Self-organization underlies developmental robustness in plants

Abstract

Development is a self-organized process that builds on cells and their interactions. Cells are heterogeneous in gene expression, growth, and division; yet how development is robust despite such heterogeneity is a fascinating question. Here, we review recent progress on this topic, highlighting how developmental robustness is achieved through self-organization. We will first discuss sources of heterogeneity, including stochastic gene expression, heterogeneity in growth rate and direction, and heterogeneity in division rate and precision. We then discuss cellular mechanisms that buffer against such noise, including Paf1C- and miRNA-mediated denoising, spatiotemporal growth averaging and compensation, mechanisms to improve cell division precision, and coordination of growth rate and developmental timing between different parts of an organ. We also discuss cases where such heterogeneity is not buffered but utilized for development. Finally, we highlight potential directions for future studies of noise and developmental robustness.

1. Introduction: Development and developmental robustness are self-organized

It has been 100 years since the groundbreaking discovery by Spemann and Mangold[1], a milestone in developmental biology. They found that the transplantation of dorsal blastopore lip to the ventral side of an embryo induced neural fate, and subsequently, a second body axis, in the host tissue. The discovery of induction showed that development happens through self-organization, where patterns or structures are not imposed, but emerge from interactions between components of a smaller scale which by themselves do not have the ability to form these patterns or structures[2,3]. In this example, the number and position of body axes arise from spatiotemporally patterned interactions between the ectoderm and the involuting mesoderm.

Similarly, plant development also relies on cell-cell interactions across spatial domains. A classic example is maintenance of stem cell homeostasis in the shoot apical meristem[4–11] (Fig. 1A). Stem cells are nested within the central zone (the medioapical domain), which proliferate and displace themselves away into the peripheral zone where lateral organs initiate. Stem cell identity in the central zone is maintained by the organizing center beneath. Specifically, WUSCHEL (WUS) is synthesized in the organizing center and traffics through plasmodesmata to the central zone to maintain stem cell fate[6,7]. WUS also activates the expression of CLAVATA3 (CLV3), encoding a small secreted signaling peptide[9,12]. CLV3 diffuses through the extracellular cell wall (apoplast) back to the organizing center to repress WUS expression, thus forming a negative feedback loop which maintains a defined size of the stem cell niche[4,5]. It has recently been shown that the size of the stem cell niche is regulated non-autonomously by lateral organs and systemic signals from the root, highlighting that plants rely on organ-organ interactions for proper development[13,14]. Another common example is the formation of supracellular patterns during plant development. In the elongating hypocotyl, supracellular alignment of the cortical microtubule and global growth along the proximodistal direction relies on cell-cell adhesion, highlighting that these patterns are not instructed but instead arise from cell-cell interactions[15,16] (Fig. 1B). Overall, these examples illustrate that development is a self-organizing process which relies on cell-cell interactions among each other and across spatial domains.

Fig. 1. Development and developmental robustness are self-organized.

(A-B) Examples of self-organization during development. (A) Size of the shoot apical meristem is maintained by cell-cell interactions across spatial domains, mediated by WUS and CLV3 that form a negative feedback loop. Meristem size is also regulated by systemic signals from lateral organs and the root. (B) In the epidermis of the elongating hypocotyl, tissue-wide alignment of cortical microtubules is mediated by cell-cell adhesion, without which the pattern no longer forms.

(C) Development is strikingly robust (iii) despite stochasticity in gene expression (i) and heterogeneity in cell growth and division (ii). The tan protein in (i) represents RNA polymerase, and exp. indicates gene expression via transcriptional bursts.

(D-E) Self-organized systems are more robust against noise. (D) In a top-down system where organ size is regulated by an external ruler, variability in the accuracy of this ruler would result in variable organ size. (E) In a self-organized system where organ size is regulated by a diffusing morphogen produced at the base of the organ, increasing the variance in morphogen production rate does not affect the morphogen gradient or organ size, thanks to morphogen diffusion (brown arrows).

The cellular nature of development, however, presents a unique challenge. Within a cell, key developmental regulators are often present in low copy number, which binds to low copy number target genes (2 for a single-copy gene in a diploid genome) to regulate their expression[17,18]. As a result, gene expression is stochastic and subject to molecular noise[19–21]. In addition, great heterogeneity exists among cells in terms of size, shape, and the activity of growth and division, even within the same organ[22,23]. Despite such heterogeneity, organs often achieve the same size, shape, number, and arrangement. For example, two wings in a fruit fly attain the same size and shape, which is important for flight[24–26]; a Brassicaceae flower develops four petals arranged orthogonally, which forms a stereotypical cruciferous display that attracts pollinators[27,28]. Such phenomenon, where organ-level development is highly reproducible despite stochasticity at the molecular and cellular levels, is termed developmental robustness[29–31] (Fig. 1C). How developmental robustness is achieved has been a fascinating topic for decades and is recently selected as one of the most attractive open questions in plant cell biology[32].

Following the argument that development is a self-organized process, developmental robustness should also be self-organized, arising from cellular properties rather than controlled globally. Intuitively, such a self-organized system provides an easy way to achieve robustness. In a top-down system, for example, where size of an organ is controlled by a ruler, the ruler can be subject to stochastic measurement errors, which would directly create variability in organ size (Fig. 1D). In contrast, in a bottom-up, self-organized system, for example, organ size is controlled by the range of a growth-promoting morphogen secreted at the base of the organ. Fluctuations in morphogen production rate of a single cell can be averaged out among the population of morphogen-producing cells, without affecting the range and distribution of the morphogen (Fig. 1E). Decreased growth rate of a single cell can be buffered by increased growth rate of nearby cells to attain the same final organ size that matches the range of the morphogen. Other examples of robustness provided by self-organization include cell competition in animals, where cell-cell signaling eliminates cells with aberrant genotype or hormone signaling activity[33,34]; and apical dominance in plants, where removal of the apical bud induces the outgrowth of axillary bud to ensure the robustness of flower production against damage[35,36]. Thus, self-organization provides robustness to development which buffers the noise or defect of a single cell or component, unless in the rare case where all cells or components becomes defective simultaneously. Indeed, single mutations that disrupt developmental robustness are often those that affect central cellular processes that occur in all cells, such as chromatin remodeling[37–40], RNA metabolism[41–45], protein synthesis[46–49], and protein folding[50,51]. Disruption of these ubiquitous processes often does not lead to a general growth reduction, but instead cause tissue-specific loss of developmental robustness. For example, during sepal initiation from the floral meristem, disruption of global protein translation does not result in generally smaller sepals but instead increases the variability in sepal size, shape, number, and positioning[49,52]. This again highlights that developmental robustness at the organ level is an emergent property conferred by individual cells.

In this review, we focus on how plants achieve developmental robustness in a self-organized manner. We first discuss the potential sources of noise, including stochastic gene expression and heterogeneity in cell growth and division. We then summarize the cellular mechanisms that buffers against these sources of noise, which has been under extensive study in recent years. Finally, we discuss cases where the stochasticity is not buffered, but instead utilized for developmental benefits. We hope that this review will stimulate research interest in the cellular, mechanistic origins of developmental robustness in plants as well as other systems.

2. Origins of stochasticity

2.1. Stochastic gene expression

Within an organ, gene expression is highly variable among cells, as revealed by single-cell RNA-seq[53] or a fluorescent reporter[54]. For example, in the abaxial epidermis of a developing Arabidopsis sepal, the concentration of a transcription factor, ARABIDOPSIS THALIANA MERISTEM L1 LAYER (ATML1), can vary more than three folds among different cells at a given time point. Moreover, ATML1 concentration within a given cell can fluctuate more than two folds when observed through time[54]. Similarly, imaging of promoter reporters reveals that activity of 35S and UBQ10, two promoters thought to drive ubiquitous gene expression, can have varying activity among cells of the same organ and also fluctuate through time[21] (Fig. 2A).

Fig. 2. Origins of stochasticity.

(A-B) Stochastic gene expression. (A) Stochastic gene expression can be revealed using a single reporter, which is expressed at different levels in different cells and may fluctuate over time. (B) Dual reporters driven by the same promoter can reveal the origins of stochastic gene expression. Under extrinsic noise where different cells receive different amount of external signal, the two reporters should vary concomitantly, between cells and through time. Under intrinsic noise such as stochasticity in transcriptional bursting, the two reporters will vary independently.

(C-F) Heterogeneity in cell growth and division. (C) Cellular growth rate is highly heterogeneous, which may arise from heterogeneity in turgor pressure and cell wall stiffness, two key determinants of cellular growth rate. (D) Cellular growth orientation is heterogeneous, which may arise from stochasticity in the orientation of cortical microtubules. (E) Cell division activity is heterogeneous. Fast-dividing cells can be adjacent to non-dividing cells. (F) Cell division can be imprecise, in terms of the location and orientation of the new cell wall.

The origins of stochastic gene expression among cells have been classified as extrinsic and intrinsic[19,21,55]. Extrinsic factors are the ones acting on the whole cell, such as the amount of RNA polymerase a cell has (which affects its capability of transcription); cell size, cell cycle stage, and cell ploidy; and upstream signals that a cell receives. All these factors can be different from cell to cell, which contributes to the stochasticity of gene expression. However, even in a hypothetical cell population where extrinsic factors are uniform, gene expression is still stochastic, due to intrinsic molecular noise. Such intrinsic factors include stochastic transcriptional bursting[20,56] and chromatin on/off state[57,58]. They may vary for the same locus in different cells and also among different gene copies in the same cell. Extrinsic vs. intrinsic noise can be distinguished by imaging the same cells carrying two different fluorescent reporters driven by the same promoter (Fig. 2B). Under the influence of intrinsic noise, the expression of these two reporters should be uncorrelated, whereas extrinsic noise would vary their expression concomitantly. This dual-reporter system has been used to study gene expression noise in E. coli, yeast, and Arabidopsis[19,21,55]. It was found that the level of extrinsic noise is several folds higher than intrinsic noise for many promoters including 35S, highlighting that cell-to-cell variability in physiological state is a major contributor to variability in gene expression. In conclusion, gene expression is highly variable between cells within individual organs, influenced by extrinsic or intrinsic factors.

2.2. Variability in cellular growth rate and direction

One would expect that stochastic gene expression leads to stochasticity in cellular behavior such as growth. Indeed, live imaging reveals that there is a considerable amount of variability in the growth rate and direction among cells of an organ[22,23,59–61]. For example, in the developing abaxial epidermis of an Arabidopsis sepal, within 24 hours, some cells do not grow at all, while others can grow two to three folds in area; in the cells that grow, their growth directions vary from proximodistal to mediolateral and anywhere in between[22,23] (Fig. 2C, D). Interestingly, growth rate can be highly heterogeneous even for different parts of the same cell[23,59].

It is easy to attribute the cell-to-cell variability in growth to global instructive cues. For example, in the developing sepal, a distal-to-proximal wave of maturation mediated by reactive oxygen species has been used to explain the global gradient in epidermal growth rate in sepals of stage 8 and beyond[22,62]. In leaves, in addition to the proximodistal gradient, cells in the blade region grows faster than cells in the midrib/petiole region[61]. Additionally, specific cell types, such as the stomatal lineage or the trichome, have different growth rates than their pavement cell neighbors[61,63]. However, adjacent cells of the same cell type at the same position within an organ can have vastly different growth rates as well[22,23,60] (Fig. 2C). Such difference could originate from the stochastic expression of growth regulators within cells of the same type. Interestingly, a recent study shows that increased noise in gene expression could lead to increased heterogeneity in cell growth, further supporting that these two sources of stochasticity may be interconnected[64].

In addition to stochastic gene expression, there are biophysical causes of cell growth heterogeneity. The plant cell is inflated by turgor pressure from the inside, encased in a viscoelastic yet plastic cell wall. Cell wall expansion under inflation underlies cellular growth, whereas growth direction is determined by mechanical anisotropy of the cell wall[65,66]. Such mechanical anisotropy, in turn, is guided by the asymmetry in the orientation of cellulose microfibrils in the cell wall, which are synthesized by cellulose synthases whose tracks are oriented by cortical microtubules[66]. Therefore, the orientation of cortical microtubules largely dictates the direction of maximal resistance to stress, which is often perpendicular to the maximal direction of growth. For cell-to-cell variability in growth rate, atomic force microscopy shows that cell wall stiffness can greatly differ between neighboring cells in Arabidopsis leaf, meristem, and in Marchantia gemma[52,67–69]. A recent study found that turgor pressure is also heterogeneous between cells in the shoot apical meristem of Arabidopsis, which is in turn affected by heterogeneity in cell size, topology, and cell wall thickness[68]. Therefore, cell-to-cell variability in growth rate may have its roots in the variability of cell wall stiffness and turgor pressure (Fig. 2C). For variability in growth direction, numerous studies have found that there is considerable heterogeneity in cortical microtubule orientation among cells of the shoot apical meristem[60], sepal epidermis[62,63], and hypocotyl[70]. It is therefore likely that heterogeneity in microtubule orientation may result in heterogeneity in cell wall anisotropy, which translates into heterogeneous growth direction when the cell walls are inflated (Fig. 2D).

2.3. Variability in cell division activity and precision

In addition to growth, cells are also highly heterogeneous in division, specifically whether and how fast they divide. For example, when the developing leaf and sepal epidermis is live-imaged, it was found that some cells do not divide at all between time points while others can divide to generate three or more daughter cells[22,23,61] (Fig. 2E). In the abaxial sepal epidermis, some cells endoreduplicates without divisions to generate giant cells while adjacent small cells keep dividing to remain small[71,72]. For cells that divide, the size at which they enter the cell division cycle is also highly variable, having two-fold differences within the same sepal^69,71. One may hypothesize that such heterogeneity in cell division activity within an organ may be attributed to effects of stochastic molecular noise: for example, stochastic expression of cell-cycle-related genes or imprecision of the cell cycle checkpoint.

Position and orientation of the cell division plane can also be variable (Fig. 2F). In the central zone of the shoot apical meristem, there is no cell type differentiation and therefore symmetric cell divisions are expected. However, a recent study revealed asymmetric cell divisions there, some generating daughter cells that have more than two-fold differences in size[73]. Similar variability exists in cell division plane orientation. In root cell files where anticlinal divisions are expected, cell division orientation can deviate towards the periclinal direction by as much as 20 degrees[74]. These deviations to precise cell division planes have been argued to be intrinsic to the stochastic nature of microtubule dynamics, whereby a cell senses its own shape to determine the orientation of cell division[75].

In summary, cells within an organ display high levels of heterogeneity in cell growth and division, which likely has roots in stochastic molecular processes such as gene expression and microtubule dynamics. We next review how such heterogeneity is buffered to achieve robust organ development.

3. Self-organized mechanisms to achieve developmental robustness

As discussed above, robust organ development is a self-organized phenomenon which relies on properties of each individual cell and their mutual interactions. We now discuss how developmental robustness can be achieved by controlling cell properties and cell-cell interactions. We will highlight recent advances in understanding how stochasticity in gene expression and heterogeneity in cellular growth and division can be buffered to achieve robust organ size, shape, and patterning.

3.1. Buffering stochasticity in gene expression

Developmental robustness can be achieved by actively decreasing noise in gene expression (Fig. 3A). The Polymerase-Associated Factor 1 Complex (Paf1C) is a conserved complex that regulates diverse aspects of gene expression, including chromatin accessibility, transcriptional elongation, and post-transcriptional mRNA regulations[76]. Earlier, it was found that association of Paf1C to Polymerase II in yeast is important for reducing gene expression noise[77]. A similar role of Paf1C in reducing gene expression noise has been recently found in plants[64]. Mutation of VIP3, which encodes a Paf1C subunit, increases stochastic gene expression. This increased gene expression noise results in local growth rate heterogeneity, causing local mechanical conflicts, which dilutes the global mechanical cues required for robust organ shape[62]. The increased gene expression variability also creates irregular patches of reactive oxygen species (ROS), which normally forms a precisely timed and localized propagating wave that controls the robust termination of organ growth[22]. As a result, the vip3 mutant shows increased variability in the size, shape, and curvature of sepals[64]. Thus, the Paf1C complex denoises gene expression, which reduces local variability in mechanical and chemical signals, and promotes the robust control of organ development by global mechanical and chemical gradients[64]. Similarly, Paf1C has been shown to promote robustness in phyllotaxy, the number of cotyledons and floral organs, and termination of the floral meristem, by reducing variability in the expression of key regulators[78,79]. Overall, these studies suggest that denoising of gene expression by Paf1C may represent a universal mechanism of developmental robustness.

Fig. 3. Self-organized mechanisms to achieve developmental robustness.

(A-B) Buffering stochasticity in gene expression. (A) The Paf1C complex denoises gene expression to achieve developmental robustness at the organ level. Without Paf1C, gene expression becomes variable, and developmental robustness of the organ is disrupted. (B) Opposing small RNA gradients buffer against stochastic gene expression to form sharp boundaries between opposing cell fates.

(C-D) Buffering stochasticity in cellular growth. (C) Large spatial correlation of growth rate results in variable tissue shape (top). In contrast, when spatial correlation of growth rate is reduced, i.e., fast-growing cells are interspersed among slow-growing cells, spatial averaging of growth rate is achieved, and tissue shape remains robust at a larger scale (bottom). A plausible mechanism is growth rate buffering by neighboring cells. White arrows indicate fast cell expansion, and black arrows indicate neighboring cell deformation. (D) Temporal averaging of growth rate is achieved when cellular growth rate fluctuates above and below average over time, so that cellular growth rate is homogeneous within a longer time period. A plausible mechanism is mechanical shielding by neighboring cells, mediated by rearrangements of cortical microtubules.

(E-F) Buffering stochasticity in cell division. (E) Organ size and shape is robust against changes in cell division rate because global growth gradients remain unchanged. (F) Imprecision in cell division is reduced by the preprophase band (PPB), which marks the future division site. After division, KRP4, a cell cycle inhibitor, further corrects imprecision in cell division by serving as a ruler of cell size upon cell cycle entry.

(G-H) Mechanisms of developmental robustness at the whole organ level. (G) During sepal (S) development, growth rates of the abaxial (Ab, blue) and adaxial (Ad, red) sides need coordination to achieve robust sepal shape and curvature. Overgrowth of the adaxial side causes premature outward bending, while overgrowth of the abaxial side causes buckles. FM, floral meristem. (H) During sepal initiation from the floral meristem, robustly concentrated auxin signaling (yellow dots) ensures coordinated initiation timing of the four sepal primordia, leading to sepal primordia of equal size. In the drmy1 mutant, the incipient outer sepal forms auxin maxima and initiates first, while these processes are usually delayed for the inner and lateral sepals. The uncoordinated initiation timing causes increased size variability for sepals within the same flower and inability to close.

Noise in gene expression can also be buffered by small RNAs, providing robustness to development (Fig. 3B). An example is the maintenance of abaxial-adaxial polarity in leaves. Leaf blade outgrowth at the margin relies on the juxtaposition between abaxial and adaxial cell fates. These fates are determined respectively by abaxial identity genes such as KANADI and AUXIN RESPONSE FACTOR (ARF) 2/3/4, and adaxial identity genes such as HD-ZIPIII and AS2[80]. Expression of HD-ZIPIII genes is inhibited by miR165/166, which are expressed on the abaxial epidermis and diffuses to form a gradient along the abaxial-adaxial axis. Similarly, an opposing gradient of tasiR-ARFs, short interfering RNAs targeting ARF genes, along the adaxial-abaxial axis restricts their targets, ARF2/3/4, to the abaxial side. Genetic disruption of these small RNA gradients disrupts abaxial-adaxial cell fate boundary. Importantly, instead of a uniform shift of the boundary towards either side, the resultant boundary is irregularly shaped, with adaxial fate cells interspersed on the abaxial side, and vice versa[81]. This indicates a loss of robust fate determination, potentially revealing the effect of noisy gene expression. It was then shown by a computational model that a molecular network including these small RNAs and their targets can provide robustness to abaxial-adaxial boundary positioning against up to ~10% noise (coefficient of variation) in gene expression[82]. Thus, these studies suggest that the two opposing small RNA gradients buffer against stochastic gene expression and promote the formation of a sharp abaxial-adaxial boundary which is crucial for leaf flatness. In summary, cells buffer against stochastic gene expression by a range of cell-autonomous (Paf1C-mediated denoising) and non-cell-autonomous (small RNA gradients) mechanisms.

3.2. Buffering stochasticity in cellular growth

In addition to noisy gene expression, noise in cellular growth also needs buffering to achieve robust organ size, shape, and number. Such buffering can be achieved by spatiotemporal averaging[22,23] (Fig. 3C, D). In the abaxial epidermis of a wild-type Arabidopsis sepal, cellular growth rate is highly heterogeneous among cells. However, fast- and slow-growing cells do not form separate clusters, but instead are interspersed among each other, such that growth rate is relatively homogeneous at a larger scale, a process called spatial averaging (Fig. 3C). Similarly, growth rate of a cell varies through time, and a period of fast growth can be followed by a period of slow growth, such that growth rate is relatively constant at a larger time scale, which is termed temporal averaging (Fig. 3D). Such spatiotemporal averaging allows development of robust sepal size and shape despite great heterogeneity in cellular growth rates. Similarly, principal direction of cellular growth is also highly heterogeneous across the sepal epidermis, but is averaged out through time to achieve uniform proximodistal growth at the whole organ level. In the ftsh4 mutant, a mitochondrial defect results in high levels of ROS in patches of cells, creating long-range correlations of growth rate and direction that are highly variable from patch to patch. The increased spatial scale of correlation reduces the effect of spatial growth averaging and disrupts the robustness in sepal size and shape. A related mechanism is growth compensation. In a recent study, growth of cell clones derived from the same progenitor cells were tracked over time. It was found that clones derived from smaller cells may subsequently grow faster, in a way similar to temporal growth rate averaging, which reduces variability in clonal size[83]. Thus, growth averaging is one of the common mechanisms buffering the heterogeneity in cell growth and promoting organ development robustness.

What is the nature of growth averaging? Why are slow-growing cells adjacent to fast-growing cells, and why do fast-growing cells slow down at the next time point? There is increasing evidence of local growth buffering which may underlie the observed growth averaging (Fig. 3C). For example, in the epidermis of developing leaves, sepals, stamens, and carpels, the stomatal lineage undergoes rapid cell expansion relative to the pavement cells. This rapid expansion is buffered by the neighboring cells, which yield and slow down in growth[61]. Similarly, in the leaf epidermis, trichome cells expand slowly in the X-Y plane after they bulge out, and this slow growth is buffered by fast-growing neighboring pavement cells[61]. Thus, growth buffering among cells neighboring to each other provides a mechanism for spatial growth averaging.

On the other hand, local mechanical shielding provides a mechanism for temporal growth averaging (Fig. 3D). During sepal development, trichome cells grow rapidly before they bulge out, but their growth in the X-Y plane slows down after they bulge out[61,63]. What underlies this change in growth rate that allows temporal averaging? A recent study found that the fast-growing trichome cell compresses the neighboring pavement cells. The compression served as a mechanical cue which reorients the cortical microtubules in neighboring cells circumferentially around the trichome cell[63]. As mentioned above, oriented cortical microtubules guide the anisotropic deposition of cell wall material[66]. The increased circumferential cell wall strength in neighboring cells slows down the growth of the trichome cell in the X-Y plane[63]. In the katanin mutant with reduced mechanical response, rapid growth of trichome cells is sustained over time, which reduces temporal growth rate averaging and increases variability in mature sepal shape[63]. Interestingly, in the spiral2 mutant where microtubules respond to mechanical cues faster, mature sepals also show increased variability in shape[63]. Further computational modeling and experimental evidence suggests that while a moderate amount of mechanical response can increase growth rate homogeneity, a further increase in the strength of mechanical response can amplify growth differences and promote growth rate heterogeneity[60,70,84]. Therefore, an optimal level of mechanical response in neighboring cells is important to decrease the growth rate of a fast-growing cell to achieve the right amount of temporal growth rate averaging.

In summary, variability in growth rate and direction can be buffered by changes in growth rate and mechanical properties of neighboring cells, which achieves spatiotemporal growth averaging important for robust organ size and shape. This again highlights that developmental robustness emerges from local cell-cell interactions in a self-organized manner.

3.3. Buffering stochasticity in cell division

In contrast to cell growth, heterogeneity in cell division activity seems to have little effect on the robustness of organ size and shape. As discussed above, the Arabidopsis sepal epidermis is composed of giant cells which endoreduplicate, interspersed among small cells which divide[71,72]. Earlier studies suggest that sepal development is robust against such heterogeneity in cell division activity[71,85]. Genetically increasing the proportion of giant cells decreases the number of cells per sepal, resulting in almost no change in mature sepal size[85]. Recently, it was found that changing global cell division activity by changing the level of a cell cycle inhibitor LGO does not alter global growth gradients in the epidermis of the developing sepal, and thus has no effects on mature sepal size and shape[23] (Fig. 3E). This is in contrast to the ftsh4 mutant, where patches of cells with aberrant growth rate and direction increases variability in mature sepal size and shape[22,23]. Interestingly, although almost all sepal epidermal cells in plants overexpressing LGO in the epidermis (pATML1::LGO) are giant cells, different sub-regions along the proximodistal axis of these giant cells still follow the global proximodistal growth gradient as in wild-type sepals[23]. These studies suggest that changes in cell division activity have little effects on the robustness of organ size and shape, because they are often compensated by concomitant changes in cell number and also masked by unaltered global growth patterns.

Cells also have mechanisms to buffer against imprecisions in division plane positioning and orientation (Fig. 3F). In the Arabidopsis shoot apical meristem, during cell division, daughter cells inherit the same amount of KIP-related protein 4 (KRP4), a cell cycle inhibitor. Cell expansion dilutes KRP4 concentration, eventually triggering the next cell division. In asymmetric cell divisions where division plane positioning is biased by noise, the larger daughter cell begins with a lower KRP4 concentration, reaching the KRP4 threshold sooner and dividing sooner, whereas the smaller daughter cell requires longer periods of growth to reach the same size that sufficiently dilutes KRP4 to trigger cell division. Thus, the stochasticity in cell division symmetry is buffered against by the equal inheritance of KRP4, which serves as an internal scale to ensure robust cell size at the time of division[73].

On the other hand, the preprophase band contributes to the precision of cell division plane orientation (Fig. 3F). The preprophase band is a structure formed in plant cells prior to division and marks the future division plane. It has recently been found that a mutation that disrupts the preprophase band increases the variability of cell division plane orientation[74]. These mutant plants also show greatly increased variability in the number of cotyledons and floral organs. This study suggests that, in plant cells where errors in division orientation cannot be corrected by movement of daughter cells, the formation of the preprophase band which marks the future division plane is important for its robust orientation, which in turn is important for the robustness of plant development in general.

In summary, changes in cell division activity are compensated for by changes in cell number and masked by unchanged global growth patterns, and noise in cell division plane is either reduced by structures formed pre-division or compensated for by post-division growth.

3.4. Mechanisms of developmental robustness at the whole organ level

We have seen noise buffering mechanisms such as spatiotemporal growth averaging and small RNA-mediated denoising of gene expression, where cell-cell interaction self-organizes to give rise to developmental robustness at the organ-level. Self-organization can also occur from interaction between different domains of an organ or different organs of the same individual to achieve developmental robustness, as evidenced by earlier work in Drosophila or mice[86–88]. Recent work in plants reveal similar coordination of growth and key developmental events which is important for robustness.

First, growth between different parts of an organ needs coordination (Fig. 3G). Recent work shows that growth coordination between the abaxial and adaxial epidermis of an Arabidopsis sepal is important for its robust shape and curvature[89,90]. In mutant screens for altered sepal shape and curvature, two mutants were recovered which disrupts the spatial expression pattern of two abaxial-adaxial polarity genes and reduces growth coordination. The vip4 mutant shows abnormal outward bending of the sepal. The vip4 adaxial epidermis overexpresses ARF3 (an abaxial fate determinant) which increases auxin signaling and downregulates a cell wall-stiffening enzyme. As a result, the adaxial epidermis is softened and overproliferates compared to the abaxial side, causing the outward bending[89]. On the other hand, the as2–7D mutant shows lumps and folds on the abaxial epidermis of the sepal. The as2–7D mutant overexpresses AS2 (an adaxial fate determinant) on the abaxial epidermis, which causes it to soften and overgrow compared to the adaxial side. In addition, cellular growth direction of the abaxial epidermis deviates mediolaterally, compared to the adaxial epidermis which grows proximodistally. As a result of this loss of coordination of growth rate and direction, the abaxial sepal epidermis buckles and forms ridges and valleys[90]. Therefore, coordinated growth rate and direction between the abaxial and adaxial sides of a sepal, mediated by precise spatial expression patterns of polarity genes, is important for developing robust organ shape and curvature.

Second, timing of key developmental events needs to be coordinated between different organs of the same individual. In an Arabidopsis flower, four sepals initiate from the floral meristem and grow to the same mature size to fully enclose and cover the inner floral organs before anthesis. Recent studies suggest that both the coordinated initiation and coordinated maturation of the four sepals are important for robustness in mature sepal size[22,49,52,91]. During wild-type sepal maturation, an arrest front of high ROS accumulation propagates proximally from the sepal tip, inhibiting cell proliferation and promoting differentiation[22,62], similar to the morphogenetic furrow in the developing Drosophila eye[92]. This arrest front is triggered at the same time and propagates at the same speed for different sepals in the same flower, coordinating sepal maturation and resulting in sepals of the same mature size. In mutants for the mitochondrial protease FtsH4, sepals accumulate ROS variably instead of a uniform, coordinated wave. Different regions of the same sepal and different sepals of the same flower accumulates different levels of ROS. As a result, maturation timing is uncoordinated within the same sepal and between sepals, resulting in loss of robustness in mature sepal shape and size[22].

Coordination of initiation timing is also important for robust mature sepal size (Fig. 3H). In wild-type flowers, the four sepals initiate from the floral meristem within a short time window of 12 hours, allowing them to grow to the same mature size. In a mutant screen for variable mature sepal size, the drmy1 mutant was recovered. In drmy1, the timing of sepal initiation is highly variable and uncoordinated, with some sepals initiating 18–36 hours later than others. The uncoordinated sepal initiation timing results in variable amounts of outgrowth for sepal primordia within the same flower. Smaller sepal primordia in drmy1 flowers remain small throughout development, resulting in mature sepals of variable sizes that are unable to fully cover the inner floral organs[49,52]. Further, it was found that the coordinated initiation of four sepals in wild-type flowers originate from the four focused auxin maxima in the floral meristem prior to sepal initiation. In contrast, drmy1 floral meristems form focused auxin maxima at some places and diffuse, noisy bands of auxin signals at others, which originates from defects in TARGET OF RAPAMYCIN (TOR) signaling and protein synthesis[49,52]. These diffuse bands of auxin signaling require additional time to coalesce into distinct, focused auxin maxima, which delays sepal initiation at these places[91]. Thus, the coordinated formation of four focused auxin maxima underlies the coordinated initiation timing of the four sepals in wild type, which is critical for robust mature sepal size.

Further, CUP-SHAPED COTYLEDON (CUC) genes, encoding boundary-specific transcription factors[93,94], have been shown to interact with auxin to modify developmental timing and affect robustness[91]. In wild-type floral meristems, CUC1 is expressed surrounding auxin maxima. It increases the polar auxin transport towards these auxin maxima and accelerates sepal initiation. In cuc1 mutants, auxin maxima are reduced, and sepal initiation is uniformly delayed. Interestingly, introducing the cuc1 mutation into drmy1 restores four focused auxin maxima and the initiation of four sepals, as opposed to 3–5 in the drmy1 single mutant. Live imaging reveals that a general delay in sepal initiation caused by the cuc1 mutation, and thus reduced auxin maxima, allow additional time for noisy auxin signaling to canalize to the correct pattern. Thus, a general delay in developmental timing allows canalization of morphogen signals which is important for developmental robustness. This hypothesis fits the previous observation that upregulating CUC protein accumulation by disrupting the negative regulation by miR164 reduces developmental robustness in floral organ number, phyllotaxy, carpel fusion, and leaf serration[95–98].

In summary, to achieve robust organ development, coordination is needed in different parts of the same organ and different organs of the same individual. Timing of key developmental events, such as organ initiation and maturation, needs to be tightly controlled for robust organogenesis. These whole-organ-level mechanisms of developmental robustness operate on top of local, cell-cell-interaction-based mechanisms that we reviewed earlier, and they likely also arise from local cell properties (such as protein synthesis rate, in the example of coordinated sepal initiation[49]) which demands further study.

4. Noise amplification and its developmental function

In some cases, heterogeneity in gene expression or cellular growth and division needs to be buffered to achieve robust organ development. In others, such heterogeneity is amplified and serve important developmental roles[99].

4.1. Amplification of stochastic molecular noise necessary for developmental decisions

It has long been known in the stem cells field that stochastic gene expression drives “probabilistic differentiation”, which is a way to achieve division of labor in a population of otherwise identical cells[100]. A famous example is the differentiation of trophectoderm, epiblast, and primitive endoderm lineages in the early mouse embryo, which is first driven by stochastic expression of fate determinants followed by cell sorting[101–104]. A similar mechanism has recently been found during the differentiation of the anterior vs. posterior somite compartments in the human segment organoids[105]. In plants, stochastic gene expression has been implicated during the differentiation of sepal epidermal cells[54] (Fig. 4A). In Arabidopsis, the abaxial sepal epidermis is composed of giant cells, which are large, endoreduplicated cells important for proper sepal curvature and opening, interspersed among small cells[71,72]. It was recently found that stochastic fluctuation of the concentration of a transcription factor ATML1 dictates the giant cell fate. Specifically, ATML1 concentration fluctuates through time and is also heterogeneous among cells. When ATML1 concentration passes a threshold by chance during G2 phase, the cell endoreduplicates to become a giant cell. In contrast to animals where cells sort to form clusters of different fates after stochastic fate determination, stochastic fluctuations of the giant cell specifier ATML1 leads to random distribution of giant cells scattered throughout the abaxial epidermis of the sepal[54].

Fig. 4. Noise amplification and its developmental function.

(A) Stochastic fluctuations of ATML1 level mediate the differentiation of giant (G) vs. small (S) cells. Cells in which ATML1 expression passes a threshold (Thres) by chance during the G2 phase are more likely to endoreduplicate to become giant cells. (B) Stochastic molecular noise seeds Turing pattern that underlies trichome specification in leaves. (C) Molecular noise is utilized to increase variability in seed germination timing. ABA promotes its own synthesis and degradation at the same time, which generates stochastic noise in its level. ABA is involved in a bistable switch with GA, which amplifies noise and creates a range of germination timing. Variability in germination timing is a bet-hedging strategy to ensure survival of the whole seed population. (D) Amplification of local growth rate differences help the formation of sharp boundaries between the shoot apical meristem (SAM) and the initiating primordia (P). (E) Amplification of growth direction differences mediates out-of-plane deformation of the ventral corolla in the snapdragon flower to form the characteristic fold.

Another example is the stochastic determination of trichome fate in the leaf epidermis[106] (Fig. 4B). Trichome fate specification involves more than 10 regulators in a Turing-like activator-inhibitor and activator-depletion network[107]. However, a purely deterministic Turing model produces evenly distributed trichomes, unlike the real biology where trichome distribution can be somewhat noisy[108]. A recent theoretical study resolved this paradox by introducing molecular noise (i.e. stochastic fluctuations in the production and degradation of the activator and the inhibitor molecules) to the Turing model. When molecular noise is included, the model produced more realistic distribution of trichomes compared with deterministic models[106]. This study supports the idea that stochastic molecular noise seeds the formation of Turing patterns in the leaf epidermis. Whether noisy trichome distribution pattern is merely a byproduct of molecular noise that has not been optimized by evolution or is actually beneficial compared to evenly distributed trichomes remains an open and intriguing question.

In addition to its role in cell fate determination, stochastic molecular noise can also be utilized to make whole-plant developmental decisions. A textbook example is the silencing of FLC during vernalization[57,58]. In this case, stochastic, digital switching of FLC expression in each individual cell is utilized to encode a continuum of vernalization memory at the whole plant level. In addition to vernalization, stochastic molecular noise is also utilized during seed germination[109,110] (Fig. 4C). Seed germination timing is controlled by two plant hormones, ABA and GA, which interact antagonistically within the seed radicle[111]. Theoretical and experimental work shows that a feedback mechanism involving ABA promoting its own biosynthesis and degradation at the same time can amplify stochastic molecular noise to generate fluctuating ABA levels[109]. Another study revealed that a bistable switch involving ABA and GA antagonistically regulating the levels of germination inhibitors can further amplify noise to create great variability in germination timing, in a population of genetically identical seeds[110]. Such variability ensures that, in fluctuating environmental conditions, by chance some seeds will germinate at the most favorable environmental condition, ensuring the survival of the seed population. In summary, in the cases mentioned above, stochastic molecular noise is not buffered but instead utilized in cell differentiation or whole-plant decision making, providing fitness benefits. It is intriguing to ask how common each scenario is during development, where noise is either buffered or utilized.

4.2. Amplification of local growth differences during organ morphogenesis

Local growth differences can also be amplified to serve a developmental function (Fig. 4D, E). When the Arabidopsis shoot apical meristem produces lateral organs, sharp boundaries consisting of two cell files separate the new lateral organ from the shoot apical meristem. Such sharp boundaries rely on heterogeneous growth – boundary cells grow slowly compared to neighboring, non-boundary cells. Such heterogeneity, in turn, is amplified from the initial growth rate differences typical of the peripheral zone of a WT shoot apical meristem. This amplification relies on a strong cellular response to mechanical stress mediated by the reorientation of cortical microtubules[60,70,84] (Fig. 4D). In the katanin mutant where this mechanical response is weakened, growth rate is more homogeneous between boundary and non-boundary cells. As a result, blunt boundaries consisting of three or more cell layers form[60]. Thus, amplification of growth rate differences in WT promotes boundary formation and lateral organ initiation.

In addition to heterogeneity in growth rate, heterogeneity in growth direction can also be utilized for morphogenesis. During the development of a snapdragon flower, the ventral corolla folds adaxially to form a characteristic wedge. Mutation of the ventral identity gene DIVARICATA (DIV) abolishes the wedge, replacing it with small bulges. These out-of-plane deformations are the result of heterogeneity in growth direction – cells that grow proximodistally are adjacent to those growing mediolaterally. This heterogeneity generates small bulges in the div mutant, and, when amplified by DIV, generates large wedges in WT flowers[112] (Fig. 4E). Overall, in these examples, heterogeneity in growth rate and direction is not averaged out but rather amplified to generate diverse organ shapes.

5. Conclusion

Development is a self-organized phenomenon, building upon individual cells and their mutual interactions. Such self-organization presents a unique challenge: How do cells cope with the inherent heterogeneity in gene expression, growth, and division to achieve robust development? There is no external ruler or blueprint which dictates how an organ or organism develops, and thus developmental robustness must also be achieved through self-organization. In this paper, we have reviewed mechanisms in which cells buffer against heterogeneity to achieve developmental robustness. In summary:

• Stochastic gene expression can be buffered at the transcriptional or post-transcriptional levels, cell-autonomously or non-cell-autonomously.

• Heterogeneity in growth rate and direction can be buffered by spatiotemporal averaging, where neighboring cells mechanically interact to homogenize cellular growth.

• Cell division imprecision can be reduced by pre-division or post-division mechanisms to improve organ-level developmental robustness, whereas changes in cell division activity without affecting cellular growth has little impact on organ-level robustness.

• Coordination of growth rate and developmental timing is also needed, between parts of an organ or different organs of the same individual, to achieve developmental robustness.

We also note that noise can be amplified to serve important developmental functions, suggesting that a multicellular collective finds a balance between buffering noise to achieve robustness vs. utilizing noise for development.

Developmental robustness is an intriguing research topic on which we encourage more studies, both on the origin of cellular stochasticity and on the mechanisms of achieving robustness. Methodologically, we encourage researchers to consider the following.

First, most of previous research has been observing noise or heterogeneity, which is correlated with how cells respond to such noise or heterogeneity to infer mechanisms of developmental robustness. For example, how neighboring cells respond to a fast- or slow-growing specialized cell type (trichomes or stomata) has been used to infer how cells would respond to a fast- or slow-growing neighbor, but the conclusions may only pertain to the investigated cell type and may not be easily generalizable. In the future, it will be advantageous to adopt ways to modulate growth rate in a more manually controllable manner. For example, a Cre-Lox recombination system or optogenetics tools can be used to create patches of cells expressing a cell wall stiffness regulator, to precisely modulate growth rate in these cells and observe behavior in neighboring cells. It would help address the following questions: (1) How do cells respond to general growth heterogeneity? (2) At what developmental stage is noise buffering most critical? (3) What is the extent of noise that can be buffered by a proposed mechanism?

Similarly, to investigate the role of stochastic gene expression during development or mechanisms that buffer it, mutations (such as vip3[64]) or drug treatments (such as idoxuridine[56]) that increase stochastic gene expression can come handy. On the other hand, we caution the interpretation of the effects of these mutations and drugs, because (1) many manipulations that change gene expression variability also change the mean expression at the same time, and (2) these manipulations may affect genome-wide gene expression noise, so the phenotypic effect may not be easily attributed to increased gene expression variability of a single gene-of-interest. To better study the effects of individual gene expression stochasticity, researchers may create constructs of a gene of interest with altered gene expression variability, such as by altering promoter structure in the construct[113].

Thirdly, previous studies on the mechanisms of developmental robustness have largely relied on mutant screening, i.e., by identifying mutants with altered developmental robustness. Similar to traditional functional analysis, this approach suffers from genetic redundancy. One would imagine that key mechanisms that govern developmental robustness are likely genetically redundant and cannot be revealed by such mutant screens. We encourage the use of high-throughput sequencing, mass spec, and imaging approaches to complement traditional genetic analysis of mechanisms that promote developmental robustness.

Lastly, we suggest that mechanisms of developmental robustness are best studied using a combination of experimental observations, perturbations, and mathematical modeling. As mentioned above, observations of development establish correlation, while causality requires perturbations that change the level of noise or the proposed noise-buffering mechanism. On the other hand, mathematical modeling reduces the study system to its simplest form, and tests what components and interactions are minimally sufficient for the proposed mechanism. In addition, because development and developmental robustness are self-organized processes, the model may yield predictions that are not straightforward when individual parts of the system (e.g., cells) are separately considered. These predictions can then be experimentally tested. However, we also caution that mathematical models need to be based on biologically relevant assumptions and parameters and be as simple as possible, as one can potentially develop a model with unconstrained assumptions and numerous tunable parameters to make any desirable predictions.

We look forward to future studies of developmental robustness. We hope that they further reveal how cell-cell interactions overcome cellular heterogeneity to produce organs of robust size, shape, and patterning in a self-organized manner.

Highlights.

• Cells in an organ are heterogeneous in gene expression, growth, and division

• Noise in gene expression is buffered transcriptionally and post-transcriptionally

• Noise in cellular growth is buffered by spatiotemporal averaging

• Noise in cellular division is buffered by pre-division and post-division mechanisms

• Coordination of growth rate and developmental timing is also crucial for robustness