Transcriptional program controlled by the floral homeotic gene AGAMOUS during early organogenesis

The genetic control of floral organ identity is one of the most remarkable examples of how regulatory genes determine plant structure (reviewed by Ferrario et al., 2004; Zik and Irish, 2003a). A flower starts its development as a group of undifferentiated cells (the floral meristem), which arises on the flank of the shoot apical meristem. The floral meristem gives rise to organ primordia, which develop into each of the four types of floral organs: sepals, petals, stamens and carpels. The identity of these organs is specified by homeotic genes, most of which encode MADS-domain transcription factors. The homeotic genes are expressed in different but partially overlapping domains in the floral meristem, and the specific combination of homeotic genes active in each organ primordium directs the development of its organ type. These partially overlapping expression domains are set up by genes that are active in the meristem, but subsequently the expression and function of the homeotic genes is maintained throughout organ development.

The molecular basis for the combinatorial action of homeotic genes may be that in each case, the corresponding proteins are assembled into a different protein complex. For example, stamen development requires combination of the homeotic genes AGAMOUS (AG), APETALA3(AP3), PISTILLATA (PI) and at least one of the SEPALLATA (SEP1, SEP2 and SEP3) genes, whereas carpel development occurs when AG and SEP are expressed, but not AP3/PI (Bowman and Meyerowitz, 1991; Honma and Goto, 2001; Jack et al.,1992; Krizek and Meyerowitz,1996; Pelaz et al.,2000; Yanofsky et al.,1990). Based on protein-protein interactions in yeast and on co-immunoprecipitation, it has been proposed that stamen development is directed by a protein complex in which SEP3 bridges the interaction between AG and the AP3/PI heterodimer; similarly, the direct interaction between SEP3 and AG in yeast and suggests that these two proteins associate to control carpel development (Honma and Goto,2001).

Presumably each of the complexes containing homeotic proteins selects a different set of downstream target genes that participate in the development of a specific organ type, although the exact composition of these complexes in vivo, and how they select different target genes, remains unknown(Jack, 2001). To understand how the activity of homeotic genes is combined and translated into the patterns of cell division and differentiation that actually shape the floral organs, it is necessary to identify these downstream targets. However, very little is known about the genes that function downstream of the floral homeotic genes.

Genetic analysis has revealed some intermediate regulatory genes that control specific aspects of floral organ development. For example, AG activates SPOROCYTELESS (SPL), which controls sporogenesis in both stamens and carpels (Ito et al.,2004). SUPERMAN (SUP) controls cell proliferation in stamen and carpel primordia and its expression depends on AG, AP3 and PI (Sakai et al., 2000; Sakai et al.,1995). The SHATTERPROOF genes (SHP1 and SHP2) are required in the carpel margins for differentiation of the dehiscence zone, where later the fruit splits open to release the seeds(Liljegren et al., 2000). SPATULA (SPT) controls cell differentiation at the carpel margins and in the transmitting tract (the tissue that guides the growth of pollen tubes towards the ovules) (Bowman and Smyth, 1999; Heisler et al., 2001), and CRABS CLAW (CRC) participates in directing the development of tissues derived from the abaxial side of the carpel primordium (e.g. the outer epidermis)(Eshed et al., 1999).

A more comprehensive view of gene expression in floral organs came from transcript profiling experiments comparing wild type and homeotic mutants(Wellmer et al., 2004; Zik and Irish, 2003b). These experiments revealed hundreds of genes that are preferentially expressed in different organs, but these were mostly expressed at late stages of development and were probably only indirectly dependent on the floral homeotic genes (Wellmer et al., 2004). To fully understand the program of gene expression controlled by the floral homeotic genes, it is necessary to know how it unfolds from organ initiation to maturity. Here, we report the results of a global analysis of the program of gene expression triggered by AG, from the onset of organogenesis to early stages of reproductive organ development.

Plants were grown on a mix of vermiculite:soil:sand at 18°C with 16-hour light/8-hour dark cycles. All mutants (ag-3, ap1-1, ap1-1 cal-1 and ag-3 ap1-1) and AGGR were in a Lerbackground, which was used as the wild type.

Dexamethasone (Sigma, stock solution 10 mM in ethanol) was used at a final concentration of 10 μM in Silwet L-77 0.015%, applied directly on the inflorescence tips; for mock treatments, the solution contained the same amount of ethanol (0.1%) and Silwet L-77. After treatment, RNA was extracted from inflorescence apices and stored at -70°C until activation of AGGR was confirmed (2 weeks later).

Plants were fixed in 2.5% glutaraldehyde in phosphate-buffered saline (PBS)at 4°C overnight, dehydrated in an ethanol series, critical-point dried in liquid CO₂, sputter-coated with gold palladium, analysed and photographed with a Philips XL 30 FEG SEM.

Total RNA was extracted using TRI reagent (Sigma) according to the manufacturer's instructions. For array hybridisation, the RNA was cleaned up with RNeasy columns (Qiagen) and precipitated to increase final concentration.

Gene Chip arrays were hybridised as in the manufacturer's protocol(Affymetrix). To calculate P-values for increase or decrease in expression, the Wilcoxon signed-rank test(Hubbell et al., 2002; Liu et al., 2002) was applied to each pair of chips after normalisation across all probe sets, using Micro Array Suite 5.0 (Affymetrix). To calculate fold differences in expression, raw expression levels were imported from Micro Array Suite 5.0 into Gene Spring 5.1 (Silicon Genetics) and normalised first to the fiftieth percentile of each chip, then across all chips before further analysis.

Total RNA (2 μg) was treated with RNase-free DNase, and first strand cDNA was synthesised using oligo(dT) primer (Invitrogen) and Superscript RT(Invitrogen). Aliquots of the cDNA were used as template for PCR with gene specific primers (see Table S3 in supplementary material).

RNA was hybridised in situ (Fobert et al., 1996), using digoxigenin-labelled probes transcribed with T7 polymerase from linearised plasmid (pGEM-T easy, Promega) containing 3′cDNA fragments. Colour detection was performed with BCIP/NBT according to the manufacturer's instructions (Boehringer).

To produce AG protein, the AG ORF was PCR-amplified from pCIT1516 vector(Yanofsky et al., 1990) and cloned into pRSET-A (Invitrogen). BL21(DE3) pLysE cells were transformed with the construct, and His-AG proteins were expressed under the control of the T7 promoter. To prepare recombinant His-AG, inclusion bodies were purified using the BugBuster HT Protein Extraction Reagent (Novagen), according to the manufacturer's instructions, dissolved in dialysis buffer (20 mM Tris, 100 mM KCl, 1 mM EDTA, 1 mM DTT, 1 mM PMSF, 12% glycerol, pH 8.0) containing 6M urea,dialysed overnight against the same buffer without urea and stored at-20°C.

Probes were made from complementary oligonucleotides (see Table S3 in supplementary material), annealed in 20 mM Tris (pH 8.0), 50 mM NaCl, 1 mM EDTA, labelled with ³²P by filling in with DNA polymerase I (Klenow fragment), and gel-purified prior to use. DNA-binding assays and gel electrophoresis were essentially as described previously(Riechmann et al., 1996).

The procedure was adapted from Ito et al. and Wang et al.(Ito et al., 1997; Wang et al., 2002). Inflorescence tissue (∼1 g) of Col-0 plants was fixed with 1% formaldehyde in MC buffer [10 mM sodium phosphate (pH 7.0), 50 mM NaCl, 0.1 M sucrose) for 1 hour under vacuum. Fixation was stopped with 0.125 M glycine, followed by three washes with MC. The tissue was ground in liquid nitrogen, the powder was suspended in M1 buffer [10 mM sodium phosphate (pH 7.0), 0.1 M NaCl, 1 M 2-methyl 2,4-pentanediol, 10 mM β-mercaptoethanol, Complete™Protease Inhibitor Cocktail (Roche Diagnostics GmbH, Mannhein, Germany)], the slurry was filtrated through 55 μm mesh and centrifuged at 1000 g for 10 minutes. Subsequent steps were at 4°C unless indicated otherwise. Filtration and centrifugation were repeated twice, then the pellet was washed five times with M2 buffer (M1 buffer with 10 mM MgCl₂, 0.5% Triton X-100) and once with M3 buffer (M1 without 2-methyl 2,4-pentanediol). The nuclear pellet was resuspended in 1 ml Sonic buffer [10 mM sodium phosphate (pH 7.0), 0.1 M NaCl, 0.5% Sarkosyl, 10 mM EDTA, Complete™ Protease Inhibitor Cocktail (Roche Diagnostics GmbH), 1 mM PMSF]. Chromatin was solubilised on ice with a probe sonicator (MSE,Soniprep 150) by 25 cycles of 15-second pulses of half maximal power with 30 seconds cooling time between pulses. After sonication, the suspension was centrifuged (microcentrifuge, top speed) for 5 minutes and the supernatant was mixed with one volume of IP buffer [50 mM Hepes (pH 7.5), 150 mM KCl, 5 mM MgCl₂, 10 μM ZnSO₄, 1% Triton X-100, 0.05% SDS]. The solubilised chromatin was pre-adsorbed overnight with 7.5 μl antiserum against CLAVATA3 (CLV3) (sc-12598, Santa Cruz Biotechnology, Santa Cruz, CA) (used as AG-negative serum due to the lack of pre-immune serum). After centrifugation, the supernatant was mixed with 40 μl of protein G-Sepharose [Sigma, 50% slurry in 10 mM Tris (pH 7.5), 150 mM NaCl] and incubated on a rotating wheel for 1 hour. After centrifuging, the supernatant was equally divided over two tubes with 2.5 μl AG antiserum(sc-12697, Santa Cruz Biotechnology) or 2.5 μl CLV3 serum(control). After 1 hour on a rotating wheel and centrifugation, the supernatant was mixed with 20 μl protein G-Sepharose (Sigma) before incubation for another hour on the rotating wheel. The protein G-Sepharose beads were washed five times with 1 ml IP buffer for 10 minutes at room temperature. Elution with 0.1 M glycine, 0.5 M NaCl, 0.05% Tween-20 (pH 2.8)was as described (Wang et al.,2002). The eluate was treated with 1 μl RNase A (10 mg/ml) and proteinase K (to final of 0.5 mg/ml). After overnight incubation, a second aliquot of proteinase K was added and incubated at 65°C for 6 hours. After phenol/cloroform, then chloroform extraction, DNA was precipitated with 2.5 volumes of ethanol, one-tenth volume 3M NaAc (pH 5.4) and 1 μl glycogen,and resuspended in 10 μl of 10 mM Tris (pH 8.0).

ChIP PCR was performed to reveal if a specific DNA fragment was enriched in the immunoprecipitated DNA sample compared with the pre-immune DNA sample. Primers were designed around the consensus AG binding sites and control primers were made for regions lacking the consensus AG binding site. Template ChIP DNA was diluted, amplified for 35 to 40 cycles (see Fig. 5), and analysed on a 1.5%agarose gel, followed by scanning with a Molecular Imager FX-PRO Plus (Bio-Rad Laboratories, Hercules, CA). The primer sequences were (5′ to 3′):

CRC, TGGATGCATGAATAATGGGTAG and CGTGGACTAGAAATAATGAGACGA;

AP3, CGGAGCTCCGTTAATAAATTGACG and TTTGGTGGAGAGGACAAGAGA;

AP3 exon 7, AACATGTTTTGGTGAATTAGGAA and GCACCAGCAAACCTTTTAGC;

GA4, TTGTCCCTTTATATACGCATTAATCA and GAGACCAAGAGGAGGCAAAA;

AG, TGGTCTGCCTTCTACGATCC and CAACAACCCATTAACACATTGG;

SEP3, CGGCCATATCCACTTTTACG and TTTTTGGGATAATTTTACTTTCCAC; and

EIF4A1 control, TCTTGGTGAAGCGTGATGAG and GCTGAGTTGGGAGATCGAAG.

To follow changes in gene expression after stamen and carpel initiation, we generated plants with AG under external control. Plants were transformed with a construct in which the 35S promoter directed a fusion between AG and part of the rat glucocorticoid receptor (GR), as reported previously(Ito et al., 2004); for simplicity, we will refer to the 35S:AG-GR construct as AGGR. In the loss-of-function ag-3 mutants, AGGRrescued development of stamens and carpels only when the plants were treated daily with the steroid dexamethasone (DEX), confirming that the AG-GR fusion could replace AG function (Fig. 1A-D).

To focus on early organogenesis, AGGR was combined with the ap1-1 and cal-1 mutations. AP1 and CAL act redundantly to specify floral meristem identity. The double mutant accumulates indeterminate lateral meristems that fail to initiate floral organs(Kempin et al., 1995)(Fig. 1E,G), although the defect can become less severe late in development, allowing flowers to form(Ferrandiz et al., 2000). Expression of AG under the 35S promoter in ap1-1, cal-1 plants restored robust stamen and carpel development(Mizukami and Ma, 1997). In AGGR, ap1-1, cal-1 plants, DEX treatment induced stamen and carpel formation, whereas mock-treated controls remained meristematic(Fig. 1E,F). A single DEX treatment was sufficient for full stamen and carpel development, which followed a time course comparable to wild-type development(Smyth et al., 1990)(Fig. 1F). However, in AGGR, ap1-1, cal-1 plants that were also homozygous for the ag-3 mutation, organ development required daily DEX treatments (not shown), implying that a single DEX treatment initiated stamen and carpel development that was subsequently sustained by the endogenous AG.Thus, although an artificial construct was used to trigger organogenesis,subsequent development was controlled by the endogenous gene, followed the normal time course and yielded fully functional organs.

In plants treated in parallel with solution lacking DEX, organogenesis was not seen, whereas the frequency of DEX-induced organogenesis after a single DEX treatment ranged from between 30% and 100% of plants in different experiments. Individual DEX-treated plants showed an all-or-nothing response(i.e. either robust organ induction in all treated inflorescence apices, or no induction). AGGR was still expressed in plants that failed to initiate organs in response to DEX (not shown), so transgene silencing was unlikely to be the cause of the variable organ induction. The all-or-nothing response suggested that organ induction was a bistable switch (see Discussion).

To screen for genes whose expression changed in ap1-1, cal-1meristems after AG activation, we used the Arabidopsis ATH1 high-density oligonucleotide array (Affymetrix). Three time points were chosen after a single DEX treatment of 35S::AG-GR, cal-1, ap1-1 meristems:one day, when no morphological changes were visible, three days, when the earliest signs of organ primordia were seen, and seven days, when stamen and carpel primordia were recognisable (Fig. 1G-I). To ensure that the samples came from plants in which organogenesis had been induced, a few treated meristems were left on each plant and organ development was checked after two weeks. For each time point,two independent samples with AGGR activated were compared with two mock-treated controls, giving four possible combinations of treatment versus control. Genes up- or downregulated were defined independently for each time point as those with a statistically significant change in all four treatment/control pairs (Wilcoxon signed-rank test, P<0.05)(Hubbell et al., 2002; Liu et al., 2002) and a mean change of at least twofold.

Using these filtering criteria, 149 of the 22,810 genes represented on the array were upregulated in at least one of the three time points(Fig. 2A, and Tables S1 and S2 in supplementary material). Based on their predicted molecular function, the majority of these genes fell into three classes: unknown function (50),DNA-binding proteins (38) and metabolic enzymes (30)(Fig. 2C). The set of upregulated genes contained most of the known genes with a specific role in stamen and/or carpel development, including AG itself(Yanofsky et al., 1990), AP3 (Jack et al.,1992), PI (Goto and Meyerowitz, 1994), SEP1, SEP2 and SEP3(Pelaz et al., 2000), SUP (Sakai et al.,1995), CRC (Bowman and Smyth, 1999) and SHP1, SHP2(Liljegren et al., 2000). JAGGED (JAG) (Dinneny et al., 2004; Ohno et al.,2004), which controls the development of leaves and floral organs,was also activated, in addition to four (At2g01520, At3g04960, At4g21590,At5g57720) out of ten uncharacterised genes whose expression correlated with that of floral homeotic genes during floral induction(Schmid et al., 2003). Thus our array experiment independently detected many of the genes expected to function downstream of AG, based on previous genetic and array-based experiments.

The set of downregulated genes was smaller (16 on day 1, 9 on day 3, 43 on day 7; Fig. 2B, and Tables S1 and S2 in supplementary material) and included only one gene with a well-known role in floral development. UFO is expressed in meristems, functions upstream of the floral homeotic genes to set the pattern of AP3expression and is only expressed at the earliest stages of reproductive organ development (Ingram et al.,1995; Lee et al.,1997; Levin and Meyerowitz,1995). Accordingly, UFO appeared among the genes that were repressed at day 7 of organ development.

A set of 1453 genes expressed mostly at relatively late stages in specific floral organs has been identified by comparing the transcripts in wild-type and homeotic mutant flowers (Wellmer et al., 2004). The overlap between these genes and our list of AG-regulated genes is relatively small (20 of the 149 AG-activated genes and six of the AG-repressed genes; see Tables S1 and S2 in supplementary material), suggesting that the transcriptional program in early organogenesis is distinct from that in late organs.

To confirm independently of the array data that we have identified genes controlled by AG, we focused on genes that were activated at multiple time points after AG induction. A set of twelve genes were upregulated on day 1 or 3 and then remained activated until day 7(Fig. 3A). This set includes four well-known regulators of stamen or carpel development (AP3, CRC, AG,SEP3), and two genes implicated in the biosynthesis of the growth regulator, gibberellin: GA4 encodes an enzyme that catalyses the production of bioactive gibberellin(William et al., 2004; Williams et al., 1998) and ATH1 encodes a homeodomain protein proposed to regulate gibberellin biosynthetic genes (Garcia-Martinez and Gil, 2001). The remaining six genes encode a B3 domain protein(At3g17010), a zinc-finger protein (At1g13400) related to SUP, a homologue (At3g11000) of a protein implicated in somatic embryogenesis in carrot (Schrader et al.,1997), a predicted bifunctional nuclease (At4g21590), a WD-domain protein (At1g47610) and a protein (At1g02190) similar to CER1, which is involved in the synthesis of epicuticular wax and in pollen development(Aarts et al., 1995).

Activation of all 12 genes in cal-1, ap1-1, AGGR plants was verified by RT-PCR using a new set of RNA samples collected 1, 3 and 7 days after treatment (Fig. 3B). Genes controlled by AG should also be active during stamen or carpel development in wild-type flowers. This has already been shown for AP3, AG,SEP3 and CRC; for other genes in the set, expression was analysed by RNA in situ hybridisation (Fig. 4). At4g21590 was expressed in the centre of the floral meristem,in a pattern similar to that of AG, and continued to be expressed at later stages of stamen development (Fig. 4A,B). At3g17010 and At1g13400 were expressed in emerging stamen primordia and later in part of the developing carpels; expression of At3g17010, but not At1g13400, remained high in the sporogenous tissue of stamens and in the carpel ovary (Fig. 4C-E). Both genes implicated in gibberellin biosynthesis were expressed at very low levels in developing stamens: ATH1 expression was seen in the early organs, while GA4 was only detectable in the stamen filaments (Fig. 4F-H). Expression of At3g11000, At1g47610 and At1g02190 was below detection levels by in situ hybridisation. In all in situ hybridisation experiments, sense control probes showed only uniform background signal (not shown).

We next tested whether the 12 genes in the `core' set contained AG binding sites. We scanned sequences upstream of the start codon for the CArG box bound by AG in vitro, TT(A/T/G)CC(A/T)₆GG(A/T/C)AA(Shiraishi et al., 1993),accepting a maximum of two nucleotide mismatches, except when the mismatches eliminated either the CC or GG sequences flanking the A/T core. This level of stringency was calibrated using the well-characterised CArG boxes present in the AP3 promoter and in the second intron of AG(Hill et al., 1998; Hong et al., 2003; Tilly et al., 1998). Of the remaining 10 genes, eight had at least one CArG box match within 3 kb upstream of the start codon (Fig. 5A);in all cases, binding to AG was confirmed in vitro(Fig. 5B).

One caveat of detecting AG binding sites is that the frequency of CArG boxes in Arabidopsis genes is high: our search criteria detected at least one match in 49% of 27,186 upstream 3 kb sequences(www.arabidopsis.org/cgi-bin/patmatch/nph-patmatch.pl). The likelihood of finding a match in eight out of ten genes, however, is relatively low (4.8%, assuming binomial distribution and 49% likelihood for any single gene). Thus our subset of 12 genes was enriched for AG binding sites. A comparable enrichment was not seen for the complete set of AG-activated or repressed genes (matches were found in the upstream 3 kb sequences for 61% of the upregulated genes and 56% of downregulated genes),possibly because the complete set includes indirect AG targets.

Another caveat of the in vitro binding results is that multiple MADS domain proteins recognise similar sequences in vitro(Riechmann et al., 1996), so the CArG boxes might be targeted in vivo by MADS domain proteins other than AG. To confirm binding to AG in vivo, we used chromatin immunoprecipitation(ChIP) for a subset of genes of particular interest: AG, AP3, SEP3and CRC (which suggested that AG activated itself and most of the other regulators of stamen and carpel identity); and GA4(which suggested that another role of AG is to promote gibberellin biosynthesis). Fragments of these genes containing the in vitro-detected AG binding sites were enriched in immunoprecipitates obtained with antibodies against AG, but not with an unrelated antibody(Fig. 5C). By contrast,fragments that lacked AG binding sequences, such exon 4 of EIF4A1(Fig. 5C) and exon 7 of AP3 (not shown), were detected to the same background levels with both antibodies. Thus AG interacted in vivo with predicted regulatory sequences of AG, AP3, CRC, SEP3 and GA4.

The activation of AP3 by AG was not predicted by previous genetic and molecular analysis, particularly because AP3 is expressed normally in ag mutants (Jack et al.,1992) (Fig. 6B). This, however, could be due to redundant activation by AP1(Lamb et al., 2002; Ng and Yanofsky, 2001), which is normally repressed by AG in the centre of the floral bud(Gustafson-Brown et al., 1994)and could take over AP3 activation in the innermost organs of ag mutant flowers. To test this idea, we compared AP3expression in the ag-3 mutant, in the ap1-1 mutant and in the double mutant (Fig. 6). In the ag-3 mutant, stamens and carpels are replaced by additional whorls of sepals and petals (Yanofsky et al., 1990) (Fig. 6A). As expected, AP3 expression was readily detected in stage 3 buds and persisted throughout the development of both normal and ectopic petals of the ag-3 mutant(Jack et al., 1992)(Fig. 6B). In the ap1-1 mutant, petals are mostly absent and sepals are replaced by leaf-like organs that often subtend ectopic flowers(Mandel et al., 1992)(Fig. 6C). In this mutant, AP3 expression was normal in stage 3 and continued throughout stamen development (Fig. 6D). Like ag-3, the ag3-3, ap1-1 double mutant flower produced an indeterminate number of organs, which were leaf-like and subtended secondary flowers (Fig. 6E), similar to the first whorl organs of ap1-1. In the double mutant, early AP3 expression showed the normal pattern in both the primary and secondary flowers, while expression in later organ development was abolished(Fig. 6F).

We conclude that early AP3 expression did not require AGor AP1, and was probably due to activation by other regulatory genes,such as LEAFY in combination with UFO(Lamb et al., 2002; Parcy et al., 1998). Maintenance of AP3 expression in later stages of floral development,however, required either AG or AP1.

Positive auto-regulatory loops are a common device to stabilise expression patterns that arise from transient inputs during development(Davidson et al., 2002). Our results suggest that AG, AP3, PI and SEP3 are part of such an auto-regulatory loop: in ap1-1, cal-1 plants, transient AG activation was sufficient to trigger self-maintaining stamen and carpel development, during which AG, AP3, PI and SEP3 were activated; in addition, AG interacted directly with the AG, AP3 and SEP3 genes in vitro and in vivo.

Previously, auto-regulation of floral homeotic genes was known only for AP3 and PI, and their orthologues in snapdragon, DEF/GLO. In early buds, these genes are activated independently of each other, and, where they overlap, a positive-feedback loop is established that maintains their expression during petal and stamen development (Jack et al.,1994; Schwarz-Sommer et al.,1992). Activation of AP3 by AP3/PI is likely to be direct, whereas activation of PI requires an intermediate protein synthesis step (Honma and Goto,2000). In the case of AP3, the auto-regulatory loop is required only in stamens: AP3 expression is still maintained in the sepal-like organs that replace petals in the pi-1 mutant(Jack et al., 1992). This is an important point, because it shows that AP3 expression can be uncoupled from the organ identity directed by AP3/PI, and therefore the absence of AP3 expression in the developing organs of ag-3, ap1-1 double mutants was not a trivial consequence of the fact that these organs were neither petals nor stamens. The requirement of AG to maintain AP3 expression when this role cannot be fulfilled by AP1 supported the idea that AG also participates in AP3/PI regulation.

The co-ordinated regulation of AG, AP3, PI and SEP3 would be expected if, as proposed by recent models, these proteins function together in the same protein complexes (Honma and Goto, 2001; Jack,2001). In particular, if the predicted protein complexes are correct, the AP3/PI auto-regulatory loop should also require either AG or AP1 (which has also been proposed to form a complex with AP3/PI and SEP during petal development)(Honma and Goto, 2001). Our results confirmed this prediction.

However, if AG can only function when complexed with other MADS-domain proteins, then initiation of organogenesis by AGGR must have relied on partner proteins already present in the cal-1, ap1-1 meristems. One possibility is that a low level of AG-independent expression of SEP,AP3 and PI genes provided the required partners. This initial expression could be controlled by the same mechanism that activates these genes independently of each other in early wild-type buds. The need to establish a regulatory loop to amplify initially limiting levels of its partners may be the reason why a single activation of AGGR in cal-1, ap1-1 meristems resulted either in no response, or in robust organogenesis in an apparently random fashion.

In addition to AP3, CRC was strongly activated by AG. This was not expected because of the genetic evidence that CRC can function in the absence of AG. In the ag-1, ap2-2, pi-1triple mutant, in spite of the loss of AG function, the floral organs develop several carpelloid features, such as stigmatic cells and ectopic ovules. In this background, loss of CRC function caused a clear reduction of these carpelloid features, showing that CRC does not require AG to direct carpel development(Alvarez and Smyth, 1999). Our results suggest that although independently activated, CRC expression is reinforced by AG. Previous genetic results suggest that this reinforcement may be mutual: loss of CRC weakens AGfunction, causing the heterozygous ag-1/AG plants, which normally have a wild-type phenotype, to show a partial ag loss-of-function phenotype (Alvarez and Smyth,1999). It remains to be tested whether this occurs because CRC also activates AG, participating in the auto-regulatory loop.

We also saw that, at least in the cal-1, ap1-1 backgound, AG activated its own transcription. This could be inferred independently of the array experiments, from the fact that the endogenous AG was required for organogenesis in cal-1, ap1-1 plants after transient activation of AGGR, and was supported by the chromatin immunoprecipitation results. One difficulty with the idea that AG auto-regulates, however, is that AG is still expressed in the inner organs of ag-1 mutant flowers(Gustafson-Brown et al.,1994). Thus if AG activates itself during normal development, this activity must be redundant. As discussed above, if CRC participates in the AG regulatory loop, then CRC activity might account for the continued AG expression in ag flowers.

Combined with the published data, our results suggest a model for how AG and other floral organ identity genes are coordinately regulated(Fig. 7). In stamen development, AG, AP3, SEP3 and PI are initially expressed independently of each other. Where their expression overlaps, the predicted AG/SEP3/AP3/PI MADS protein complex (Honma and Goto,2001; Jack, 2001)maintains and amplifies their expression. In carpel development, the predicted AG/SEP3 complex may establish a similar feedback loop, which also reinforces CRC expression.

A link between gibberellin and homeotic genes has been shown previously via regulation of LEAFY (LFY), which activates homeotic genes in the early stages of floral development(Blazquez et al., 1998). More recently, gibberellin has been reported to activate floral homeotic genes at later stages of development, when LFY is no longer active(Yu et al., 2004). Our results suggest that the reverse is also true, that is, homeotic genes positively regulate the gibberellin pathway. GA4 is part of a small family of genes that encode GA3-β-hydroxylases, which catalyse the last step in the biosynthesis of gibberellin and have a regulatory role in the pathway(Hedden and Phillips, 2000; Itoh et al., 1999; Talon et al., 1990), so GA4 activation suggested that AG induced gibberellin biosynthesis during organogenesis.

Activation of GA4 by AG may be another branch of the homeotic gene autoregulatory loop. There may be, however, additional functions for gibberellin in floral organogenesis. Another gibberellin biosynthetic gene, encoding GA20-oxidase, is repressed by genes that maintain undifferentiated cells in the meristem, and activated in the leaf primordia that emerge from the meristem (Hay et al.,2002; Sakamoto et al.,2001). This suggests that gibberellin may have a more general role in the transition from meristem identity to organogenesis. This idea seems inconsistent with the fact that organ emergence is normal in gibberellin-deficient mutants, both during the vegetative phase and in flowers(the floral defects in ga1-3 become visible only at later stages of development) (Goto and Pharis,1999; Wilson et al.,1992). However, even severe mutants such as ga1-3, still produce low levels of gibberellin (Hedden and Phillips, 2000), which might be sufficient for the proposed functions in early organ development. Although it is not clear what these functions might be, the known role of gibberellin in controlling cell growth and division (Yang et al.,1996) suggests that it might play a role in the localised changes in growth that drive the emergence of organ primordia from the meristem. If this is true in the floral organ primordia, then gibberellin could be part of the link between homeotic genes and the cellular behaviour that shapes floral organs.

From the predicted protein functions of the 149 genes that were upregulated by AG, two prominent features emerged (Fig. 2). First, genes expected to function in transcriptional control were over-represented (26%), compared with their total frequency in the genome(5.9%) (Riechmann and Ratcliffe,2000). The fraction of regulatory genes increased over the time course from 13% (day 1) to 28% (day 3) to 34% (day 7). This contrasts with more mature organs, where the frequency of regulatory genes was 5.5%, similar to their representation in the genome(Wellmer et al., 2004), and suggests that up to 7 days after organ initiation much of the program of gene expression downstream of AG was concerned with refining patterns of gene expression. Complex cascades of transcription factors, as seen in early development of Drosophila and sea urchin, also cause delayed responses to initial inputs and have been proposed to function as timing devices during development (Rosenfeld and Alon, 2003).

Second, of the 36 predicted DNA-binding proteins that were upregulated at day 7, 53% belonged to two transcription factor families (10 B3 domain, PFAM profile PF02362, and 9 MADS domain, PF00319). MADS domain proteins play a prominent role in floral development and the diversification of this family correlates with the evolution of plant reproductive structures(Theissen et al., 2000). Our data suggests that the B3 domain family has undergone a comparable diversification of roles in reproductive development.

Developmental genetics has identified many regulatory genes whose expression determines where and when a specific structure or organ develops. The problem of understanding how regulatory gene expression is translated into complex multicellular structures is universal, and has led to a number of attempts to describe the gene expression programs controlled by these regulators (Furlong et al.,2001; Livesey et al.,2000; Michaut et al.,2003). Like other global descriptions of changes in gene expression during development, however, our view of gene expression under AG has two limitations. First, it is unlikely to be complete, because we cannot guarantee that genes with very low or localised expression were not missed,and because of the difficulties associated with detecting downregulation if it occurs only in a subset of the cells. Second, the set of genes controlled by AG probably cannot be organised within a single network of interactions,because they may represent the overlap of multiple programs of gene expression that run in parallel in different regions and cell types of organ primordia.

In spite of these limitations, the list of genes controlled by AG will provide a basis for the functional analysis of intermediate regulators of early organogenesis, and will provide target promoters that are needed to test current models for the molecular basis of how homeotic genes act combinatorially.

We thank James Hadfield (JIC) for help with the Affymetrix chip experiments and John Walshaw (JIC) for bioinformatics help. C.G.-M. was funded by a Marie-Curie fellowship (QLK5-CT-2001-52), M.M.R.C. received a fellowship from the FCT, Portugal (SFRH/BPD/3604/2000), S.d.F. was supported by the Netherlands Proteomics Centre (NPC) and R.S. received a grant from the Biotechnology and Biological Sciences Research Council (BBS/B/04234).