Stability in Advanced Gezira Population of Sorghum (Sorghum Bicolor (L.) Moench) at Drought Porne Environments in Sudan

Experimental Material

One hundred and twenty S1 families were taken at random from an advanced random mating Gezira sorghum population(G S P-1) developed and improved for six cycles using S1family selection, at Rain-fed Crop Research Centre for Arid and Semi-Arid areas (RCRCASA) in the University of Gezira, Wad Medani Sudan. The experiments were conducted during two seasons (2004-2005) at four rain- fed environments namely; Gedarif University farm at Gedarif northern marginal environment (200-300 mm) in 2004, Rahad rain-fed area (300-400mm) in 2004, Gedarif Research Station at northern Gedarif in 2005 and Kasamor at Gedarif middle environment (500-600mm) in 2005. One hundred and twenty S1 families were taken at random from an advanced random mating Gezira sorghum population (G S P-1); each family was grown in one row 5.0 m long, 0.6 m between rows and 0.2 m within row spacing. After 3 weeks from sowing, plants were thinned to 2 plants/hole. Weeds were controlled manually when necessary. No fertilizer or other inputs were added. At harvest, a 3.0 m length in the middle of each row was marked as experimental unit and all data were then based on 3.0 m of row length. Harvesting and threshing were done manually.

Experimental Layout

The design used in this study was a modified randomized complete block (replications-in-block) design; the hundred and twenty S1 families were divided into 6 sets of 20 families each. Each block contained two replications of the same set from the population. Blocks were assigned at random, and replications were assigned at random within each block. The 20 families, representing a group from the population, were assigned at random to each replication. A modified randomized complete block (replications-in-block) design has been used because it was more efficient than a blocks-in- replication design for controlling the experimental error and it allows loss of whole blocks only, if necessary, rather than the loss of whole or an entire replication.

Parameters Studied

i. Days to 50% flowering (Bloom): The number of days from planting to the date when approximately 50% of the plants in the row were at half-bloom (had their flowers open). ii. Grain yield (kgh-1): Panicles were harvested by hand from a three-meter section in the center of each row, where total panicle weight and total grains weight were measured and used to estimate grain yield and threshing percent Weight of actual grain yield was taken to estimate total grain yield in kgh-1. Harvested area = 3.0 x 0.6 =1.8 m².

Estimated grain yield (kgh-1) =

Yield from harvested area 10000 1.8 1000 × ×

iii. Plant height (cm): The average of heights, taken at random from each family was measured from the soil surface base of the plant to the tip of the panicle for representative plants in each plot.

Statistical Approaches

The least squares method was used in genetic variability analyses, utilizing the Statistical Analysis System (SAS) as outlined by Jane T Helwig, et al. [4], while the IRRISTAT software was used to conduct the AMMI analysis [5] for stability or G×E interaction. According to Gauch, et al. [6] and Nichit, et al. [7] as in the following:

Analysis of variance of data for each trait combined over different environments: A combined analysis for each trait in each environment was performed. The additive linear model assumed was: Y ijkl = u +Li+bj+ Lbij + rijk + fil + Lfijl + eijkl. Where: Yijkl: the observation on the lth family at the kth replication within the ith block in the ith location (environment). U: the overall mean of all families. Li: the random effect associated with the ith environment; I = 1, 2,…4. bj: the random effect associated with the jth block; I = 1,2,…,6. Lbij: the random effect of the interaction between the ith environment and jth block. rijk: the random effect of the kth replication within the jth block in the ith environment. fil: the random effect associated with the lth family in the jth block .

Lfijl: the random effect resulting from the interaction of the lth family in jth block with the ith environment. eijkl: the random error effect associated with the plot containing the lth family in the kth replication within the jth block in the ith environment.

As mentioned, the various effects in the model were computed as deviations about the mean within which they were nested, so that the sum of the deviations about the mean adds to zero. The form of the analysis of variance pertinent to the assumed model is given in Table 1.

*, , * are the levels of significance 0.05, 0.01, and 0.001 respectively. Table 1: Mean squares for the combined analysis of variance for three traits in120 families sorghum population evaluated at four low input areas in Sudan.

The coefficient of variation (CV) was calculated as: C.V = (Me ½)/overall mean x 100.

Since the design had nested features, the effects of the parameters in the model were computed as deviations about the next mean up in the hierarchy, i.e., the block effect was computed as the deviation of the individual block from the grand mean, replication effect as the deviation of the replicate value from the block mean in which the replicate was located, and the family effect was calculated as the deviation of the observed family value from the block mean in which it is nested. Genetic differences among families nested with blocks were tested by the null hypothesis: Ho: ơ2 f/b = O.

The interaction between the families and environments: It was computed using the F-test as: F: Msf/b/Me with b (s-1) (f-1), s b (f-1) (r-1) degrees of freedom. Where: s is for site (environment). The various parameters and their SE’s for the combined environments data from each of the populations were estimated as follows: Familles X Environment Interactions Variance ( ) ,2 sf/b sf/b e o = M -M /r, and its SE $$ \mathrm {S E} \left(\mathrm {o} _ {\mathrm {s f / b}} ^ {\prime}\right) = \left[ \left(\frac {1}{\mathrm {s} ^ {2} \mathrm {r} ^ {2}}\right) \left(\frac {2 \mathrm {M} _ {\mathrm {s f / b}} ^ {2}}{\mathrm {b} (\mathrm {f} - 1) (\mathrm {s} - 1) + 2} + \frac {2 \mathrm {M} _ {\mathrm {e}} ^ {2}}{\mathrm {s b} (\mathrm {f} - 1) (\mathrm {r} - 1) + 2}\right) \right] ^ {1 / 2} $$              +         Stability Analysis ( ) ( )( ) ( )( ) Combined analysis of data generated from four production environments (Combination of location and years) was carried out for estimation of stability parameters for grain yield.

The Additive Main Effect and Multiplicative Interaction (AMMI) Analysis

It was carried out to show the stability and pattern of adaptation sorghum families to the four environments. AMMI analysis fits additive effects due to genotypes (G) and environments (E) and by usual additive analysis of variance procedure and then fits multiplicative effects for genotype × environment interactions (GE) by principal components analysis (PCA). The IRRISTAT software was used to conduct the AMMI analysis [5] as in the following; Equation of AMMI model:

Yij = µ +gi +ei + ∑ λn αin γjn + Rij

Where: Yij: is the grain yield of the ith genotype in the jth environment. µ: is the grand mean. gi: is the deviation of the genotype mean from the grand mean . ei: is the deviation of the environment mean from the grand mean. λn: is the eigenvalue of the nth PCA. αin and γjn: are the genotype and environmental interaction principal components eigenvectors(PCAg and PCAe, respectively) for axis n . N: is the number of IPCA retained in the model. Rij: is the residual.

Environmental and genotype PCA scores are expressed as unit vector time the square root of λn. The multiplicative part of the model is obtained by PCA (αin and γjn). The principal advantage of the AMMI is that the interaction can be modeled by only one or two PCA-axes.

To analyze genotype-environmental interaction and adaptation graphically, AMMI –bi-plot with the PCA1 scores was plotted against the mean yield (main effect). Genotypes or environments that appear almost on a perpendicular line have similar means and those that fall almost on a horizontal line have similar interaction patterns. Genotype (or environments) with large PCA1 scores (either positive or negative) have high interactions, whereas Genotypes (or environments) with PCA1 scores near zero have small interactions. To further explain the GE and adaptation a biplot between the PCA1 scores and PCA2 scores was drawn. The AMMI expected yield of any genotype and environmental combination can be calculated from the biplot as indicated by Zobel, et al. [8]. The interaction part is simply the genotype PCA1 score times the environmental PCA1 score. Genotype and environments with PC1 scores of the same sign produce positive interaction effects, whereas combinations of PC1 scores of opposite signs have negative interactions.

The Additive Main Effect and Multiplicative Interaction (AMMI) Analysis

The parametric approach such as mean yield over- environments, genotypic coefficient of variability, genotype variance, the ecovalence, Shukla’s [11] stability variance (interaction variance) and Eberhart and Russell [12] stability parameters(regression coefficient or slope and deviation from regression) give only the individual aspect of stability but cannot provide an overall picture of the responses. So it is difficult to reconcile all of these assessments into a unified conclusion because genotype response to environments is multivariate. Consequently, nonparametric approach (multivariate) has been proposed to overcome univariate problems associated with parametric approach [13]. Multivariate analysis such as AMMI analysis groups genotypes or environments in a qualitative manner according to their similarity of performance rather than quantitative manner of the stability parameters. In AMMI analysis the genotype response to environment is multivariate, AMMI analysis involves the clustering analysis to classify genotypes under the most adapted sites for them depending on the AMMI principal components scores similar signs in genotypes and sites (genotypes having PCA scores < 0 responded positively to environments, Those having PCA scores < 0 and the reverse is true for families that had PCA scores > 0). Also AMMI models integrate the usual additive analysis of variance (ANOVA) for the additive effect with the principal component analysis (PCA) for the multivariate effects [6, 7].

AMMI analysis of variance model: In this study, AMMI analysis of variance indicated that grain yield was significantly affected by environment (E), genotype (G) and genotype x environment interaction (GE), which explained 86.7%, 2.4% and 10.9% of the total variation (E + G + GE), respectively (Table 2). This result indicated that environment component represented the largest amount of total variation, while variations due to genetic component and GE interaction were considerably low. The partitioning of GE interaction through AMMI model analysis revealed that the four multiplicative terms first (PCA1), second (PCA2) and third (PCA3) principal components, were significant factors that captured 81.3%, 9.8 % and 8.9 % of variation due to GE interaction sum of squares, respectively. Together they accounted for 100% of GE interaction sum of squares (i.e. PCA1 represented 8.9 % out of 10.9% of total sum of squares). Hence, most of the variation was explained by the first principal component (PCA1) and it was the most informative as shown in Table 3.

*, , *are the levels of significance 0.05, 0.01 and 0.001 respectively. Table2: AMMI analysis of variances of environment (E), genotypes(G) and genotypes x environment interaction(GEI) on grain yield (kgh-1).

Df, degree of freedom; SS: sum of square; MS: mean square, Efficiency%, percentage of GEI sum of squares, and * significance at 0.001 probability level, N.s, not significant. Table3:** AMMI analysis of variances of environment (E), genotypes (G) and genotypes x environment interaction (GEI) on grain yield (kgh-1) and the partitioning of GEI into AMMI scores.

Non-significant differences among families x environments interaction indicating the stability of population as a whole over the four targeted environments, whereas the G x E interaction partitioning into AMM1 principle-components has shown highly significant differences (p ≤ 0.01) due to APC1 that is 81.3% out of GEI, (which is equal 8.9% SStot ) as in Table 3.

A large variation among the studied genotypes for grain yield and their interaction to the environment was determined. According to the highest average grain yield (yield potentialities) the best environment was KAS (2180.4) that ranked 1st followed by Rahad environment (1967.2), GRS environment (1194.8) and the lowest environment in yield was UG that obtained 449.47 kg h-1. Based on AMMI biplot G and E having PCA values close to zero have small interaction effects, whereas those having large positive or negative PCA absolute values largely contribute to GE interaction [14, 15]. Hence UG was the least interactive among the four environments, while Rahad was the most interactive, because environment in Rahad exhibited the largest absolute value of PCA score (+0.9082460), whereas the smallest score was shown by UG (-0.3451480) as shown in Table 5.

This indicated the relative ranking of genotypes were more stable at UG than at Rahad, making it too difficult to recommend a genotype for Rahad. This may be due to the fact that Rahad area was more marginal. Also a large variation among the studied families for grain yield kg ha-1 was explained in Table 4. Where family MD90 (99), obtained 1604.9 as highest yield (it ranked 1st in order of the total 120 families in the population) and SB191(116) has got 977.4 as lowest yield (among selected families) and it ranked number 120 in order (of the total 120 families in the population) Tables 4-6.

Slope: slopes of Regression of families means on environmental index .Indicated slopes significantly different from the slope for overall regression, which is -20.48. Ms-tx: contribution of each family to interaction Means Squares. Ms-Reg: contribution of each family to Regression components of the families’ x environment interaction. Interaction Components of Ms-Dev: Deviations from Regression. R2 (%): Squared correlation between residuals from the main effects model and the environmental index. cp1, cp2, cp3: 1st ,2nd and 3rd AMMI principal components respectively. Table 6:** Scores best families stability, regression and AMMI principal component.

The families from 93 to 116 were selected as superior from the total of 120 families population based on yield potential, because they ranked at the top and/or they have specific adaptability to certain environment over the others (Not adapted to other). Also some satisfied all the basis of stability measurements in comparison. Among the selected families family SB74 (101) revealed the smallest absolute PCA1score (5.837), indicating its least variability in interaction, while SB196 (110) showed the largest score (15.55), pointing out its highest variability in interaction (Table 5). Hence, depending on high yield potentiality the high yielding families were MD90(99), MD395(98), MD68(97) and MD106(94) those ranked as the 1st four families, while the lowest yield was obtained by SB191(116) family that ranked 120th (Table 4). On the other hand depending on the AMMI residuals, additive effects and multivariate scores the families order for stability as flowing; SB74(101) -5.837, MD7 (95) -6.286, MD122 (93) -6.544, MD415 (96) -6.934, MD106 (94) -6.955, SB179 (103) -6.98, MD68 (97) -7.036, Ed133 (102) -7.11, MD90 (99) -7.644, MD412 (100) -7.7, 104(SB186) 7.804 and MD395 (98) -7.827 respectively (Table 5).

The AMMI cross site analysis: The AMMI cross site analysis of main and PCA1 effect of both G and E on yield explained that sorghum families having PCA scores < 0 responded positively (adapted) to environments, Those having PCA scores < 0, and the reverse is true for families those having PCA scores > 0 [16] in wheat. Hence the families with the lower PCA scores are the SB74 (101) -5.837, MD7 (95) -6.286, MD122 (93) -6.544, MD415 (96) -6.934, MD106 (94) -6.955, SB179 (103) -6.98, MD68 (97) -7.036, Ed133 (102) -7.11, MD90 (99) -7.644, MD412 (100) -7.7, SB186 (104) -7.804 and MD395 (98) -7.827 (Table 5). This indicates their least variability contribution in the GEI and therefore, they would have less variability across environments. These best families all responded positively or adapted to UG (-0.34518), Kas(-0.51204300), and negatively to Rahad (+0.9082460) and G.R.S (+0.4553300) as in Table 5. AMMI biplot analysis: AMMI model is found to be the best predicting model, a graphical display of the GE interaction. PCA1 and their main effects should be useful for revealing favorable pattern in genotype response across environments [10]. The AMMI biplot of the main and PCA1 effects of both G and E on grain yield explained 98 % of the treatment sum of squares, with 2.4%, 86.7% and 8.9% due to genotype, environment and PCAI sum of squares, respectively (Table 3 and Figure 1). Wheat genotypes that had PCAI scoring > 0 responded positively (adapted) to environments, that had PCAI scoring > 0 (i.e., their interaction is positive) but responded negatively to environments that had PCAI scoring < 0 and the reverse applied for genotypes that had PCAI scores < 0 [16]. Consequently all the families from 93- 116 responded negatively therefore, they are adapted to environments UG obtained -0.345 and environments KAS that got -0.512, (Their GE interaction with negative sign), but they responded negatively to environment Rahad that obtained + 0.908 and GRS which obtained + 0.456 Table 5.

Families are represented by values from 1-120, Environments are represented by letters.

The families with a lower absolute PCA1 score such as 101, 95, 93, 96, 94, 103, 97, 102, 99, 100, 104 and 98 would produce a lower absolute GE interaction effect and would have a less variable yield across sites than families with a higher absolute PCA1 score such as 116, 115 and 110 Table 5.

AMMI biplot of the first two principal component axes is a powerful way of detecting important sources of GE effects [8]. This analysis represents stability of the cultivars across environments in terms of principal component analysis. It is used to identify broadly adapted cultivars that offer stable performance across sites, as well as cultivars that perform well under specific conditions. In this study, the first two principal component axes (PCA1 and PCA2) in the ordination (biplot) analysis explained a large proportion of the variation 91.1% of the total GE sum of squares (Figure 2). The AMMI biplot displays similar genotypes or environments near to each other (have small vectors angles) and dissimilar items are farther apart (have large vectors angles). Accordingly, the environments UG with Rahad and Kas with GRS, as an example, were similar to each other in the way they discriminate among genotypes (Figure 2). Environment UG and Rahad showed relative similarity (appeared near to each other) in the way they discriminate among genotypes. UG had the widest vector angles (appeared at far distant) from Kas and GRS, indicating its extreme dissimilarity from them. UG was noticed as a unique and the lowest yielding environment. On the other hand, family SB191 (116) appeared at a far distance from the other families and environments, reflecting the different characteristics of this family in that it had poor performance and consistently lower yielding than the average at all environments. In contrast the families SB4 (111), SB180 (114), SB45 (113) and SB22 (112), for example, showed similar performance across the production environments or appeared near to each other (Figure 2).

Families are represented by values from 1-120, Environments are represented by letters.

The analysis of the genotype and environment parameters resulting from AMMI helps to describe the interaction effects of the genotypes and environments. On AMMI biplot, genotypes and environments having PCA values close to zero (near the origin) have small interaction effects, whereas those having large positive or negative PCA values (distant from zero) largely contribute to GE interaction [14, 15]. Hence, the families SB4 (111), SB180 (114), SB45 (113) and SB22 (112) were the most interactive, while SB191 (116) was the least interactive. Entries yield relatively better in sites having PCA values of the same sign, but not in sites with opposite sign. Genotypes that are farther along in the positive direction of the environment vector are higher yielding and vice-versa [17]. Hence the family 119 was higher yielder than the family 116 (Figure 2). Acute angles between any two vectors indicate positive associations (i.e. they influence the genotypic relative performance in similar manner), 90˚ indicates negative associations [17]. Hence Rahad has positive associations with UG and GRS (due to acute angles between Rahad and each of them), whereas negative associations were detected between Rahad and Kas (angle > 90˚) as shown in figure 2. On the other hand, environment UG, KAS and GRS appeared at a far distance from the origin (large PCAI score); hence, they had large interaction effect, whereas Rahad had small interaction effects (Figure 2 and Table 5). Hence in this investigation, visual observations of AMMI biplot analysis enable to identify genotypes and testing environments that exhibited major sources of GE interaction as well as those that were stable. Similar results were reported in wheat by Thomason and Phillips [18].

Comparing the effectiveness of regression and AMMI analysis for analyzing GE interaction, it was found that PCAI in AMMI accounted for the GE sum of squares by 81.3%, while regression analysis accounted for GE by 22.9 % (Table 2). Hence, AMMI analysis was superior to the regression techniques in accounting more effectively partitioning of the interaction sum of square. The same results were reported in wheat [7, 15, 19]. Cornelius [20] showed that regression analysis and AMMI analysis with one PCA have the same model form, differing only in fitting procedure.