Multicellular organisms develop as 4 dimensional spatiotemporal arrays to ultimately form cohesive organs and tissues with the appropriate size, composition, and function. Critical to this process is spatial and temporal patterning. The concept of spatial patterning, which posits that progenitor cells adopt different fates based on their position within a given tissue, was first hypothesized and investigated by embryologists around 70 years ago [1], [2], [3]. Secreted cues often act as spatial patterning signals to instruct different cell fates based on positional information. A classic example of this is observed in the developing neural tube, where the position of progenitor cells within a dorso-ventral gradient of morphogens is integrated to determine neuronal cell type output [4], [5]. Beyond spatial patterning, temporal patterning is also paramount during development to control when a given progenitor cell population gives rise to a specific combination of cell types. Together, the integration of temporal and spatial patterning mechanisms ensures that developing tissues differentiate in an organized manner [6]. In this review, we focus on the mechanisms underlying temporal patterning in the developing mammalian retina.

The mechanisms of temporal patterning were extensively studied in the Drosophila ventral nerve cord (VNC), in which neuroblasts give rise to neuronal subtypes in a strict chronological sequence. Classical work showed that VNC neuroblasts rely on transcription factors collectively referred to as ‘temporal identity factors’ (TIFs), which are sequentially expressed as a result of feedforward and feedback regulatory loops and are necessary and sufficient to generate progenies associated with each TIF expression window [7], [8], [9], [10], [11], [12]. Specifically, the temporal sequence containing hunchback (Hb), krüppel (Kr), pdm1/2 (Pdm), castor (Cas) and grainyhead (Grh) was shown to orchestrate neuroblast output without affecting cell cycle exit. More recently, other temporal cascades involving different transcription factors have also been identified in the Drosophila medulla [13], which is part of the visual processing center in the brain, suggesting that TIF cascades are a general mechanism used to control temporal patterning and diversify neuronal output.

But how do TIFs differ from classical fate-determining transcription factors? TIFs do not promote the production of specific cell fates, but rather are necessary and sufficient to confer competence to generate a cell type or combination of cell types appropriate for the temporal window in which they are expressed. The exact identity of the cell type produced is not controlled by TIFs per se, but by fate-determining cues (intrinsic or extrinsic) specific to any given lineage (Fig. 1). The sequential activation and repression of different TIFs allows progenitors to alter their response to fate-controlling cues over time, thereby ensuring the appropriate chronological order of cell birth. This concept is clearly illustrated by the findings that the same TIF cascade operates in different lineages to control chronological production of distinct cell types in Drosophila [14], [15].

Like Drosophila, several regions of the developing mammalian CNS rely on temporal patterning to establish the vast arrays of cell types that compose a fully functional nervous system [16], [17], [18]. The retina, an extension of the CNS, is an excellent working model to address questions related to temporal patterning because: i) it is easy to access for experimental manipulations; ii) it contains a manageable number of cell types organized in histologically distinct layers, facilitating their identification; iii) retinal progenitor cell (RPC) outputs can be traced by various cell lineage tracing assays; iv) the retina is not essential for survival, allowing genetic manipulations that might be lethal if carried out in other regions of the CNS; and v) the various retinal cell types are produced in a clear chronological order. Extensive studies using H3 thymidine assays have indeed observed that retinal cell types are generated in a sequential yet partially overlapping order that is conserved between mammalian species [19], [20], [21], [22]. These observations suggest that molecular mechanisms operate in RPCs to change their temporal identity as development proceeds [19], [23]. While the transcription factor code regulating cell fate decisions in the retina has been extensively studied and reviewed in the past 15 years [21], [24], [25], [26], the mechanisms regulating temporal patterning have received less attention and will thus be the focus of this review. A summary of the factors and their function discussed below is presented in Table 1.

While extrinsic cues can modulate RPC proliferative behavior and influence RPC fate decisions [27], [28], results from classical heterochronic transplantation experiments suggest that RPCs are intrinsically restricted to generate specific cell types at any given time, regardless of their surrounding environment [29], [30], [31], [32]. Additionally, it was demonstrated that dissociated RPCs cultured at clonal density in an arbitrary environment can give rise to a clonal population that is indistinguishable from that observed in situ both in terms of size and composition [31]. Furthermore, general cell birth order is reproduced in such cultures, consistent with observations in Drosophila neuroblast culture assays, in which the sequential expression of TIFs (hb – kr – pdm – cas – grh) and neuronal birth-order is recapitulated [10], [33]. Together, these results support a model in which intracellular programs largely drive retinal histogenesis.

But are these programs pre-determined or plastic? Evidence suggests that both mechanisms may be contributing. On the one hand, Olig2-expressing mouse RPCs have been shown to generate a specific combination of cell types at their last division; either a cone and a horizontal cell at early stage, or a rod and an amacrine cell at late stage. These results indicate that some pre-programming occur in RPCs, at or just before the final division [34], [35]. On the other hand, lineage tracing experiments demonstrate that RPC lineages are highly variable [36], [37], [38], [39], suggesting that a purely deterministic model is unlikely to fully explain retinal development in vertebrates [38], [40], [41], [42]. Despite the variability in lineage composition, the result is consistent, producing a retina with invariable proportions of various cell types. How is this achieved? Live imaging experiments in fish and rats have shown that stochastic mechanisms may be at play, in which equipotent RPCs undergo proliferative or differentiative divisions according to probabilities that change over time [36], [38], [39] (Fig. 1). This raises the question of which molecular mechanisms operate to bias RPCs to divide or differentiate or to generate an early or late-born cell type. One possibility is that TIFs function to bias the probability to generate specific combination of cell types during their window of expression [24], [36], [42], [43] (Fig. 2).

The discovery of TIF cascades in Drosophila neuroblasts has prompted the investigation of whether this logic may be used in vertebrate CNS development. In recent years, a picture has emerged to suggest that at least some parts of the Drosophila temporal cascade are indeed conserved. We discuss these mechanisms below.

Ikzf1 (previously known as Ikaros), is the mammalian homolog of Drosophila hb, showing high conservation in the zinc finger regions. IKZF1 is expressed in early-stage RPCs [44], suggesting that it might share functional similarities with Drosophila hb to control early temporal identity. Consistently, Ikzf1 knockout mice show reduced numbers of early-born retinal cell types, such as retinal ganglion cells (RGCs), amacrine and horizontals cells, though production of cone photoreceptors, another early-born cell type, is unaffected. Conversely, lineage tracing experiments showed that Ikzf1 misexpression in late RPCs is sufficient to induce production of cell types that are normally restricted to early stages, such as horizontal, amacrine, and RGCs, without altering RPC proliferation [44], [45]. Thus, Ikzf1 is sufficient and at least partially required for early-born cell type generation. This is consistent with the idea that mammalian TIFs may function permissively in RPCs, perhaps by changing the probability to generate specific cell types associated with the temporal window in which the factor is expressed. The molecular mechanism by which Ikzf1 functions in RPCs remains unclear. It was previously reported, however, that IKZF1 associates with members of the polycomb repressive complex 2 (PRC2) during lymphocyte development [46], suggesting that chromatin remodelling may be involved. As there are five other IKZF family proteins, and some of them are expressed in the mammalian retina [44], it is also possible that Ikzf1 cooperates with other members of the Ikzf family to control temporal progression in RPCs. Consistently, it was recently reported that Ikzf4 cooperates with Ikzf1 to control the timely production of cones at early stages of retinal development, whereas Ikzf4 functions independently of Ikzf1 at late stages to control glial cell production [47].

The Pou domain TFs, Pou2f1 and Pou2f2, are the mammalian homologs of Drosophila nub/pdm. Recent evidence has revealed that Ikzf1 induces Pou2f1 expression, which in turn represses Casz1 expression, a late TIF (see below), likely to ensure that RPCs do not prematurely switch to a late temporal identity [48]. Expression of Pou2f1 or Pou2f2 in late-stage RPCs is sufficient to promote cone genesis at the expense of late-born cell types such as rods, bipolar and Müller glia, without influencing proliferation. Conversely, inactivation of Pou2f1/2 reduces production of early-born cones and horizontal cells. Importantly, Pou2f1 was found to upregulate Pou2f2, which then represses the rod-promoting factor Nrl to favour cone production in photoreceptor precursors [48].

It will be interesting to evaluate to what extent Pou2f1/2 can also interact with other factors known to regulate cones, horizontals, and amacrines fates such as Onecut1/2 [49] or Foxn4 [50], both discussed below, and investigate whether they could be part of the same network or act through independent programs. For instance, Onecut1/2 dKO retinas present a two-fold increase in Pou2f2 mRNA levels, suggesting that these factors might act in separate pathways [49].

Casz1 is the mammalian homolog of the Drosophila late TIF cas [51]. In the mammalian retina, CASZ1 is expressed in mid-to-late-stage RPCs, as well as in rods and a small subset of amacrine cells in the adult retina [43]. Casz1 KO mice exhibit an increased production of early-born retinal cell types such as horizontal cells, amacrines and cones, a decrease in late-born rods, and an increase in the latest-born Müller glia [43]. When overexpressed in early-stage RPCs, Casz1 has the opposite effect and promotes late-born rod and bipolar cells while repressing early born fates and Müller glia [43]. Interestingly, Ikzf1 was shown to repress Casz1 in early RPCs [52], like cas represses hb in Drosophila neuroblasts [7], [53]. Thus, Casz1 appears to function as a bona fide TIF, as it is present in all mid/late-stage RPCs, does not alter the proliferation capacity of RPCs, and is sufficient and necessary, at least partly, to generate cell types produced during its expression window [43].

Recent work has shed light into the molecular mechanisms behind Casz1 activity. CASZ1 was found to physically interact with the nucleosome remodeling and deacetylase (NuRD) and polycomb repressor complexes (PRC) [52]. Importantly, CASZ1 was found to require both the NuRD and PRC to promote late temporal identity transitions, suggesting that chromatin remodelling is involved in RPC temporal identity changes. Based on these results, it was proposed that Casz1 functions by altering the epigenome, allowing cell fate determinants to induce specific neural cell types appropriate to the Casz1 late temporal window [52]. Thus, in contrast to the classical cascade seen in Drosophila where the TIFs regulate each other as part of a molecular clock, mammalian Casz1 appears to act through epigenetic modifiers to shift progenitor competence. Moreover, the NuRD and PRC are also known to interact with IKZF1 in mammalian lymphocytes [46], suggesting that the NuRD and polycomb repressor complexes may generally function with TIFs in the retina. Recent evidence is consistent with this idea and suggests that IKZF1 alters the epigenetic landscape when expressed in mouse embryonic fibroblasts [54]. It will thus be intriguing to investigate whether NuRD and polycomb repressor complexes have temporally defined binding partners at different stages of retinogenesis, and if such interactions are sufficient and necessary for temporal patterning.

While homologs of Drosophila of TIFs contribute to temporal patterning in the developing mammalian retina, other factors also play a part, as discussed below, suggesting a more complex regulation of temporal patterning in mammals than observed in flies.

Individually, Onecut1 or Onecut2 are both key cell fate determinants for horizontal cell generation [55], [56]. However, evidence has emerged from double knockout experiments that the Onecut factors may work together to regulate early temporal identity, conferring competence in RPCs to generate RGCs, cones, horizontal and amacrine cells [56]. It remains unclear, however, how Onecut1/2 control temporal identity. There is evidence that the Onecut 1 promoter contains Pax6 binding motifs, suggesting that it is a downstream target of Pax6 [49], which is known to confer multipotency in RPCs [57]. However, definitive evidence to support a role for Onecut 1/2 as bona fide TIFs is still lacking. It is unknown, for example, whether ONECUT1/2 are sufficient to confer early temporal competence. It will be important to determine if ectopic expression in late-stage RPCs can induce early born cell fates. Interestingly, recent work has implicated a role for the Onecut family in regulating early-born cell fates in other CNS regions such as spinal cord, hindbrain, midbrain, and forebrain, where it is similarly expressed during early developmental stages [58].

Foxn4 is a member of the forkhead/winged helix TF family that was first identified as a cell fate determinant, promoting the genesis of amacrine and horizontal cells [59]. More recently, evidence was provided to suggest that Foxn4 acts as a TIF, conferring competence in RPCs to generate amacrine, horizontal, and rod/cone photoreceptors, while inhibiting competence to generate RGCs [50]. Foxn4 overexpression in early-stage RPCs was shown to promote amacrine and horizontal cell production, as well as transiently increase rod and cone genesis [50]. At early stages, loss of Foxn4 reduced the number of photoreceptor cells produced. Since no changes in photoreceptor cell number was observed in adult retinas, it is likely that loss of Foxn4 simply delays photoreceptor genesis. Interestingly, RNA-seq analysis from Foxn4 KO retinas showed that rod genes are upregulated prematurely, suggesting that Foxn4 regulates the timing of rod gene expression [50]. The exact mechanisms by which Foxn4 regulates temporal identity remain unclear, but it may involve the regulation of Ikzf1 and Casz1, as these were repressed and upregulated, respectively, upon Foxn4 overexpression [50].

The Nuclear Factor I (Nfi) genes are a family of transcription factors recently described as initiators of gliogenesis and late-stage neuron generation in the mouse spinal cord and retina [60], [61]. Comparative analyses of different CNS regions indicate that Nfia/b/x are generally expressed during late stages of development and play a role in late temporal identity. In the spinal cord, Nfia and Nfib bind near the NeuroD2 gene and are required for the generation of NeuroD2-positive neurons, which arise specifically at late stages of development [58]. In the retina, Nfia/b/x are similarly required for the generation of late-born bipolar cells and Müller glia [62]. The gene regulatory networks of Nfi factors were recently investigated using single cell ATAC-seq and single cell RNA-seq analysis, where Nfia/b/x were shown to promote late-stage RPC competence through changes in chromatin accessibility patterns, specifically in the Nfi DNA binding motif [60]. Using antibodies that recognize all three Nfi factors, Chip-seq analysis identified over 13,000 peaks directly regulated by Nfia/b/x. Most of these peaks were found in open chromatin and enriched in late-stage RPCs and Müller glia differentially accessible regions [60]. Importantly, ectopic expression of Nfia/b/x in early-stage RPCs reduces the number of early-born RGCs while increasing the number of late-born rods, bipolar and Müller glia. It is also sufficient to induce glial gene expression in early RPCs [60]. Conversely, Nfia/b/x triple conditional knockouts show fewer late-born rods at P2, while at P14, Müller glia genes are downregulated, and bipolar neurons are absent [60]. Together, these results support a model in which Nfi family members contributes to confer late temporal identity in RPCs [60], [62].

Part of the miRNA processing pathway, DICER is an enzyme required for the full processing of miRNAs [63] and was shown to play a critical role in the early-to-late transition of RPC competence states [64]. Dicer KO mouse retinas show an increase in RGCs and lack of Müller glial cells. Intriguingly, RGCs are generated at postnatal stages in the Dicer KO retina, well past their normal birth window. Thus, Dicer might be required to control the RGC generation window and to open the Müller glia genesis window. Since DICER is a key element of the miRNA processing pathway, it is likely that this transition is due to temporally regulated miRNA activity. However, as DICER is an enzyme and may have other functions beyond miRNA processing [65], it will be important to evaluate whether DICER gain-of-function during early developmental stages is sufficient to restrict RGC production and favour a premature early-to-late temporal identity transition, potentially leading to precocious production of Müller glia. Nonetheless, recent discoveries have described a variety of miRNAs that appear to regulate RPC temporal competence, as discussed below.

The first evidence that miRNAs may be involved in temporal patterning was highlighted by the activity of noncoding miRNAs targeting xOtx2 and xVsx1 mRNAs during late-stages of Xenopus laevis (African clawed frog) retinogenesis, affecting the generation of rods and bipolar cells [66]. During rodent retinal development, early RPCs and late RPCs have distinct transcriptional profiles [60], and also differ in terms of their miRNA profile [67], but whether these different miRNAs are a cause or a consequence of temporal identity states remains unclear. Indeed, global microarray analysis at a variety of different developmental timepoints (E15, E18, P1, P5, P12, and adult) revealed that, out of the 408 known murine miRNAs, 138 of them are expressed in the retina at some point [67].

Important in the expression of miRNAs is the relationship with genomic clusters. Genomic clusters are miRNA genes that are less than 10 kB apart from each other in the genome that are co-expressed under certain conditions, such as a specific temporal window. At early stages between E15-E18, 8 miRNAs are part of a genomic cluster with the highest expression in the developing murine retina: miR-106a, miR-130b, miR-17, miR-18a, miR-19a, miR-20a, miR-20b, miR-93 [67]. At late stages, 5 miRNAs are part of a genomic cluster with the highest level of expression: miR-129–3p, miR-143, miR-29a, miR-29b, miR-29c [67]. Furthermore, the rate of change of expression of individual miRNAs during development is staggering. mir-29b expression increases 100-fold, whereas mir-18a expression decreases 100-fold during mouse retinal development [67]. Although evidence is missing for most miRNAs to support the idea that they act as a bona fide TIFs, there is substantial evidence suggesting that this is the case for let-7 and miR-29c. let-7 is a heterochromatic miRNA that was first identified to facilitate the neuron-to-glia switch occurring during late stages of retinogenesis [68]. Accordingly, overexpression of let-7, together with miR-125 and miR-9 accelerates developmental time in the retina, alongside a marked increase in rods and late RPCs, and a reduction of early-born RGCs [68]. Knock down of all three miRNAs during embryonic stages increases early born cell types. However, knock down of let-7, miR-125 and miR-9 starting at P1 only increases RGC production, suggesting that there is a limited timeframe during which RPCs can respond to miRNA regulation.