the sister cell. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm. Arrowheads point to the dynamic basal process of the non‐canonical sister cell.

G. Schematic comparison of the outcome of Atoh7+ progenitors in control and Prdm1a morphants.

Source data are available online for this figure.

Full size image

Prdm1a morphant retinas showed very similar size to controls at 24 and 36 hpf, stages at which the retina is mostly populated by progenitor cells (Fig 2C and D, two independent experiments, N and P‐values in Table 2).

Table 2 Retinal diameter measurements for control and Prdm1a‐MO conditions at 24, 36, 48 and 72 hpf.
Full size table

However, at 48 hpf, the morphant retinas started to be smaller than controls (Fig 2A and D, N = 9 embryos for control and Prd1ma MO) and at 72 hpf showed severe microphthalmia (Fig 2B and D, N = 11 embryos control and N = 14 embryos Prdm1a MO). Further, an overall reduction in retinal thickness, mostly due to shrinkage of the ONL and of the ganglion cell layer (GCL), was observed (Fig 2C and E, N = 4 embryos (control) and 6 embryos (Prdm1a morphant)). As Prdm1a is specifically expressed in the PRpr population and not in progenitor cells (Fig EV2D), this is likely a consequence of abnormal PRpr emergence. To test whether apoptosis could account for the shrinkage of especially the GCL, we performed active caspase‐3 staining at 72 hpf in embryos injected with control or Prdm1a morpholino. This revealed that similarly to control embryos, a small amount of dying cells were present in all the different retinal layers, without a bias towards the GCL (Fig EV2E and F, N = 6 embryos per condition, 1 experiment). Therefore, cell death itself was most likely not responsible for the shrinkage of the RGC layer. It is possible that the shrinkage of the RGC layer is due to the fact that Prdm1a knockdown reduces the overall amount of Atoh7+ progenitors, and therefore the number of Atoh7+ neurons, without affecting the production of Atoh7− neurons (BCs and part of ACs).

Together, these results show that overall tissue architecture is compromised upon inhibition of PRpr emergence.

We next set out to understand how the outcome of Atoh7+ progenitor divisions changed upon inhibition of PRpr emergence. Possibilities included a fate switch to other neuronal fates, differentiation into a different cell type or progenitors skipping this division completely and directly generating one neuron. Twenty Atoh7+ progenitor divisions in three embryos were followed in the early neurogenic window in Prdm1a morphants as established in controls. This revealed that in 2/20 cases Atoh7+ cells did not divide but differentiated directly, once generating an RGC (Fig EV2G) and once generating an inhibitory neuron (AC, Fig EV2H). One division produced an RGC and a PRpr as seen in controls. Most divisions (17/20), however, generated an RGC or an inhibitory neuron as seen in controls, and a sister cell that initially showed PRpr‐like unipolar morphology during basal migration (Fig 2F and Movie EV2; preprint: Rocha‐Martins et al, 2021). Afterwards, this cell extruded a dynamic basal process and positioned itself apically (Fig 2F and G), a phenomenon never observed in wild‐type embryos. Upon > 15 h of imaging, these cells did not acquire any previously observed neuronal morphology and their basal process did not establish a basal attachment (Fig 2F, compared with Fig 1B–E).

Thus, interference with the deterministic part of the lineage resulted in lineage and tissue morphology defects, possibly due to the production of a non‐canonical sister cell of an unknown state from Atoh7+ progenitor divisions.

Lineage topology is not affected by interference with the probabilistic lineage branch

Interference with the emergence of PRprs led to severe defects in lineage progression. To understand whether similar constraints in progenitor competence and potency occurred in the probabilistic branch of the lineage, we used knockdown approaches to suppress the emergence of RGCs, inhibitory neurons or both. We used established morpholinos against the two bHLH pro‐neural transcription factors Atoh7 (prevents RGC fate specification (Pittman et al, 2008)) and Ptf1a (prevents amacrine and horizontal cell fates (Jusuf et al, 2011); Table 1). Despite the absence of one or two retinal populations, retinal thickness did not significantly change in Atoh7 and Ptf1a morphants (Fig 3A and B, and Table 3). This is consistent with the minimal defects previously observed at the tissue level in Atoh7 and Ptf1a morphants (Randlett et al, 2013; Almeida et al, 2014). Analysis of the thickness of different layers showed that the observed conserved retinal size could result from size changes at the level of the different neuronal layers: in Atoh7 morphants, a significant reduction in GCL thickness was counteracted by a thickness increase in the inner nuclear layer (INL) and ONL (mean, SD and P‐values for the different conditions in Table 3); and in Ptf1a morphants, the absence of ACs and HCs led to a reduction in the thickness of the INL but showed increased GCL and ONL thickness (Fig 3B and Table 3).

Figure 3
figure 3

Probabilistic lineage branch shows flexibility upon interference with fate probabilities

A. Retina at 72 hpf in (top left) control, (top right) Atoh7 knockdown, (bottom left) Ptf1a knockdown and (bottom right) Atoh7 and Ptf1a knockdown. Atoh7+ cells (magenta), inhibitory neurons (yellow) and photoreceptors (cyan). Scale bar 50 μm, 20 μm in close‐up panels.

B. Layer thickness analysis in control and morphant embryos measured at 72 hpf. N = 9 embryos (control), 7 embryos (Atoh7 morphant), 7 embryos (Ptf1a morphant) and 7 embryos (Atoh7 + Ptf1a morphant). P‐values for thickness measurements are found in Table 3. Mixed‐effects analysis with Bonferroni correction. Mean and SD are indicated, as well as single values.

C. Number of PH3+ cells per retina in control and morphant conditions at 28, 32 and 36 hpf. N = 4 to 10 embryos per condition, mixed‐effects analysis with Dunnett's correction. All comparisons are statistically non‐significant. Mean and SD are indicated, as well as single values.

D. Schematic comparison of the outcome of Atoh7+ progenitor divisions in control and morphants.

E. Proportions of PRpr sister cell fates during early neurogenesis. Mean and 95% CI are indicated. For Atoh7 morphants, n = 48 divisions and N = 4 embryos. For Ptf1a morphants, n = 39 divisions and N = 4 embryos. For Atoh7 + Ptf1a morphants, n = 62 divisions and N = 4 embryos.

F. Proportions of PRpr sister cell fates during late neurogenesis. Mean and 95% CI are indicated. For Atoh7 morphants, n = 47 divisions and N = 4 embryos. For Ptf1a morphants, n = 41 divisions and N = 5 embryos. For Atoh7 + Ptf1a morphants, n = 50 divisions and N = 4 embryos.

G. Montage of neurogenic progenitor division upon Atoh7 and Ptf1a knockdown, generating two PRprs (cyan dots). Dashed line indicates the apical side. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm.

H. Montage of neurogenic progenitor division upon Atoh7 and Ptf1a knockdown, generating a BC (blue dot) and a PRpr (cyan dot). Dashed line labels the apical side, yellow arrow points at BC apical process and white arrows point at BC basal process. atoh7:GFP‐CAAX (Atoh7, grey). Scale bar 10 μm.

Source data are available online for this figure.

Full size image
Table 3 Layer thickness measurements for control, Atoh7‐MO, Ptf1a‐MO and Atoh7 + Ptf1a‐MO conditions.
Full size table

Even in double Atoh7/Ptf1a morphants, in which RGCs, ACs and HCs were missing, only a minor reduction in retinal thickness was observed, mainly due to shrinkage of the GCL (Fig 3B and Table 3), as previously proposed (Randlett et al, 2013). This showed that upon interference with the probabilistic branch of the linage, retinal thickness and lamination were generally maintained even in the absence of one or more neuronal cell types. This apparent robustness in tissue thickness is accompanied by changes in the thickness of single layers, suggesting changes in the proportions of neuronal output of progenitor divisions.

One previously suggested possibility that could explain the changes in neuronal proportions was that, in the absence of pro‐neural factors such as Atoh7, progenitors would go through an additional round of cell division to produce later‐born neurons (Kay et al, 2001; He et al, 2012; Boije et al, 2015). If this was the case, a higher number of progenitors dividing at the onset of neurogenesis would be expected. This would lead to changes in the number of progenitor divisions and the lineage branching structure, here referred to as lineage topology. To understand whether lineage topology changed upon knockdown of the pro‐neural transcription factors Atoh7 and Ptf1a, we used the mitotic marker PH3 and assessed the number of PH3+ cells at 28, 32 and 36 hpf in single Atoh7 and Ptf1a morphants and Atoh7/Ptf1a double morphants. Interestingly, no major difference in the number of mitotic progenitors was observed in the four conditions over development (Fig 3C).

We then followed single Atoh7+ progenitors divisions in the three morphant conditions (the Atoh7 morpholino blocks only the expression of the protein, not the reporter) (n = 290 divisions, N = 24 embryos) and found that these divisions never produced an additional Atoh7+ progenitor (Fig 3D), indicating that the absence of the pro‐neural factors Atoh7 and Ptf1a did not delay neuronal production. Further, plotting the developmental time of each neurogenic division showed that the temporal distribution of neurogenic divisions was similar in control and single morphant conditions, with only a slight delay in the double morphants (Fig EV5A). This suggested that knockdown of pro‐neural genes neither affects the topology of progenitor lineages nor the timing of neurogenic entry.

Further, we found that a higher number of committed precursors cells were produced upon knockdown of Atoh7, Ptf1a or both (see below Fig 3D–F). This is consistent with the increase in BrdU‐positive cells previously reported in the Atoh7 morphant condition after the onset of neurogenesis at 38 hpf (Kay et al, 2001). A higher number of committed precursors in the different knockdown conditions (see below) would indeed increase BrdU incorporation as these cells have relatively long cell cycles, with an average of 10.5 h for PRpr and 13 h for HCpr (Amini et al, 2019), compared to 5 h for progenitor cells.

Neurogenic progenitors restrict competence without losing potency

To understand how the outcome of Atoh7+ divisions changed upon interference with the likelihoods of different neuronal fates, we analysed the possibilities and constraints of fate decisions of what would usually be Atoh7+ neurogenic progenitors in different morphants (as in two conditions the Atoh7 protein is not produced, we in this context refer to these as “neurogenic progenitors”). Interestingly, independently of the morphant condition (n = 290 divisions, N = 24 embryos), all neurogenic progenitor divisions produced one PRpr and a neuronal sister cell (Fig 3D), as seen in controls. Thus, while the deterministic PRpr fate remained unchanged, the sister cell acquired one of the other available fates with different probabilities (Fig 3D–F). This is consistent with the previously proposed “fate switch” hypothesis (Jusuf et al, 2011; Almeida et al, 2014). Differently from what was seen in controls, however, symmetric divisions producing two PRpr (and consequently four PRs) occurred in all morphant conditions (Fig 3D and G, Movie EV3). This could explain the increased thickness of the PR layer observed in all morphants (Fig 3B). Furthermore, in double morphants in which neither RGC nor IN fates were available, divisions producing one PRpr and one BC (Fig 3D and H, Movie EV3) appeared. As BCs usually do not emerge from these neurogenic progenitor divisions (Fig 1G), these results indicated that neurogenic progenitors generally have the potency to produce BCs but are not competent to generate them when Atoh7 or Ptf1a are expressed. This, together with the broad spectrum of possibilities for early neurogenic progenitor fate decisions, indicated that these progenitors remain multipotent throughout lineage progression, while their competence is restricted by the expression of Atoh7 and Ptf1a.

In wild‐type, the likelihood that the probabilistic branch of Atoh7+ progenitor divisions acquired certain neuronal fates changed over development. RGCs were more likely produced at early stages while inhibitory neurons became more prevalent later (Fig 1H and I). To understand whether and to what extent these proportions were affected in the different morphant conditions, the outcome of neurogenic progenitor divisions was analysed over time. In Atoh7 morphants (n = 95 divisions, N = 8 embryos), we observed an increase in the proportions of division giving rise to IN, from 19.2% in control to 41.4% ((CI = [15.9%, 56.8%]), Fig 3E) at early developmental stages. In addition, and unobserved in controls, divisions giving rise to a second PRpr (58.586%, CI = [43.2%, 84.1%], Fig 3E and G) appeared. During later neurogenesis stages, the proportions of divisions producing an IN remained invariant compared to controls (57.1% in control vs. 60.6%, with CI = [42.0%, 78.8%] in Atoh7 morphant, Fig 3F), but the proportions of divisions that would have produced RGCs in controls, now produced a second PRpr (38.1%, with CI = [21.2%, 53.2%]; Fig 3F).

In Ptf1a morphants, (n = 80 divisions, N = 9 embryos), the proportion of divisions that would have produced an IN in controls produced a second PRpr (12.6% with CI = [1.9%, 19.7%]) at early stages (Fig 3E), while the production of RGCs remained invariant compared to controls (80.8% in control vs. 87.4% with CI = [80.3%, 98.1%] in Ptf1a morphant, Fig 3E). However, at later stages, the proportions of RGCs increased with respect to controls (from 42.9 to 77.7% with CI = [67.2%, 96.1%]), indicating that the window of RGC formation is extended in this condition (Fig 3F). This prolonged production of RGCs could also explain the observed increase in RGC layer thickness (Fig 3B).

In double morphants (n = 112 divisions, N = 8 embryos), the increase in PRpr production (Fig 3D–F) led to an ONL with slightly different morphology at 72 hpf (Fig EV3A–B′). Cells were still attached to the apical side but nuclei were more densely packed into the tissue with respect to control (Fig EV3C and D), giving the layer an almost pseudostratified appearance. At later stages (84 hpf), PRs were more elongated than in controls, but aligned in the ONL (Fig EV3E). Despite the different morphology, photoreceptors differentiation seemed unaltered as shown by the expression of the zpr‐1 marker (Fig EV3E). Furthermore, bipolar cells were produced throughout the entire neurogenic window (Fig 3F and H). This is interesting for two reasons: (i) they were never produced by the Atoh7 lineage in control embryos (Fig 1G) and (ii) they usually start emerging only at later developmental stages from 45 hpf onwards (Weber et al, 2014; Engerer et al, 2017). Previous studies hypothesized that in zebrafish, BCs arise from a different lineage that does not express Atoh7 (Vitorino et al, 2009). Analysis of a double transgenic line expressing Atoh7 and Vsx1 (a BC marker (Vitorino et al, 2009; Weber et al, 2014)) confirmed that BCs do not express Atoh7 in wild‐types (Fig EV4A and B). As we previously showed that asymmetric divisions that produce one Atoh7+ progenitor also produce one Atoh7‐negative progenitor (Atoh7−, Figs 1A and EV4C; Nerli et al, 2020), we asked whether BCs could arise from this Atoh7− sister progenitor lineage. Notch inhibition using the gamma‐secretase inhibitor LY411575 starting at 24 hpf is known to induce symmetric divisions that produce two Atoh7+ progenitors (Fig EV4C), depleting the pool of Atoh7− sister progenitors (Nerli et al, 2020). Thus, in this condition, less bipolar cells should emerge. Inhibiting Notch before neurogenesis onset at 24 hpf indeed resulted in a dramatic reduction in Vsx1+ BCs and an increase in Atoh7+ neurons in the BC layer (Fig EV4D). To exclude that Notch inhibition affects BC fate per se, Notch was inhibited from 45 hpf onwards, a stage at which BC precursors already emerged (Weber et al, 2014; Engerer et al, 2017). This treatment did not affect the BC population (Fig EV4D, bottom panel), showing that BC fate decisions were not generally impaired upon Notch inhibition. Thus, BCs indeed seem to arise from the Atoh7‐negative sister cell in the wild‐type (Fig EV4E) and therefore from a different, neurogenic lineage.

Figure 4
figure 4

A simple theoretical gene regulatory network reproduces temporal changes in progenitor competence

A. Distribution of each Atoh7+ division in time (event plots) from 28 hpf in Atoh7 morphants, Ptf1a morphants and double Atoh7/Ptf1a morphants. Coloured violin plots indicate different fates and grey violin plots show all neurogenic divisions analysed. Time in hpf.

B. Probability estimation from lineage data in (A) for each fate outcome at the time at which an Atoh7+ division occurs in Atoh7 morphants, Ptf1a morphants and double Atoh7/Ptf1a morphants. Time in hpf. Mean (dark line) and 95% confidence interval (thick transparent stripe) are plotted.

C. Simulation scenario A run using a simple GRN tested based on the hypothesis formulated based on the probability estimation plots in (B). Mean (dark line) and 95% confidence interval (thick transparent stripe) are plotted.

D. Simulation scenario B run using a simple GRN in which inhibitory effects of two or more TFs are summed (green lines), and tested based on the hypothesis formulated based on the probability estimation plots in (B). Mean (dark line) and 95% confidence interval (thick transparent stripe) are plotted.

E. Probability estimation from lineage data in Fig 1I for each fate outcome at the time at which an Atoh7+ division occurs in controls. Time in hpf. Mean (dark line) and 95% confidence interval (thick transparent stripe) are plotted.

F. Prediction of fate shares in the control situation using Simulation Scenario A (left) and Simulation Scenario B (right). Mean (dark line) and 95% confidence interval (thick transparent stripe) are plotted.

Source data are available online for this figure.

Full size image

Together, these results indicated that neurogenic progenitor competence and potency are extended in morphant conditions when compared to controls. Thus, these neurogenic progenitors can in principle generate all retinal neuronal fates and these fates can be produced even outside their canonical temporal window. These experiments also reinforced the finding that deterministic and probabilistic fate decisions co‐exist in the same lineage, as the deterministic PRpr production was not altered in any of the morphant conditions.

A simple theoretical gene regulatory network can explain temporal changes in progenitor competence

Our transcription factor (TF) knockdown experiments revealed an unexpectedly broad spectrum of possibilities for fate decisions in the probabilistic branch of the lineage. Previous studies suggested that interactions between the pro‐neural TFs in the retina in the form of gene regulatory networks (GRN) influence the proportions of different neuronal fates within a clone (Vitorino et al, 2009; Jusuf et al, 2011). We thus set out to use the lineage data obtained from the knockdown experiments to develop a minimalistic stochastic model capturing possible interactions among Atoh7, Ptf1a and Prdm1a. This model used the morphant data to probe whether a simple GRN based on interactions between these transcription factors could explain the temporal changes in progenitors' competence observed in control conditions.

Event plots were generated for Atoh7, Ptf1a and Atoh7/Ptf1a morphant conditions (Figs 4A, and EV5A and B) and the distribution of fate outcomes was renormalized time point wise by the distribution of all neurogenic division events. This allowed us to calculate the probability of different neuronal fates for a given division occurring at a certain time, which revealed the fate share changes over development (Fig 4B). We found that when Ptf1a is knocked down, the likelihood of RGC production remained constant over time. Conversely, the production of INs was only slightly changed upon Atoh7 knockdown. We thus hypothesized that, as previously proposed by Jusuf et al (2011), Ptf1a could have an inhibitory effect on Atoh7, while Atoh7 had no inhibitory effect on Ptf1a. As bipolar cells emerged from Atoh7+ divisions only when both Atoh7 and Ptf1a were depleted (Fig 3E, F and H), we further hypothesized that expression of either of these genes was sufficient to inhibit bipolar cell transcription factors such as Vsx1. Furthermore, as Prdm1a has been shown to inhibit bipolar cell genes in the mouse retina (Brzezinski et al, 2010), we implemented this inhibition into our GRN. The fact that a second PRpr is generated in all morphant scenarios (Fig 3D–G) made us add the assumption that PR genes were inhibited by the expression of Atoh7 and/or Ptf1a. As the knockdown of Atoh7 or Ptf1a or both did not alter overall lineage topology (Figs 3C and D, and EV5A) but only affected fate decisions at the terminal point of the lineage, our model neglected the number of progenitor divisions that occurred before neurogenesis onset.

All above‐stated assumptions were implemented in a phenomenological stochastic model to test whether a hypothetical GRN based on the morphant lineage data could predict the temporal dynamics of fate decisions observed in control conditions. In the model, noise came from the base levels of transcription factor expression (Elowitz et al, 2002; Johnston & Desplan, 2008; Raj & van Oudenaarden, 2008; Raj et al, 2010) and the inhibitions were modelled considering the presence of a threshold level of expression of each factor. If the expression level of a TF was above the threshold, then effective inhibition took place on its target TFs (Huang et al, 2007; Hersbach et al, 2022). To model the temporal evolution of fate probabilities from the experimental data, we considered that TF levels change during development, as previously shown (Jusuf & Harris, 2009). Interestingly, assuming that only Ptf1a expression varies over time (Jusuf & Harris, 2009) was sufficient to completely reproduce the temporal changes of fate probabilities (see Materials and Methods for details).

By considering this GRN and implementing a simple effective inhibition (Fig 4C), the simulation recapitulated the double morphant condition well (compare Fig 4C right panel with B right panel) but failed to recapitulate the fate outcome and the changes in fate shares over time in the single morphant conditions (compare Fig 4C middle and left panels with B middle and left panels). In particular, the simulation showed the presence of divisions generating a BC also in single morphants for Atoh7 and Ptf1a, an outcome not observed in our experiments. Further, the simulation predicted that BCs and a second PRpr could be generated in the control situation (Fig 4F, Simulation A), scenarios not found in our experimental dataset (Figs 4E and 1G, and EV4).

When the inhibition mode of the model was amended to a joint inhibition, that is, considering all inhibitors of a given target to sum their levels to reach the threshold, while keeping all other parameters constant (Fig 4D), the model faithfully recapitulated the fate shares and the temporal dynamics of fate decisions in all morphant conditions (compare probabilities estimation in Fig 4D with B). Further, and most importantly, the simulation based on this second GRN also correctly predicted the temporal changes of fate shares in the control condition (Fig 4F, Simulation B), including the lack of bipolar cells production and the absence of symmetric divisions generating two PRprs (compare Fig 4E with F, Simulation B).

Thus, a simple GRN based on the fate probabilities observed in our knockdown experiments can recapitulate the temporal changes in neuronal proportions in control retinas. This suggests that seemingly complex temporal shifts in progenitor competence observed during neurogenesis could originate from a simple gene regulatory pattern in a noisy environment. In principle, this means that such a simple GRN could be sufficient to achieve the observed fate shares during development from multipotent neurogenic progenitors.

Discussion

In this study, we investigated the possibilities and constraints of fate decisions of retinal neurogenic progenitors during zebrafish development.

Analysis of the output of single progenitor divisions revealed that two different types of fate decisions, deterministic and probabilistic, co‐exist in the same lineage. The probabilistic fate decision produces different fates with different probability distributions, while the deterministic lineage branch always gives rise to one PRpr. We further find that upon interference, the two lineage branches respond differently depending on which branch is affected. Upon interference with PRpr emergence, the deterministic lineage branch does not produce physiologically observed fates. In contrast, interference with the more probabilistic lineage branch led to an even broader spectrum of fate outcomes than observed in the control scenario and to changes in fate probabilities. Through stochastic modelling of the fate outcome of the probabilistic branch in perturbed conditions, we could propose a simple GRN that predicts the temporal changes in progenitor competence in controls. This GRN was based on the observed changes in fate probabilities in the morphant experiments and explained the temporal production of different retinal neurons in the right proportions during development. It is thus tempting to speculate that the observed lineage variability could be important for robust tissue development.

To date, fate decisions in the vertebrate CNS have mostly been investigated at the clonal or population level. Depending on the techniques used and interpretation of the data, fate decisions were considered either mostly driven by stochasticity (Gomes et al, 2011; He et al, 2012; Boije et al, 2015; Llorca et al, 2019) or by deterministically encoded programmes (Rapaport et al, 2001; Hippenmeyer et al, 2010; Cepko, 2014; Gao et al, 2014). Our quantitative analyses of single progenitor divisions and their neuronal output over development show that both deterministic and stochastic fate decisions can co‐exist in the same progenitor lineage. In the case of the Atoh7+ neurogenic progenitors, PRprs are always and continuously produced during development in one branch of the lineage, while the sister cell acquires different fates with time‐dependent probabilities. This indicates the presence of stereotypic lineage patterns among the general stochasticity in fate decisions: while neuronal fate decisions in one branch of the lineage follow a probability distribution that changes over time, the other branch always and reliably produces one PRpr with a probability of 1. This adds to the previously proposed stochastic model, in which all neuronal fates are decided according to a set of different probabilities (He et al, 2012; Boije et al, 2015).

Furthermore, these previous models assumed that retinal progenitors can produce neurons through P‐D divisions (progenitor‐differentiated cells). Our data, however, show that neurogenic Atoh7+ progenitors in the retina divide and produce neurons and committed precursors (D‐CP division), but never another progenitor P cell (Fig 1). It is important to note that while also committed precursors undergo one more division, they differ from progenitor cells as they have a different proliferative capability (only one additional cell division), morphology and fate potential (unipotent; Godinho et al, 2007; Suzuki et al, 2013; Weber et al, 2014; Engerer et al, 2017). Thus, committed precursors contribute minimally to the overall clone size meaning that indeed canonical P‐D divisions do not exist in the lineages described in this study.

Our work thereby improves the interpretation of previous data on the final clonal size both in wild‐types and morphant conditions (He et al, 2012; Boije et al, 2015). It would be interesting to investigate whether the absence of P‐D divisions and the presence of lineage patterns that generate committed precursor cells is specific to the zebrafish retina or whether it is also found in other areas of the CNS and/or other species. Fate‐restricted progenitors have been found in mouse and chicken retinas (Boije et al, 2009; Hafler et al, 2012) and in the neocortex (Noctor et al, 2004; Hevner, 2019) and they might arise from lineage patterns similar to those found in this study. The presence of a different lineage topology might also entail that a different fate decision mechanism is employed with respect to other tissues where P‐D divisions occur, like the Drosophila CNS (Brand & Livesey, 2011).

When we probed the possibilities and constraints of the deterministic and probabilistic fate decisions, interference with the deterministic PR fate affected overall tissue integrity and lineage topology. Interestingly, upon interference with the probabilistic fate decision, neurogenic progenitors often divided symmetrically and generated a second PRpr, a type of division observed only in the absence of Atoh7 or Ptf1a or of both. We speculate that the PR fate could be a “default state” for neurogenesis outcome and the GRN involving Atoh7, Ptf1a and Prdm1a might override this default programme to specify different neuronal types. This is an attractive idea also in evolutionary terms as PR cells are widespread throughout evolution (Arendt, 2003) and are thought to be the first retinal cell types that evolved (Arendt, 2003; Lamb et al, 2007). It is therefore possible that the other cell types evolved “on top” of a general PR programme as proposed by Constance Cepko (2014) and Arendt (2003). Currently, this is, however, purely speculative.

Interestingly, in contrast to the deterministic branch, interference with the probabilistic branch of the lineage showed a widened spectrum of fate possibilities for neurogenic progenitors. Here, neurons that in controls were mainly produced at early stages were also born later. This finding is in line with previous reports in the RP2/sib neuroblast lineage in Drosophila, where it was shown that progenitors exhibit temporal plasticity and can give rise to early lineages at later stages (Gaziova & Bhat, 2009). Furthermore, the reverse is also true as in double Atoh7/Ptf1a morphants, BCs were seen to arise as sister cells of PRprs earlier than they would usually appear (Fig 3E). The occurrence of BCs is also interesting as this is a fate that was never generated from Atoh7+ progenitors in controls. Together, this indicates that neurogenic progenitors are multipotent and can generally produce any neuronal fate at any time. This is strikingly similar to what was recently observed in the mouse embryonic ganglionic eminence, where progenitor cells can generate all the inhibitory neuronal fates until their last cell division (Bandler et al, 2022). Our findings could imply that competence windows can be more flexible than previously thought (Cepko, 2014).

Overall, the fact that the deterministic and probabilistic lineage branches show different fate outcome possibilities and different reactions to interference suggests that different molecular mechanisms act to specify the PRpr versus the sister cell fate. Modelling the fate probabilities observed upon TF knockdown revealed a network of TF interactions that follow these rules of fate decisions for the sister cell of the PRpr. This simple GRN (Fig 4D) featuring Atoh7, Ptf1a and Prdm1a was able to recapitulate temporal changes in the probabilities of generating different fates in the morphant conditions and to accurately predict the changes in progenitor competence in control conditions. We are aware that the scenarios that allowed us to induce fate probability changes were rather artificial, as we suppressed one or more neuronal populations. However, the fact that the lineage reacted to such perturbations with an unexpectedly broad spectrum of fate possibilities makes us speculate that such lineage flexibility could underlie the presence of plasticity in fate decisions during normal development. As modelling the changes in fate probabilities predicted the temporal changes in progenitors' fate decisions in the control scenario, fate plasticity could exist to ensure progenitors' competence changes during development. However, it is important to note that the model presented here is only one possible scenario. Further, this model does not reveal whether the included TFs act directly or indirectly, nor whether further TFs are involved. Thus, additional investigations involving ChiP‐sequencing and timely controlled overexpression studies are needed.

In contrast to the probabilistic fate decisions, the deterministic PR fate decision does not seem to rely on the same GRN, as interference with Prdm1a expression does not result in a fate switch (Fig 2F and G). As a PRpr is reproducibly produced from each Atoh7+ division, one possibility is that fate determinants specifying the PR fate are asymmetrically inherited by the cell that becomes a PRpr. This is a widespread strategy employed by different systems in which asymmetric fate decisions occur, like in the Drosophila neuroblast (Chia et al, 2008; Brand & Livesey, 2011) or the mouse retina (Cayouette et al, 2001; Cayouette & Raff, 2003; Kechad et al, 2012). Transcriptomics analysis combined with advanced live imaging techniques will reveal key regulators of this fate decision.

Overall, our study contributes to the understanding of possibilities in lineage decisions in the retina and revealed different degrees of flexibility for different types of fate decisions. Studies like ours, combining live imaging with theoretical modelling, are important to understand the temporal dynamics of fate decisions that produce an organ of the correct size, cellular proportions and connectivity and will allow to reveal core principles of reproducible brain formation.

Materials and Methods

Zebrafish husbandry

Wild‐type zebrafish were bred and maintained at 26°C. Embryos used for experimental work were raised at 21, 28.5 or 32°C in E3 medium supplemented with 0.2 mM 1‐phenyl‐2‐thiourea (PTU, Sigma‐Aldrich) from 8 h post‐fertilization (hpf) to prevent pigmentation. Medium and PTU were changed daily. Animals were staged in hpf according to Kimmel et al (1995). All animal work was performed in accordance with European Union Directive 2010/63/EU, the German Animal Welfare Act and the Portuguese Legislation (Decreto‐Lei n° 113/2013).

Transgenic lines

Tg(atoh7:gap‐GFP) and Tg(atoh7:gap‐RFP) zebrafish transgenic lines were used to identify Atoh7+ progenitors and Atoh7+ neurons (Zolessi et al, 2006). The Tg(crx:gap‐CFP) line was used to visualize PRs and PRpr (Almeida et al, 2014). To visualize all different neurons and for transcriptomics, the triple transgenic line Tg(crx:gap‐CFP), Tg(atoh7:gap‐RFP) and Tg(ptf1a:Gal4/UAS:gap‐YFP) (Almeida et al, 2014) was used.

DNA injections

DNA was injected at one‐cell stage to mosaically label progenitors and neurons in the retina. One nanolitre of atoh7:GFP‐CAAX (Zolessi et al, 2006; Icha et al, 2016a) plasmid was injected in each wild‐type or Tg(atoh7:gap‐RFP) embryo in controls and in morphant conditions to label Atoh7+ progenitor cells and their progeny.

Photoconversion

The Crx:H2B‐Dendra construct used to label crx+ cells for photoconversion was assembled using Gateway cloning (Thermo Fisher Scientific) based on the Tol2 kit.

The pME H2B‐Dendra was combined with the 5′ entry clone containing the Crx promoter (a kind gift from Rachel Wong) into the destination vector pTol2 + pA + cmlc:eGFP R4‐R3 (Kwan et al, 2007).

Fifteen to 30 ng/μl of the plasmid were co‐injected with 2 ng of p53 MO in one‐cell stage embryos. Embryos were raised until 42 hpf, mounted in agarose in 35 mm glass‐bottom Petri dishes (Greiner Bio‐One) and imaged with spinning‐disk confocal. Two to three cells per embryo were photoconverted. Isolated cells expressing Crx:H2BDendra were chosen for photoconversion according to their proximity to the apical side. After photoconversion, embryos were taken out of the agarose and grown for 24 h in a 12‐well plate in E3 medium supplemented with 0.2 mM 1‐phenyl‐2‐thiourea. After 24 h, embryos were mounted in agarose again for imaging, and photoconverted cells were assessed.

FACS sorting for transcriptomics

For RNAseq of PRprs, RGCs and INs, retinas from a triple transgenic line containing cell‐type‐specific reporters [Tg(crx:gap‐CFP), Tg(atoh7:gap‐RFP), Tg(ptf1a:Gal4/UAS:gap‐YFP)] (Almeida et al, 2014) were dissociated mechanically. FACS was performed using FacsAria Fusion. The gates were adjusted for autofluorescence/background fluorescence using single transgenic and wild‐type embryos. A minimum of 10,000 live cells were sorted per experiment from a pool of 25 retinas. Different neurons were sorted thanks to the expression of a different combination of markers (Fig EV2A). For transcriptome analysis, 500 cells were sorted per population from a pool of 25 retinas directly into the buffer for extraction.

RNAseq of FACS‐sorted zebrafish retinal neurons

For each of the five biological replicates, RNA extraction of FACS‐sorted PRpr, RGC and INs was performed according to previously published protocol (Picelli et al, 2013) by the Deep Sequencing Facility at the Genome Center of the Technische Universität Dresden. Data and methodological details (reverse transcription, cDNA amplification, library preparation, sequencing and data processing) are accessible through GEO series accession number GSE194158 at NCBI.

Morpholino experiments

To knock down specific genes, the following amounts of morpholinos were injected per embryo into the yolk at one‐cell stage: 2 ng p53 MO, 5′‐GCGCCATTGCTTTGCAAGAATTG‐3′ (Gene Tools, (Robu et al, 2007)); 4 ng Atoh7 MO, 5′‐TTCATGGCTCTTCAAAAAAGTCTCC‐3′ (Gene Tools, (Pittman et al, 2008)); 10 ng Ptf1a MO1, 5′‐CCAACACAGTGTCCATTTTTTGTGC‐3′ (Gene Tools, (Jusuf et al, 2011)); 10 ng Ptf1a MO2, 5′‐TTGCCCAGTAACAACAATCGCCTAC‐3′ (Gene Tools, (Jusuf et al, 2011)); 5 ng Prdm1a MO, 5′‐TGGTGTCATACCTCTTTGGAGTCTG‐3′ (Gene Tools, (Lee & Roy, 2006; Liu et al, 2012)) and standard control MO 5′‐CCTCTTACCTCAGTTACAATTTATA‐3′ in equal volume. This Prdm1a (also called Blimp‐1) morpholino effectively blocks splicing of the pre‐mRNA of prdm1a, as confirmed by RT–PCR (Lee & Roy, 2006). Furthermore, the prdm1a morphant phenocopies the mutant narrowminded, characterized by defects in neural crest and sensory neuron development. This phenotype is due to a point mutation in the gene prdm1a and injection of the prdm1a mRNA rescued the mutant and the morphant phenotype, confirmed by RNA in situ hybridization (Hernandez‐Lagunas et al, 2005).

Drug treatments

The Notch inhibitor LY411575 (Sigma‐Aldrich) was dissolved in DMSO and used at a concentration of 10 μM. An equal volume of DMSO was used for controls. Embryos were dechorionated and up to 10 embryos were placed in a well of a 24‐well plate, then incubated at 28°C in the dark in E3 medium. The treatment windows are specified in the figure and the figure legend.

In vivo labelling of proliferative PRprs

To label proliferative PRprs, 48 hpf embryos were incubated at 4°C for 1 h in E3 supplemented with 500 μm of EdU (ClickiT‐Alexa 488 fluorophore kit, Invitrogen) in 10% DMSO. After incubation, embryos were washed twice with E3 and immediately fixed overnight in 4% PFA. After antibody staining, incorporated EdU was detected according to manufacturer's protocol. Embryos were stored in PBS at 4°C until imaging.

Immunofluorescence

All immunostainings were performed on whole‐mount embryos fixed overnight in 4% paraformaldehyde (Sigma‐Aldrich) in PBS at 4°C as previously described (Icha et al, 2016a).

Embryos were washed five times for 10 min in PBS‐T (Triton X‐100 in PBS) 0.8%. For permeabilization, embryos were incubated with 0.25% Trypsin–EDTA in PBS on ice for different times depending on the developmental stage (10 min for 24, 28 and 36 hpf, 12 min for 42 hpf, 15 min for 48 hpf and 17 min for 72 hpf). Embryos were then kept on ice for 30 min in PBS‐T 0.8%. Blocking was performed with 10% donkey or goat serum in PBS‐T 0.8% for 3 h at room temperature or overnight at 4°C.

Embryos were incubated with the following primary antibodies for 72 h at 4°C: GFP (50430‐2‐AP, Proteintech) 1:100, Histone H3 (phospho S28, ab10543, Abcam) 1:500, zpr‐1 (AB_10013803, zirc) 1:750, RFP booster (Chromotek rba594) 1:200, active caspase‐3 (BD‐Biosciences 559565) 1:200 and zn‐5 (Zirc, ZDB‐ATB‐081002‐19) 1:20.

Embryos were then washed five times for 30 min with PBS‐T 0.8% and then incubated for 48 h with the appropriate fluorescently labelled secondary antibody (Molecular Probes) at 1:500 and DAPI 1:1,000 (Thermo Fisher Scientific). Finally, embryos were washed four times for 15 min with PBS‐T 0.8% and stored in PBS at 4°C until imaging.

Image acquisition

Confocal microscopy

Fixed samples were imaged with a laser‐scanning microscope (Zeiss LSM 880 Airyscan inverted or Zeiss LSM 980 Airyscan2 inverted, equipped with two PMT and one GaAsP) using the 40×/1.1 C‐Apochromat water immersion objective (ZEISS). Samples were mounted in 0.6% agarose in glass‐bottom dishes (MatTek Corporation) and imaged at room temperature. Serial sections were acquired every 1 μm with ZEN 2011 (black edition) or Zeiss's ZEN Blue v3.0.

In vivo light sheet fluorescence Microscopy (LSFM)

Imaging started at 28 hpf for the early neurogenic window experiments and at 36 hpf for the late neurogenic window. Embryos were manually dechorionated and mounted in ~1‐mm‐inner‐diameter glass capillaries in 0.6% low‐melting‐point agarose as previously described (Icha et al, 2016b). The sample chamber was filled with E3 medium containing 0.01% MS‐222 (Sigma) and 0.2 mM PTU (Sigma). Imaging was performed on a Zeiss Lightsheet Z.1 microscope equipped with two PCO Edge 4.2 sCMOS cameras (max 30 fps with 2048 × 2048 pixels—pixel size 6.5 μm) and with a 20×/1,2 Zeiss Plan‐Apochromat water immersion objective. Imaging was performed at 28.5°C. Z‐stacks spanning the entire retinal epithelium (70–100 μm) were acquired with 1 μm optical sectioning every 5 min for 15–24 h with double‐sided illumination mode. The system was operated using the ZEN 3.1 software (black edition).

To confirm the effect of the morpholino injections in live imaging experiments, the absence of the optic nerve was assessed after live imaging of the Atoh7 MO condition. Embryos in which the optic nerve was present were discarded from the analysis. Embryos injected with Ptf1a MOs were grown at 28.5°C until 72 hpf after imaging, fixed in PFA and immunostained against HuC/D (A‐21271, Thermo Fisher). Embryos displaying HuC/D staining in the inner nuclear layer were discarded from the analysis.

Spinning disk imaging for photoconversion

Photoconversion experiments were performed using an Andor spinning disk confocal microscope composed of an Andor IX 83 stand and a CSU‐W1 scan head (Yokogawa) with Borealis upgrade, equipped with a DMD Andor Mosaic module and a 405 nm photomanipulation light source. The microscope was operated via the Andor iQ software version 3.6. Embryos were embedded in 0.8% of low‐melting‐point agarose in E3 medium supplemented with 314 μg/ml of MS‐222 and 0.1 M Hepes (pH 7.25) in 35 mm glass‐bottom Petri dish (Greiner Bio‐One). The dish was filled with E3 supplemented with 120 μg/ml of MS‐222. Z‐stacks were acquired using Olympus UPLSAPO objective 60× 1.3 SIL and Andor iXon 888 Ultra with Fringe suppression.

Quantitative analysis

Images from live imaging experiments were cropped and averaged in ZEN Black and/or Fiji (Schindelin et al, 2012) and were corrected for drift using the Fiji plugin “Manual Drift Correction” (https://imagej.net/Manual_drift_correction_plugin) created by Benoit Lombardot (Scientific Computing Facility, Max Planck Institute of Molecular Cell Biology and Genetics, Dresden, Germany).

Cell fate assignment

To identify the fate of the daughter cells of Atoh7‐expressing progenitors, single cells were followed until they reached their final position or until they acquired a distinct morphology (Fig 1B). Morphology was assessed by the use of the membrane marker atoh7:GFP‐CAAX in combination with the transgenic line Tg(atoh7:gap‐RFP). PRpr fate was assigned based on unipolar morphology during migration and columnar morphology after reaching the apical side (Suzuki et al, 2013; Weber et al, 2014; preprint: Rocha‐Martins et al, 2021), their bidirectional migration pattern (preprint: Rocha‐Martins et al, 2021) and final apical positioning. RGC fate was assigned based on basal position, basal somal translocation (Icha et al, 2016a) and the emergence of a basal axon (Fig 1B and C). Inhibitory neuron (ACs and HCs) fate was assigned based on multipolar migration mode (Appendix Figs S1 and S2; Chow et al, 2015; Amini et al, 2019, 2022), absence of basal axons and by final positioning (Fig 1D and E, basal position for ACs and apical position for HCs). Bipolar cell fate was assigned by bipolar morphology and nuclear positioning in the INL (Weber et al, 2014; Engerer et al, 2017; Fig 3H).

Analysis of cell migration

To generate the trajectories for each cell type in Fig 1B and for HCs and ACs in Appendix Figs S1 and S2, cells were cropped with ZEN 3.1 (Zeiss) and processed in Fiji as described above. Cells were tracked in 2D in maximum‐projected sub‐stacks by following the centre of the cell body in Fiji using the semi‐automated ImageJ plugin MtrackJ (Meijering et al, 2012). Tracking started at birth of each cell and ended after the cell reached its final position in the tissue.

Retinal size measurements

Retinal size measurements were performed manually using the Fiji line tool on 72 hpf retinas from the Tg(ptf1a:Gal4; UAS:YFP; atoh7:gap‐RFP; and crx:gap‐CFP) line, in controls and in all morphant conditions as illustrated in Fig 2C.

The largest retinal diameter was measured on three different z‐planes per retina, and the average measurement was plotted for each replicate.

Retinal thickness was measured on three different z‐planes of the central region of the tissue, and the average measurement was plotted for each replicate. The outer nuclear layer (ONL) was assigned as the distance between the apical side and the outer plexiform layer; the inner nuclear layer (INL) as the distance between the inner plexiform layer and the outer plexiform layer; and the ganglion cell layer (GCL) was assigned as the layer between the inner plexiform layer and the most basal position of the retina.

Photoreceptors cell count

Photoreceptor cells were manually counted at 72 hpf in wild‐type and Atoh7 + Ptf1a double morphant embryos. A ROI of 67.37 × 67.37 μm in size was drawn and PRs were manually counted using the Fiji Multipoint tool in Tg(crx:gap‐CFP), Tg(hsp70:H2B‐RFP) double transgenic embryos. Ten z‐planes per embryo were measured in the nasal, central and temporal sides of the retina. To measure PRs cell density, the number of PRs was divided by the length of the apical side, measured with the Segmented Line Tool in Fiji.

PH3+ cells count

To count the number of PH3+ cells in whole retinas, the Tg(hsp70:H2B‐RFP) was used to identify the boundaries of the tissue. Then, stacks covering the whole tissue were acquired using the laser scanning confocal. Images were imported in Imaris and the mask tool was used to isolate the retinal tissue from the rest of the tissues in the image. The spot detection tool was used to count the number of PH3+ cells with these parameters: diameter = 8 μm (x–y dimensions) and PSF modelling = 16 μm. Threshold for detection of the PH3+ nuclei was adjusted manually to ensure that all the PH3+ nuclei were counted but was usually kept around 1,800.

Statistical analysis

All statistical tests used are indicated in the figure legend, as well as the definitions of error bars. All tests used were two‐sided and 95% confidence intervals were considered. P‐values are indicated in the figure legends, as well as sample sizes, or Table 2 for experiments in Fig 2D or Table 3 for experiments in Fig 3B. Data were analysed using GraphPad Prism 6 or Python 3. Statistical analysis was performed using GraphPad Prism 6 and Julia 1.7.2. Information about the exact libraries, together with their versions, used for the Event plots and the in‐silico stochastic model are detailed in the Project.toml and Manifest.toml files in the respective Git repositories, as per standard dependency handling procedures of the Julia language:

Event plots

Event plots were produced from the raw experimental data using a custom data analysis and plotting pipeline developed by the authors, which is available at https://git.mpi‐cbg.de/nerli/lineage‐analysis. The code comes together with the data files, allowing for easy and complete replication of the analysis and plots that appear in this work.

The data analysis consists of the following steps:

  • Consider the time at which each division occurs as the time of the fate decision event and categorize these events by the fate acquired by the sister cell of the PRpr.

  • Compute the kernel density estimation (KDE) of the distribution of events over time for each fate type and for the total distribution of divisions.

  • Rescale the KDE of each fate by the KDE of all divisions to obtain an estimation of the probability of acquiring the different fates dependent on the time the division occurs.

  • Use the bootstrap resampling method (Efron, 1979) to obtain uncertainties of the KDEs and of the fate probabilities.

In silico stochastic model of cell fate decision

To evaluate the potential of candidate GRNs to reproduce our experimental data of the probabilistic branch of the lineage, we designed and implemented a simple in silico phenomenological model of stochastic fate decisions. The code is available at https://git.mpi‐cbg.de/bianucci/transcription‐factors‐interactions.

Since we know from the lineage analysis data that each division produces one PRpr and one sister cell whose fate is stochastically determined, we modelled only one fate decision per division. Also, as the TFs Atoh7, Ptf1a and Prdm1a are responsible for the specification of the different neuronal fates, we modelled their expression as random variables, with time‐dependent expected values. Their expression levels were normalized with respect to an effective threshold, that is, a level at which they start to carry out their inhibitory function and produce a downstream effect.

The simulation of this model relies on drawing N = 100 fate decisions for each time point to estimate the shares of different fates, this was then repeated for R = 100 times to compute the mean and confidence interval of these shares. The TF levels were modelled as normally distributed, with a time‐dependent mean value and a constant standard deviation. Being the threshold set at the arbitrary value of 1.0, in the model we set the mean expression levels of Atoh7 and Prdm1a at the values of 1.1 and 1.05 respectively. The mean levels for these TFs were kept constant over time. The mean expression level for Ptf1a was instead time dependent (Jusuf & Harris, 2009) and increased according to a logistic function with lower and upper asymptotes at 0.9 and 1.1 respectively.

The standard deviation was constant over time for all TFs and was computed by multiplying a fixed coefficient of variation (CV = 0.14) by the time‐averaged expression level of each TF.

Finally, the GRN was modelled as a decision rule occurring at the terminal branching point of the lineage (Fig 1G), taking the TF levels as input and producing a fate choice as output. Any inhibition in the GRN was translated into a precedence rule, that is, a TF that inhibits a second one is able, if its level is above the threshold, to determine the acquisition of a certain fate, while the inhibited TF is ignored. The precedence rule for Scenario A (Fig 4C and F) is illustrated by the following diagram:

Table 4
Full size table

To achieve inhibition of the target TF in additive inhibitions, it was enough that the sum of the levels of the inhibitors involved was above the threshold. The precedence rule in the additive inhibition case (Scenario B, Fig 4D and F) is illustrated by the following diagram:

Table 5
Full size table

In summary, the fate decision model consists of (i) drawing independent and normally distributed stochastic TF expression levels, (ii) applying the decision rule and recording the resulting fate, (iii) repeating this stochastic decision for 100 divisions taking place at any given time point and (iv) again repeating the whole process for 100 times, creating a synthetic dataset that can then be analysed in the same way as the experimental data.

Disclosure and competing interests statement

The authors declare that they have no conflict of interest.