Loss of seven-up from Drosophila R1/R6 photoreceptors reveals a stochastic fate choice that is normally biased by Notch

Although many cell fate choices are biased by extracellular signals or inherited determinants, there is increasing evidence that some occur stochastically. For example, only a subset of cells within an E. coliculture infected with lambda phage are lysed, and only a subset of cells within a B. subtilis culture become competent to accept DNA; this variability is derived from the noise inherent in cellular processes such as transcription and translation, causing genetically identical cells to express different levels of the same gene product(Arkin et al., 1998; Ozbudak et al., 2002; Elowitz et al., 2002; Suel et al., 2006; Maamar et al., 2007). It is thought that the diversity generated by these random binary choices might enhance the survival of unicellular organisms in unpredictable environments. Stochasticism is also important for cell fate specification during the development of multicellular organisms. During N-mediated lateral inhibition,for example, stochastic differences in the ability of two initially equivalent cells to provide and receive N signals are amplified by feedback loops,ultimately causing the cells to adopt different fates (reviewed by Simpson, 1997). Stochasticism also generates the scattered distribution of sensory neuron types within receptive fields (Wernet et al.,2006; Bell et al.,2007; Serizawa et al.,2003; Shykind,2005; Lomvardas et al.,2006). We have identified a normally hidden stochastic fate choice that occurs during the development of the Drosophila eye; our work suggests that stochastic binary choices may underlie even fate choices that are induced by signals or lineage.

The fly eye is an important paradigm of fate specification(Voas and Rebay, 2004). Each of its ∼750 ommatidia contains eight photoreceptor neurons (R1-R8) that are specified by reiterative cell-cell signaling. First, proneural gene expression is restricted to regularly spaced cells by N-mediated lateral inhibition, generating the founder R8 neurons. Next, each R8 recruits neighboring cells as the R2/R5 and R3/R4 neuron pairs. The surrounding undifferentiated cells then undergo a final round of mitosis [the `second mitotic wave' (SMW)], generating the pool of equipotent cells from which the R1/R6 neuron pair is then recruited, followed by R7 and the non-neuronal cone and pigment cells.

In the adult, R1-R6 are thought to function as motion-detectors: they have large light-sensing rhabdomeres, express the broadly tuned Rh1 rhodopsin, and send their axons to the first optic ganglion, the lamina (reviewed by Cook and Desplan, 2001). R7 and R8 probably sense color: they have small rhabdomeres, express more narrowly tuned rhodopsins and send their axons to two different layers within the second optic ganglion, the medulla (Cook and Desplan, 2001). Despite the divergent mechanisms by which R7 and R8 are initially specified, recent work has shown that these fates are closely related: both require the redundant zinc-finger transcription factors Spalt major (Salm) and Spalt related (Salr) (together referred to as Sal)(Domingos et al., 2004a; Mollereau et al., 2001), and R7s lacking the homeodomain transcription factor Prospero (Pros) adopt a mixture of both R7 and R8 fates (Kauffmann et al., 1996; Cook et al.,2003).

In this paper, we focus on the binary fate choice faced by members of the R1/R6/R7 equivalence group: R1/R6/R7 precursors in which N is activated adopt the R7 fate but otherwise become R1/R6s(Tomlinson and Struhl, 2001; Cooper and Bray, 2000). Previous work has suggested that N might promote the R7 fate by repressing expression of the Svp orphan nuclear hormone receptor(Mlodzik et al., 1990; Hiromi et al., 1993; Begemann et al., 1995; Kauffmann et al., 1996; Cooper and Bray, 2000; Tomlinson and Struhl, 2001). In this paper, however, we demonstrate that svp mutant R1/R6 precursors stochastically and with approximately equal likelihood adopt either the R7 fate or the R8 fate. This result has two broad implications. First, it reveals an unexpected stochastic binary choice between the R7 and R8 fates. We show that svp mutant R1/R6s transition from an initially mixed R7/R8 fate to discrete R7 or R8 fates, suggesting that the stochasticism and bistability of this choice is derived from mutual negative feedback between the two programs. Second, our results suggest that if N normally represses Svp in R7 precursors, a second pathway must then prevent them from stochastically adopting the R8 fate. We provide evidence that N itself both represses Svp,exposing the stochastic R7 versus R8 choice, and tips this otherwise stochastic choice towards the R7 fate. This type of gene regulatory strategy has not previously been described.

z4 and z5 were identified in a screen of mosaic animals in which ∼15% of retinal cells derived from the SMW were homozygous for randomly mutagenized right arms of chromosome 3; mosaic animals were generated using GMR-FLP and MARCM (Lee et al., 2001). Mutations were induced with ethylmethane sulfonate by standard methods (Ashburner,1989). z4 and z5 failed to complement the lethality of svp^e22(Mlodzik et al., 1990; Begemann et al., 1995). The svp^e22 null allele was used for all data presented.

Homozygous wild-type, svp^e22, or pros¹⁷ R1s, R6s and R7s were created by GMR-FLP-induced mitotic recombination between FRT82chromosomes (Lee et al.,2001). Homozygous cells were labeled by the MARCM technique(Lee and Luo, 1999) with either act-Gal4 UAS-Synaptotagmin (Syt)-GFP (axon terminals) or act-Gal4 UAS-mCD8-GFP (cell bodies). The pros¹⁷chromosome was obtained from T. Cook (Cincinnati Children's Hospital). Rh3-lacZ, Rh4-lacZ, Rh5-lacZ and Rh6-lacZ stocks were obtained from C. Desplan (New York University). Mosaic animals lacking R7s were males hemizygous for a sev, GMR-FLP chromosome; as previously hypothesized (Mlodzik et al.,1990), svp mutant R1/R6s were unaffected by loss of sev. Individual homozygous wild-type or svp^e22R3s, R4s and MCs were created by ey-FLP^3.5 (from Iris Salecker, NIMR, London) and labeled by MARCM with act-Gal4 UAS-mCD8-GFP. Wholly wild-type or svp^e22 mutant eyes were created by ey-FLP^3.5 and the EGUF/hid method(Stowers and Schwarz, 1999). sev-N^act was obtained from G. Doughts (MGH, Boston).

Fixation was in 4% PLP at room temperature for 25 minutes (adult retinas)or 20 minutes (all other tissue), and antibody staining was as described by Lee et al. (Lee et al., 2001). Confocal images were collected on a Leica SP2 microscope and analyzed with Leica or ImageJ software.

We obtained mouse anti-Chaoptin (24B10; 1:200), mouse anti-Elav (9F8A9;1:10) and rat anti-Elav (7E8A10; 1:5) from the Developmental Studies Hybridoma Bank; rabbit anti-Rh1 (1:1000) from D. Ready (Purdue University); mouse anti-Rh3 (1:10), anti-Rh4 (1:10), anti-Rh5 (1:10) and anti-Rh6 (1:50) from S. Britt (UCHSC, Denver); rabbit anti-Rh4 (1:300) and anti-Rh6 (1:1000) from C. Desplan (New York University); rabbit anti-Salm (1:500) from R. Barrio(Universidad Autónoma de Madrid); guinea pig antiSens (1:1000) from H. Bellen (Baylor College of Medicine); mouse anti-Pros (mR1A, 1:1000) from C. Doe (University of Oregon); and mouse anti-Svp (1:500) from Y. Hiromi via C. Doe; chicken anti-GFP (1:500) and anti-β-gal (1:800) from Abcam(Cambridge, MA); and rabbit anti-GFP (1:1000), phalloidin conjugated to Alexa Fluor 555 (1:10) and all secondary antibodies [goat IgG coupled to Alexa Fluor 488, 555 or 633 (1:250)] from Molecular Probes (Eugene, OR).

To identify genes required for layer-specific targeting of R axons, we performed a genetic screen using mosaic animals in which ∼15% of the cells derived from the SMW, including R1s, R6s and R7s, were homozygous for a randomly mutagenized chromosome (Lee et al., 2001). We isolated two lethal and non-complementing mutations, z4 and z5, that caused mutant R1 or R6 axons to terminate in the R8 target layer (Fig. 1B, arrowhead). When we used a sevenless (sev)mutation to remove all R7s (Tomlinson and Ready, 1986; Mlodzik et al.,1990), we found that a similar number of mutant R1 or R6 axons terminated in the R7 target layer (Fig. 1D). z4 and z5 are alleles of svp (see Materials and methods); we therefore performed all analyses (including Fig. 1B,D) using the svp null allele e22.

svp encodes several isoforms of an orphan nuclear hormone receptor and controls fate in a variety of cell types, including photoreceptors,neuroblasts, cardioblasts and kidney cells(Mlodzik et al., 1990; Kanai et al., 2005; Lo and Frasch, 2001; Kerber et al., 1998). Within the fly eye, svp is expressed specifically in the R3/R4 and R1/R6 neuron pairs, where it was previously reported to repress the R7 cell fate(Mlodzik et al., 1990). We found that about half (48%; 72/151) of svp mutant R1/R6 axons that reached the medulla terminated precisely in the R8 target layer and half (45%;68/151) in the R7 target layer (Fig. 1F), suggesting either that svp mutant R1s or R6s are transformed into R7s whose axons fail to target correctly or that some svp mutant R1s or R6s adopt R8 fates.

To test whether svp mutant R1s or R6s can indeed adopt R8 fates,we examined their phenotype in adult retinas. As previously reported, most svp mutant R1s and R6s (∼95%) lost expression of the R1-R6-specific rhodopsin Rh1 (Zuker et al., 1985) (data not shown) and, like R7 and R8, elaborated small rhabdomeres in the center of the ommatidium(Fig. 2)(Mlodzik et al., 1990). Previous work demonstrated that svp mutant eyes contained an increased number of cells expressing the R7-specific rhodopsin Rh4, but the R8 rhodopsins had not yet been identified(Mlodzik et al., 1990). Because available antibodies did not allow us to distinguish all four R7 and R8 rhodopsins simultaneously, we examined R8 rhodopsin expression (Rh5 and Rh6) and R7 rhodopsin expression (Rh3 and Rh4) in separate experiments. We found that 52% (71/136) of svp mutant R1s and 53% (73/138) of svp mutant R6s had small rhabdomeres that contained R8 rhodopsins(Fig. 2A-E); the nuclei of svp mutant R1s and R6s expressing R8 rhodopsins were, like those of R8s, located proximal to wild-type R1, R6 and R7 nuclei (data not shown). These results suggest that both svp mutant R1s and R6s can adopt R8 fates, including expression of R8 rhodopsins, proximal positioning of the nucleus and axon termination in the R8 target layer. In the same experiment,44% (60/136) of svp mutant R1s and 44% (60/138) of svpmutant R6s had small rhabdomeres that did not contain R8 rhodopsins(Fig. 2E). We hypothesized that these may instead express R7 rhodopsins. Indeed, in a separate experiment we found that 42% (28/67) of svp mutant R1s and 41% (26/63) of svp mutant R6s had small rhabdomeres that contained R7 rhodopsins(Fig. 2F-J) and nuclei that were not displaced proximally (data not shown). Together, our results support a model in which approximately half of svp mutant R1s and R6s express R8 fates and half express R7 fates.

Neither the svp mutant R1s and R6s that expressed R8 rhodopsins nor those that expressed R7 rhodopsins formed a regular spatial pattern (data not shown), suggesting that the decision between the two fates was stochastic. To test whether R1 and R6 might somehow influence one another's decision, we examined ommatidia in which both the R1 and R6 neuron lacked svp. When staining for R8 rhodopsins, we observed approximate 1:2:1 ratios of ommatidia in which both svp mutant R1 and R6 expressed R8 rhodopsin,ommatidia in which only one expressed R8 rhodopsin, and ommatidia in which neither expressed R8 rhodopsin (data not shown); we observed the same ratios when staining for R7 rhodopsins (data not shown). We therefore conclude that svp mutant R1s and R6s make their decisions to express R7 or R8 rhodopsins independently of one another. The lack of correlation between fate transformations within an ommatidium provides additional evidence that this fate decision is not influenced by the ommatidium's position in the retina. Because the effects on R1 and R6 were indistinguishable, we henceforth refer to them as R1/R6s.

Although these experiments were consistent with a model in which R1/R6s adopt R8 or R7 fates stochastically and with approximately equal likelihood,they did not directly address whether svp mutant R1/R6s cleanly adopt either R7 or R8 fates, or might instead adopt partial or mixed R7/R8 fates as pros mutant R7s do (Cook et al.,2003). To test this, we first determined whether svpmutant R1/R6s co-express R7 and R8 rhodopsins. We examined expression of pairwise combinations of one R7 rhodopsin and one R8 rhodopsin (Rh3 and Rh6;Rh4 and Rh5; and Rh4 and Rh6; we did not examine Rh3 with Rh5 because working antibodies made in different organisms were not available). We found that only 7% (12/172) of svp mutant R1/R6s that expressed Rh4 or Rh6 expressed both; 0% (0/136) of svp mutant R1/R6s that expressed Rh4 or Rh5 expressed both, and 6% (7/109) that expressed Rh3 or Rh6 expressed both. By contrast, 25% (32/129) of pros mutant R7s that expressed Rh4 or Rh6 expressed both. These results suggest that most svp mutant R1/R6s do not adopt mixed R7/R8 fates. We note that the decision by svp mutant R1/R6s (unlike that made by wild-type R8s) to express Rh5 or Rh6 does not strictly correlate with the presence or absence, respectively, of Rh3-expressing R7s within the same ommatidium (data not shown)(Chou et al., 1996; Papatsenko et al., 1997).

We next assessed svp mutant R1/R6 axon targeting. As described above, we observed that >90% of svp mutant R1/R6 axons cleanly target either the R7 or R8 target layer(Fig. 1D,F), consistent with a discrete transformation into either an R7 or an R8. By contrast, we have found that most pros mutant R7 axons (62%, 72/118) have synaptic bouton-like thickenings at both R8 and R7 target layers(Fig. 1E,F). To determine whether the choice of axon target of a svp mutant R1/R6 is appropriately correlated with the rhodopsin it expresses, we used Rh-lacZ transgene reporter constructs that reflect normal rhodopsin expression patterns (Papatsenko et al.,2001); whereas rhodopsin proteins are restricted to the retina,β-galactosidase expression is detectable in the axon, allowing us to score target selection and Rh-lacZ expression simultaneously. In wild type, 30% of R7s express Rh3 and 70% express Rh4, while 30% of R8s express Rh5 and 70% express Rh6 (Zuker et al.,1987; Montell et al.,1987; Fryxell and Meyerowitz,1987; Fortini and Rubin,1990; Chou et al.,1996; Papatsenko et al.,1997; Huber et al.,1997; Chou et al.,1999). We found that 67% (94/141) of svp mutant R1/R6s that target the R8 layer expressed Rh6-lacZ and 47% (30/64) expressed Rh5-lacZ (Fig. 3A-D,I); only 12% (10/83) expressed Rh3-lacZ and 3%(2/62) expressed Rh4-lacZ (Fig. 3I), indicating a tight correlation between selection of the R8 target layer and R8 rhodopsin expression(Fig. 3I). Similarly, we found that 64% (106/166) of svp mutant R1/R6s that target the R7 layer expressed Rh4-lacZ and 39% (99/255) expressed Rh3-lacZ(Fig. 3E-H,I); only 2%(4/172) expressed Rh5-lacZ and 11% (47/447) expressed Rh6-lacZ, indicating a tight correlation between selection of the R7 target layer and R7 rhodopsin expression(Fig. 3I). Together, our results suggest that most svp mutant R1/R6s either express R7 rhodopsins and select the R7 target layer or express R8 rhodopsins and select the R8 target layer; we therefore conclude that they adopt either the R7 or the R8 fate and not a mixed R7/R8 fate.

svp is also expressed in the R3/R4 neuron pair. We therefore examined whether removal of svp from these cells might similarly cause a stochastic transformation into R7 or R8. Because the genotypes of R3 and R4 are unaffected by GMR-FLP, we used ey^3.5-FLP to induce mutant clones(Nern et al., 2005; Newsome et al., 2000). We found that 76% (96/126) of svp mutant R4s in otherwise wild-type ommatidia had small rhabdomeres that lacked Rh1; approximately half expressed R8 rhodopsins and half R7 rhodopsins (Fig. 4B,D,F,H), suggesting that svp mutant R4s adopt the R7 or R8 fate with equal likelihood. Although 77% (54/70) of svp mutant R3s in otherwise wild-type ommatidia retained a large outer rhabdomere that expressed Rh1, the remainder adopted either R7 or R8 fates with equal likelihood (Fig. 4D,H), as did half of the svp mutant mystery cells (MCs) that adopted R neuron fates (data not shown). These results suggest that svp mutant R3s,R4s and MCs can also stochastically adopt either the R7 or the R8 fate. We noticed that most svp mutant R3s with large outer rhabdomeres were in ommatidia with reversed chirality, indicating that they had adopted the R4 fate (Fig. 4C,G); this reversal in chirality has been previously observed, although not explicitly attributed to a transformation of svp mutant R3s into R4s(Fanto et al., 1998; Domingos et al., 2004b).

svp mutant R1/R6 and R3/R4 neurons were previously reported not to adopt R8 fates because they fail to express the larval R8 marker Bride of Sevenless (Hiromi et al.,1993). This suggested to us that svp mutant R neurons might not adopt R8 fates upon initial recruitment into the ommatidium but instead undergo a later transformation. We therefore sought to define the developmental timecourse of svp mutant R neuron transformation. Because of Gal80 protein perdurance, the GMR-FLP/MARCM technique does not result in GFP-labeling of mutant cells until ∼15 hours after R neuron recruitment. We therefore used ey^3.5-FLP and the EGUF/hid method (Stowers and Schwarz,1999) to create entirely svp mutant eyes and first confirmed that they contained the transformations predicted from our analysis of single clones; indeed, we observed an average of 2.2 R7s (n=112)and 2.5 R8s (n=232) per ommatidium in adult svp mutant retinas. We therefore analyzed wholly svp mutant larval eye discs for the expression of three early R7- and R8-specific markers: Salm, Pros and Senseless (Sens). The redundant Sal transcription factors are expressed in R7 and R8, where they prevent adoption of R1-R6-like fates(Domingos et al., 2004a; Mollereau et al., 2001); Salm is also briefly expressed in R3 and R4(Domingos et al., 2004b). The Pros transcription factor is expressed in R7, where it prevents co-expression of R8 fates (Kauffmann et al.,1996; Cook et al.,2003); and Sens is a zinc-finger transcription factor expressed in R8s, where it prevents adoption of R2/R5-like fates(Frankfort et al., 2001). We co-stained with antibodies against the neuron-specific marker Elav to distinguish R cells from non-neuronal cone cells, which also express Pros and Salm. We found that svp mutant R1/R6s expressed Salm by row 14-15 posterior to the morphogenetic furrow (Fig. 5B-B″, solid circles), and that approximately half expressed Pros by row 15 (Fig. 5D-D″, solid circles), and approximately half expressed Sens by row 15-16 (Fig. 5B-B″,D-D″, solid circles). We also found that, as previously observed, svp mutant R3/R4s failed to turn off Salm expression (Fig. 5B-B″,dashed circles) (Domingos et al.,2004b), but that while some expressed Pros(Fig. 5D-D″, dashed circles), a similar number instead expressed Sens(Fig. 5B-B″,D-D″,dashed circles). These results indicate that svp mutant R1/R6s and R3/R4s are directed towards R7 or R8 fates shortly after their recruitment during larval development.

We were surprised to find that many svp mutant R1/R6s and R3/R4s expressed both Pros and Sens (Fig. 5D-D″, arrowhead and arrows). We have never detected co-expression of Pros and Sens in wild-type R neurons at any stage(Fig. 5C,E). That most adult svp mutant R neurons do not adopt mixed R7/R8 fates suggested either that this initially mixed Pros/Sens expression is later resolved into mutually exclusive Pros or Sens expression, or that R neurons expressing both Pros and Sens can nonetheless adopt discrete R7 or R8 fates. To distinguish between these possibilities, we examined Sens, Pros and Elav expression 25 hours after puparium formation and found that <5% of Sens- or Pros-expressing R neurons expressed both (six out of the 157 Sens-positive and 105 Pros-positive R neurons found in 54 ommatidia); an example is shown in Fig. 5F-F″ (arrow). These results indicate that most Pros/Sens co-expression is ultimately resolved,presumably allowing those R neurons that express Pros to become R7s and those that express Sens to become R8s.

Our discovery that svp mutant R1/R6s stochastically become either R7s or R8s has important implications for the normal specification of wild-type R7s. The R1/R6/R7 precursor cells have previously been shown to form an equivalence group (Tomlinson and Struhl, 2001; Cooper and Bray,2000); however, R7s fail to express Svp and yet never adopt R8 fates. Why?

The equivalence of the R1/R6/R7 precursors is broken by Dl/N signaling: R1 and R6 redundantly use Dl to activate N in the next recruited precursor,causing it to become an R7 (Tomlinson and Struhl, 2001; Cooper and Bray,2000). Loss of N from R7 precursors causes them to adopt the R1/R6 fate, whereas ectopic N activation in R1/R6 precursors causes at least some to adopt the R7 fate (Tomlinson and Struhl,2001; Cooper and Bray,2000) and to lose expression of a svp enhancer trap(Kauffmann et al., 1996). To confirm that N is sufficient to repress Svp in R1/R6/R7 precursors, we directly examined Svp protein expression in larval eye discs of sev-N^act animals, in which the constitutively active N intracellular domain is expressed in R1/R6, R3/R4, R7 and cone cells(Basler et al., 1989; Bowtell et al., 1989; Fortini et al., 1993; Tomlinson and Struhl, 1999; Tomlinson and Struhl, 2001);indeed, sev-N^act caused a loss of Svp from R1 and R6 (Svp expression in R3/R4s was unaffected) (Fig. 6B), suggesting that the activation of N in wild-type R7s is sufficient to explain their lack of Svp.

How do wild-type R7s avoid stochastically adopting the R8 fate despite their lack of Svp? One possibility is that activated N might itself repress both svp expression and R8 fates by parallel mechanisms. However, it is also possible that some N-independent property of the R7 precursor, not possessed by the R1/R6 precursors, might block R8 fates. To determine whether ectopic N activation in R1/R6 is sufficient to prevent adoption of R8 fates despite causing loss of Svp, we examined Sens expression in sev-N^act larval discs. sev-N^act was previously shown to transform at least some R1/R6s into R7s or cone cells, but R8 markers were not examined (Tomlinson and Struhl, 2001; Cooper and Bray, 2000). We found no examples of ectopic Sens expression by sev-N^act R1/R6s (or by any cell other than R8) (0/161 ommatidia) (Fig. 6B,C,C″). Instead, nearly all sev-N^actR1/R6s expressed Salm and Pros, and most co-expressed Elav, indicating a transformation into R7s (Fig. 6B-C″). A small number expressed Salm and Pros but not Elav,consistent with adoption of the cone cell fate, as has been previously observed (Cooper and Bray,2000); this is probably caused by higher than normal levels of N,which antagonizes Ras (Rohrbaugh et al.,2002) (reviewed by Doroquez and Rebay, 2006), transforming cells that would otherwise adopt the R7 fate into cone cells (Fortini et al., 1992; Tomlinson and Struhl, 2001) (Fig. 6C,C′, arrow). To confirm that N activation is sufficient to repress Sens in R1/R6s lacking Svp, we examined the effect of sev-N^act on svp mutant L3 eye discs created by ey^3.5-FLP and the EGUF/hid method(Stowers and Schwarz, 1999). As expected, all sev-N^act; svp mutant R1/R6s expressed Salm and Pros but not Sens, indicating a transformation into R7s(Sens expression in sev-N^act; svp mutant R3/R4s was unaffected) (Fig. 6D-D″). We conclude that N activation is sufficient to repress both Svp and Sens in R1/R6 precursors, allowing them to adopt the R7 fate. We therefore suggest that N normally directs R7 precursors to adopt the R7 fate by repressing Svp expression, thereby exposing the stochastic choice between the R7 and R8 fates, and in parallel preventing adoption of the R8 fate(Fig. 7).

We have found that loss of svp causes R1/R6 precursors stochastically to adopt either the R7 or the R8 fate with approximately equal likelihood. svp mutant R1/R6s are transformed shortly after their recruitment as photoreceptors but transition from a mixed R7/R8 fate to discrete R7 or R8 fates. Precursors in which N is activated fail to express svp and yet exclusively adopt the R7 fate. Our work therefore reveals a novel branched pathway in which N both represses svp, thereby exposing an underlying stochastic choice between the R7 and R8 fates, and also tips this stochastic choice towards the R7 fate(Fig. 7).

We have found that svp mutant R1/R6s (and R3/R4/MCs) misexpress Salm, which is sufficient to downregulate the R1/R6-specific transcription factor BarHI and to induce expression of Pros in larval, Sens in pupal, and R7 and R8 rhodopsins in adult R neurons(Domingos et al., 2004a). Retinas lacking both svp and sal resemble those lacking sal only: all R rhabdomeres have an R1-R6-like morphology and express Rh1 (Domingos et al., 2004b),indicating that the transformation of svp mutant R neurons into R7s and R8s requires misexpression of Sal. These results suggest that Svp normally represses expression of Sal, which otherwise turns on R7 and R8 determinants,including Pros and Sens (Fig. 7), although Sal is likely not the only regulator of R7 and R8 fate that is repressed by Svp, as Sal has only a limited effect on axon connectivity (Domingos et al.,2004a; Mollereau et al.,2001).

How might misexpression of Sal (and other determinants) cause discrete yet stochastic adoption of either the R7 or R8 fate? Theoretical models suggest that stochasticism can arise from intrinsic noise within the underlying gene regulatory network (reviewed by Kaerns et al., 2005; Raser and O'Shea, 2005); this noise is particularly influential in networks whose gene products exist at very low levels. Bistability (i.e. `switch-like' behavior between two discrete fates) can arise from feedback within the network (reviewed by Ferrell, 2002). We have found that some svp mutant R1/R6s initially express both the R8 marker Sens and the R7 marker Pros but later express only one or the other. This result supports a model in which their switch-like ability to adopt exclusively R8 or exclusively R7 fates is derived from mutual negative feedback between the two programs; the winning fate can emerge more quickly (for example, in those cells that never appear to co-express Pros and Sens) or more slowly. We hypothesize that expression of R7 determinants and R8 determinants in svp mutant R1/R6s initiates approximately simultaneously, such that negative feedback between the fate pathways begins when the relevant molecules are at low levels. The intrinsic variability caused by low molecule numbers might therefore cause the observed stochasticism of the final fate choice, as well as the variability in the length of time required for a final choice to emerge.

One obvious candidate is pros, which prevents R7s from also adopting R8 fates. However, we have found that Pros misexpression is not sufficient to prevent svp mutant R1/R6s from adopting the R8 fate(data not shown), and others have found that Pros cannot repress Sens in wild-type R8s (Cook et al.,2003). These results suggest that as yet unidentified genes (`gene X' in Fig. 7) mediate the negative feedback that turns off Sens, thereby reinforcing R7 fate in svp mutant R1/R6s. Pros instead directly represses Rh5 and Rh6 expression (Cook et al., 2003)and, we have shown here, prevents axon targeting to the R8 recipient layer.

By contrast, we have found that forced Sens expression is sufficient to cause svp mutant R1/R6s, as well as wild-type R7s, to express R8 rhodopsins and project axons that terminate in the R8 target layer (data not shown). This result suggests that Sens can mediate the negative feedback that reinforces the R8 fate in svp mutant R1/R6s(Fig. 7) and that wild-type R8s never adopt R7 fates because they already express a substantial amount of Sens by the time they turn on Sal.

Theoretically, N might either promote the R7 fate or repress the R8 fate,but because N has a well-known earlier role in repressing R8 fates (reviewed by Frankfort and Mardon, 2002; Hsiung and Moses, 2002), we favor the latter model. In particular, during the initial specification of R8 neurons by lateral inhibition, N activation prevents cells from becoming R8s by turning on expression of the Enhancer of Split complex[E(Spl)-C] (Ligoxygakis et al.,1998); during N-mediated lateral inhibition in sensory organ development E(Spl)-C genes have been shown to repress Sens expression directly (Jafar-Nejad et al.,2003). A parsimonious model of R7 specification is therefore that activated N in the R7 precursor represses both Svp and Sens in parallel to induce adoption of the R7 fate (Fig. 7).

We thank R. Barrio, H. Bellen, S. Britt, C. Desplan, C. Doe, G. Doughts, Y. Hiromi, D. Ready, L. Zipursky, T. Cook, I. Salecker, S. Stowers, the Bloomington Drosophila Stock Center, and the Developmental Studies Hybridoma Bank for flies and antibodies. We also thank C. Doe, K. Guillemin, R. Herman and T. Stevens for discussion and comments on the manuscript. This work was funded by a Burroughs-Wellcome Career Development Award to T.G.H. and by an institutional NIH NRSA Genetics Training Grant (5-T32-GM007413) to A.C.M.