Effect of genotype and environment on branching in weedy

Transcription

1 Molecular Ecology (2006) 15, doi: /j X x Effect of genotype and environment on branching in weedy Blackwell Publishing Ltd green millet (Setaria viridis) and domesticated foxtail millet (Setaria italica) (Poaceae) ANDREW N. DOUST and ELIZABETH A. KELLOGG University of Missouri-St Louis, Department of Biology, One University Boulevard, St Louis, MO 63121, USA Abstract Many domesticated crops are derived from species whose life history includes weedy characteristics, such as the ability to vary branching patterns in response to environmental conditions. However, domesticated crop plants are characterized by less variable plant architecture, as well as by a general reduction in vegetative branching compared to their progenitor species. Here we examine weedy green millet and its domesticate foxtail millet that differ in the number of tillers (basal branches) and axillary branches along each tiller. Branch number in F 2:3 progeny of a cross between the two species varies with genotype, planting density, and other environmental variables, with significant genotype environment interactions (GEI). This is shown by a complex pattern of reaction norms and by variation in the pattern of significant quantitative trait loci (QTL) amongst trials. Individual and joint analyses of high and low density trials indicate that most QTL have significant GEI. Dominance and epistasis also explain some variation in branching. Likely candidate genes underlying the QTL (based on map position and phenotypic effect) include teosinte branched1 and barren stalk1. Phytochrome B, which has been found to affect response to shading in other plants, explains little or no variation. Much variation in branching is explained by QTL that do not have obvious candidate genes from maize or rice. Keywords: barren stalk1, branching, foxtail millet, genotype environment interaction, QTL, teosinte branched1 Received 2 June 2005; revision accepted 4 October 2005 Introduction Plant architecture is a result of the initiation and differential elongation of growth axes (branches) (Bell 1991). These processes determine the eventual shape of the plant and allow for dynamic reshaping of its architecture under the influence of environmental stimuli such as shading, mechanical damage, and resource limitation (Thomas 2000). Production and differentiation of branches occur throughout the vegetative and reproductive phases of growth in most plants, and strongly influence resource acquisition, competitive ability, and reproductive success. In grasses, vegetative branches form at various levels above the ground (Gould & Shaw 1983; Clark & Fisher 1987). Most grasses form branches from the short basal nodes on the primary stem (culm). These are commonly Correspondence: Andrew N. Doust, Fax: ; adoust@umsl.edu known as tillers and may produce adventitious roots. Some grasses also produce branches in the axils of leaves along the stem (cauline axillary branches, called here simply axillary branches). Often these axillary branches only elongate after the meristem of the main culm has transformed from a vegetative meristem to an inflorescence meristem (Perreta & Vegetti 2004). Further axillary branches can be initiated on previously initiated axillary branches, leading to a much-branched bushy plant. The number and placement of the branches affect the distribution of the leaves and thus influence both light acquisition and shading of other plants. In many grasses, particularly annuals, all branches terminate in an inflorescence, so the number of branches that are initiated and grow may directly influence the number of seeds the plant produces. Branch production is controlled by a complex interplay of environmental inputs, hormonal responses, and genetic activity (McSteen & Leyser 2005). Light, particularly the red far-red ratio, affects plant architecture via signals mediated 2006 Blackwell Publishing Ltd

2 1336 A. N. DOUST and E. A. KELLOGG by the phytochromes, particularly phyb (Schmitt et al. 2003), and plant hormones such as auxin (Dharmasiri et al. 2005; McSteen & Leyser 2005). In maize, genes such as barren stalk1, knotted1, and teosinte branched1 affect axillary meristem formation, maintenance, and outgrowth, respectively (Sinha et al. 1993; Sinha & Hake 1994; Hubbard et al. 2002; Gallavotti et al. 2004). The signal transduction pathways that link all these components are beginning to be elucidated in a few model systems, including rice and maize (Takano et al. 2001; Sawers et al. 2002, 2004, 2005), but much remains to be learned about how environment and genetics interact in agricultural crops and weeds. The developmental plasticity shown by vegetative branching responses in weeds under different growing conditions is an important part of the weedy strategy (Perrins et al. 1992; Dekker 2003). Such genotype environment interactions (GEI) have been found to be important in both vegetative and reproductive architecture in a number of species (Pigliucci 1998; Juenger et al. 2000; Schlichting 2002, 2003; Schmitt et al. 2003; Ungerer et al. 2003; Koornneef et al. 2004; Weinig & Schmitt 2004; Weinig 2005). In many weedy grasses, axillary branches continue to be initiated as long as favourable conditions for plant growth continue, allowing for continuous flowering and seed production (Dekker 2003). Domestication effectively reverses some of the branching characteristics of weeds, and in domesticated grasses such as maize or millet has led to a sharp reduction in vegetative branching to a few tillers and axillary branches (Doebley & Stec 1991; Doust et al. 2004). In our recent work, we have investigated differences in branching patterns between a serious weed of temperate croplands, Setaria viridis (green millet) (Dekker 2003), and its domesticated derivative, Setaria italica (foxtail millet) [treated as subspecies by several authors, including Wang et al. (1995) and Mabberley (1987)]. The sister relationship between these two species has been confirmed by chromosomal fluorescent in situ hybridization (FISH) and ribosomal, nuclear and chloroplast sequence data (Benabdelmouna et al. 2001; Doust & Kellogg 2002; Doust, Penly & Kellogg, unpublished). It is likely that foxtail millet was domesticated from a native ruderal variant of green millet that would have displayed variable branching, but that would have been under minimal selection for weedy responses to human farming activities. Thus, the differences seen between the two species today can be viewed as the result of divergence from a common stock under different selection pressures. The morphological changes seen between the two species include a shift from many tillers and axillary branches, each terminating in a short inflorescence with relatively few orders of branching (typical weed architecture), to few tillers and no axillary branches, with long inflorescences that have many orders of branching (typical domesticate architecture) (Harlan 1992; Doust & Kellogg 2002; Doust et al. 2004, 2005). The loss of axillary branches in domesticated foxtail millet correlates with a decrease in the time over which inflorescences develop and mature, a characteristic that is important for efficient management and harvesting (Harlan 1992). This effectively reverses the broad dispersal period characteristic of the weedy progenitor. At the genotypic level, plasticity might be achieved by having multiple genes affecting a particular trait, with each gene responding to environmental inputs in a different way, or by having alternative suites of genes activated in different environments. Such genes can be identified using a quantitative trait loci (QTL) approach to locate regions of the genome that are differentially affected by the environment. Different combinations of alleles at these environmentally responsive regions should lead to different reaction norms in the plants that bear them. Our previous studies (Doust et al. 2004, 2005) identified loci that were minimally affected by environment, and we interpreted these as the likely sites of selection during domestication. Here we return to some of the same data, but consider loci with a significant interaction between genotype and environment; these are important for the study of plasticity. In addition, we consider the possibility that domestication loci might also be considered weediness loci. In this interpretation, the S. italica allele at a locus may have been artificially selected for domestication, but the S. viridis allele at the same locus may have been selected for weediness. Materials and methods Mapping, plant growth, and phenotype evaluation We used an F 2 mapping population derived from a cross between foxtail millet (Setaria italica) and green millet (Setaria viridis) that had previously been used to construct a genetic map, using 257 restriction fragment length polymorphism (RFLP) probes from rice, foxtail millet, pearl millet and wheat (Devos et al. 1998; Wang et al. 1998). As detailed previously, we used 119 of these markers for our QTL analysis, chosen to cover the genome at approximately 10-cM intervals (Doust et al. 2004,2005). An additional 23 RFLP markers from maize plus 6 known genes were also added to the original map. We grew F 3 offspring selfed from 120 of the original 127 F 2 plants in four separate trials, trials 1 and 2 at high density with 5 representatives per family, and trials 3 and 4 at low density with 15 representatives per family (trials 3 and 4 were labelled 1 and 2 in Doust et al. 2004). Soil, fertilizer, water, and day length were standardized, but natural light intensity and average temperatures differed slightly among the trials, as trials 1 and 3 were grown in early summer and trials 2 and 4 in late summer. In the low density trials, replicates of families were planted one to a pot, and

3 ENVIRONMENTAL EFFECTS ON BRANCHING IN MILLETS 1337 randomized with respect to position in the greenhouse. In the high density trials, the five representatives of each family were planted in a single pot, and therefore were not randomized in position in the greenhouse, and are not true replicates. Thus for analyses that compared high and low density trials, we used least square mean values only. Means were also used for QTL analyses, because these are the best approximation to the phenotypes of the F 2 plants used to construct the genetic map (Wang et al. 1998). Plants were harvested after the seeds had ripened. We counted the number of tillers (branches coming from the base of the plant) and the number of axillary branches (in the axils of culm leaves). In general these two classes of branching were easy to distinguish, although a few plants produced multiple axillary branches from new tillers late in the growing period. This seems to be a phenomenon separate from the production of tillers and axillary branches during normal growth, but could not be distinguished in this analysis because all measurements were done at harvesting. Preliminary observations suggested that, during normal growth, tillers are produced before flowering whereas axillary branches appear after the inflorescence terminating the main culm has been produced. Data exploration We examined the shape of the frequency histograms for the trait values in each trial and found evidence of two distributions for most trait/trial combinations. In particular, there was a large excess of plants that had no axillary branches, resulting in an extreme positive skew to the data in all trials. As well, in the high density trials, many of the plants had only a single tiller. Previous observations (Doust et al. 2004) indicated that the foxtail millet parent probably lacks axillary meristems and therefore cannot produce axillary branches. Therefore, the large number of plants that do not have axillary branches may be the result of segregation of a gene responsible for this phenotype in the foxtail millet parent. However, lack of axillary branching may also be the effect of crowding (more families had no axillary branching in high density than in low density trials). It is also possible that some replicates in a family could have axillary branches while others lack them, because of incomplete segregation in these F 3 families. Therefore, to analyse quantitative variation in axillary branch number on only the subset of families that we could be confident had the potential to express such variation, we took low density trial 4 as our guide and pruned families that had means for axillary branch number that were less than one. This takes account of the possibility that some families were segregating for the lack of axillary branches, while keeping in the analysis families that had no axillary branches at high density (possibly as the result of crowding) but had axillary branches at low density. After pruning, the data set for quantitative analysis consisted of 79 families. The frequency distributions of the pruned data set approached normality more closely than that of the full data set. Residuals obtained from a univariate analysis of each trait, using trial as the factor, indicated that a log transformation would equalize the variances amongst the four trials for axillary branch number. Tiller number in the low density trials was normally distributed, but we could not find any satisfactory transformation that would convert tiller number in the high density trials to a more normal distribution. Outliers were identified for each trait/trial combination by examining boxplots and searching for extreme values using the Descriptive Statistics Explore command (SPSS 2005). Analyses were run with and without outliers, and differences noted where present. Genetic correlations and genotype by environment interactions Genetic correlations were calculated for both the full and reduced data sets for the low density trials, but the small variances and extreme non-normality of trait distributions of tiller number in the high density trials precluded statistical analysis. Genetic correlations of tiller and logtransformed axillary branch number in each trial examine the extent to which the two traits might be under common genetic control (pleiotropy). Genetic correlations were also calculated for each trait between trials (e.g. tiller numbers in trial 3 vs. those in trial 4). These correlations indicate the extent to which the same set of genes underlies variation in that trait amongst trials. Genetic correlations were estimated as cov [trait1, trait2] /σ trait1 σ trait2 (eqn 1) where cov [trait1, trait2] is the covariance of the cross-product of the correlation between the two traits, and σ trait1 and σ trait2 are the square roots of the variance components for each trait (Robertson 1959; Falconer & Mackay 1996). Variance components were estimated using restricted maximum-likelihood estimation (SPSS 2005). Reaction norm plots were also assembled for the four trials, using all 120 families, to highlight both general trends between trials, as well as genotype environment interactions. Analysis of variance Two sets of analyses were performed. One set compared log-transformed axillary branch numbers (LOGAXB) between high and low densities and amongst the four trials using mean values for each family and the reduced data set

4 1338 A. N. DOUST and E. A. KELLOGG of 79 families. It was not possible to satisfy conditions for analysis of variance with the two high density tiller trials, so tiller number (TILL) was not analysed in this set of analyses. A nested GLM anova was used to partition variance into sources attributable to density [D], F 2:3 family [F], trial nested within density [T(D)], and the interaction between these. The sources of variation in the model can be represented as y = µ + D + T(D) + F + D F + T(D) F + E (eqn 2) where µ represents the overall mean of the experiment, D is a fixed effect, and T(D), F, D F, and T(D) F are random effects. T(D) F cannot be tested in this experimental design because the use of a single mean value for each family results in too few degrees of freedom. Thus, this source of variation was included in the error, E. Family and trial are treated as random effects because the families available for analysis are only a random selection of all possible genotypes that could be produced from this cross, and because the timing of the four trials was not planned. Because both trial and family are random factors in this analysis, the nested design does not allow an F-test to test the effect of density by itself (Quinn & Keough 2002, p. 318). Instead, a quasi F-test is presented in the nested analysis to test the effect of density, using a combination of the mean squares of all of the interaction terms as a denominator (Quinn & Keough 2002; SPSS 2005). A second set of analyses used data only from the low density trials, where values for individual replicates could be used. Here, the model can be represented as y = µ + T + F + T F + E (eqn 3) Both TILL and LOGAXB were analysed for the low density trials. These two variables were also used as dependent variables in a multivariate anova. Analyses were done using the GLM anova programs in SPSS (SPSS 2005). Variance components were estimated using restricted maximum-likelihood estimation (SPSS 2005), and used to calculate the percentage of variance explained by each model effect. QTL analyses (detailed below) suggested three candidate loci; these were analysed for TILL and LOGAXB in the low density trials using mean values for each family. The three loci replace the family effect, and were treated as fixed effects. A full factorial model was fitted, using trial and the three gene loci and all appropriate interactions between them. LOGAXB was also analysed in both high and low density trials, using a similar nested model as in the first anova above. Appropriate interactions are all those for which all allele combinations were present (that is, to measure an interaction between gene1 and gene2, it is necessary to have present amongst the 79 families all possible combinations of the three allelic states possible for each gene). After the first analysis, nonsignificant interactions were dropped from the model and the analyses re-run. These analyses had unequal numbers of samples in each group, so error mean squares for each of the terms were weighted to remove this bias (SPSS 2005). QTL analyses Means for all 120 families were used in the QTL analyses, as we wanted to explore all possible genetic regions (QTL) underlying variation in phenotypic traits. The small number of families used makes it likely that not all QTL regions can be identified and that the size of the effect of the identified QTL regions may be exaggerated (Beavis 1998). Our previous report on tillering and axillary branching incorporated data only from the two low density trials analysed individually (Doust et al. 2004). Here we present new data on the high density trials and a joint analysis of all data on tillering and axillary branching. In trials 1, 2 and 4, the parental and hybrid ranges were examined for evidence of transgressive segregation; parents of the cross were not incorporated into the planting design in trial 3. QTL were detected using composite interval mapping (CIM), as implemented in QTL Cartographer (Basten et al. 2002). Background markers were selected at P = 0.05, and five background parameters were included as cofactors in each CIM model. Tests were made at 2.0-cM intervals, with a window size of 10 cm. QTL in separate trials were considered to be identical if their 1-LOD support intervals overlapped, and the sign of their additive effects was the same. Joint analysis of each trait for all four trials taken together was analysed using the module JZmapqtl (Basten et al. 2002). The joint analysis allows an estimate of GEI effects between the trial values for each trait, and provides a measure of both the main and interaction effects of detected QTL. Significance thresholds for QTL were calculated by 1000 permutations of the original data, using the same parameter settings as for the original analysis (Churchill & Doerge 1994; Doerge & Churchill 1996). We calculated significance levels both genome-wide and chromosome-wide (both at P < 0.05). Identification of QTL based on multiple chromosome-wide significance levels will increase type I error compared to the genome-wide level because nine different tests are being performed (one for each chromosome), but will also increase the probability of identifying more true QTL (Cheverud 2001). We considered occurrence of QTL significant at the chromosome-wide P < 0.05 level in more than one trial as evidence that a real QTL had been detected. The program epistacy (Holland 1998) was used to identify digenic epistatic interactions. There are ( )/2

5 ENVIRONMENTAL EFFECTS ON BRANCHING IN MILLETS 1339 Fig. 1 Histograms of tiller number and axillary branch number in the four trials (using means from all 120 families). The continuous and dashed arrows show the positions of the means for foxtail and green millet, respectively, in trials 1, 2, and 4. HD, high density; LD, low density. Fig. 2 Scatter plots of tiller number vs. axillary branch number in each of the four trials (using means from all 120 families). TILL, tiller number; AXB, axillary branch number; HD, high density; LD, low density. comparisons that are made in this analysis, and a Bonferroni correction to the P < 0.05 experiment-wide significance level gives a per-comparison significance level of (Rieseberg et al. 2003a). To identify possible candidate genes from maize and rice, we used markers mapped on maize and foxtail millet or rice and foxtail millet to define intervals on the maize or rice maps that correspond to QTL regions in foxtail millet. MaizeGDB (Lawrence et al. 2004) and Gramene (Ware et al. 2002a, b) were used to identify genes that had mutant phenotypes or putative functions that might affect some aspect of branching. Several of these were analysed by anova, as detailed above. Results Phenotypic distribution of traits Both parents adjusted their number of tillers to planting density, but only green millet modified the number of axillary branches, producing fewer axillary branches at high than at low density (Fig. 1). Conversely, foxtail millet never produced axillary branches. In the three trials in which the two parents were grown (trials 1, 2 and 4), green millet consistently had more tillers than foxtail millet (Fig. 1). At high density, both green and foxtail millet produced fewer tillers than at low density. Our observations indicate that an increase in branching in either parent always increases the amount of seed produced, even if the inflorescences are somewhat smaller on laterproduced branches. Hybrids in trial 2 for tillering and trials 1, 2 and 4 for axillary branching showed transgressive segregation, where the trait values for the hybrids had values that exceeded the range between the two parents. However, no transgressive segregation was seen for tillering in trial 1 or 4, where the parental values were at the extremes of the range of values found for the hybrid population. In the F 2:3 families both tiller number and axillary branch number were much lower at high density than at low density. Genetic correlations The phenotypic relationships between tiller and axillary branch number in each trial indicate that there is a strong relationship in trial 3 but almost no relationship in the other three trials (Fig. 2). Genetic correlations between tiller and axillary branch number were examined for both

6 1340 A. N. DOUST and E. A. KELLOGG Fig. 3 Scatter plots of relationships for each trait between all possible pairs of trials (using means from all 120 families). the full data set (FDS) and pruned data set (PDS) in each of the two low density trials (3 and 4), and were significant for trial 3 (FDS r = 0.40, P < 0.001; PDS r = 0.38, P < 0.001) but not for trial 4 (FDS r = 0.10, NS; PDS r = 0.05, NS). This indicates that there is a significant overlap in genetic control of the two traits in one trial but not in the other. Scatter diagrams of each trait for pairs of trials indicate that there is a stronger relationship between trait values in the two low density trials than in the two high density trials (Fig. 3). Genetic correlations were not calculated for the high density trials but the relationship between tiller numbers in trials 3 and 4 was significant (FDS r = 0.72, P < 0.001; PDS r = 0.85, P < 0.001), as was that between logtransformed axillary branch numbers (FDS r = 0.64, P < 0.001; PDS r = 0.60, P < 0.001). This indicates that genetic control of variation in each trait is, to a large extent, similar between the two low density trials. To explore responses to environment, we examined variation in the reaction norms of families amongst trials (Fig. 4). A common trend in reaction norms is evident for tiller number, with low mean values at high density and higher mean values at low density. The reaction norms for Fig. 4 Plots of reaction norms for means of traits across the four trials. All reaction norms are plotted as light grey lines. As examples of the range of reaction norms observed, several were arbitrarily chosen and highlighted with thicker black lines, together with standard error bars for each trial. four exemplar families, chosen to show the range of variation in reaction norms, show the same general trend, although they also illustrate changes in both the magnitude of effects and in rank order (crossing of lines). The reaction norms for axillary branch number show a complex pattern of crossing reaction norms, with an especially large response in low density trial 3 (Fig. 4).

7 ENVIRONMENTAL EFFECTS ON BRANCHING IN MILLETS 1341 Factor d.f. MS F P value VC %V D NS T(D) *** F *** D F NS E Table 1 Nested anova for density (D), trial (density) [T(D)], and family (F) and their associated interactions for LOGAXB (using family means) VC, variance component; %V, percentage of total variance explained. Significance levels: *P < 0.05; **P < 0.01; ***P < Error term for MS D = MS T(D) + MS D F MS E ; Error term for MS T(D) = MS E ; Error term for MS F = MS D F ; Error term for MS D F = MS E. Table 2 anova for trial (T), family (F) and trial by family (T F) for TILL and LOGAXB (using replicates within families), for low density trials. Significance levels and abbreviations as for Table 1 Factor d.f. MS F P value VC %V Tiller number T * F *** T F ** E Axillary branch number (log) T *** F *** T F *** E Error term for MS T = MS T F MS E ; Error term for MS F = MS T F ; Error term for MS T F = MS E. Analysis of variance density, trial and family A nested anova testing for the effect of density, trial(density) and family, using family means, on LOGAXB found significant effects for trial(density) and family, but no significance for density (Table 1). However, the percentage of the variance explained by density is large, and it is possible that the design of the experiment, with only two density levels (and thus only 1 degree of freedom for the comparison) did not allow for adequate testing of density responses. The two anovas performed on TILL and LOGAXB in the low density trials used all available replicates, giving greater power to separate error variation from variation due to trial, family or their interaction (Table 2). All factors and their interaction had significant effects on both TILL and LOGAXB. However, the percentage of variation explained by the factors is much smaller for TILL than for LOGAXB, with error variation accounting for over 80% of the variation in TILL. A manova was also performed, using TILL and LOGAXB as the dependent variables. This was significant for the multivariate Pillai s Trace test as well as for the two univariate F-tests (SPSS 2005) (results not shown). QTL analyses In previously published analyses on vegetative branching at low density, we reported QTL as worthy of interest if they were significant at the chromosome-wide level of P < 0.05 and were found in more than one trial (Doust et al. 2004, 2005). In the present study we are interested in both QTL with fixed differences between the species (QTL replicated across environments) and those that differ across environments (GEI). For this reason we have looked at QTL significant at the chromosome-wide P < 0.05 level in both individual trials and across all four trials in joint analyses. Individual analyses of the separate trials found varying numbers of QTL: for tiller number, 5, 10, 6, and 7 QTL were found in trials 1 to 4, respectively, of these 6 were also significant at the genome-wide P < 0.05 level (Table 3). These represent 16 QTL regions, as a number of QTL amongst the trials map to the same genomic region (Table 3, Fig. 5; note that, for clarity, only the stringent genome-wide significance line is shown in the figure). Analyses of the individual trials for log-transformed axillary branch number found 25 QTL significant at the chromosome P < 0.05 level (5, 8, 8, and 4 QTL in trials 1 to 4, respectively), of which 8 were also significant at the genome-wide P < 0.05 level. These represent 16 QTL regions (Table 3, Fig. 5). Segregation distortion is severe in one of these regions, on chromosome VIII, and moderate in another, on chromosome IX (Wang et al. 1998). This distortion may affect our ability to place QTL precisely in these regions. Several QTL regions identified for tillering and axillary branching appear to be controlling both traits, such as those at chromosome positions III-8, V-5, V-10, VI-6, and IX-14. The QTL region on chromosome VI is found for axillary branching in all four trials but is only found in high density trials for tillering. The QTL region on chromosome III is detected for tillering in trials 2 and 4, and for axillary branching in trials 1 and 3. These QTL may reflect multiple closely linked genes, each affecting one trait. Alternatively, these QTL might reflect the action of genes that are generally involved in branching, but that are detected in tillering or axillary branching depending on growing conditions.

8 1342 A. N. DOUST and E. A. KELLOGG Table 3 QTL (chromosome and closest marker to LOD peak) detected at chromosome-wide significance levels (P < 0.05) in individual trials. QTL in bold are also significant at the genome-wide level of P < 0.05, corresponding to loci shown as significant in Fig. 5. HD, high density; LD, low density; spring, planted late spring; summer, planted midsummer; TILL, tiller number; LOGAXB, log transformed axillary branch per tiller number; A, additive effect; D, dominance effect; R 2, explained phenotypic effect (%). A positive sign for additive or dominance effects indicates an increase and a negative sign a decrease in the trait value Trait Trial 1 Chrom HD A Spring Trial 2 D R 2 Chrom HD A Summer Trial 3 D R 2 Chrom LD A Spring Trial 4 D R 2 Chrom LD A Summer D R 2 TILL I I I II II II II III III III IV IV V V V V V VI VI VII VII VII VII IX IX IX IX IX LOGAXB I II II II III III IV V V V V V V VI VI VI VI VII VIII VIII VIII VIII IX IX IX Joint analyses for each trait across the four individual trials found 8 significant QTL for tillering and 5 for axillary branching at the chromosome-wide P < 0.05 level. Of these, one QTL for tillering and one for axillary branching were also significant at the genome-wide level of P < 0.05 (Table 4, Fig. 5). In two cases, QTL for tillering and for axillary branching appear to be in the same genomic region (VI-7/8, IX-14), suggesting that individual QTL may have effects on more than one trait. Approximately half of the joint QTL for each trait had significant GEI effects, indicating that the effect of the QTL is significantly influenced by the environment. However, differences between trials observed in the individual analyses are not always discernable as GEI effects in the joint analyses. Nine QTL for tiller number and nine for axillary branch number were neither replicated between trials nor found in the joint analysis. Some of these may represent false positives but others are likely to represent loci that are affected by the environment, and thus provide a window into environmental plasticity. There is considerable variation in the pattern of LOD peaks between the individual trials (Fig. 5), and the pattern for the joint trials appears to incorporate many of the patterns seen amongst the individual trials. Figure 5 also shows that the trials differ substantially in how many LOD peaks approach significance, with the two high density trials having fewer QTL significant at the genome-wide P < 0.05 level. The amount of phenotypic variation explained by the QTL was generally high for all traits and trials. Individual QTL were

9 ENVIRONMENTAL EFFECTS ON BRANCHING IN MILLETS 1343 Table 4 QTL detected at the chromosome-wide level of P < 0.05 by the joint analysis of the four individual trials. QTL in bold are also significant at the genome-wide level of P < 0.05, corresponding to loci shown as significant in Fig. 5. GEI, genotype by environment interactions; *, significant at the chromosome-wide level of P < 0.05; NS, not significant; other abbreviations are as for Table 3 Trait Chrom. position Main effect GEI A D TILL IV-2 * NS IV-7 * NS V-11 * * VI-8 * * VII-2 * NS VIII-4 * NS IX-14 * * IX-16 * * LOGAXB II-7 * * VI-7 * * VII-1 * * IX-11 * NS IX-14 * NS only of moderate size, although one QTL for axillary branching on chromosome VI explained between 16% and 32% of the variation. QTL that were replicated across trials explained varying amounts of phenotypic variation in those trials, sometimes varying by a factor of two but sometimes, as for tillering on chromosome V-10/11, varying by as much as a factor of 6. This is not surprising, as the proportion of variation explained depends in part on the total variation and the amount explained by other loci, both of which differ among the trials. The additive effects of QTL for both traits were a mixture of positive and negative effects (Tables 3 and 4). Separate QTL with additive effects of differing sign for a particular trait indicate that each parent contains a mixture of alleles for that trait, some acting to change the phenotype towards that of one parent and others acting to change the phenotype towards the other. The additive effects of these QTL, if all alleles were of the same sign, are not sufficient to explain the range in variation seen in the F 2:3 hybrids, indicating that dominance and epistatic interactions are important in generating phenotypic variation. Dominance effects are often as large as additive effects for both tillering and axillary branching, although in other cases the effects observed are below the power of this experimental design to detect unambiguously (Lander & Botstein 1989). The dominance effects for vegetative branching traits are substantially greater than those reported previously for inflorescence branching in this mapping population (Doust et al. 2005). Digenic epistasis was detected between some markers associated with QTL for tillering and axillary branching, but also between markers not associated with QTL (Table 5). Comparative mapping There is good colinearity between the millet genome and that of other cereals (Devos et al. 1998, 2000), so that comparison of species is useful for suggesting candidate genes. We looked particularly for genes that were in the identified QTL regions and that were known to control branching in other species. One well-studied gene that controls outgrowth of tillers and axillary branches in maize is teosinte branched1 (tb1), which has been associated with QTL controlling vegetative branching in maize teosinte crosses, and which has been hypothesized to suppress axillary meristem elongation (Doebley et al. 1997; Hubbard et al. 2002). We hybridized a maize cdna clone of tb1 to the F 2 mapping filters for millet, and placed the gene on Setaria chromosome IX, in the same region as several QTL for branching. Another gene which is associated with QTL on chromosome V is barrenstalk1 (ba1). We confirmed this placement by hybridization of a foxtail millet DNA clone of this gene, and placed it between markers psm768 and rgc385. This corresponds to QTL regions found for tillering in both individual trials 2, 3, and 4 as well as the joint analysis. We had initially suspected that phytochrome B would be a candidate for one of our QTL, as this has been shown to control changes in branching and stem elongation that are associated with plant crowding. However, only a single minor QTL was found for tillering in trial 2 in the region on chromosome IX where phytochrome B maps, accounting for only 6% of the variation. Trait Trial Locus1 Position QTL? Locus2 Position QTL? R 2 TILL 1 RGC147 II-9 RGC950 VIII TILL 1 PSF470 I-4 RGC389 VII-1 J, Y 0.25 TILL 1 PSM713 V-4 Y RGC601.2 V-12 J 0.26 TILL 1 PSM713 V-4 Y RGR642 VIII-3 J 0.26 LOGAXB 1 PSM371 III-12 RGR1943 VI LOGAXB 1 RGR1943 VI-12 RGR830 III LOGAXB 1 PSM671 VI-6 Y PSM713 V-4 Y 0.16 LOGAXB 3 RG83.6 VI-1 RGC950 VIII LOGAXB 3 PSF163 VIII-10 UGT737 VI LOGAXB 3 PSF63.1 V-3 PSM671 VI-6 Y 0.16 Table 5 Digenic interactions calculated using epistacy (Holland 1998). Results only reported if the overall probability of interaction is P < (Bonferroni correction to give experiment-wide error rate of P < 0.05). Locus1 and locus2 are the two markers for which an interaction is being tested; QTL? recognizes those loci that are associated with a QTL for that trait and trial combination; Y, presence of QTL in individual analyses; J, presence of QTL in joint analysis; R 2, proportion of variation explained by the digenic interaction. Trait abbreviations as for Table 3

10 1344 A. N. DOUST and E. A. KELLOGG Fig. 5 Plots of LOD scores at 10-cM intervals across the nine chromosomes of the genome in the four individual analyses and in the joint analysis of tiller number and log-transformed axillary branch number. The plot is constructed as if the chromosomes have been laid end to end. The LOD score curve in the joint analysis composed of closed diamonds represents the main effects while the curve made of open circles represents GEI. The straight dashed line in trials 1 to 4 indicates the genome-wide significance level of P < 0.05, as do the open and compact dashed lines in the joint analyses. In the joint analyses, the significance of the main effect is given by the compact dashed line and that of the GEI effect by the open dashed line. N.B. This is a more stringent significance cut-off than the chromosome-wide significance level applied in Tables 3,4. The QTL region on chromosome VI, which accounts for up to 30% of the variation in axillary branch number, has no obvious candidate genes from other grass species. Analysis of variance density, trial, and the three major QTL We used tb1 on chromosome I, ba1 on chromosome V, and hhu33, a marker closely associated with the large QTL on chromosome VI, to represent the three major QTL for tillering and axillary branching. A nested anova, testing for the effect of density, trial (density), and the three markers and their interactions on LOGAXB found that density was not a significant factor, but that trial(density) and the three QTL markers were significant. Density ba1 and trial(density) hhu33 were also significant. Interestingly, as already noted above, density explains a large proportion of the variation, but has only one degree of freedom and thus fails to achieve significance. Trial(density) and hhu33 also explain large proportions of the variation, suggesting that these factors are important in determining LOGAXB variation (Table 6).

11 ENVIRONMENTAL EFFECTS ON BRANCHING IN MILLETS 1345 Table 6 Nested anova for density (D), trial(density) [T(D)], ba1, tb1, hhu33 and selected interactions for LOGAXB (using family means). Significance levels and abbreviations as for Table 1 Factor d.f. MS F P value VC %V D NS T(D) ** ba *** tb *** hhu ** D ba * T(D) hhu * E Error term for MS D = 0.87 MS T(D) MS E ; Error term for MS T(D) = 0.84 MS hhu33 T(D) MS E ; Error term for MS ba1, MS tb1, MS D ba1, MS hhu33 T(D) = MS E ; Error term for MS hhu33 = 0.95 MS hhu33 T(D) MS E. Table 7 anova for trial (T), ba1, tb1, hhu33 and significant interactions for TILL (using family means), in the low density trials. Significance levels and abbreviations as for Table 1 Factor d.f. MS F P value VC %V T NS ba ** tb NS hhu NS ba1 tb * E Error term = MS E. VC estimate set to zero, because term is redundant. Analysis of TILL in the low density trials shows significant effects for ba1, and ba1 tb1, but no significant effects for tb1, hhu33, trial or other interactions (Table 7). Analysis of LOGAXB in the low density trials differs from the analysis including high density trials in that hhu33 is no longer significant, although trial and trial hhu33 are still significant. Ba1 and tb1 are also still significant (Table 8). Hhu33 is not significant, even though it explains as much variation as ba1 and tb1. This is because of the inclusion of the interaction term between trial and hhu33, which is used as the error term to test the significance of hhu33. Almost 50% of the variation in LOGAXB is explained by variation between the two (low density) trials. Discussion Phenotype of the weed vs. the domesticate Our data show that the weed green millet is more responsive to environmental differences than its domesticated Table 8 anova for trial (T), ba1, tb1, hhu33 and significant interactions for LOGAXB (using family means), in the low density trials. Significance levels and abbreviations as for Table 1 Factor d.f. MS F P value VC %V T * ba *** tb ** hhu NS T hhu ** E Error term MS T = 0.84 MS T hhu MS E ; Error term MS ba1, MS tb1, MS T hhu33 = MS E ; Error term MS hhu33 = 0.95 MS T hhu MS E. descendant foxtail millet. At low density, green millet produces more than three times as many axillary branches as it does at high density, whereas foxtail never produces axillary branches. Although both species produce more tillers at low density, the increased number in green millet is more than double the increase observed in foxtail millet. These results confirm the observations of Dekker (2003) that green millet has highly plastic responses to shading and resource availability, whereas foxtail millet exhibits less plasticity to environmental conditions. Many of the environmentally labile changes seen in green millet should have strong fitness benefits, especially those increasing the duration of flowering and thus the amount of seed produced. In contrast, the lack of sensitivity to environmental changes seen in foxtail millet is likely the result of human selection for ease of cultivation and harvesting. Human selection for a reduction in axillary branching in foxtail millet is evident in the absence of axillary branches at either planting density. This, together with our failure to find any visible axillary branch primordia in the leaf axils of foxtail millet, even with the aid of a microscope, suggests that axillary branch primordia are either initiated but arrested at a very early stage or that they are not initiated at all. These two hypotheses suggest different candidate genes: in maize ba1 controls axillary meristem initiation while tb1 controls the elongation of the axillary branches. The pattern of expression of orthologous copies of these genes in foxtail and green millet is currently under study. Plasticity and its genetic basis We only calculated genetic correlations in the low density trials (3 and 4), because of the extreme non-normality of tiller distribution at high densities. Tiller and axillary branch number had significant genetic correlations in trial 3 but not in trial 4. This, coupled with the results from the QTL analyses and our observations on the timing of tiller and axillary branch production, suggested that the two

12 1346 A. N. DOUST and E. A. KELLOGG forms of branching are actually different phenomena that might be expected to be under separate genetic control. This has also been reported in pearl millet (Pennisetum glaucum), a species that is more closely related to foxtail millet than either maize or rice (Poncet et al. 2000). Tillers are initiated early in the life cycle of the plant, whereas axillary branches only elongate after the inflorescence terminating the main culm has initiated. We also observed that some genotypes only produced axillary branches on the main culm, whereas other genotypes produced axillary branches on other axillary branches, leading to multiple orders of branching. Although we made no quantitative measurements of overall plant size, there appeared to be no relationship between plant size and number of branches in any particular trial. Rather, genotypes with multiple orders of branching could be either relatively small or large compared to other genotypes with few or no branches. Thus there is potential for considerable phenotypic plasticity in branching, dependent on particular allele combinations. Genetic correlations between tiller numbers in the two low density trials were high, as were those for axillary branch numbers. This indicates that the same suite of genes underlies the phenotypes in at least these two low density trials. This is not surprising for TILL, but is more so for LOGAXB, because of the large differences seen in the reaction norms between these two trials. The QTL analysis provides further insight into the correlations observed. The correlations between trials presumably are reflected in the 13 QTL detected by the joint analysis (Table 4). However, many QTL were found in individual trials but not in the joint analysis, and these may reflect the activity of additional loci that are significantly affected by environment. Environmental effects are also evident in the QTL identified by the joint analysis, as indicated by significant GEI, as well as by their detection in only a subset of trials. These environmentally responsive regions of the genome are found on I-9/10, II-4, II-7, III-8/ 9, IV-7, V-4/5/7, V-9/10/11, VI-6/8, VII-1, VIII-1/2, VIII- 10/11, IX-3/4, IX-14, and IX-16 (Tables 3 and 4). We observed transgressive segregation in trials 1, 2 and 4, in which some hybrid families had mean tiller number and/or axillary branch number exceeding the range seen between the two parents. Not all trait/trial combinations showed transgressive segregation, so that, for example, the green millet parent in trial 4 actually had more tillers than any of the hybrids. Both tillering and axillary branching in all trials had QTL with both positive and negative additive effects, the favourable recombination of which could lead to transgressive segregation. Transgressive segregation is commonly found in QTL studies (Rieseberg et al. 2003b), and our data are no exception. Dominance and epistatic effects also contributed to the observed variation in some trials, indicating that they too varied between environments. Dominance effects in most cases were substantial and could account for differences in the hybrid populations over and above those attributable to the combination of additive effects. Significant epistatic effects were also found in trial 1 for both tillering and axillary branching, and in trial 3 for axillary branching. These effects were found both between QTL positions identified in the individual and joint analyses, as well as between marker positions outside the QTL regions. This raises the intriguing possibility that some regions of the genome do not by themselves contribute to changes in phenotype, but can do so in specific combinations with other regions. It will be necessary to construct RILs with combinations of these particular regions to test this hypothesis. We show that the F 2:3 families have different reaction norms, as predicted for any species likely to disperse into heterogeneous environments (Sultan & Spencer 2002). Each family could have a distinct combination of alleles for each of the 14 environmentally responsive regions, which would immediately lead to high variability in the magnitude and possibly the direction of response. We found that several of the QTL for tillering and axillary branching occurred in the same genomic region, and it may be that these regions contain genes that pleiotropically affect both types of branching. Two of these regions, found in individual and joint analyses for both tillering and axillary branching, are the lower region of chromosome IX (near markers 14 and 16 and the mapped location of teosinte branched1), and the region near the centromere on chromosome V (near markers 10 and 11 and the mapped location of barren stalk1). The anovas provide their own insight into the cause of variability in the hybrid population. We deliberately removed a portion of the families from the anovas in order to improve the tractability of the data, and to remove what appeared to be a subpopulation of families within the trial that never branched. This was presumably because of the inheritance of a homozygous factor from the foxtail millet parent, which itself never has axillary branches. Even so, there was still an extreme positive skew to the distribution of tiller values in the high density trials, which is likely the result of severe suppression of branching in crowded conditions. Difference in planting density is obviously not the only environmental factor at work in this experiment, because differences between trials at the same density were often as large as those between different densities. This is consistent with the unexpected result from the individual QTL trials that the pattern of QTL at either density differs substantially between trials, and that trials at one density may share replicated QTL regions with trials from the other planting density. These observations are supported by the anova results, where the nested analysis showed that axillary branching was affected much more by differences between trials and between genotypes than by density.