Stochasticity

Stochasticity and Cell Fate

Richard Losick Department of Molecular and Cellular The Biological Laboratories Harvard University 16 Divinity Ave. Cambridge, Massachusetts 02138

Claude Desplan Center for Developmental Genetics Department of Biology New York University 1009 Silver Center, 100 Washington Square East New York NY 10003

“I, at any rate, am convinced that He does not play dice.” Albert Einstein, 1926

Summary Fundamental to living cells is the capacity to differentiate into subtypes with specialized attributes. Understanding the processes by which cells acquire their fates has been, and continues to be, a major challenge in the field of . How cells adopt a particular fate is usually thought of as being deterministic, and indeed in the large majority of cases it is. That is, cells acquire their fate by virtue of their lineage or their proximity to an inductive signal from another cell. In an increasing number of cases, however, and in organisms ranging from bacteria to humans, the fate that cells adopt is emerging as a stochastic process in which cells choose one or another pathway of differentiation

1 Stochasticity without any apparent regard to their environment or their history. Stochasticity has several important mechanistic requirements as we will discuss. We will also speculate on why stochasticity is advantageous, and even critical in some circumstances, to the survival of the individual, the colony, or the species.

Stochasticity in Bacteria and Flies Classic systems for the study of development offer numerous examples of in which cell fate is not left to chance. For example, the bacterium Caulobacter crescentus gives rise to a motile, swarmer cell and a sessile, stalked cell every time it divides [1]. The generation of progeny with distinct cells fates is hard wired into the bacterium’s cell cycle. Likewise, the capacity of the yeast Saccharomyces cerevisiae to switch mating types [2] and of the fly Drosophila melanogaster to generate neurons and glial cells is governed by intrinsically asymmetric processes of cell division [3]. Also not left to a roll of the dice is the decision to become a photoreceptor in the fly eye; in this case, cell fate is determined by the proximity of a precursor cell to a signaling peptide emitted by another nearby cell [4]. In striking contrast are the examples of cell fate determination illustrated in Fig. 1: entry into the state of competence by the bacterium Bacillus subtilis and the generation of alternative color vision photoreceptors in D. melanogaster. Although these two systems could not be more different, they have in common that the choice of cell fate is made stochastically. The left-hand panel shows a field of B. subtilis cells that contains a fusion of DNA encoding the Green Fluorescent to the promoter of a gene that is under the control of the DNA-binding protein ComK, the master regulator for entry into the state of competence [5, 6]. Competence is a specialized cell state involving the expression of about 100 genes. In the competence state, growth ceases and the cells become capable of taking up DNA from the environment and incorporating it into the chromosome by recombination. About twenty percent of the cells are active for ComK and the rest are not. Each cell makes a binary choice between these two states randomly [7-10]. Presumably, entering the state of competence

2 Stochasticity imparts a fitness advantage to B. subtilis that outweighs the cost of producing cells that temporarily stop growing. Interestingly, whereas the choice to entry into competence is made stochastically, exit from competence and resumption of growth occurs after a relatively fixed period of time [11]. Thus, competence exhibits both non-deterministic and deterministic features. The right hand panel shows the retina of Drosophila melanogaster, which has a compound eye composed of multiple unit eyes known as ommatida. Each ommatidium consists of eight photoreceptor cells [12]. In each ommatidium a stochastic choice is made in one of the photoreceptors (called R7) to become one of two possible cell types [13]. Once this choice is made, the R7 cell instructs the photoreceptors lying directly underneath it in the ommatidium (called R8) to express either a blue-sensitive or a green-sensitive rhodopsin photopigment, as shown with antibodies against the two rhodopsins [14]. Here too, the choice is made randomly. A nearest-neighbor analysis shows that the chance of belonging to one fate is not influenced by the decision made by the neighbors: each ommatidium makes its choice independently [15]. For both the bacteria and the ommatidia, the choice is not simply the equivalent of flipping a coin. Instead, it is biased: for the bacteria, the ratio of competent to non-competent cells is about 20:80 (actually, in bacteria that have not been domesticated in the laboratory, the proportion of competent cells is much lower) whereas, for the ommatidia, the ratio of the blue to the green subtype is 30:70. Interestingly, the 30:70 ratio of blue to green ommatidia is conserved between Drosophila and the house fly (Musca) in spite of over 120 millions years of independent evolution [16]. This observation suggests that the precise ratio has an important biological meaning. What is this meaning and what are the mechanisms underlying these stochastic choices [9]?

Noise and bistability Stochastic choices in cell fate require both a means to generate biological noise and mechanisms to stabilize decisions reached in response to it. Noise can arise from multiple sources, such as variations in the activity of individual genes within a cell or from cell-to-cell variations in overall metabolic activity, or

3 Stochasticity from fluctuating levels of an external signal. For example, a given cell might enter into the state of competence (see above) as a response to noise in the intrinsic transcriptional activity of the gene for the competence regulatory protein ComK [9]. Noise alone is insufficient to create binary switches between alternative cell fates. Fluctuations due to noise are generally small and transient; what is also needed are mechanisms to amplify these fluctuations and then to stabilize one choice or another. Systems of this kind are said to be bistable, that is, the system persists in one or another alternative state for prolonged periods of time [17]. Bistable systems often exhibit a kind of memory known as hysteresis. Put simply, hysteresis indicates that when a switch is thrown in one direction in response to a signal, it does not readily switch back when that signal is removed (much as when ferromagnetic material becomes magnetized in response to a magnetic field, it remains magnetized when the field is removed). Bistability ensures that once the switch is thrown, the circuit remains locked in a stable state. Bistability can be achieved by simple positive autoregulatory loops (Fig. 2A) or double negative loops (Fig. 2B), or by more complex circuits, comprising several intermediary loops (Fig. 2C, see below) [18]. A classic example of a bistable system is the alternative lytic and lysogenic states of the bacterial virus lambda [19]. The virus is locked into lytic or lysogenic modes by mutually antagonistic repressors that inhibit each other’s synthesis. When one repressor takes over, even weakly, the system switches for long periods of time in one direction (Fig. 2D). Bistability also requires mechanisms to render the switch hypersensitive, allowing a rapid and dramatic response once a threshold has been attained. In phage lambda, this is achieved by cooperative DNA-binding interactions among repressor molecule. For competence, production of the regulatory protein (ComK) is controlled by a positive feedback loop in which ComK stimulates its own synthesis [5, 7]. In this case, hypersensitivity is also achieved by cooperative DNA binding of ComK to its promoter. What these bistable systems have in common is that a regulatory switch is poised on a knife edge and can flip in one direction or the other when pushed by noise.

4 Stochasticity

As mentioned above, stochastic choice is not simply the equivalent of flipping a coin. It is often biased in one direction or the other. In the case of competence, the basis is favor of non-competence is probably related to the basal level of transcription of the ComK gene [9]. A high basal level would allow more cells to attain the threshold necessary to set the positive feedback loop in motion whereas a low basal level would bias against high level production of ComK. Bistability need not be governed by cell autonomous mechanisms. Sometimes, as we shall see, bistability is achieved by interactions between cells (Fig. 4) [20].

Cell autonomous choices Why is choosing cell fate stochastically advantageous? We address this question first in the case of stochastic choices that are made cell-autonomously. Perhaps the most attractive explanation comes from studies of stochastic switches in bacteria. It is well known that bacteria respond to adverse environmental conditions by inducing the expression of adaptive genes. Stochasticity allows bacteria to deploy different types of specialized cells in anticipation of possible adverse changes in the environment. A striking example is the persister state, which is observed in E. coli and many other bacteria, including pathogenic bacteria [21, 22]. Populations of E. coli cells are found to contain a tiny subpopulation of cells that have temporarily entered non-growing or slow growing states in which they can elude the action of antibiotics, which often require bacterial cells to be in a state of active growth to cause killing. Thus, when a population of E. coli cells is treated with, for example, the antibiotic ampicillin, the persister cells survive by virtue of their quiescence. Cells that exit the persister state after the antibiotic treatment has ended resume growth. An appealing interpretation of this phenomenon is that E. coli is hedging its bets against the future possibility of encountering antibiotics. If it waited to respond until after the antibiotic was present, it would be too late to adapt and the entire population would die. It is believed that persisters contribute to the difficultly of sterilizing patients infected with a pathogenic bacterium by antibiotic treatment,

5 Stochasticity especially if the drug treatment is terminated prematurely. The mechanism that causes cells to enter the persister state, and to do so stochastically, is believed to involve two-gene modules that encode a toxin and an antitoxin. These two- gene modules are ubiquitous in bacteria and persistence is thought to be due to an imbalance between the abundance of the toxin and its antitoxin. Normally, the antitoxin is in excess and it neutralizes the toxin. However, when the toxin is in excess, cell growth is arrested but the cells are not killed. Rather, they are simply in a state of stasis. Another example of apparent bet hedging is swimming and chaining in B. subtilis. Bacterial cells in steady-state, exponential phase growth are a mixture of unicellular, motile cells and long chains of non-motile cells [23]. The swimming cells are active for the D, which governs motility and the production of enzymes (autolysins) that allow newly divided cells to separate from each other. Conversely, the chains of non-motile cells are inactive for D. Just how the cells spontaneously interconvert between the D-ON and D-OFF states is not known. Nonetheless, the choice, once again, is not a coin flip. Rather, in undomesticated bacteria, the switch is strongly biased in the direction of swimming. What is the biological significance of the alternative swimming and chaining states? An appealing possibility is that the swimmers are nomadic cells in search of new food sources whereas the chains are sessile cells that exploit the current niche. Thus, B. subtilis would appear to be hedging its bets against the likelihood that its current food source will be exhausted while at the same time taking full advantage of existing food. In this regard, swimming and chaining provides an interesting parallel to the swimming and stalked cells produced by C. crescentus, where the goal of having nomadic and sessile cells may be the same. But, as mentioned above, the production of swimming and sessile cells in C. crescentus is hard-wired into the cell cycle and a single cell of each type is obligatorily produced each round of the cell cycle. Perhaps the best adaptation to the aquatic habitat of C. cresentus is a 1:1 ratio of each cell type whereas the soil habitat of B. subtilis might favor a higher proportion of motile cells.

6 Stochasticity

In a population of animals, as in bacteria, individuals might choose to explore new territories as a way for the group to hedge its bets, even if it at a high cost for the explorers. However, when it comes to cell fate in metazoans, interpretations other than bet hedging must be invoked to explain stochastic choices because all cells depend on one another. Consider the case of olfactory receptors in mammals [24]. As for most sensory systems, only one type of olfactory receptor protein is produced in any given olfactory receptor neuron so as to avoid the sensory confusion that would occur if the same cell were to express more than one olfactory receptor gene. As the genome of the mouse devotes 4 percent of its protein-coding sequences to olfactory receptors, representing 1,000 genes, the task of achieving this sensory exclusion is formidable. To meet the challenge, each neuron chooses to express one olfactory receptor gene in a stochastic manner. It also prevents expression of all other olfactory receptor genes in this cell, and at the same time it informs the brain of its decision so as to allow it to process olfactory information [25, 26]. Thus, only one of the 1,000 olfactory receptor genes (actually, only one of 2,000 copies, each gene being present in two alleles in a diploid organism) is randomly activated in any one cell and all others are maintained in the off state (Fig. 3A). Here, the explanation for using stochasticity is one of economy: a regulatory circuit designed to choose among 2,000 genes in a directed manner would need to be extraordinarily complex and would involve large numbers of regulatory genes. The olfactory receptor decision is made in a cell autonomous manner (although it has to be communicated to the target neuron of the olfactory receptor neuron in the brain) [25, 26], but its molecular mechanism remains poorly understood. An elegant model was proposed based on a transcriptional enhancer (the H locus) that could associate with, and thereby activate only one of the olfactory receptor genes in the genome [27]. However, this model (at least in its specific form) was recently challenged by the finding that a knock out mutation of the H locus has little or no consequence on olfactory receptor choice [28].

7 Stochasticity

Nonetheless, a similar mechanism apparently explains the stochastic choice in the distribution of green and red cones in the human retina, which express the M and L opsin genes, respectively [29, 30]. The M and L genes are located near each other, with their own promoters. However, a unique Locus Control Region (LCR) located upstream of both genes is required for their expression, but it can only activate one gene at a time [31]. When the LCR connects to the L gene through DNA looping, the connection is stabilized and the cell becomes an L cone for the life of the cell: the M gene cannot be expressed. If the LCR associates by chance with the M gene, the M gene is expressed and the L gene is off (Fig. 3B). Given the diploid nature of mammalian cells, how does the cone cell ensure that only one gene (M or L) is expressed? The answer is that the LCR-L-M cluster is located on the X chromosome. In humans and other mammals, only one X chromosome is expressed in females due to X chromosome inactivation and males, of course, have only one X chromosome. Interestingly, the system has a built-in way to control the proportion of M/L cones: the gene closest to the LCR has more chances to be chosen by the LCR! It could be argued that the opsin system involves only a choice between two genes and therefore cannot be considered a case of stochasticity in cell fate choice. However, the choice of M or L gene expression profoundly affects the fate of the cone cell; cone cells have a defined connectivity to other neurons that instructs the brain as to the type of signal it receives [32]. A parallel can be made between the human color vision system and the visual system of the fly described above. As described above, R7 color photoreceptor cells exist in two types. Each type expresses a distinct rhodopsin gene, either rh3 or rh4, which encode molecules that are sensitive to different hues of UV light. The rh3 and rh4 genes are not clustered on the chromosome near a common LCR. Rather, the basis for stochasticity is attributed to the expression of a transcription factor called Spineless [13]. Somehow, the regulatory protein is only present in a subset of R7 cells and directs these cells to express rh4 rather than rh3. Just how Spineless becomes expressed exclusively in a subset of R7 cells is not understood. The choice between rh3 and rh4 is made cell autonomously in R7, but it has a profound consequence on a

8 Stochasticity neighboring photoreceptor cell called R8, which is instructed to produce a green opsin if R7 is expressing rh4 and a blue opsin if it is expressing rh3 (see Fig. 1) [14, 33, 34]. What is the meaning of stochasticity in the choice of photoreceptor cells in the eye of the fly or of a human? Because the retina in these two very different eyes is composed of many photoreceptors of different types, stochasticity is a simple mechanism to distribute two kinds of photoreceptors (in a particular ratio) across a large field and to avoid repetitive patterns that might limit the ability of the eye to perceive corresponding patterns in the visual field. Fishes, which live in murky waters where regular patterns would not be seen, distribute their color photoreceptors in a very regular pattern where no stochastic decision is made [35].

Non-autonomous choices In the preceding examples, a cell decides its fate stochastically in a manner that is independent of other cells. In some instances of cell autonomous decision making, the choice the cell makes influences the fate of other nearby cells. Nonetheless, the original cell fate decision was made independently of its neighbors. But not all stochastic decisions are cell autonomous. In some cases the decision is the result of back and forth interactions between two (or more) cells. In animals, the simplest system where non-cell autonomous decision making is seen is the choice between the Anchor Cell (AC) and the Ventral Uterine Cell (VU) fates in the nematode C. elegans [20]. Two neighboring precursor cells of the gonad can choose either fate. The two cells are the products of two parallel lineages that arose from a common ancestor several divisions earlier. However, small differences in the cell cycle of cells in these lineages lead one or the other of the two precursors to be born first. The first- born cell is biased to become the VU cell but it does not make this decision alone. Rather, the decision-making process involves inhibitory lateral interactions between the two cells via the Lin-12 signaling pathway (known as the Notch pathway in flies and vertebrates), as we now explain.

9 Stochasticity

Lin-12 is a receptor for the ligand Lag-2. The Lag-2 signal stimulates the activity of the Lin-12 pathway, resulting in the production of additional Lin-12 receptor. Thus, stimulating a cell with the Lag-2 ligand causes it to become hypersensitive to that ligand. Meanwhile, high levels of Lin-12 activity decrease the production of the ligand (Fig. 4A). Therefore, a cell that is activated for Lin-12 has diminished capacity to stimulate its neighbor [20]. As with the paradigm of bistable processes that are noise driven, stochasticity in birth order (developmental noise) tips the switch in one direction or the other. This bias is then amplified and locked in by lateral actions between the two cells. The first born exhibits somewhat higher Lin-12 activity than its neighbor and hence has diminished levels of the Lag-2 ligand, perhaps because, being born first, it is stimulated by cryptic Lag-2 signals coming from other cells. Lag-2 signaling from the second-born neighbor results in yet higher levels of Lin-12 and yet lower levels of ligand in the first cell [20]. This sets up a self-reinforcing cycle of lateral inhibition in which the first born cell achieves higher and higher levels of Lin-12 and the second born cell, not receiving any stimulation from its neighbor, has lower and lower Lin-12 activity. High Lin-12 activity leads to the VU fate and low activity to the AC fate. Lateral inhibition is also the basis for non-autonomous cell fate determination in the epidermis of the fly Drosophila melanogaster. One cell in a pro-neural cluster of equivalent cells becomes a neuroblast and it must do so to the exclusion of all the other cells in the cluster, which become epidermal cells [3, 36]. Flies use the same system as worms to achieve this (Fig. 4A). Notch is the Lin-12 equivalent in flies and its ligand is called Delta, the equivalent of worm Lag-2. The neuroblast fate arises stochastically by transcription noise leading to a very small increase in the capacity of one cell in the cluster to produce more Delta and hence stimulate the Notch receptor pathway a little more in all of its neighbors. This signaling stimulates Notch production in the neighbors, increasing their sensitivity to Delta and, as in the AC/VU example, setting up a self-reinforcing cycle (Fig. 4A). Meanwhile, the cell that, due to noise, exhibited an elevated signaling capacity attains a state of low Notch activity and hence becomes a neuroblast. Whereas in the nematode, loss of Lin-12 (Notch) function

10 Stochasticity leads to the AC fate, in flies, loss of Notch function leads to the neuroblast fate. All cells in the cluster are competent to become a neuroblast since killing the neuroblast, and thereby relieving lateral inhibitory signaling, allows another random cell to start the bistable loop again and to adopt the neuroblast fate [3, 36]. An equivalent example of non-cell-autonomous decision making is not known in bacteria. But the phenomenon of cannibalism combines stochastic decision making with reciprocal intercellular interactions [37]. When grown under conditions of nutrient limitation, B. subtilis enters an elaborate developmental process that culminates in the formation of a dormant spore. Entry into sporulation is governed by the regulatory protein Spo0A, whose activation is governed by a bistable switch [37-39]. Thus, only some cells in the population (about half) are ON for Spo0A and the others OFF. The Spo0A-ON cells produce a toxin and a killing factor that kill the Spo0A-OFF siblings. The dying siblings, in turn, release nutrients that limit further Spo0A activation in the Spo0A-ON cells, thereby arresting sporulation or even reversing it. This phenomenon has been interpreted as bet hedging: Uncertain as to whether they are experiencing a temporary shortage of nutrients or the onset of a prolonged famine, the bacteria stall for as long as possible before committing to spore formation, even at the expense of fratricide. In the Notch signaling systems, intercellular interactions reinforced alternative cell fate decisions. In contrast, in the cannibalistic bacterial system, the reverse is true as the remaining cells are delayed in committing to the spore fate.

Bistable switches that are hard wired by upstream events Not all bistable switches are driven by noise. In many cases, bistability is used to lock a cell in one or another fate but the decision is not left to chance. For instance, in the fly eye, two photoreceptors named R3 and R4 are derived from seemingly identical cells. Once again competition for Notch activation leads to a critical distinction between the R3 or R4 fates, and this distinction is crucial to promote the correct orientation of the ommatidium [40, 41]. However, in each of the 800 ommatidia, it is always the cell closer to the equator that becomes R3,

11 Stochasticity the polar one becoming R4 (Fig. 4B). Thus, all ommatidia rotate in the same orientation North of the equator and in the other orientation in the Southern (ventral) domain. This is because superimposed on circuitry that in other contexts (e.g. the choice between VU and AC fates in worms and neuroblast commitment in flies) is noise-driven, are gradients of other signaling (e.g. Wnt) that drive the decision to the R4 fate [40-42] . The Wnt protein is at its highest concentration at the North and South poles and at its lowest at the equator. Interestingly, it is not the absolute value of Wnt that matters. Rather, it is the relative difference in the level of signaling perceived, directly or indirectly, between the precursors of R3 and R4 that determines the outcome [40]. Thus, for each ommatidium, the precursor cell closest to the pole (where Wnt levels are higher) becomes R4 and the one closest to the equator (where Wnt is relatively lower) R3. Another example of bistability in which the outcome of the switch is hard wired is the establishment of left-right asymmetry between the two neurons (ASE) that sense either Na+ or Cl-`in C. elegans [18]. The switch consists of a complex regulatory loop in which a micro RNA (mir-273) inhibits translation of the mRNA for a transcription factor (Die-1), which, itself turns on the synthesis of another micro RNA (lsy-6) (Fig. 2C). Closing the loop, lsy-6 blocks the synthesis of the transcription factor (Cog-1) that is responsible for directing mir-273 synthesis. The left and right fates of ASE are specified by Die-1 and Cog-1, respectively (Fig. 2C). The ASE switch has the same logic as the double- negative loop that governs the alternative lytic and lysogenic states of phage lambda (Fig. 2D). Thus, when Cog-1 is ON, the synthesis of mir-273 blocks the production of the transcription factor (Die-1) for the opposite cell fate (Fig. 2C) (just as one lambda repressor blocks the synthesis of the other repressor). Conversely, when Die-1 is ON, it determines the right-hand fate and induces the synthesis of lsy-6 that prevents the accumulation of the transcription factor Cog- 1. In contrast to the stochasticity that drives the phage lambda double-negative loop, the choice between the left-hand (ASE-L) and right-hand (ASE-R) fates is instructed by the lineage of the two neurons; ASE-R is always on the right and ASE-L always on the left [18]!

12 Stochasticity

Why then have a bistable system? Perhaps, the ASE system derives from an ancestral worm that made the choice between the right- and left-hand fates stochastically. If so, only half of the ancestral animals would have had both ASE- R and ASE-L [43, 44]. If having a given neuron on the left, or on the right, proved advantageous, the system might have evolved through ‘genetic assimilation’ into directional asymmetry, in which it is always the same cell type that is on the right, and the other on the left [45]. Even though upstream signals dictate the outcome in the contemporary nematode, the circuitry of what once was a noise-driven switch might have been maintained in evolution as a way to lock in the decision robustly.

Conclusions Most organisms exhibit characteristics that are reproducibly inherited from generation to generation, which strongly implies that development is hard wired. However, certain developmental decisions are left to chance, sometimes out of necessity (when the choices are too many to be tightly controlled), or sometimes when it benefits the community to hedge its bets. In yet other cases, particular developmental outcomes are imposed on systems that are otherwise intrinsically stochastic. Nature knows how to make deterministic decisions, but, in contrast to Einstein’s view of the universe, She also knows how to leave certain decisions to a roll of the dice when it is to her advantage.

13 Stochasticity

Figure Legends Figure 1. Stochastic distribution of cell fates in bacteria and in insect photoreceptors: Left panel: Fluorescence micrograph showing cells of B. subtilis containing a fusion of the coding sequence for GFP to the promoter for a gene under the control of the competence regulator ComK. The cells were visualized with a red stain. The green fluorescence reveals the subpopulation of cells that are ON for ComK. Right panel: Photograph showing a whole adult Drosophila retina whose R8 photoreceptors were stained with antibodies against the green-sensitive photopigment Rh6 (green) and the blue-sensitive photopigment Rh5 (blue).

Figure 2. Regulatory circuits exhibiting bistability Panels A and B illustrate two kinds of regulatory circuits that can exhibit bistability. Shown in A is a positive feedback loop in which an activator (as in the example of the activator of competence ComK) stimulates the transcription of its own gene. Hypersensitivity is achieved by cooperativity among activator molecules in binding to the promoter region for the gene (not illustrated). Shown in B is a double-negative regulatory circuit in which two repressors (as in the example of the phage lambda CI and Cro repressors) antagonize the transcription of each other’s gene. Hypersensitivity is achieved by cooperativity among repressor molecules in binding to operator sites in DNA. Panel C illustrates an example of a double-negative regulatory circuit that governs the alternative neuronal ASE-L and ASE-R fate in C. elegans. In this case, the two transcriptional regulators (Cog-1 and Die-1) antagonize each other’s synthesis indirectly through the action of the micro RNAs lsy-6 and mir- 273, which block the translation of the mRNAs for Cog-1 and Die-1, respectively. Neurons have the ASE-L fate when Die-1 levels are high and Cog-1 levels are low (left-hand cartoon) and the ASE-R fate when the opposite is the case (right- hand cartoon). Panel D illustrates the case of the classic example of the double-negative circuit (see panel B above) governing the alternative lytic and lysogenic states of phage

14 Stochasticity lambda. When the lambda repressor CI is at high levels it represses the gene for the Cro repressor and genes involved in lytic growth (left-hand cartoon). Hence the phage is held in the dormant, lysogenic state. Conversely, when Cro is at high levels it represses the gene for CI under which condition genes involved in lytic growth are freely expressed (right-hand cartoon).

Figure 3: Cell-autonomous cell fate decisions Panel A illustrates cell-autonomous stochasticity in a mouse olfactory neuron. The neuron expresses one olfactory receptor gene (red) to the exclusion of all others (blue, brown, dark or light green, yellow or pink), including the other allele of the ‘red’ gene. The olfactory neuron somehow instructs its target neuron in the olfactory bulb of its choice (dashed arrow). Panel B illustrates cell-autonomous stochasticity in an old world primate color vision cone photoreceptor. The choice of a cone photoreceptor to become M (green-sensitive) or L (red-sensitive) depends on the ability of a single Locus Control Region (LCR) located upstream of the L and M genes to contact one of the two genes. If the LCR contacts the M gene, the cone becomes an M cone, and similarly for the L gene. This ensures that only one gene is expressed in each cone. As the LCR-M-L cluster is located on the X chromosome, only one copy is present in males and only one is active in females, due to X-chromosome inactivation.

Figure 4: Non-cell-autonomous cell fate decisions Panel A illustrates lateral inhibition by the Notch (Lin-12) regulatory system in which a stochastic decision by one cell prevents its neighbor(s) from making the same decision. Two neighboring epidermal cells of Drosophila start with the same potential to become neuroblasts, both initially exhibiting low Notch activity (N+/-) (left-hand cartoon). Variations in gene expression in the precursor cells leads one cell (dark pink nucleus) to increase production of the Notch ligand Delta (red lollipop) and to decrease production of the Notch receptor (blue Y) (right-hand cartoon). This asymmetry sets in motion a self-reinforcing cycle in which one cell (N-) becomes less and less sensitive to the Delta ligand and more

15 Stochasticity and more active in producing ligand whereas the other cell (N+++) becomes more and more sensitive to ligand but less active in producing it. The N- cell becomes a neuroblast while the N+++ cell remains an epidermal cell. Panel B illustrates a Notch-Delta regulatory switch that is biased in one direction by gradients of signaling molecules. Two neighboring photoreceptor cells, R3 and R4, in the fly compete as in (A) to acquire their cell fate. High Notch leads to the R4 cell fate while low Notch leads to the R3 fate. Pairs of R3/R4 precursors are in a gradient of a signaling molecule (e.g. wingless, green). In each pair, the cell positioned at the polar side receives more signal than its more equatorial neighbor, thus biasing it to becoming R4. The decision is then reinforced by lateral inhibition: all equatorial cells become R3 and all polar cells become R4.

Acknowledgements We thank J. Hahn and D. Dubnau for the image of Fig. 1A, J. Blau, D. Dubnau, M. Elowitz, O. Hobert, R. Johnston, E. O’Shea, A. Schier, L. Shapiro, G. Suel, A. Tomlinson and G. Yuan for comments on the review. This work was supported by NIH grant GM18568 to RL and EY13010 to CD.

References 1. Holtzendorff, J., et al., Oscillating global regulators control the genetic circuit driving a bacterial cell cycle. Science, 2004. 304(5673): p. 983-7. 2. Rine, J., et al., Pattern of switching and fate of the replaced cassette in yeast mating-type interconversion. Cold Spring Harb Symp Quant Biol, 1981. 45 Pt 2: p. 951-60. 3. Campos-Ortega, J.A., Asymmetic division: dynastic intricacies of neuroblast division. Curr Biol, 1997. 7(11): p. R726-8. 4. Freeman, M., Reiterative use of the EGF receptor triggers differentiation of all cell types in the Drosophila eye. Cell, 1996. 87(4): p. 651-60. 5. Chen, I., P.J. Christie, and D. Dubnau, The ins and outs of DNA transfer in bacteria. Science, 2005. 310(5753): p. 1456-60. 6. van Sinderen, D., et al., comK encodes the competence transcription factor, the key regulatory protein for competence development in Bacillus subtilis. Mol Microbiol, 1995. 15(3): p. 455-62. 7. Maamar, H. and D. Dubnau, Bistability in the Bacillus subtilis K-state (competence) system requires a positive feedback loop. Mol Microbiol, 2005. 56(3): p. 615-24. 8. Smits, W.K., et al., Stripping Bacillus: ComK auto-stimulation is responsible for the bistable response in competence development. Mol Microbiol, 2005. 56(3): p. 604-14.

16 Stochasticity

9. Maamar, H., A. Raj, and D. Dubnau, Noise in gene expression determines cell fate in Bacillus subtilis. Science, 2007. 317(5837): p. 526-9. 10. Suel, G.M., et al., An excitable gene regulatory circuit induces transient cellular differentiation. Nature, 2006. 440(7083): p. 545-50. 11. Suel, G.M., et al., Tunability and noise dependence in differentiation dynamics. Science, 2007. 315(5819): p. 1716-9. 12. Ready, D.F., T.E. Hanson, and S. Benzer, Development of the Drosophila retina, a neurocrystalline lattice. Dev Biol, 1976. 53(2): p. 217-40. 13. Wernet, M.F., et al., Stochastic spineless expression creates the retinal mosaic for colour vision. Nature, 2006. 440(7081): p. 174-80. 14. Mikeladze-Dvali, T., et al., The Growth Regulators warts/lats and melted Interact in a Bistable Loop to Specify Opposite Fates in Drosophila R8 Photoreceptors. Cell, 2005. 122(5): p. 775-87. 15. Bell, M.L., J.B. Earl, and S.G. Britt, Two types of Drosophila R7 photoreceptor cells are arranged randomly: a model for stochastic cell-fate determination. J Comp Neurol, 2007. 502(1): p. 75-85. 16. Franceschini, N., K. Kirschfeld, and B. Minke, Fluorescence of photoreceptor cells observed in vivo. Science, 1981. 213(4513): p. 1264-7. 17. Ferrell, J.E., Jr., Self-perpetuating states in signal transduction: positive feedback, double-negative feedback and bistability. Curr Opin Cell Biol, 2002. 14(2): p. 140-8. 18. Johnston, R.J., Jr., et al., MicroRNAs acting in a double-negative feedback loop to control a neuronal cell fate decision. Proc Natl Acad Sci U S A, 2005. 102(35): p. 12449-54. 19. Ptashne, M., On the use of the word 'epigenetic'. Curr Biol, 2007. 17(7): p. R233- 6. 20. Greenwald, I., LIN-12/Notch signaling: lessons from worms and flies. Genes Dev, 1998. 12(12): p. 1751-62. 21. Lewis, K., Persister cells, dormancy and infectious disease. Nat Rev Microbiol, 2007. 5(1): p. 48-56. 22. Balaban, N.Q., et al., Bacterial persistence as a phenotypic switch. Science, 2004. 305(5690): p. 1622-5. 23. Kearns, D.B. and R. Losick, Cell population heterogeneity during growth of Bacillus subtilis. Genes Dev, 2005. 19(24): p. 3083-94. 24. Buck, L. and R. Axel, A novel multigene family may encode odorant receptors: a molecular basis for odor recognition. Cell, 1991. 65(1): p. 175-87. 25. Mombaerts, P., Odorant receptor gene choice in olfactory sensory neurons: the one receptor-one neuron hypothesis revisited. Curr Opin Neurobiol, 2004. 14(1): p. 31-6. 26. Mombaerts, P., et al., Visualizing an olfactory sensory map. Cell, 1996. 87: p. 675-686. 27. Lomvardas, S., et al., Interchromosomal interactions and olfactory receptor choice. Cell, 2006. 126(2): p. 403-13. 28. Fuss, S.H., M. Omura, and P. Mombaerts, Local and cis effects of the H element on expression of odorant receptor genes in mouse. Cell, 2007. 130(2): p. 373-84.

17 Stochasticity

29. Jacobs, G.H., Primate photopigments and primate color vision. Proc Natl Acad Sci U S A, 1996. 93(2): p. 577-81. 30. Nathans, J., The evolution and physiology of human color vision: insights from molecular genetic studies of visual pigments. Neuron, 1999. 24(2): p. 299-312. 31. Smallwood, P.M., Y. Wang, and J. Nathans, Role of a locus control region in the mutually exclusive expression of human red and green cone pigment genes. Proc Natl Acad Sci U S A, 2002. 99(2): p. 1008-11. 32. Makous, W., Comment on "Emergence of novel color vision in mice engineered to express a human cone photopigment". Science, 2007. 318(5848): p. 196; author reply 196. 33. Chou, W.H., et al., Identification of a novel Drosophila opsin reveals specific patterning of the R7 and R8 photoreceptor cells. Neuron, 1996. 17(6): p. 1101-15. 34. Chou, W.H., et al., Patterning of the R7 and R8 photoreceptor cells of Drosophila: evidence for induced and default cell-fate specification. Development, 1999. 126(4): p. 607-16. 35. Stenkamp, D.L., et al., Temporal expression of rod and cone opsins in embryonic goldfish retina predicts the spatial organization of the cone mosaic. Invest Ophthalmol Vis Sci , 1996. 37(2): p. 363-76. 36. Doe, C.Q. and J.B. Skeath, Neurogenesis in the insect central nervous system. Curr Opin Neurobiol, 1996. 6(1): p. 18-24. 37. Gonzalez-Pastor, J.E., E.C. Hobbs, and R. Losick, Cannibalism by sporulating bacteria. Science, 2003. 301(5632): p. 510-3. 38. Veening, J.W., L.W. Hamoen, and O.P. Kuipers, Phosphatases modulate the bistable sporulation gene expression pattern in Bacillus subtilis. Mol Microbiol, 2005. 56(6): p. 1481-94. 39. Chung, J.D., et al., Gene expression in single cells of Bacillus subtilis: evidence that a threshold mechanism controls the initiation of sporulation. J Bacteriol, 1994. 176(7): p. 1977-84. 40. Wehrli, M. and A. Tomlinson, Epithelial planar polarity in the developing Drosophila eye. Development, 1995. 121(8): p. 2451-9. 41. Axelrod, J.D., et al., Differential recruitment of Dishevelled provides signaling specificity in the planar cell polarity and Wingless signaling pathways. Genes Dev, 1998. 12(16): p. 2610-22. 42. Yang, C.H., J.D. Axelrod, and M.A. Simon, Regulation of Frizzled by fat-like cadherins during planar polarity signaling in the Drosophila compound eye. Cell, 2002. 108(5): p. 675-88. 43. Chuang, C.F., et al., An innexin-dependent cell network establishes left-right neuronal asymmetry in C. elegans. Cell, 2007. 129(4): p. 787-99. 44. Troemel, E.R., A. Sagasti, and C.I. Bargmann, Lateral signaling mediated by axon contact and calcium entry regulates asymmetric odorant receptor expression in C. elegans. Cell, 1999. 99(4): p. 387-98. 45. Palmer, A.R., From symmetry to asymmetry: phylogenetic patterns of asymmetry variation in animals and their evolutionary significance. Proc Natl Acad Sci U S A, 1996. 93(25): p. 14279-86.

18 Figure 1 A B

C Cog-1 Cog-1 lsy-6 mir-273 lsy-6 mir-273

Die-1 Die-1

ASE-L ASE-R ASE-L ASE-R

D Cro Cro

CI CI

lytic lytic lysogenic Figure 2 A

B

L M L cone LCR

L M M cone LCR

Figure 3 A neurolast epidermal N+/- N+/- N- N+++ (or Lin-12)

Y Y Y Y Y Y Y Y Y Y

Y

Y Y Y

Delta (or Lag-1)

B R3/R4 precursors e l o

Specified R3/R4 p equator R3 R4 R3 R4 R3 R4 R3 R4 R3 R4

Figure 4