APETALA2 functions as a temporal factor together with BLADE-ON-PETIOLE2 and MADS29 to control flower and grain development in barley

Sex in angiosperms occurs in the flower, a complex structure of multiple organs, the coordinated growth of which leads to production of seed. Double fertilisation of the embryo sac within the ovary generates two filial seed tissues, the embryo and endosperm, which proliferate and expand while enclosed by layers of maternal ovary tissue, including the proximal nutritive nucellus and the more distal integuments that form the protective seed coat (Wilkinson et al., 2018). The grass flower or ‘floret’ ovary is additionally encircled by the ovary wall or pericarp, which, with the underlying maternal and filial tissues, grows into the grain fruit or ‘caryopsis’. The caryopsis isotopically expands following fertilisation and then elongates and swells to fill the cavity between opposing floret hulls (Brinton and Uauy, 2019). Several genes that regulate ovary, nucellus, integument, pericarp and/or hull development also influence grain size (Song et al., 2007; Yin and Xue, 2012; Brinton et al., 2017; Ren et al., 2018; Zhao et al., 2018; Wilkinson et al., 2019). However, the mechanism used by these genes to influence grain size is poorly understood. This problem has major translational significance in cultivated cereal grasses, the endosperm of which provides more calories than any other source to the human diet.

Learning more about the genetic and functional networks in the floret and grain may help to dissect this problem. Each floret arises from a floret meristem made by the spikelet, the basic reproductive unit of grasses. Floret meristems form opposing lemma and palea hulls which enclose the stamens and single-ovary carpel (Schrager-Lavelle et al., 2017). Small sacs called lodicules develop between the lemma and stamens, which enlarge at anthesis (pollen shed), prising open the lemma to facilitate pollen transfer (Kellogg, 2015). Molecular patterns controlling floret morphogenesis appear to be partially conserved, with the ABCDE gene combinatorial model proposed to explain flower development in dicots such as Arabidopsis thaliana (Schrager-Lavelle et al., 2017). In the Arabidopsis flower, activity of class ‘A’ genes defines the outer perianth (sepals), ‘AB’ the inner perianth (petals), ‘BC’ the stamens, ‘C’ the carpel and ‘CD’ the ovule, with the ‘E’ class genes contributing to all other class functions (Theißen et al., 2016; Irish, 2017). Other than the ‘A’ class APETALA2 (AP2) gene, the ABCDE genes encode MADS-box transcription factors. Orthologous MADS-box expression in rice suggests that the palea and lemma are analogous to sepals (Lombardo and Yoshida, 2015), whereas the lodicules are highly derived petals (Yoshida, 2012). Lemma formation also represents a commitment of the axillary meristem to floret fate (Arber, 2010; Kellogg, 2015). Thus, grass perianth development involves decisions about organ identity and spikelet versus floret fate.

Specific genes establishing floret and organ identity in grasses are not completely resolved, especially in temperate cereals such as barley and wheat that develop unbranched ‘spike’ inflorescences. However, the A-class APETALA2-like (AP2L) genes and their regulation by microRNA172 (miR172) appear to play a central role. In spikes, spikelets directly attach to the spike axis at alternating nodes flanked by two bract-like glumes. In wheat, miR172-resistant alleles of TaAP2L5 (the Q domestication gene) or TaAP2L2 cause glume-to-lemma transformations, ectopic florets and compact spikes, whereas overexpression of miR172, or loss of TaAP2L5 or TaAP2L2 function, leads to longer spike internodes, glume-like lemmas and reiterations of empty glumes (Simons et al., 2006; Sormacheva et al., 2015; Debernardi et al., 2017, 2020; Greenwood et al., 2017). These phenotypes suggest that TaAP2Ls promote lemma identity and floret establishment, a role partially conserved in rice based on loss-of-function phenotypes of the SUPERNUMERARY BRACT (SNB) and INDETERMINATE SPIKELET1 (IDS1) AP2L-genes (Lee et al., 2007; Lee and An, 2012; Ji et al., 2019). In barley, a gain-of-function HvAP2 allele called Zeo1.b disrupts miR172-directed HvAP2 transcript cleavage, elevating HvAP2 transcript levels and leading to small, non-swelling lodicules, semi-dwarfism and dense spikes (Houston et al., 2013; Patil et al., 2019). We do not know whether HvAP2 fulfils similar roles as TaAP2L5 or TaAP2L2 in floret establishment or organ identity. The only gene currently identified that controls floret organ identity in barley is the LAXATUM.A/HvBLADE-ON-PETIOLE2 (HvBOP2) gene encoding a transcription factor necessary for lodicule identity and repression of spike internode elongation (Jost et al., 2016). The regulatory mechanism(s) of HvBOP2 and its relationship to HvAP2 or the ABCDE model are unknown.

AP2L genes also regulate post-fertilisation development across different plants. In Arabidopsis, AtAP2 restricts integument cell expansion, seed size and seed mass (Modrusan et al., 1994; Jofuku et al., 1994; Ohto et al., 2005) and in rice, SNB limits grain size and weight, a role associated with constraining floret hull cell expansion (Jiang et al., 2019; Ma et al., 2019). Whether AP2L genes also control post-fertilisation development in temperate cereals is unexplored. Here, we identify and characterise a new HvAP2 allele in the barley Bowman Near-Isogenic Line (BWNIL) mutant population (Druka et al., 2011) and generate additional alleles by gene editing to reveal both conserved and novel roles for HvAP2 in floret and grain development.

To learn more about HvAP2, we screened the BWNILs for loss-of-function HvAP2 alleles, focusing on mutants with introgressions that overlapped HvAP2 (Druka et al., 2011; Houston et al., 2013). We selected the BW381 line containing the gigas1.a locus (Greek for giant), originally isolated in the cultivar Golden Melon (Tsuchiya, 1962; Franckowiak, 1995; Druka et al., 2011; Table S1). BW381 (hereafter called gigas1.a) showed multiple elongated features compared with the recurrent parent cultivar Bowman (Fig. 1; Table S2). Spike internodes were longer (P=0.036), causing less dense, or laxatum, spikes, glumes were 20% longer (P=0.028) and lemmas and paleas were 50% longer (P≤0.001), shifting the gigas1.a spikelet shape from wedge to lance-shaped (Fig. 1A-D,F,G; Fig. S1A-C; Table S2). Adaxial lemma epidermal cell length in gigas1.a was increased 52% compared with Bowman (P≤0.001, Fig. S1D), suggesting that increased cell elongation explained longer gigas1.a lemmas, although these cells were also 16% wider compared with Bowman (P≤0.001; Fig. S1E). In Bowman and most cultivated barleys, lemma and glume tips have thin projections called awns; the lemma awn is longer, with a distinct boundary from the lemma (Fig. 1C,E). Glume awns in gigas1.a were 50% longer compared with those of Bowman (P<0.001; Fig. 1C,E,G; Table S2) yet lemma awns in gigas1.a were 23% shorter (P≤0.001), with a less distinct lemma-awn boundary (Fig. 1C,E,F; Table S2), two glume awn-like features. Lodicules in gigas1.a developed ectopic distal lamina (Fig. 1H,I), associated with extreme open-flowering (Fig. S1F), and were decorated with glume-like hairs (Fig. 1J; Fig. S1G,H). We observed that gigas1.a stigmas had fewer, shorter papillae branches in contrast to feathery ‘plumose’ stigmas of Bowman (Fig. 1K), which may contribute to reduced seed set in gigas1.a, as previously reported (Tsuchiya, 1962). Grain length in gigas1.a mirrored the longer hulls and increased by 47% compared with Bowman (P≤0.001; Fig. 1L-N; Table S2). Pericarp epidermal cells in gigas1.a were only 16% longer compared with Bowman, as well as 23% wider (P≤0.001; Fig. S2B,C), suggesting that changes in grain length likely involves increases in cell size and cell number. However, gigas1.a thousand grain weight (TGW) increased only 7%, as gigas1.a grain was also narrower and thinner compared with Bowman (P≤0.001; Fig. 1M; Fig. S2; Table S2), although its lemma width was unchanged (Fig. S1A). Caryopses in gigas1.a were darker than Bowman (Fig. 1M), suggesting increased proanthocyanidins in the seed coat (Aastrup et al., 1984). We re-examined Zeo1.b for additional phenotypes, finding that Zeo1.b glumes often transform into lemmas (Fig. 1O) and that Zeo1.b grain is 8% wider compared with Bowman (P≤0.001; Fig. S2; Table S2). Although we did not observe differences in cell length or width in the adaxial lemma of Zeo1.b, the pericarp cells were 52% wider compared with Bowman (P≤0.001; Fig. S2C), suggesting that HvAP2 promotes medial cell expansion in the barley pericarp. Taken together, elongated glume-like organs, expanded cells, lax spikes, open-flowering and longer grain of gigas1.a contrasted with the closed-flowering, compressed growth, glume-to-lemma transformation and wider grain of Zeo1.b.

Given the opposing phenotypes to Zeo1.b, we speculated that gigas1.a may be a loss-of-function HvAP2 allele. To clone the locus, we first placed the BOPA2 markers associated with the gigas1.a introgression on the physical map (Mascher et al., 2017), which located a Golden Melon introgression on 2H between 710843099 bp and 758851055 bp (Fig. S3). Testing gigas1.a genomic DNA on the barley 50K iSelect SNP Array (Bayer et al., 2017) identified a slightly larger area on 2H from 710163110 bp to 760762651 bp (Fig. 2A; Table S3). SNPs starting after 729506693 bp and ending between 730687131 bp and 730852717 bp on 2H were present in Golden Melon and Bowman, but appeared as missing in gigas1.a, and demarked an area encompassing HvAP2 (HORVU2Hr1G113880.23) and six other high-confidence genes (Fig. 2A; Tables S3,S4). Using gigas1.a genomic DNA, we successfully amplified two genes outside this region but could not amplify HvAP2 or two other genes in this region (Fig. S3). We detected HvAP2 transcripts in Bowman and Zeo1.b spikes but not in gigas1.a (Fig. 2B). Collectively, our evidence suggests that a deletion on 2H in gigas1.a removed HvAP2 and at least six other genes.

To our knowledge, gigas1.a was the only available gigas1 allele. To confirm that gigas1.a phenotypes do not result from the deletion of genes other than HvAP2, we targeted the HvAP2 gene using CRISPR/Cas9 gene editing. We transformed Golden Promise with two binary vectors containing the barley codon optimised Cas9 (bcoCas9) sequence and one guide RNA sequence targeting HvAP2 sequences upstream of those encoding the first AP2 DNA-binding domain (Fig. 2C). Screening 18 independent T₁ transformants identified two new HvAP2 alleles, hvap2-1 and hvap2-2, with 39 bp and 40 bp deletions, respectively, in the first exon (Fig. 2C). The hvap2-1 deletion (175 bp to 214 bp relative to the coding sequence start) removed 13 amino acids (residues 60-73) before the first AP2 domain in the predicted protein but kept the remaining sequence in frame (Fig. 2C). We observed no obvious morphological differences in hvap2-1 compared with Golden Promise (Fig. 2D-F). The hvap2-2 deletion (175 bp-215 bp) removed the same 13 amino acids but also caused a frame shift and a premature stop codon in exon 5 (368 bp), predicted to significantly impair HvAP2 function (Fig. 2C). The hvap2-2 mutant largely phenocopied gigas1.a, with longer lemmas (P≤0.001), lax spikes (P=0.005) and long, slender grain compared with Golden Promise (Fig. 2D,E; Fig. S4; Table S3). We note that, in this experiment, gigas1.a had marginally wider lemmas compared with Bowman (P<0.05; Fig. S4A). Lodicules in hvap2-2 were larger, extended and swollen compared with the non-swelling lodicules of Golden Promise (Fig. 2F), a cleistogamous cultivar with the Zeo2 allele (Houston et al., 2013). Green organs with a hybrid composition of stamen-like filaments and smooth and hairy bract-like regions replaced the lodicules in 3% of hvap2-2 florets (Fig. 2G,H), a more severe loss of lodicule identity compared with gigas1.a, potentially reflecting the different cultivar backgrounds. Overall hvap2-2 closely phenocopied gigas1.a, corroborating that gigas1.a phenotypes result from deletion of HvAP2 alone. Thus, we propose that HvAP2 defines the boundary between the glume and lemma (outer perianth), promotes floret perianth identity, increases stigmatic papillae branching and widens grain, while also restricting longitudinal growth of spike internodes, spikelets, floret organs and grain.

We compared early spikelet development in Bowman, Zeo1.b and gigas1.a to understand when and how HvAP2 influences organ identity. We staged floret organ development using the Waddington (WD) stages (Waddington et al., 1983). WD4 stage Zeo1.b spikelets exhibited wider glumes than Bowman, leading to two interlocking lemma-like organs (Fig. 3A) which later overgrew the floret lemma (Fig. 3B). Lodicule primordia emerged similarly in Bowman, gigas1.a and Zeo1.b at WD4. By WD5.5, WD7 and later, Bowman lodicules displayed distinct proximal cushion and distal fringe tissues (Fig. 3C; Fig. S5). In contrast, gigas1.a lodicules were flatter with distal extensions whereas Zeo1.b lodicules remained small, lacked cushions and formed hairs (Fig. 3C; Fig. S5). These data suggest that HvAP2 and its miR172-regulation between WD4 and WD7 influence early spikelet differentiation.

AtAP2 controls Arabidopsis floral organ development in part through regulating MADS-box gene expression (Drews et al., 1991; Yant et al., 2010; Dinh et al., 2012). We hypothesised that HvAP2 also controls floral development by modulating target gene(s) expression. We selected candidate HvMADS-box genes based on their predicted function, AP2-like binding motifs in their regulatory regions, expression in relevant tissues and whether they were differentially expressed in our earlier Zeo1.b microarray study (Figs S6,S7; Table S5; Patil et al., 2019). Our top candidate was HvMADS1 (HORVU4Hr1G067680.2). The HvMADS1 orthologue in rice, OsMADS1 [LEAFY HULL STERILE (LHS1)] promotes lemma and lodicule identity and differentiation, transforms glumes into lemmas when overexpressed and causes glume-like lemmas and elongated bract-like lodicules when downregulated (Prasad et al., 2001, 2005). HvMADS1 is expressed in differentiating spikelets, lemmas and lodicules (Fig. S7) and co-expressed with HvAP2 in RNA-seq datasets of spikelet development (Digel et al., 2015). We sampled developing spikes in Bowman, gigas1.a and Zeo1.b genotypes, normalising expression to Bowman. HvMADS1 expression increased in all genotypes between WD3.5 and WD4.0, suggesting this temporal pattern is independent from HvAP2. Overall, HvMADS1 transcripts were little changed in gigas1.a, so factors other than HvAP2 likely contribute to HvMADS1 expression in gigas1.a. However, we detected increased HvMADS1 expression in Zeo1.b WD4 and WD5.5 spikes compared with Bowman (Fig. 3D; Fig. S8). In situ hybridisation with an antisense HvMADS1 probe gave a strong signal in young and older Zeo1.b spikelets compared with Bowman, especially within developing glumes, lemma/palea, lodicule and stamen primordia (Fig. 3E,F). These data support the suggestion that ectopic HvAP2 promotes HvMADS1 expression, potentially explaining the glume-to-lemma transformations in Zeo1.b.

Other potential HvAP2 targets include HvMADS2 (HORVU3Hr1G091000.8) and HvMADS4 (HORVU1Hr1G063620.2), the orthologue of which, the ‘B’ class PISTILLATA (PI) gene, is a direct target of Arabidopsis AtAP2 (Krogan et al., 2012). HvMADS2 expression increased in gigas1.a at WD5, and HvMADS4 expression showed no HvAP2-dependent differences (Fig. S8). We also examined the expression of HvMADS3 (HORVU3Hr1G026650.1) and HvMADS58 (HORVU1Hr1G029220.1), two AGAMOUS-like genes, the orthologues of which in Arabidopsis are direct targets of AtAP2 (Zhao et al., 2007; Yant et al., 2010; Ripoll et al., 2011). HvMADS3 was lower in Zeo1.b at WD5 and HvMADS58 expression was lower in gigas1.a at WD4 compared with Bowman and Zeo1.b (Fig. 3D; Fig. S8). Reduced HvMADS58 expression could contribute to gigas1.a lodicule and pistil phenotypes, as OsMADS58 is essential for carpel and lodicule identity in rice (Yamaguchi et al., 2006; Dreni et al., 2011). Altogether, our comparative gene expression analyses link HvAP2-dependent changes in floral organ development with a specific MADS-box gene misexpression. Whether this relationship is direct is unknown.

To learn when HvAP2 alters grain parameters, we tracked changes in pre-anthesis ovary and caryopsis dimensions at days post anthesis (DPA). Depth was measured along the dorsal:ventral axis. Compared with Bowman, gigas1.a ovaries were 12% longer, 20% wider, 24% deeper (all P<0.001) and 70% lighter (P<0.01; Fig. 4A,B; Fig. S9; Table S6). Post-fertilisation, gigas1.a caryopses became 36% and 55% longer at 10 and 30 DPA, respectively, (P<0.001) and remained narrower during most growth (P<0.05; Fig. 4A,B; Table S6). Caryopses in gigas1.a were 9-16% shallower at 15-25 DPA (P<0.05) and 10-50% lighter (P<0.01) compared with Bowman until final stages (Fig. S9; Table S6). Pre-anthesis, Zeo1.b ovaries were 16% deeper (P<0.001), 6% shorter (P<0.001) and 36% lighter (P<0.01) and, following fertilisation, Zeo1.b caryopses at 5 DPA were 30% shorter (P<0.05) and afterwards showed no clear length or depth trend but became progressively heavier (Fig. 4A; Table S6). Zeo1.b caryopses were 20% wider (P<0.001) and 10% heavier (P<0.01) than Bowman by 30 DPA (Table S6). Overall, HvAP2 activity positively correlated with wider and heavier grain during grain fill, whereas loss of HvAP2 function lengthened and narrowed grain.

We examined transverse sections of developing caryopses to explore potential causes of HvAP2-dependent differences. Bowman caryopses showed a dumbbell-shaped embryo sac with dorsal indentations at 5 DPA, whereas the gigas1.a embryo sac was rectangular-shaped and the Zeo1.b embryo uniformly oval (Fig. 4C). Variation in embryo sac shape correlated with differences in the lateral and dorsal mesocarp provascular strands, structures that supply nutrients to the pericarp before degenerating (Fisher, 1990). Caryopses from gigas1.a displayed provascular strands, Bowman showed remnants of provascular strands and Zeo1.b lacked any trace of provascular strands (Fig. 4C), suggesting that HvAP2 accelerates provascular degeneration, potentially influencing the shape of the expanding embryo sac. The nucellar projection, a specialised structure that differentiates from the nucellus, also differed amongst alleles. Flanked by a pigment strand and vascular bundle, the nucellar projection transports nutrients from maternal to filial tissues before undergoing programmed cell death (PCD), leaving a large cavity (Radchuk et al., 2006; Thiel et al., 2008; Domínguez and Cejudo, 2014; Lu and Magnani, 2018). At 10 DPA, nucellar projection degeneration was least noticeable in gigas1.a, whereas Zeo1.b showed more breakdown compared with Bowman, with corresponding differences in cavity size (Fig. 4D). Thus, HvAP2 may promote breakdown of nucellar tissues. We also detected differences in the integuments across HvAP2 alleles. Grass ovules have outer and inner integuments, and the outer integument degrades rapidly following fertilisation (Kellogg, 2015). At 5 DPA, Bowman and Zeo1.b showed the two-layered inner integument, whereas gigas1.a had inner integuments plus multiple layers of enlarged integument-like cells, which may be a non-degraded outer integument (Fig. 4E). By 10 DPA, the inner integuments of Bowman and Zeo1.b were crushed, along with most of the nucellar epidermis, to form the seed coat (Fig. 4F). At 10 DPA, the inner integuments and nucellar epidermis of gigas1.a also compressed into the seed coat, but the extra integument-like layers persisted, separated from the pericarp by a prominent cuticle (Fig. 4F). Altogether, loss of HvAP2 function inhibited or profoundly delayed the degradation of multiple maternal tissues while gain of HvAP2 function promoted the degradation of multiple maternal tissues. We propose that HvAP2 is a crucial negative regulator of maternal tissue growth and survival during grain development.

We asked whether HvAP2 expression correlated with its proposed roles in the grain. HvAP2 transcripts localise to lodicule primordia as well as the lemma and glume (Nair et al., 2010; Anwar et al., 2018) but the HvAP2 expression pattern in grain has yet to be reported. We detected HvAP2 mRNA in the lemma and pre-anthesis ovary, and concentrated expression in the vascular bundle, pigment strand and nucellar projection at 2 DPA (Fig. 4G; Fig. S10). By 5 and 10 DPA, HvAP2 transcripts persisted in remaining nucellar tissues, seed coat and pericarp (Fig. 4G). Taken together, HvAP2 transcripts accumulate in tissues altered owing to HvAP2 allelic variation.

We used qPCR to examine expression levels of potential downstream genes in the developing grain. We were particularly interested in HvMADS29 (HORVU6Hr1G032220), a gene almost exclusively expressed in young grain (Fig. S7) and the orthologue of a B-sister gene (the closest relatives of B-class genes) in rice called OsMADS29. Downregulation of OsMADS29 in rice inhibits degradation of all maternal tissues, including the nucellus and nucellar projection, leading to shrunken seeds (Yin and Xue, 2012; Yang et al., 2012). HvMADS29 expression was reduced to 37% of the Bowman level in gigas1.a grain at 5 DPA (Fig. 4H). Although the ‘C’ class HvMADS3 and HvMADS58 are also highly expressed in grain (Figs S7,S11), neither were differentially regulated in gigas1.a 5 DPA grain (Fig. S12). Jasmonate (JA) positively regulates nucellus PCD in tomato (Schubert et al., 2019) and is associated with PCD in barley grain (Sreenivasulu et al., 2006). Our previous work suggested that HvAP2 promotes JA-associated gene expression to suppress growth (Patil et al., 2019). Here, we show that transcripts encoding two JA-induced proteins (JIP), JIP23 and JIP60, expressed during barley grain development (Figs S7,S11), were strongly downregulated in gigas1.a caryopses at 5 DPA compared with Bowman (Fig. 4H). In sum, lower levels of HvMADS29 and JIP transcripts agree with the reduced maternal PCD in gigas1.a and is consistent with being putative targets of direct or indirect regulation by HvAP2.

To assess the relationship between HvAP2 and HvMADS29 in greater detail, we addressed the role of HvMADS29 in caryopsis development. RNA-seq data from unfertilised pistils and developing grain staged by days after pollination (DAP) (Aubert et al., 2018) indicated that HvMADS29 expression increases during pre-anthesis stages before peaking at around 9 DAP and subsequently decreasing during grain development (Fig. 5A). Before fertilisation, HvMADS29 expression overlaps with HvAP2 (in addition to HvMADS1, HvMADS3 and HvMADS58), although HvMADS29 is subsequently maintained during grain development when HvAP2 levels decrease (Fig. S11). In situ hybridisation showed HvMADS29 expression in the developing ovule, predominantly in the nucellus, integuments and embryo sac (Fig. S11), in agreement with previous microarrays showing HvMADS29 expression in maternal tissues (Thiel et al., 2008). After fertilisation, HvMADS29 showed weak expression in the integuments and strong signals in the vascular bundles overlying the nucellar projection and in peripheral vascular bundles (Fig. 5B), overlapping with the pattern observed for HvAP2.

To directly demonstrate the importance of HvMADS29 for caryopsis development, we edited the HvMADS29 gene using the CRISPR/Cas9 system (Ma et al., 2015). Screening of 25 T0 transformants in Golden Promise revealed five lines (hvmads29) with a 1 bp deletion in HvMADS29, leading to a frame shift, early stop codon, and predicted complete loss of function (Fig. 5C). Caryopses of hvmads29 were malformed and shrivelled compared with Golden Promise, with differences in ovule development already evident at anthesis (Fig. 5D; Fig. S12). In particular, the two layers of inner integument cells showed similar morphology to Golden Promise but with an overall misshapen structure, possibly due to reduced nucellus growth. In addition, the outer integument cells in hvmads29 were abnormally enlarged and occasionally exhibited three cell layers rather than the two cell layers detected in Golden Promise (Fig. S13). Defects in caryopsis development were particularly obvious at 5 DPA, when hvmads29 grain appeared thinner and unfilled, suggesting defective transport tissues including the nucellar projection and vascular bundle (Fig. 5E). Transverse sections of developing caryopses revealed that 5 DPA Golden Promise contained a well-developed nucellar projection, and a vascular bundle from the ventral pericarp joining the nucellar projection (Fig. 5F). Notably, the phloem cells intensely stained by Calcofluor White highlighted the vascular bundle in a semi-circle pattern. By 10 DPA, the nucellar projection increased in size and differentiated various cell types. In hvmads29, the nucellus failed to degrade, the nucellar projection did not differentiate, no transport tissue was formed and the vascular bundle was shrunken with fewer phloem fibres, compared with wild type (Fig. 5F). Importantly, hvmads29 mutants failed to produce endosperm or viable seed. Collectively, the putative hvmads29-2 null allele showed several tissue-specific defects similar to gigas1.a in terms of defective integument development and an under-developed nucellar projection. Along with the qPCR, RNA-seq and in situ data, this provides additional evidence that HvAP2 and HvMADS29 both influence the development and differentiation of transport tissues, particularly in the nucellar projection. We speculate that HvAP2 may contribute to upstream control of HvMADS29 expression, but this relationship remains to be tested.

Although HvMADS29 and HvMADS1 may mediate some HvAP2 functions, we know little about other factors interacting with HvAP2 beyond miR172. Deletion of the transcription factor-encoding gene HvBOP2 in laxatum.a (lax.a, BW419) causes lax spikes, elongated narrow paleas and lemmas, an impaired lemma-awn boundary, skinny grain, and transforms lodicules into stamens, leading to open-flowering (Jost et al., 2016; Fig. 6A-I). We noted that lax.a also develops needle-like glumes, elongated narrow ovaries and shorter lemma awns (P<0.04; Fig. 6C,D; Fig. S14). As lax.a and gigas1.a share some phenotypes, we speculated that HvAP2 and HvBOP2 may interact. To explore this hypothesis, we generated gigas1.a lax.a and Zeo1.b lax.a double mutants (Table S13). Most double mutant phenotypes suggested independent, additive roles of HvAP2 and HvBOP2. The gigas1.a lax.a double mutant displayed more extreme open-flowering than either parent, needle-like lax.a glumes equivalent in length to gigas1.a (P=0.86), lemma length equivalent to lax.a (P=0.31) and lemma awns shorter than either parent (P<0.02), and with a more impaired awn-lemma boundary, as well as narrow ovaries lacking stigmatic papillae and extremely narrow grain (P≤0.001; Fig. 6A-I; Figs S14, S15; Table S7). Zeo1.b lax.a lemmas had shorter awns (P≤0.001) similar to Zeo1.b, and defective awn-lemma boundaries similar to lax.a (Fig. 6E,F; Fig. S14; Table S7). The Zeo1.b lax.a mutant also retained Zeo1.b glume-to-lemma transformation but these organs were extremely thin with longer awns, indicating that loss of HvBOP2 influences lemma morphology regardless of position, (P≤0.001; Fig. 6C,D; Table S7). Grain of gigas1.a lax.a appeared narrower than either parent, suggesting an additive effect (P≤0.05; Fig. 6I; Fig. S15). Interestingly, the Zeo1.b lax.a mutant showed both lax.a elongated lemmas while also displaying Zeo1.b shortened grain (P<0.05; Fig. 6I,J; Fig. S15, S16), suggesting that these two traits can be uncoupled. However, Zeo1.b lax.a double mutants also showed striking epistasis in other features. Double mutants showed Zeo1.b-like spike density (P≤0.001) and a recovery of lodicule identity with Zeo1.b-like morphology (Fig. 6A,B,G,H; Figs S14, S15, Table S6), whereas Zeo1.b/+ lax.a plants showed lodicule/stamen mosaic-like structures (Fig. S16). Thus, we propose that HvAP2 may act downstream of HvBOP2 in the control of spike density, grain length and lodicule identity. To explore the molecular nature of this interaction, we analysed HvAP2 and Hvmir172 expression in Bowman and lax.a mutant plants. We detected no difference in HvAP2 mRNA levels in entire seedlings at 2 weeks after planting or developing spikes; however, we detected slightly elevated levels of HvmiR172 in lax.a as well as lowered levels in Zeo1.b, suggesting that HvAP2 suppresses levels of HvmiR172 and that HvBOP2 may enhance HvmiR172 expression (Fig. S17).

Cereal yield depends upon multiple factors including floret number, coordinated growth of floral organs, fertilisation and relocation of maternal nutrients to the developing seed (Brinton et al., 2017; Wilkinson et al., 2019; Sakuma and Schnurbusch, 2020; Ren et al., 2020; Paul et al., 2020). Targeted modification of individual factors is challenging because they typically have both direct and indirect impacts on grain size, number and quality (Wang et al., 2012; Xie et al., 2015; Si et al., 2016; Bull et al., 2017; Wilkinson et al., 2018; Li et al., 2019). One way to dissect pleiotropic effects is to characterise key regulatory modules and genes that contribute to tissue-specificity. Here, we show via analysis of the gigas1.a, hvap2 and Zeo1.b mutants that HvAP2 is a major regulator of barley reproductive development that influences tissue-specific factors in the flower and seed (Fig. 7).

Effects on the perianth from gain and loss of HvAP2 function alleles (Figs 1 and 2) and restored lodicule identity in Zeo1.b lax.a (Fig. 6) show that HvAP2 promotes perianth organ identity and defines the outer perianth boundary, suggesting that HvAP2 participates in the commitment to floret fate. Barley lacking HvmiR172 expression and wheat with strong overexpression of TaAP2L5 show a complete conversion of glumes to florets (Brown and Bregitzer, 2011; Debernardi et al., 2017; Greenwood et al., 2017; Song et al., 2019). As homeotic lemmas in Zeo1.b did not enclose floret organs, we suggest that gain of HvAP2 function in Zeo1.b is insufficient for ectopic florets and instead gives rise to ‘sterile lemmas’ – intermediate organs normally found in cultivated rice, and in wheat with moderate overexpression of TaAP2L5 (Debernardi et al., 2017; Greenwood et al., 2017; Song et al., 2019). Our interpretation supports previous suggestions that glumes, sterile lemmas and lemmas develop on an ontogenetic gradient that determines the fate of their axillary meristem, from glumes subtending spikelet meristems, to empty sterile lemmas, and finally to fertile lemmas subtending floret meristems (Lee et al., 2007; Chuck et al., 2008; Lee and An, 2012; Song et al., 2019; Debernardi et al., 2020). We propose that the gradient shifts towards fertile lemma identity with increasing AP2L function, making AP2L genes master regulators of floret establishment in grasses.

Control of floral fate by AP2Ls likely involves the LHS1-like subclade of the ‘E’ class SEPALLATA genes considered central for the evolution of the floret-bearing grass spikelet (Malcomber and Kellogg, 2004). The rice APL2s SNB and IDS1 promote the expression of the LHS1-like OsMADS1 which confers perianth organ identity and accelerates the transition from spikelet to floret meristem fate (Jeon et al., 2000; Prasad et al., 2001, 2005; Ohmori et al., 2009; Lee and An, 2012; Khanday et al., 2013; Dai et al., 2016). HvAP2-dependent changes in HvMADS1 expression (Fig. 3D,F; Fig. S5B), suggest that HvMADS1 also promotes perianth and floret identity in barley. Elevated JA-signalling in Zeo1.b (Patil et al., 2019) may also contribute to HvAP2 regulation of HvMADS1 as JA-signalling upregulates OsMADS1 expression in rice (You et al., 2019). OsMADS1 inhibits miR172 accumulation and possibly directly regulates AP2Ls (Khanday et al., 2016; Dai et al., 2016). We propose that HvMADS1 and HvAP2 co-expression (Digel et al., 2015), and our data showing HvAP2-responsive HvMADS1 expression, reflect positive feedback that may coordinate identity switches and hormone signalling to ensure a sharp transition from spikelet to floret fate. However, as substantial HvMADS1 levels persist in gigas1.a (Fig. 3D), factors besides HvAP2 must also upregulate HvMADS1 expression.

AP2 was first described in Arabidopsis as a class ‘A’ gene conferring sepal and petal identity (Bowman et al., 1989, 1991; Kunst et al., 1989; Drews et al., 1991). Weak ap2 alleles develop stamenoid petals and stronger ap2 alleles show carpel-like transformation in the sepal and petal whorl, phenotypes associated with loss of ‘B’ and ‘E’ class function and expanded ‘C’ class function (Kunst et al., 1989; Drews et al., 1991; Jack et al., 1992; Modrusan et al., 1994; Goto and Meyerowitz, 1994). In wheat, loss of both TaAPL2 and TaAPL5 function leads to carpel-like structures on lodicules, consistent with increased ‘C’ class MADS3/MADS58 (TaAG1/TaAG2) expression and reduced ‘B’ class gene expression at WD3.5-WD4.25 stages (Debernardi et al., 2020). We did not observe carpelloidy in gigas1.a or hvap2-2 lodicules, which instead showed bract and filament-like transformations (Fig. 1J-L; Fig. 2F-H). This may reflect reduced HvAP2 activation of HvMADS1, as SNB/OsIDS1-dependent expression of OsMADS1 is important for lodicule formation development (Jeon et al., 2000; Prasad et al., 2001, 2005; Lee and An, 2012), rather than HvAP2 regulation of B or C class genes. However, Zeo1.b WD5.5 spikelets showed reduced HvMADS3 mRNA abundance (Fig. S8). In rice, ectopic expression of OsMADS3 converted lodicules to stamens (Kyozuka and Shimamoto, 2002), an identical phenotype to lax.a (Jost et al., 2016; Fig. 6G). Thus, Zeo1.b-dependent reductions in HvMADS3 expression may help suppress homeotic stamen identity in Zeo1.b lax.a lodicules (Fig. 6G). Thus, we propose that HvAP2 may control lodicule differentiation through regulating both ‘C’ and ‘E’ class genes. Nonetheless, lodicules in gigas1.a or hvap2-2 usually retained lodicule features, so other genes must confer lodicule identity either along with HvAP2 or when HvAP2 function is impaired, consistent with redundant control of lodicule identity amongst wheat AP2Ls (Debernardi et al., 2020). In Arabidopsis, AtBOP promotes AtAP2 function via the miR172-AP2 network (Khan et al., 2015). Although we did not detect changes in HvAP2 expression in lax.a, we speculate that HvBOP2 could regulate HvAP2 at a protein level or that HvBOP2 may promote the function of other miR172-regulated HvAP2L genes to regulate lodicule identity.

Hulls are proposed to physically limit grain size in rice (Li and Li, 2016; Li et al., 2019) and multiple rice grain size quantitative trait loci control hull cell number and/or expansion (Song et al., 2007; Wang et al., 2012; Zhang et al., 2012; Si et al., 2016; Ren et al., 2016, 2018; Lyu et al., 2020; Ruan et al., 2020). Here, we found that hulls and grain in gigas1.a and hvap2-2 equally elongated (Fig. 1C,D,L-N; Fig. 2D-E; Fig. S4), suggesting that HvAP2 limits both hull and grain length. Interestingly, the mechanism differs between these tissues, with HvAP2-dependent repression of cell expansion underlying changes in lemma length, whereas HvAP2 is required to suppress both cell length and number in the pericarp. Caryopses in gigas1.a extended longer than Bowman at 10 DPA, corresponding with the timing of pericarp cell expansion in barley (Radchuk et al., 2011). As pericarp cell number and length were increased in gigas1.a, HvAP2 could limit this final cell longitudinal expansion event as well as earlier proliferation, similar to its role in the internode (Patil et al., 2019). The role of HvAP2 in ovary wall cell length control appears to be conserved in grasses, where the rice SNB shortens both hull cell length and pericarp epidermis cell length (Jiang et al., 2019; Ma et al., 2019), as well as in Arabidopsis where miR172-resistant AP2 represses cell expansion in the replum valves (Ripoll et al., 2011). However, although multiple genes, including HvAP2, may modulate grain length by influencing hull length, how the hull mechanistically limits caryopsis growth is largely unexplained. For example, do changes in grain length occur in direct response to the hull, such as surface-surface contact and/or mechanical pressure? Uncoupling of hull from grain length in the Zeo1.b lax.a double mutant suggests that HvAP2 and HvBOP2 may be key nodes in this communication. Learning the identity of putative hull ‘signals’ and how they might synchronise caryopsis development with hull proportions would advance our understanding of the control of cereal grain size.

Although increased grain length can lead to heavier grain (Zhang et al., 2012; Brinton et al., 2017), TGW increased by only 7% in gigas1.a as its grain is narrower and shallower (Figs 1,4; Figs S2,S8). Our data links HvAP2-dependent variation in grain width and depth to an altered balance between maternal versus filial growth and survival. Most strikingly, multiple maternal tissues in gigas1.a show defective and/or delayed degradation (Fig. 4), which may reduce remobilisation of nutrients and/or space to enlarge, both considered crucial for endosperm growth (Radchuk et al., 2006; Thiel et al., 2008; Domínguez and Cejudo, 2014; Wilkinson et al., 2019). Impaired maternal elimination correlates with defective endosperm development across plants (Radchuk et al., 2011; Yin and Xue, 2012; Domínguez and Cejudo, 2014; Xu et al., 2016), with nucellar degradation playing a major role in promoting endosperm growth (Lu and Magnani, 2018). We show here that HvMADS29 expression is significantly reduced in gigas1.a (Fig. 4H) and demonstrate for the first time in temperate cereals that MADS29 function is essential for nucellar differentiation and degradation, control of integument growth and endosperm development. We propose that HvAP2 controls grain width and weight in part by influencing the rate of maternal degradation via MADS29-driven processes, suggesting that coordination of filial endosperm expansion with the maternal tissue degradation and differentiation is at least partially under phase change miR172/AP2 control. In rice and Arabidopsis, the B-sister genes OsMADS29 and TRANSPARENT TESTA16 (TT16), respectively, promote the degeneration of nucellar and other maternal tissues in response to auxin produced from the endosperm following fertilisation (Yin and Xue, 2012; Yang et al., 2012; Nayar et al., 2013; Xu et al., 2016; Lu and Magnani, 2018). Whether HvAP2 directly regulates HvMADS29 and/or other activators of nucellar elimination in response to filial signals remains a pressing question.

In Arabidopsis, signals from the endosperm transform integuments into the seed coat (Figueiredo et al., 2016), a process sustained by mechanical pressure from the expanding endosperm (Creff et al., 2015). This in turn limits endosperm growth (Garcia et al., 2005), highlighting a developmental interdependency which may underlie ap2 mutant seed phenotypes. AtAP2 appears necessary to restrict integument cell expansion, promote seed coat epidermal differentiation, accelerate endosperm cellularisation and constrain endosperm cell expansion, roles linked to limits on seed weight and size, embryo size, storage protein accumulation and sugar metabolism (Jofuku et al., 1994; Ohto et al., 2005, 2009). Our data suggest that HvAP2 limits integument number in barley, potentially by promoting the degradation of the outer integument, showcasing a role for an AP2L gene in cereal integument development. This role may relate to regulation of HvMADS29, as anthesis-stage hvmads29 ovules showed abnormally enlarged cells in disorganised integument layers. In Arabidopsis, TT16 coordinates communication between the integuments and endosperm (Xu et al., 2016), promotes inner integument flavonoid deposition and differentiation (Nesi et al., 2002) and controls outer integument thickness (Fiume et al., 2017), while a recently duplicated B-sister gene, GORDITA (GOA) contributes to outer integument differentiation (Prasad et al., 2010). Although HvMADS29 is clearly not relevant to all functions of HvAP2, tissue-specific regulation of one or more of the three barley B-sister genes (Yang et al., 2012) by HvAP2 may explain the darker seeds and persistence of the outer integuments in gigas1.a, in addition to alterations in the rate of nucellar degradation. Increased proanthocyanidins in the seed coat of barley are associated with increased dormancy (Himi et al., 2012). We observed that germination of gigas1.a grain was less efficient compared with wild type, suggesting that HvAP2 may influence seed germination through its effects on the seed coat.

Changes in lemma and palea dimensions control the overall shape of the floret, influencing final grain size while lodicule size and swelling leads to open flowering. Similar to gigas1.a, spikelets of wild barley (Hordeum spontaneum) have elongated lance-shaped hulls and show open flowering compared with the wedge-shaped form of cultivated barley (Abdel-Ghani et al., 2004; Clayton et al., 2006). Although wild barley populations show large variation in seed size (Chen et al., 2004), cultivated barley grain tends to be shorter, wider and heavier (Fuller, 2007; Hughes et al., 2019) with more uniform germination (Fuller and Allaby, 2018). Our data suggest that HvAP2 controls multiple traits which differ between wild and cultivated barley. In wheat and rice, selection of allelic variation in AP2L genes was associated with improved grain traits (Xie et al., 2018; Jiang et al., 2019). It is tempting to speculate that changes in HvAP2 function and/or HvMADS regulation contributed to selection for changes in spikelet and grain during barley cultivation.

Parent cultivars and mutant germplasm are listed in Table S1. Plants were grown in the glasshouse under 16 h light/8 h dark and day/night temperatures of 18/15°C. Plants were grown in plastic pots filled with universal compost [1200 l of peat, 100 l of sand, 2.5 kg of magnesium limestone, 2.5 kg of calcium limestone, 1.5 kg of Osmocote^® Start (ICL Speciality Fertilizers), 3.5 kg of Osmocote^® Exact Standard 3-4M (ICL Speciality Fertilizers), 0.5 kg of Celcote, 100 l of Perlite, 390 g of Intercept insecticide (active ingredient: imidacloprid)]. Golden Promise and hvmads29 plants were grown in controlled environment reach-in chambers in The Plant Accelerator, the University of Adelaide, Australia, under the same conditions as all other plant materials which were grown in the UK. gDNA from the gigas1.a (BW381), Bowman and Golden Melon were genotyped using the Barley 50K SNP chip (Bayer et al., 2017).

Whole plant phenotypic measurements were taken from mature plants. Spikelet length was measured from the base of the spikelet to the lemma-awn boundary on the fourth spikelet from the base of the spike. Awns were measured from awn tip to the top of the glume or lemma body. Spike length was the length from the collar node to the top of the rachis (spike axis). Spikelet width was measured at the widest part of the lemma. Culm height was measured from the top of the soil to the collar at the base of the spike. Mature and developing Bowman, Zeo1.b and gigas1.a spikes were harvested at 21, 23, 25, 30 and 35 days after germination, their length recorded and stages assigned based on Waddington et al. (1983) (Table S8). Scanning electron microscopy was performed as previously described (Houston et al., 2013). Mature grain width and length were analysed using MARVIN-Universal (GTA Sensorik GmbH). Developing caryopses (n=5 independent replicate grains per genotype) were sampled on their respective DPA. For Golden Promise and hvmads29 samples, caryopses were collected at anthesis, 5 and 10 DPA, photographed by stereo microscope (Leica, MZ FLIII) or fixed in FAA solution, dehydrated in an ethanol series and embedded in Technovit 7100 resin (Kulzer Technique). Transverse 1.5 μm sections were stained with Calcofluor White (Sigma-Aldrich) or 1% Toluidine Blue. Sections were photographed using a Zeiss AxioImager M2 for cell wall (excitation, 335-383 nm; emission, 420-470 nm) and auto fluorescence (excitation, 538-562 nm; emission, 570-640 nm).

HvAP2 in situ hybridization was performed as previously described (Hands et al., 2012) and HvMADS29 in situ hybridisation was performed automatically using an InsituPro VSi robot (Intavis), following a standard protocol (Javelle et al., 2011). cDNA fragments of 2225 bp and 319 bp were amplified from Bowman cDNA using primers fused to the T7 promoter as a template for HvAP2 and HvMADS1 in situ probes, respectively (Table S9). Digoxigenin-labelled antisense and sense probes were transcribed using T7 polymerase (Thermo Fisher Scientific) according to the manufacturer's instructions.

CRISPR/Cas9 technology was used to generate mutations in HvAP2 at the University of Dundee (Garcia-Gimenez et al., 2020 preprint). Two guide RNAs (gRNAs) were designed (Table S9) using the Broad Institute sgRNA Designer and the Zhao Bioinformatics Laboratory pssRNAit (Noble Foundation). Each gRNA was cloned into pC95-gRNA entry vector downstream of the rice small nuclear RNA (snRNA) U6 promoter (OsU6p) by Gibson Assembly^®. Each sgRNA cassette was then released and inserted into the pBract214m-bcoCas9-HSPT expression vector which contains a bcoCas9 under the control of maize ubiquitin promoter and Arabidopsis heat shock protein 18.2 terminator. The resultant construct was transformed by electroporation into Agrobacterium strain AGL1 containing replication helper pSoup. Transformed Agrobacterium clones from each CRISPR construct were combined and co-transformed into Golden Promise immature embryos (Bartlett et al., 2008) in the FUNGEN facility at The James Hutton Institute, Dundee, UK. Transgenic plants containing CRISPR constructs were regenerated under hygromycin selection. Of the 174 T0 plants, 143 were transformed with a single gRNA from 35 separate calli and 31 plants transformed with the two gRNAs from 12 different calli. No mutations were found in the single gRNA transformation lines and one was detected in the double transformation lines. We examined 16 individuals from 18 T1 lines (originating from 10 different T0 calli) which still contained the Cas9 gene. We detected three homozygous mutations in these T1, including a 39 bp deletion (hvap2-1) in nine different lines from three different calli and a 40 bp deletion (hvap2-2) in four lines from two different calli. In the T2 generation, following segregation of Cas9, hvap2-1, hvap2-2 and Golden Promise were phenotyped (n=8 per genotype).

We used a monocot-optimised CRISPR/Cas9 system (Ma et al., 2015) to create the hvmads29 mutant at the University of Adelaide. The selected target of HvMADS29 was sequenced before the sgRNA expression cassette was amplified from vector pYLsgRNA-OsU6a and cloned into a binary vector pYLCRISPR/Cas9Pubi-H using BsaI sites as previously described (Ma et al., 2015). The CRISPR construct was transformed into Golden Promise by AGL1 as previously described (Harwood et al., 2009). A total of 25 T0 transformants were analysed in greater detail.

For HvAP2 CRISPR lines, genotyping conditions are described in the Supplementary Materials and Methods and primers are listed in Table S10. Genomic DNA (gDNA) was isolated from young leaf tissue using the Qiagen DNA easy plant mini kit. Cas9 was detected using Cas9 primers. A region spanning 1 kb around the gRNA target region was amplified using external primers followed by a nested PCR using FAM-labelled internal primers. This product was analysed using a capillary sequencer, and genotypes were determined using GeneMapper^® Software 5. Samples predicted to contain insertions and/or deletions (indels) were re-amplified without FAM-labelling and sequenced. T0 and T1 plants from HvMADS29 CRISPR transformation events were genotyped using a Phire Plant Direct PCR Kit (Thermo Fisher Scientific) to amplify a 588 bp fragment that was directly sequenced by Sanger sequencing (AGRF, Australia).

RNA extraction and qRT-PCR were performed as in Patil et al. (2019) with the following modifications: RNA was isolated from entire spikes harvested at 23, 25, 30, 35 and 40 days after sowing. cDNA was synthesised using ProtoScriptII kit (New England Biolabs) using random primers. The qPCR was normalised using RQ values calculated by the Pfaffl method 2^ (−ddCT) (Pfaffl, 2001). One replicate of Bowman at the earliest timepoint was normalised to 1.0 and each other sample replicates normalised to this value. We used ACTIN2 (HvACT2) and PROTODERMAL FACTOR7 (HvPDF7) as endogenous controls as in Patil et al. (2019). The SYBR Green Power Up kit (Thermo Fisher Scientific) was used to detect HvJIP23, HvJIP60, HvmiR172 and HvsnoR101 transcripts. Primers for qPCR are listed in Table S11.

Double mutants between gigas1.a or Zeo1.b with lax.a were generated by crossing. Double Zeo1.b lax.a and Zeo1.b/+heterozygote lax.a mutants were isolated by screening a segregating Zeo1.b/lax.a F2 population which showed the expected ratio of double homozygotes of 1:16 (Table S12). An F3 population from a Zeo1.b/lax.a F2 individual was grown and segregated as expected. The gigas1.a lax.a F2 population was screened by genotyping with CAPS markers to isolate double mutants (Supplementary Materials and Methods; Table S10).

Data were modelled in R 3.5.1 using ANOVA. Models were checked visually for normality in variance and any non-significant terms dropped from the model. Where only two genotypes were compared, a two-tailed unpaired t-test was performed. Multiple genotypes were compared using a Tukey honestly significant difference (HSD) test on the modelled data. Grain dimensions over time were analysed by ANOVA followed by a Dunn's post-hoc test.

S.D. acknowledges support from the University of Leicester and from Natalie Allcock and Ania Straatman-Iwanowska in the Electron Microscopy Facility. M.R.T. and X.Y. acknowledge technical support from Mr Chao Ma for operation of the in situ robot and plant maintenance.

The peer review history is available online at https://dev.biologists.org/lookup/doi/10.1242/dev.194894.reviewer-comments.pdf