9

SUMMARY

Biplot Analysis of Diallel Data

Diallel crosses have been used in genetic research to determine the inheritance of a trait among a set of genotypes and to identify superior parents for hybrid or cultivar development. Conventional analysis of diallel data is limited to partitioning the total variation attributable to crosses into general combining ability (GCA) of each parent and specic combining ability (SCA) of each cross. The SCA effects are just residuals not explained by the GCA effects; they are cross specic and do not provide much information on parents. The biplot approach of diallel data analysis introduced in this chapter allows a much better understanding of parents. For a given set of data, the following information can be easily visualized: 1) the GCA effect of each parent; 2) the SCA effect of each parent (not cross); 3) the best crosses; 4) the best testers; 5) the heterotic groups; and 6) genetic constitutions of parents with regard to the trait under investigation. Diallel crosses represent matings made in all possible combinations among a set of genotypes. They have been widely used in genetic research for investigating inheritance of quantitative traits among a set of genotypes. There are four types of diallel mating schemes (Grifng, 1956): 1) method 1 diallel cross with parents and reciprocals; 2) method 2 diallel cross with parents but without reciprocals; 3) method 3 diallel cross with reciprocals but without parents; and 4) method 4 diallel cross without parents and reciprocals. Reciprocals are made for the purpose of detecting any maternal effect. We conne our discussion to method 2 diallel cross, although other types of diallel crosses can be easily accommodated. Conventionally, analysis of diallel cross data is conducted to partition total genetic variation into GCA of the parents and SCA of the crosses. In this chapter, we will use two sets of diallel data (Tables 9.1 and 9.2) to demonstrate the biplot approach to diallel analysis. The rst dataset is to demonstrate the general steps and utilities of biplot analysis for diallel data. The second dataset is used to exemplify analysis of a large dataset for which a biplot of the rst principal component (PC1) vs. the second principal component (PC2) may not be adequate. A discussion of both datasets is necessary because they contain contrasting entry tester interaction or gene action patterns.

9.1 MODEL FOR BIPLOT ANALYSIS OF DIALLEL DATA

Buerstmayr et al. (1999) reported a diallel study of winter wheat resistance to Fusarium head blight (FHB). They used seven winter wheat genotypes of diverse origin and with large differences in resistance to FHB. They presented data on area under the disease progress curve and percentage of infected kernels for all seven parents and their F1 hybrids; the two measures were highly correlated. Since no reciprocal effect was observed for this trait, means of the reciprocals for individual crosses were reported. Percentages of uninfected kernels of the seven parents and their F1s as a measure of resistance to FHB, are given in Table 9.1. The GCA effects of the parents are presented in the last column of the table, which is simply the row means minus the grand mean. Thus, we see parents F and G have high GCA, A and B have low GCA, and C, D, and E have intermediate GCA. This is the only piece of information on the parents that can be obtained using the conventional analysis. All other information that can be derived by conventional methods is pertinent to crosses rather than parents. We will show that the biplot approach allows a much deeper understanding of the parents and of the trait under investigation.

2003 by CRC Press LLC

TABLE 9.1 Resistance to FHB of Seven Winter Wheat Genotypes and their F1 Hybrids, as Measured by Percentage of Uninfected Kernels Upon Inoculation
Testers Entries A B C D E F G Mean A 27.5 35.7 46.4 53.7 33.3 64.9 43.3 43.5 B 35.7 37.5 46.2 40.8 51.9 45.6 57.5 45.0 C 46.4 46.2 38.7 49.1 50.4 55.6 69.4 50.8 D 53.7 40.8 49.1 51.2 49.4 48.1 57.5 50.0 E 33.3 51.9 50.4 49.4 42.5 63.1 68.9 51.4 F 64.9 45.6 55.6 48.1 63.1 60.0 63.1 57.2 G 43.3 57.5 69.4 57.5 68.9 63.1 43.7 57.6 MEAN 43.5 45.0 50.8 50.0 51.4 57.2 57.6 50.8 GCA 7.3 5.8 0.0 0.8 0.6 6.4 6.8 0.0

Note: The genotype codes are A = Alidos, B = 81-F379, C = Arina, D = SVP7201717510, E = SVP-C87155, F = UNG136.1, and G = Ung226.1. Source: Data based on Buerstmayr, H. et al., Euphytica, 110:199206, 1999.

In a diallel cross data, a parent is both an entry and a tester. The model used for biplot analysis of diallel data is the tester-centered principal component analysis. It was labeled as Equation 4.5 and is presented again below: Yij j = gi1e1 j + gi 2 e2 j + ij where Yij is the expected value of the cross between entry i and tester j; is the grand mean; j is the main effect of tester j; gi1 and e1j are called the primary effects for entry i and tester j, respectively; gi2 and e2j the secondary effects for entry i and tester j, respectively; and ij is the residue not explained by the primary and secondary effects. A biplot is constructed by plotting gi1 against gi2 and e1j against e2j in a single scatter plot. Using GGEbiplot software, a biplot will be generated in about one second after the data are read (Figure 9.1). The parents are presented in lowercase italics when they are viewed as entries; they are in regular uppercase when viewed as testers. The biplot explained 77% (46 and 31% by PC1 and PC2, respectively) of the total variation, which, in conventional analyses, would be partitioned into GCA effects of the parents and SCA effects of the crosses. The guidelines cross at the biplot origin (0,0). What we can learn from this biplot follows.

9.2 GENERAL COMBINING ABILITY OF PARENTS

Figure 9.2 is the average tester coordination (ATC) view of the biplot brought up by the Average Tester Coordination function of GGEbiplot. The small circle represents an average tester, which is dened by the average PC1 and PC2 values of all testers. The line passing through the biplot origin and the average tester, with an arrow pointing to the average tester, is called the average tester axis or ATC abscissa and the line that passes through the origin and is perpendicular to the average tester axis is called average tester ordinate or ATC ordinate. GCA and SCA are dened here as properties of the entries. The GCA effects of the entries are approximated by their projections onto the ATC abscissa. Lines parallel to the ATC ordinate help rank entries relative to GCA effects. Thus, entries f and g (UNG136.1 and UNG226.1), both being derivatives of Sumai 3, a very resistant wheat cultivar of Chinese origin, had the highest

2003 by CRC Press LLC

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

4.4

G
3.3

2.2

c d f

1.1

P C 2

b e B D F a C g E

-1.1

-2.2

-3.3

-4.4 -4.4 -3.3 -2.2 -1.1 0 1.1 2.2 3.3 4.4 5.5 6.6

P C1 Symmetrical Scaling
FIGURE 9.1 Biplot based on the wheat FHB research data. The seven parents are in lowercase when viewed as entries and in uppercase when viewed as testers.

2003 by CRC Press LLC

GCA effects, whereas entry a (Alidos, the most susceptible parent) had the lowest GCA effect. The GCA effects of the entries are in the order of: g f > d > c > b e > a. Note that this order is highly consistent with that of the last column in Table 9.1 except that E was misplaced. We see later that E is something special; it is the best tester among the tested parents (see Section 9.5). The correlation between the GCA effects and the projections onto the ATC axis is 0.926.

9.3 SPECIFIC COMBINING ABILITY OF PARENTS

As was mentioned earlier in this chapter, in conventional diallel cross analysis, the term specic combining ability or SCA is associated with crosses. It is merely a residual that is not explained by the GCA effects and does not shed any light on the parents. The biplot, however, allows viewing of the SCA effects of the parents. If, as just demonstrated, the ATC abscissa approximates the GCA effects of the entries, then the ATC ordinate, which is orthogonal to the GCA effects, must approximate the SCA of the entries, which represents the tendency of an entry to produce superior hybrids with some, but not all, testers. Since the testers are located both above A and G and below B, C, D, and F the ATC abscissa (Figure 9.2), a greater projection toward either direction means greater SCA effects. Thus, entries g and a, below the ATC abscissa, had the greatest SCA effects or largest projections onto the ATC ordinate. Above the ATC abscissa, entries c and f had large SCA effects relative to entries b and d. Entry e had the smallest SCA effect since it had smallest projection onto the ATC ordinate.

9.4 HETEROTIC GROUPS

The tester vector view, brought up by the Show/Hide Tester Vectors function, helps group testers into heterotic groups (Figure 9.3). Since all testers have positive scores for PC1, the testers are grouped primarily by PC2 scores. Three groups of testers are obvious: A and G above the PC2 guideline, and B, C, D, and F below the PC2 guideline. Moreover, above the guideline, testers A and G interacted positively with entries b, c, d, and f; but they interacted negatively among themselves. Similarly, below the guideline, testers B, C, D, and F interacted positively with entries a and g; the four testers also interacted negatively among themselves. The PC2 guideline serves as a mirror the same thing is displayed both above and below it. This interaction pattern clearly indicates heterosis in crosses (A and G) (B, C, D, and F). Thus, A and G and B, C, D, and F are two different heterotic groups. Tester E does not seem to belong to either heterotic group (Figure 9.3). Rather, it is a tester that can effectively reveal the GCA effects of the parents, which is discussed next.

9.5 THE BEST TESTERS FOR ASSESSING GENERAL COMBINING ABILITY OF PARENTS
An ideal tester for revealing the GCA effects of entries should fulll two criteria: it should be representative of all testers and, at the same time, be most discriminating of the entries. Based on this denition, an ideal tester must be located on the ATC axis to be representative of all testers; its vector should be the longest of all testers to be most discriminating. Such a tester is indicated by the center of the concentric circles in Figure 9.4. Although such an ideal tester may not exist in reality, it can be used as a reference to compare the real testers. The concentric circles are drawn for this purpose with the hypothesized ideal tester at the center (Figure 9.4). The closer a tester is to this ideal center, the more desirable it is. Clearly, tester E was the best tester in this dataset, as it was very close to the ideal tester. On the contrary, G was the poorest tester as it is the least representative of all testers. Figure 9.4 can be brought about by a single click on the GGEbiplot function Compare Withthe Ideal Tester (Chapter 6).

2003 by CRC Press LLC

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

4.4

G
3.3

2.2

c d f

1.1

P C 2

b e B D F a C g E

-1.1

-2.2

-3.3

-4.4 -4.4 -3.3 -2.2 -1.1 0 1.1 2.2 3.3 4.4 5.5 6.6

P C1 Symmetrical Scaling
FIGURE 9.2 Average tester coordination (ATC) view of the biplot based on the wheat FHB research data. The small circle represents the average tester. Entries are in lowercase italics and testers are in uppercase.

2003 by CRC Press LLC

28

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

21

G c d b e B D F E f A

14

P C 2

-7

-14

C g

-21

-21

-14

-7

14

21

28

35

42

P C1 Tester-focused Scaling
FIGURE 9.3 The tester vector view of the biplot based on the wheat FHB research data. The seven parents are in lowercase when viewed as entries and in uppercase when viewed as testers.

2003 by CRC Press LLC

28

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

21

G c c c c B D F E c A

14

P C 2

-7

-14

C c

-21

-21

-14

-7

14

21

28

35

42

P C1 Tester-focused Scaling

FIGURE 9.4 Visual evaluation of the parents as testers. The concentric center represents the ideal tester, which is the most discriminating of entries and has no preference for mating partners, i.e., with zero specic combining ability. The seven parents are in lowercase italics when viewed as entries and in uppercase when viewed as testers.

2003 by CRC Press LLC

8 6 4 GCA scores 2 0 -2 -4 -6 -8 -10 30 40 50 60 70 Performance of F1 cross with tester E

FIGURE 9.5 Measured F1 values of each parent when crossed with tester E, plotted against the measured GCA of each parent. This gure reveals that the F1 with tester E is a good indicator of the GCA of an entry.

R = 0.69

C D

B A

2003 by CRC Press LLC

That tester E is the best tester implies that the GCA effect of an entry can be reasonably assessed by the value of its hybrid with tester E. To verify, F1 hybrid between each entry and tester E is plotted against the GCA effect of the entry (Figure 9.5). The two are highly correlated except entries b and e are off the regression line. The performance of BE was above expectation based on the overall pattern of the data.

9.6 THE BEST CROSSES

The polygon view of the biplot (Figure 9.6), brought up by the GGEbiplot function Show/Hide the Polygon, helps reveal the best crosses between the entries and the testers. The polygon consists of straight lines connecting the entries positioned furthest from the biplot origin such that all other entry markers are contained within the polygon. Lines perpendicular to each side or its extension of the polygon are drawn from the biplot origin to divide the biplot into sectors. In Figure 9.6, the biplot was divided into four sectors, with entries a, c, f, and g as the vertex entries, and are referred to as sector a, sector c, sector f, and sector g, respectively. No tester fell in the a sector, suggesting that entry a was not the best mating partner with any of the genotypes. Moreover, this suggests that entry a produced the poorest hybrids with some or all of the testers. The original data (Table 9.1) did show that parent a per se was poor and it produced the poorest hybrids with B, E, and G, which is consistent with the prediction from Figure 9.6. The original data also indicated that entry a was the best mating partners for F (Table 9.1), which is not reected in Figure 9.6. This illustration suggests entry a to be the second best mating partner of F, following entry g. Drawing a vector for tester F will facilitate visualization of this aspect. A single tester, G, fell in the c sector, indicating that entry c was the best mating partner with G. That is, the cross CG was predicted to be the best of all crosses involving G, which is consistent with the data (Table 9.1). Moreover, since parent C as a tester was not in sector c, the cross CG must be heterotic, i.e., better than both parents (CC and GG). Had tester C fallen in sector c, the combination CC, i.e., pureline C would be the best among all crosses involving C, and consequently, heterosis between C and any other parents would not be possible. A single tester, A, fell into sector f, indicating that entry f was the best mating partner for A, which is consistent with the data. Since tester F was not in the f sector, the cross FA was heterotic. Testers B, C, D, E, and F fell in the same sector, i.e., sector g, suggesting that entry g was the best mating partner for these testers. This is almost exactly the case (Table 9.1). Since G was not in the g sector, all crosses between genotype G and B, C, D, E, and F were heterotic. To summarize, the following seven F1 hybrids were among the best in this FHB study (Figure 9.6): CG in sector c; FA in sector f; and, GB, GC, GD, GE, and GF in sector g. Note that the cross CG appeared twice in the above list: in sector g, G was predicted to be the best mating partner with C, among others; and in sector c, C was predicted to be the best partner with G. Thus, C and G are mutually identied as the best partners; therefore, the cross CG should be the best of all possible combinations, which is consistent with the results (Table 9.1). In addition, tester E was almost on the line that separates sectors g and f. Therefore, F should also be a good parent for crossing with E, as can be veried from the data (Table 9.1).

9.7 HYPOTHESIS ON THE GENETIC CONSTITUTION OF PARENTS

In conventional analyses of diallel data, interpretation of the genetic constitutions of parents with regard to the trait under investigation is not attempted until the F2 generation. The biplot provides a unique means to visualize the interrelationships among parents, and thus allows hypotheses to be formulated on the genetic constitution of parents.

2003 by CRC Press LLC

28

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

21

G c d b e B D F a C g
-21 -14 -7 0 7 14 21 28 35 42

14

A f

P C 2

-7

-14

-21

P C1 Tester-focused Scaling
FIGURE 9.6 Polygon view of the biplot, showing the best crosses among all possible combinations. The best crosses are g by C, B, D, E, and F in the f sector and f by A, and E in the f sector. The parents are in lowercase italics when viewed as entries and in uppercase when viewed as testers.

2003 by CRC Press LLC

All testers had positive PC1 scores (Figure 9.7), implying that the interaction between the entries and the testers displayed by PC1 is of non-crossover type or proportional interaction (Crossa and Cornelius, 1997; Yan et al., 2000). Thus, the entry PC1 scores should approximate the GCA effects, as discussed in Section 9.2. Figure 9.7 indicates three groups relative to GCA. Group 1 contains entry a only, with the smallest GCA; Group 2 includes entries b, c, d, and e, with intermediate GCA effects and minor differences among them; and Group 3 includes entries f and g, with the largest GCA. To explain the differences in GCA, we hypothesize that Group 2 had an additive gene (A1) relative to Group 1 i.e., entry a, and Group 3 had an additional gene (A2) relative to Group 2. PC2 displays the nonproportional interactions between entries and testers, as the testers assumed different signs (Crossa and Cornelius, 1997; Yan et al., 2000). Specically, PC2 displays positive interactions between two heterotic groups: A and G as one group and B, C, D, and F as the other. If we assume that heterosis arises from the accumulation of different dominant genes, then the two groups must have different dominant FHB resistance genes that are designated as D1 and D2 (Figure 9.8). Entry e is located right on the PC2 guideline, implying that there was no nonproportionate interaction between e and either of the two heterotic groups. This can be explained by one of the three hypotheses: 1) entry e carries neither D1 nor D2; 2) entry e carries both D1 and D2; and 3) entry e carries a resistance gene that is different from both D1 and D2. The rst hypothesis is the least tenable because it cannot explain the heterosis observed between entry e and other parents. The second and the third hypotheses are equally satisfactory in explaining the heterosis between e and other parents, but the third hypothesis must invoke an additional gene. Figure 9.9 integrates Figures 9.6 to 9.8 to formulate the genotypes of the seven parents and to explain their performance as purelines and the performances of their hybrids. The hybrids between G and B, C, D, and F were among the best hybrids relative to FHB resistance because they each integrated the four resistance genes (A1, A2, D1, and D2). The same can be said for the cross AF. The formulations in Figure 9.9 allow some general discussion on the nature of a superior parent and a superior tester. Parents F and G are regarded as superior, because they had both high GCA and SCA. They had high GCA by having more resistance genes, and they had high SCA by having resistance genes different from those in the other heterotic groups. Parent E was regarded as a superior tester for identifying parents with high GCA, because it has a gene that is different from all the existing resistance genes in the other parents. Superior hybrids combine all or most of the resistance genes through one of the two pathways: 1) both parents exhibit high GCA but belong to different heterotic groups, and 2) one high-GCA parent and one superior tester. Caution must be used when reading this section, however. The hypotheses on the genetic constitutions of the parents with regard to resistance to FHB are highly speculative. These hypotheses have yet to be subjected to critical testing. Nevertheless, we are excited about the capability of the GGEbiplot to allow such hypotheses to be formulated, given the importance of hypotheses in scientic discovery.

9.8 TARGETING A LARGE DATASET

Butron et al. (1998) reported on percentage of yield loss in a 10-parent maize (Zea mays L.) diallel cross following articial infestation with corn pink stem borer (PSB), Sesamia nonagrioides. The yield loss data were converted into tolerance to PSB, measured as the yield of infested plants as percentage of noninfested plants (Table 9.2). We use this dataset to illustrate biplot analysis of a large diallel dataset.

2003 by CRC Press LLC

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

4.4

G
3.3

2.2

c d f

1.1

P C 2

b
A1

e B D F a C g
A1A2

-1.1

-2.2

-3.3

-4.4 -4.4 -3.3 -2.2 -1.1 0 1.1 2.2 3.3 4.4 5.5 6.6

P C1 Symmetrical Scaling
FIGURE 9.7 The proposed genotypes of the parents based on PC1. Parents when viewed as entries are in lowercase italics and in uppercase when viewed as testers.

2003 by CRC Press LLC

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

4.4

G
3.3

2.2

c d
D1

A f

1.1

P C 2

b e
D1D2?

E B D F

-1.1

-2.2

a
-3.3

C
D2

g
-4.4 -4.4 -3.3 -2.2 -1.1 0 1.1 2.2 3.3 4.4 5.5 6.6

P C1 Symmetrical Scaling
FIGURE 9.8 The proposed genotypes of the parents based on PC2. Parents when viewed as entries are in lowercase italics and in uppercase when viewed as testers.

2003 by CRC Press LLC

28

Model 1 PC 1 = 46% PC 2 = 31% Sum = 77%

21

G
A1D1

14

c d b e
A1D1D2?

A f
A1A2D1

P C 2

E B D F

-7

-14

a
D2

C g A1A2D2

-21

-21

-14

-7

14

21

28

35

42

P C1 Tester-focused Scaling
FIGURE 9.9 Possible genotypes of the parents based on both PC1 and PC2.

TABLE 9.2 Tolerance of Ten Corn Inbreds (Diagonal) and their F1 Hybrids to Pink Stem Borer, as Measured by the Percentage of Yield Retained after Infestation
A A B C D E F G H I J 85.8 89.1 86.3 82.0 86.6 92.4 82.9 88.1 84.2 86.0 B 89.1 88.3 79.7 72.4 89.6 78.5 97.6 84.8 83.0 89.8 C 86.3 79.7 89.9 88.3 97.2 86.0 72.2 92.5 74.9 82.6 D 82.0 72.4 88.3 -a 91.1 91.4 76.5 90.4 76.8 81.5 E 86.6 89.6 97.2 91.1 81.7 83.3 86.3 94.9 83.9 86.5 F 92.4 78.5 86.0 91.4 83.3 88.4 88.9 87.0 76.8 83.8 G 82.9 97.6 72.2 76.5 86.3 88.9 70.7 97.7 75.8 83.9 H 88.1 84.8 92.5 90.4 94.9 87.0 97.7 87.6 99.9 83.9 I 84.2 83.0 74.9 76.8 83.9 76.8 75.8 99.9 92.9 82.1 J 86.0 89.8 82.6 81.5 86.5 83.8 83.9 83.9 82.1 78.0 Mean 86.3 85.3 85.0 83.4 88.1 85.6 83.2 90.7 83.0 83.8

Note: The codes of the inbreds are: A = A509; B = A637; C = A661; D = CM105; E = EP28; F = EP31; G = EP42; H = F7; I = PB60; and J = Z77016.
a

Missing cell replaced by its column average for completing the calculation.

Source: Data based on Butron, A. et al., Crop Sci., 138:11591163, 1998.

2003 by CRC Press LLC

9.8.1 SHRINKING

THE

DATASET

BY

REMOVING SIMILAR PARENTS

The polygon view of the biplot based on the 10 10 diallel data (Table 9.2) explained 63%, 37% by PC1 and 26% by PC2, of the total variation (Figure 9.10). This is considerably smaller than that explained by the biplot for the FHB dataset (Figure 9.1). Less variation explained by the biplot implies that some predictions based on the biplot will be less accurate. Therefore, it would be a good strategy to try to reduce the data size by removing redundant parents. Figure 9.10 suggests that A and J, C and D, and E and F are pairs of parents that are similar both as entries and testers. Therefore, parents J, D, and F (alternatively, A, C, and E) can be removed from the data without losing critical information.

9.8.2 THE BEST CROSSES

Figure 9.11 is the polygon view of the biplot based on the subset of data with parents D, F, and J removed. It explains 73%, (53% + 20%), of the total variation. It indicates that entry h was the best mating partner of testers A, G, and I; entry g was the best mating partner of H and B; entry c was the best mating partner of E; and entries c and h were equally good partners of tester C. Therefore, the best hybrids for PSB resistance were: four hybrids involving H: HA, HG, HI, HC, GB; and CE. Sure enough, these are the best hybrids based on the data per se (Table 9.2).

9.8.3 GCA

AND

SCA

The ATC view brought up by the GGEbiplot function Average Tester Coordination helps in visualizing the GCA and SCA effects of the parents (Figure 9.12). The ATC axis happens to coincide with the PC1 axis. Thus, the PC1 scores of the entries approximate their GCA effects, with r = 0.953. The highest PC1 entry is h, followed by b, e, a, i, c, and g. Entry c had the highest SCA; it interacted positively with itself and E but negatively with others.

9.8.4 BEST TESTER

The best tester in this subset was parent G, as it is closest to the ideal tester represented by the center of the concentric circles (Figure 9.13). All other parents are poor as testers, either due to lack of discriminating ability (A and E), or lack of representativeness (C and I). Parents B and H are particularly poor as testers as they have negative PC1 scores, which means that parents that produce good hybrids with B and H tend to have low, rather than high, GCA. The existence of different signs for PC1 scores is an indication of large SCA or interaction effects relative to GCA or additive effects (Yan et al., 2001).

9.8.5 HETEROTIC GROUPS

The tester vector view, brought about by the GGEbiplot function Show Tester Vectors, helps visualize the heterotic groups (Figure 9.14). The testers seem to fall into three groups: H and B in group 1; G and I in group 2; and C and E in group 3. Tester A falls between groups 2 and 3 and has a short vector. Groups 1 and 2 interacted positively to produce heterosis in terms of PSB resistance. They are two heterotic groups with different dominant resistance genes. C and E in group 3 interacted negatively with groups 1 and 2, suggesting recessive resistance genes. However, within group 3, C and E interacted positively.

9.8.6 GENETIC CONSTITUTIONS

OF

PARENTS

WITH

REGARD

TO

PSB RESISTANCE

From the perspective of the entries, PC1 is the contrast of c, g, and i vs. b, e, and h, although there are differences within each group. For simplicity, we assign dominant genes R1 and R2, respectively (Figure 9.15). This assignment implies that there could be heterosis in all crosses between the two groups.

2003 by CRC Press LLC

Model 1 PC 1 = 37% PC 2 = 26% Sum = 63%

3.3

b
2.2

B i I G

1.1

g j J a E A e f F h

P C 2

-1.1

-2.2

c
-3.3 -2.2 -1.1

d
0 1.1

D
2.2

C
3.3 4.4 5.5

P C1 Symmetrical Scaling
FIGURE 9.10 Polygon view of the biplot of the 10 10 maize diallel crosses for resistance to corn pink stem borer. The ten inbreds are in lowercase when viewed as entries and in uppercase when viewed as testers.

2003 by CRC Press LLC

4.4

Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

3.3

2.2

P C 2

1.1

Ee

a H

A h

-1.1

g i
-3.3 -2.2 -1.1

b I
0 1.1 2.2 3.3

-2.2

4.4

5.5

P C1 Symmetrical Scaling
FIGURE 9.11 Polygon view of the biplot based on a subset of the maize diallel cross data; three parents (D, F, and J) were deleted from the data. Parents when viewed as entries are in lowercase italics and in uppercase when viewed as testers.

2003 by CRC Press LLC

Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

18

c
12

P C 2

E e a A H h B i I
-6 0 6 12 18 24 30

-6

-12 -24 -18 -12

P C1 Entry-focused Scaling
FIGURE 9.12 The ATC view of the biplot based on the maize diallel cross subset, showing the GCA of the entries.
20
Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

15

10

P C 2

cE c H A c B c
-15 -10 -5 0 5 10

-5

c I
15 20 25

-10

30

35

P C1 Tester-focused Scaling
FIGURE 9.13 The ATC view of the biplot based on the maize diallel cross subset, comparing testers with the ideal tester, which is presented by the concentric center. The ideal tester is the most discriminating of entries and has no preference for mating partners.

2003 by CRC Press LLC

PC2 is the contrast between c and e vs. g, i, b, and h (Figure 9.16). Interestingly, these two groups interacted negatively between groups and positively within groups, suggesting that recessive resistance genes are involved. Therefore, r3 and r4 are assigned to the two groups, respectively. Parent A is near the biplot origin, indicating lack of interaction with any of the two groups. It is, therefore, assigned the double recessive genotype r3r4. Integrating Figures 9.11, 9,15, and 9.16 gives a complete picture of the PSB resistance phenotypes of the parents and their hybrids, and possible interpretations from the perspective of genetic constitutions (Figure 9.17). Again, it should be pointed out that these interpretations are only hypotheses, which have to be critically tested.

9.9 ADVANTAGES AND DISADVANTAGES OF THE BIPLOT APPROACH

The main purpose of conducting diallel analyses is to obtain information on the parents: their genetics and their potential or probability of generating superior hybrids or pureline cultivars if used as parents. In conventional diallel-cross analyses, understanding of the parents relies solely on GCA. The GCA effects are reliable for parent evaluation only if SCA effects of the crosses are negligibly small, however. If there are considerable SCA effects, GCA alone will be of limited use. Thus, we are taught in statistics and quantitative genetics courses that if there are signicant interaction effects, the analysis should be focused on individual crosses rather than main effects of the parents, i.e., GCA. On the contrary, the biplot approach allows understanding of the parents through two dimensions: GCA and SCA. Since principal components are estimated through least-squares methods, PC1 always explains at least as much variation as the GCA effects, and any variation explained by PC2 is additional to GCA. Therefore, theoretically there is no doubt that a biplot of PC1 vs. PC2 is at least as effective as the conventional approach in achieving an understanding of the parents. This is the rst advantage of the biplot approach over the conventional method. A second obvious advantage of the biplot approach is its graphical presentation of the diallel data: A picture is worth a thousand words. From a biplot, the GCA and SCA of the parents, the best crosses, the best testers, and the heterotic groups are immediately displayed. With the help of GGEbiplot, visualization of these aspects of a dataset occurs within seconds. Indeed, the biplot approach provides a nonsubstitutable means for revealing patterns of a diallel cross dataset. The GGEbiplot software also has a function Diallel Without Parents, which generates a biplot after the parent per se values are removed. Although results with or without parents are generally close, this provides another perspective for examining the data. A third advantage of the biplot approach follows. Since the biplot displays a complete picture of the interrelationships among parents, it provides a unique opportunity or possibility to peek into the genetic constitutions of the parents using the F1 rather than the F2 generation. Although the hypotheses on the genetics of parents are preliminary at this stage, since no experimental verications have been attempted so far, we are quite condent about the relative genetic differences among the vertex parents. Research on testing hypotheses based on the biplot pattern will have great merit. Once veried, this would be a great asset for genetics research. The only disadvantage of the biplot approach we can think of is its lack of a measure of uncertainty. If two parents are located close to one another in the biplot, we are sure that they have similar genetic behaviors. We cannot quantify the difference, however. Also, we cannot be sure if the two parents are signicantly different if they look different in the biplot. Therefore, conventional statistics, such as variance analysis on GCA and SCA, are still necessary. If both GCA effects and SCA effects were nonsignicant, there would be no need to conduct biplot analysis. In reality, however, this would be very rare, since diallel crosses are expensive, and the parents are usually carefully selected. Once either GCA or SCA is known to be signicant, the lack of a measure of uncertainty will no longer be a problem, because the difference between any two parents can be

2003 by CRC Press LLC

20

Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

15

10

P C 2

eE a H A h B i
-15 -10 -5 0 5 10

-5

b I
15 20 25

-10

30

35

P C1 Tester-focused Scaling
FIGURE 9.14 The tester vector view of the biplot based on the maize diallel cross subset, showing groups of testers.
4.4
Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

3.3

2.2

P C 2

1.1

Ee
R1 R2

a H

h B i b I
-1.1 0 1.1 2.2 3.3 4.4 5.5

-1.1

-2.2 -3.3 -2.2

P C1 Symmetrical Scaling
FIGURE 9.15 The proposed genotypes of the entries based on PC1.

2003 by CRC Press LLC

4.4

Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

3.3

2.2

r3

P C 2

1.1

Ee
r3r4

a H

A h

-1.1

g i
-3.3 -2.2 -1.1

r4

b I

-2.2

1.1

2.2

3.3

4.4

5.5

P C1 Symmetrical Scaling
FIGURE 9.16 The proposed genotypes of the entries based on PC2.
4.4
Model 1 PC 1 = 53% PC 2 = 20% Sum = 73%

R1r3
3.3

2.2

P C 2

1.1

Ee
R2r3

a H
r3r4

A h R2r4 b
R2r4

-1.1

g
R1r4 R1r4

B i

G I

-2.2 -3.3 -2.2

-1.1

1.1

2.2

3.3

4.4

5.5

P C1 Symmetrical Scaling
FIGURE 9.17 Proposed genotypes of the entries based on both PC1 and PC2. Minor genes may exist to differentiate between entries g and i and between entries b and h.

2003 by CRC Press LLC

visualized with reference to the biplot size. If two parents look different, they are probably different. If they look similar, they are probably not very much different. Unfortunately, decision-making in scientic research is heavily dependent on statistical tests. If two things look similar, what is the point in trying to prove that they are different? Yes, statistical tests make us more condent about our conclusions. But the bottom line is that all decisions are subjective: it is subjective to choose among many testing methods, and it is subjective to choose a signicance level. Someone has said, the relationship between statistics and agriculture is like that between a lamp post and a drunk it is for support, not illumination. It is our belief that researchers heavily rely on statistical tests partly because they have no means of getting a complete picture of their problem, research, or dataset. In the light of a biplot, statistical tests may become less crucial for making decisions. Seasoned breeders do not make decisions based on statistical tests, not because tests are not available, but because they are not needed. In most cases, the need for a statistical test for a particular hypothesis is an indication of lack of condence by the researcher in the hypothesis; but when this is the case, statistical tests will not dramatically increase condence. Nonetheless, the GGEbiplot software is now equipped to do conventional statistics. This will allow researchers to examine their data both via the biplot and conventional way.

2003 by CRC Press LLC