6. Definition of adaptation strategy, selection environments, genetic resources and adaptive traits

6.1 Adaptation strategy

Indications based on predicted yield gains

This chapter focuses specifically on the possible indications provided by regional yield trials for breeding programmes. The use of selection theory is considered for the comparison of different strategies for the final stages of selection in terms of relative efficiency, following the outline based on predicted yield gains over a target region that derive from one cycle of selection (Annicchiarico, 2002). Its conventional use for the estimation of selection gains is inappropriate here, unless the multi-environment data available relate to a random sample of elite breeding lines (as may be the case for self-pollinated species). A fair and straightforward comparison of wide vs. specific adaptation implies the absence of substantial differences in costs between the two strategies. This can be envisaged on the basis of predicted yield gains obtained from the same number of total selection environments (e.g. 4 sites by 2 years = 8 environments, in a wide adaptation prospect, and 2 sites by 2 years = 4 environments for each of two subregions, in a specific prospect), with the further assumption of no duplication of breeding stations in the specific adaptation option (i.e. the hypothesis that a single station would centralize crossing and hybridization operations and provide each subregion with novel germplasm). The predicted gain from wide adaptation is not necessarily lower in this case, because more environments allow for greater precision in the estimation of genotypic values (mainly in relation to random variation due to GE interactions).

In general, the mean predicted yield gain across E environments can be estimated as (Falconer, 1989^[29]; Cooper et al., 1996a):

ΔG= i h² s_p

where i = standardized selection differential, h² = estimated broad sense heritability on a genotype mean basis and s_p = square root of the estimated phenotypic variance across environments. In particular:

h² = s_g²/[s_g² + (s_ge²/E) + (s_e²/E R)]

[6.1]

where s_g², s_ge² and s_e² are estimates of the components of variance for genotype, GE interaction and pooled error, and E and R = number of selection environments and number of experiment replicates. The s_p term is equal to the square root of the denominator in equation [6.1].

In the following formulae, h² values for prediction of yield gains are calculated from components of variance estimated from ANOVAs including test environments of all locations or their subsets, whereas E and R values are set by the user as hypothesized for the selection work. E, in particular, is usually different from that of the multi-environment data set. R and i are set to constant values in the assessment. Hereafter, E_A and E_B represent the number of selection environments for two subregions (A and B, respectively) in a specific adaptation scenario, whereas E_AB = E_A + E_B represents the number of selection environments which are used, in a wide adaptation scenario, for parallel selection across the subregions. P_A and P_B represent the proportion of the target region occupied by subregions A and B (P_A + P_B = 1). The proportion relates to the cropping area of each subregion, as indicated by the number of test locations assigned to subregions in the analysis of adaptation or, more precisely, by the relative size of subregions (preferably in terms of growing area for the crop) following the spatial and temporal scaling-up of results. The relative number of selection environments for each subregion should be roughly proportional to the relative extent of the subregion (e.g. if E_AB = 6 and P_A = 0.64, then E_A = 4 = 2 sites by 2 years, and E_B = 2 = 1 site by 2 years - see Fig. 6.1). An exact match between proportion of selection environments and P value for each subregion in the parallel selection for wide adaptation could be obtained by a weighted selection procedure, such as that in Podlich et al. (1999), that down-weights the contribution of environments that are over-represented, and vice versa. The average predicted gain (per unit area) provided by a wide adaptation strategy is:

ΔG_W = i h_AB² s_p(AB)

[6.2]

where h_AB² and s_p(AB) are obtained from equation [6.1] after estimating the components of variance from the combined ANOVA including all test environments, and inserting E_AB and R values as appropriate in the formula. The average predicted yield gain over the target region provided by breeding for specific adaptation (ΔG_S) arises from a weighted mean of the gains ΔG_A and ΔG_B predicted for the subregions A and B, respectively:

ΔG_A = i h_A² s_p(A)

[6.3]

ΔG_B = i h_B² s_p(B)

ΔG_S = [(ΔG_A P_A) + (ΔG_B P_B)]/(P_A + P_B) = (ΔG_A P_A) + (ΔG_B P_B)

[6.4]

where heritability and phenotypic variance values are obtained from equation [6.1] after estimating the components of variance from the combined ANOVA including only environments of the test locations grouped in subregion A (values h_A² and s_p(A)) or B (h_B² and s_p(B)), and inserting E_A or E_B and R values in the same equation (as appropriate). The procedure described for two subregions can easily be extended to three or more subregions, comparing predicted yield gains over the region in a wide adaptation scenario (formula [6.2]) with those deriving from specific selection for each subregion (formula [6.4]).

FIGURE 6.1 - Hypothetical selection locations for three adaptation strategies compared in terms of predicted yield gains

In fact, a third scenario may also be envisaged for two subregions, contemplating selection in only one subregion, so that yield gains in the other subregion derive from correlated responses in an indirect selection context (Falconer, 1989^[30]; Cooper et al., 1996a). The predicted yield gain in subregion B deriving from indirect selection in subregion A is:

ΔG_B/A = i h_A h_B r_g(AB) s_p(B)

[6.5]

where h_A and h_B are square roots of broad sense heritability values for each subregion previously estimated, together with s_p(B), through equation [6.1]; and r_g(AB) is the coefficient of genetic correlation for genotype yields between the subregions. The predicted yield gain from direct selection in subregion A may be calculated with equation [6.3]. However, comparison with the other scenarios on the basis of the same number of selection environments suggests the introduction of a difference in the current calculation of h_A² and s_p(A) values used for estimation of ΔG_A, because all selection environments would be attributed to subregion A for direct selection now. For example, if E_A = 4 and E_B = 2 in previous formulae for direct selection in each subregion, E_A = 4 + 2 = 6 environments assigned to subregion A (e.g. 3 sites by two years - see Fig. 6.1) and, hence, used for estimation of h_A² and its introduction in formula [6.3] in the present context. The average predicted gain over the region provided by breeding only for subregion A arises from a weighted mean of the gains ΔG_A and ΔG_B/A predicted for subregions A and B, respectively:

[(ΔG_A P_A) + (ΔG_B/A P_B)]/(P_A + P_B) =(ΔG_A P_A) + (ΔG_B/A P_B)

[6.6]

Extending the case of two environments envisaged by Burdon (1977) to two subregions, an estimate of r_g(AB) in equation [6.5] can be provided by the following formula:

r_g(AB) = r_p(AB)/(h_A' h_B')

where r_p(AB) is the phenotypic correlation coefficient between subregions for genotype yields (averaged across environments in each subregion), and h_A' and h_B' are square roots of the broad sense heritability on a genotype mean basis estimated for subregions A and B. The difference between h_A' and h_B' and previous estimates (h_A and h_B) is that E and R in equation [6.1] are, respectively, the actual number of test environments per subregion and experiment replicates in the analysed data set.

TABLE 6.1 - Predicted yield gain over the region of lowland northern Italy from selection of lucerne populations specifically adapted to three or two subregions relative to selection for wide adaptation (DGS/DGW ratio)

Note: See Figure 5.7 for subregion definition.

E = number of selection environments; P = proportion of the target region; h₂ = broad sense heritability; sp = phenotypic standard deviation, estimated for E environments and four experiment replicates; ΔG = predicted yield gain (over a three-year crop cycle) per selection cycle (standardized selection differential = 1.64, for 10% selection intensity applied to 20 elite populations). Estimated genetic correlations for subregion B are: r_g = 0.46 with subregion A, r_g = -0.11 with subregion C.

Source: Annicchiarico, 2002.

The relative merit of the three selection strategies exemplified in Figure 6.1 for two provisional subregions derives from the comparison of predicted yield gains according to formulae [6.2], [6.4] and [6.6]. Other more complex scenarios may be assessed by a combination of the above procedures. For example, direct selection could be devised specifically for two subregions, in either of which indirect selection is also performed for a third subregion.

As an example, different adaptation strategies for lucerne in northern Italy are compared on the basis of predicted yield gains using data of regional variety testing and the provisional definition of subregions reported in Figure 5.7. Scenarios contemplate either one selection location per subregion, or one for each of the contrasting subregions A and C (with no repetition in time of the trials). In the latter case, yield gains for subregion B in a specific adaptation prospect derive from selection in the subregion that maximizes the indirect selection gain (i.e. subregion A). Table 6.1 reports the results, as well as the information that is needed for application of previous formulae. P values are rough estimates of the relative cropping areas of the subregions. Selection for specific adaptation to three or two subregions is estimated to be at least six times more effective than selection for wide adaptation. In fact, the current comparison of predicted yield gains relates to the selection of populations which, for an open-pollinated species such as lucerne, is rather out of the context of breeding schemes. In this case, breeding schemes basically involve the selection of individual plants, although the final selection among different improved populations may also be present. Caution is required and indications should be validated by further research in terms of actual yield gains relative to phenotypic or genotypic selection. The above described procedure is completely reliable when multi-environment data and the final stages of selection are relative to inbred lines, hybrids or clones.

As anticipated in Section 2.2, comparison based on predicted yield gains tends to underestimate the potential advantage of a specific adaptation strategy if the material with specific adaptation to each subregion is under-represented in the data set. Also in this case, further comparison in terms of actual yield gains, involving the most promising specific adaptation scenario vs. the wide adaptation option, can be recommended.

Indications based on actual yield gains

Singh et al. (1992), Ceccarelli (1994) and Ceccarelli et al. (1998) provide examples of the comparison between wide and specific adaptation strategies based on actual yield gains. In the latter work, a large set of barley genotypes representative of the genetic base was evaluated in two subregions, and lines were selected across all locations (wide adaptation strategy) or across locations of each individual subregion (specific adaptation). The selected material was tested again in the two subregions. Results are partly summarized in Table 6.2. For the less favourable subregion (i.e. A), the yield of specifically selected material is, on average, about 9 percent higher than that of germplasm selected across the two subregions. Yield gains (ΔG) are currently estimated in relation to mean value of a set of control cultivars. Only specific breeding can provide genetic progress for this subregion, whereas the two adaptation strategies provide similar values of yield response and yield gains of selected material in the other subregion. The advantage of specific breeding is clear from these results. The availability of information on the relative size of the two subregions would allow to estimate the yield gain provided by the two strategies across the target region. For example, if the subregions were of equal size (P_A = P_B = 0.50), the gain from wide (ΔGW) and specific (ΔGS) breeding would derive from yield gains realized in each subregion by the two strategies:

ΔG_W = (ΔG_A P_A) + (ΔG_B P_B) = (-0.03 × 0.50) + (0.08 × 0.50) = 0.025 t/ha per cycle; and

ΔG_S = (ΔG_A P_A) + (ΔG_B P_B) = (0.03 × 0.50) + (0.08 × 0.50) = 0.055 t/ha per cycle.

The efficiency of specific breeding relative to breeding for wide adaptation would be: 0.055/0.025 = 220%.

TABLE 6.2 - Mean yield of barley lines previously selected for specific adaptation to either of two subregions or for wide adaptation to the region of northern Syria, yield ratio of specific (S) to wide (W) adaptation strategies, and actual yield gain (ΔG) relative to mean value of six control cultivars

^a Specific adaptation: independent selection in each subregion. Wide adaptation: parallel selection across subregions.

^b Mean value of control cultivars is 0.70 t/ha for subregion A and 4.31 t/ha for subregion B.

Source: Modified from Ceccarelli et al., 1998 - data for Cohort 89 in Table 5, with kind permission from Kluwer Academic Publishers.

Concluding remarks

The comparison of adaptation strategies in terms of predicted or actual yield gains has so far assumed that selection in different subregions is performed on a common pool of breeding material. The advantage of specific adaptation could be even greater if it also implied (at least to some extent) the use of a distinct genetic base for each subregion. This approach is compatible with the presence of a single institution centralizing the breeding work (Fig. 2.2), provided that distinct germplasm is produced for each subregion from specifically adapted parent material. The adaptive response of the possible parents may be assessed on the basis of the evaluation of genetic resources in each subregion, or inferred from the possession of morphophysiological traits that are known to confer specific adaptation to a given subregion.

The current emphasis on predicted or actual yield gains across the target region, which reflect the opportunities for increasing crop production at country level, is particularly useful for decisions on adaptation strategies by public breeding programmes. Alternative criteria, e.g. the size of yield gains for individual subregions in relation to the commercial interest of each area, may be preferred by private breeding programmes.

6.2 Selection environments

Multi-environment trials may also generate useful indications with regard to selection procedures, especially selection environments, for use in the framework of the selected adaptation strategy. In fact, a distinction can be made between early and late stages of selection:

• Early stages (including large samples of segregating or single-seed descent lines for self-pollinated varieties, candidate parents for improved open-pollinated populations, clones for vegetatively-propagated cultivars or inbred lines for hybrids) are likely to be performed in only one location.

• Late stages (including elite inbred lines, clones, hybrids or candidate parents of a synthetic variety) frequently involve two or more test sites, particularly when directed to the entire target region.

Locations for early selection

For early stages, the optimal site maximizes the indirect selection gain predicted across the sample population of environments represented by test sites of the target region (wide adaptation) or subregion (specific adaptation). A simple and convenient criterion for assessing the relative value of the candidate selection locations is provided by the size of the phenotypic correlation coefficient between yields of genotypes on the individual site (averaged across years, for trials repeated in time) and average yields of genotypes across the region or subregion (Braun et al., 1992). This correlation takes account of the two components (i.e. the genetic correlation between the individual location and the target population of environments; and the broad sense heritability on the location) which contribute to the differences between sites in screening ability (Pederson and Rathjen, 1981; Cooper et al., 1996a). Economic considerations may point to an existing selection site rather than one showing the highest correlation value, with a predicted loss in efficiency of selection proportional to the relative difference between the two correlation coefficients. The test site adopted for early stages of selection can conveniently be used to evaluate genetic resources which may be potential parent material (as hypothesized in Fig. 2.2 for a specific adaptation strategy).

Locations for late selection

For late stages of selection in annual crops, the inspection of variance components relative to environments of the target region or subregion can help define the optimal number of selection sites, years and experiment replicates as a function of the desired level of precision for genotype comparison. The estimated variance of a genotype mean value (s_m²) for L = location, Y = year and R = experiment replicate numbers that are hypothesized for selection is (Bowman, 1989):

s_m² = s_gl²/L + s_gy²/Y + s_gly²/LY + s_e²/RLY

Early work by Sprague and Federer (1951) and numerous subsequent reports have highlighted that increasing the number of selection environments is more beneficial than increasing the number of experiment replicates. However, selection trials for advanced breeding material hardly ever include less than two replicates. In fact, indications concerning the optimal number of replicates may be obtained from inspection of the ratio of genotypic variance to experimental error variance and its implications (Gauch and Zobel, 1996b), conveniently repeated on several individual experiments performed on a given selection location. The number of years of selection is usually kept low (often no more than two) so as not to delay the release of improved germplasm. For perennials, trials in the last stages of selection are rarely repeated in time. Therefore, breeding programme decisions mainly concern the number and type of selection locations. The number can be determined either by the above criterion on the basis of the precision required for genotype comparison, or by the estimation of predicted yield gains for the relevant region or subregion as a function of different numbers of selection sites (according to the formulae provided in Section 6.1). The latter facilitates the costs-benefits assessment, as the positive effects of yield gains on national agricultural production (public programmes) or on seed market income (private programmes) can be weighed against the associated costs. This is fully appropriate, however, if the tested genotypes can be assimilated to a representative sample of elite breeding lines.

An accurate selection of test locations is particularly important when a wide adaptation strategy is associated with relatively large variance of the GL interaction component across the region (cases 3 and 4 in Fig. 2.3). As suggested by several works (e.g. Brennan et al., 1981; Zavala-Garcia et al., 1992a; Calhoun et al., 1994; Cooper et al., 1997), parallel selection is preferable on sites which contrast for GL effects and are jointly capable of reproducing the mean responses of genotypes across the target population of environments, rather than on sites (implicitly similar for GL effects) which are capable individually of reproducing the same responses. The reason lies in the opportunity provided by selection on contrasting sites for disclosing and selecting material capable of assembling different adaptive traits of possible interest in the target environments (e.g. tolerance to various biotic or abiotic stresses). Contrasting locations can conveniently be selected following the ordination of locations according to site mean yield (joint regression), significant GL interaction PC axes (AMMI) or significant environmental covariates (factorial regression), giving priority to the most informative analytical model. Contribution analysis can help identify selection sites in pattern analysis (Shorter et al., 1977). In any case, selected locations should also possess high broad sense heritability (h₂) as estimated on a genotype mean basis:

h₂ = s_g²/(s_g² + s_gy²/Y + s_e²/YR)

where s_g² is genotype, s_gy² GY interaction and s_e² experimental error components of variance (all estimated from ANOVA for the location), and R and Y are, respectively, the number of replicates and test years (or crop cycles) hypothesized for selection. It should be noted that the estimation of h² values based on a small number of test years may be largely biased by unusually high or low within-site GY interaction or experimental error values. A reliable evaluation of candidate selection locations in this respect may require the repetition of the assessment on historical data sets (including a common set of genotypes evaluated for several years at the location). The screening ability of different sets of contrasting locations may also be compared on the basis of the phenotypic correlations coefficient between yields of genotypes averaged across candidate sets of locations and average yields of genotypes across the target population of environments (preferably excluding the test environments belonging to the evaluated set of candidate locations from the sample of target environments).

Managed/artificial environments

In some cases, selection for wide or specific adaptation may partly be carried out using managed or artificial environments. These environments, which differ in terms of one or more environmental factors strictly associated with the occurrence of GL interaction, aim to reproduce the levels of the factors characterizing either different subregions (specific adaptation strategy) or locations constrasting in GL effects (wide adaptation). The adoption of these environments in the place of ordinary selection locations may be of special interest for a reduction in costs when:

• test sites belong to areas that are remote or have little infrastructure; and

• the managed environments do not entail high implementation costs and can conveniently reproduce the adaptive responses of genotypes across the target region.

The definition of these environments may be envisaged after identifying the environmental factor(s) that are closely related to GL interaction occurrence (as described, in particular, for analysis of adaptation of regional yield trials based on AMMI or factorial regression modelling).

For example, Annicchiarico and Mariani (1996) showed that the adaptive responses of durum wheat genotypes across the target region including southern Italy and coastal areas of central Italy (reported in Fig. 5.1 for best-yielding material and strictly related to rainfall) could substantially be reproduced across two managed environments placed in a favourable, high-rainfall location. One environment was ordinarily managed, while the other was drought-stressed by elimination of a part of the rainfall through metal channels placed on the soil between rows of plants. The evaluation of advanced breeding lines across the two environments could be envisaged in both a wide and a specific adaptation prospect (widely adapted material would show high mean yield and average regression slope; specifically adapted material would show high yield in either environment), reducing in both cases the need for multilocational testing. The adoption of one rainfed and one irrigated environment on a severely drought-prone site would have been a cheaper alternative, but could not be put into practice because of the high year-to-year variation in rainfall amount characterizing the drier sites in the region.

A second example is reported by Annicchiarico (2002). Following the analysis of adaptation of lucerne varieties in northern Italy, showing that the ordination of sites in Figure 5.4 (A) is correlated positively with the soil clay content and negatively with the level of drought stress of locations (Annicchiarico, 1992), four artificial environments were created by the factorial combination of two soil types (sandy-loam and clay) by two stress levels (severe and limited) representing the range of variation of these factors in the region. The environments were accommodated at the breeding station located in subregion A of Figure 5.7. Local sandy-loam soil available on the site and clay soil imported from subregion C were used to fill large, bottomless containers in concrete laid in a field. A high or low drought stress level was reproduced by means of irrigated or rainfed cropping. The adaptive responses across locations of probe cultivars, 'Europe', 'Prosementi' and 'La Rocca', could successfully be reproduced across the four artificial environments (Fig. 5.4). These environments may be used in selecting for wide adaptation, or in selecting for subregion C while working in subregion A. As a matter of fact, the two environments 'no drought stress, sandy-loam soil' and 'stress, clay soil', reproducing subregions A and C, respectively, would be sufficient.

In other cases (e.g. Dowker et al., 1978; Cooper et al., 1997), the information on crucial environmental variables generated by analysis of adaptation has been used to define a set of managed environments with contrasting management factors (e.g. presence and extent of irrigation; level of fertilization; sowing date), in the context of a definite wide adaptation strategy. Federer and Scully (1993) propose statistical designs to select breeding material for wide adaptation across a factorial combination of two or three physical or management factors that reproduce the variation for crucial environmental variables across a target region. These designs may also be used within the framework of a specific adaptation strategy to identify genotypes that respond particularly well under specific conditions.

Additional aspects

Participatory plant breeding procedures may facilitate breeding also or exclusively for areas in which test sites are implemented with difficulty (Ceccarelli et al., 2000). In most cases, the phase of farmers' selection may conveniently follow a preliminary phase of breeder's selection across or within (Fig. 2.2) subregions, performed on a small number of contrasting sites or managed/artificial environments. A specific adaptation strategy is generally recommended in conjunction with participatory plant breeding, in order to fully exploit the potential of this approach, and better meet farmers' expectations concerning the production of novel germplasm adapted to specific cropping environments and local requirements for quality traits (Eyzaguirre and Iwanaga, 1996; McGuire et al., 1999).

For target regions where major climatic characteristics show high year-to-year variation, a wide adaptation strategy is frequently caused by, and has to cope with, large GY interaction effects within locations. Even for carefully chosen selection sites, environments in individual years may frequently misrepresent the target population of environments, thereby failing to adequately reproduce the mean responses of genotypes across the region. This may result in low selection gains, especially when using just a few selection environments (Cooper et al., 1996a). Strategies to cope with this situation are discussed by Cooper et al. (1996b). One possibility implies the preliminary classification of a large sample of target environments by pattern analysis on the basis of GE interaction effects, identifying groups characterized by a specific response of some reference genotypes and of which the relative frequency is estimated. These reference genotypes (termed probe genotypes by Fox and Rosielle [1982b]), grown alongside tested germplasm in future trials, are used to assign selection environments to one of the groups on the basis of their response. Each environment receives a weight on selection that is proportional to the frequency of its group (Podlich et al., 1999). Another strategy implies the definition of a set of managed environments that can adequately reproduce the mean responses of genotypes in the target region, mainly by means of correlation between yields of genotypes averaged across candidate sets of managed environments and average yields of genotypes across the target population of environments (Cooper et al., 1995, 1996a) or its subsets as defined by pattern analysis of GE effects (Cooper et al., 1997).

The information on site similarity for GL interaction effects, possibly combined with information on broad-sense heritability value of locations, can also be used to define optimal test sites for public institutions responsible for the definition of lists of recommended varieties or the assessment of the value for cultivation and use of newly released cultivars (Abou-El-Fittouh et al., 1969; Lin and Butler, 1988). Also in this case, all GL effects related to the lack of genetic correlation among locations are relevant for the choice of sites. For example, the site classification by cluster analysis reported in Figure 5.4 (A) and its up-scaling reported in Figure 5.7 imply one test site for each of the three subregions when evaluating candidate lucerne varieties for admission to the Italian Register of Varieties (Piano et al., 2001).

6.3 Genetic resources and adaptive traits

Genetic resources

Analysis of adaptation can produce information concerning material (tested cultivars or breeding lines) of particular interest as parent germplasm in virtue of its adaptive response, in the framework of a given adaptation strategy. In general, evidence points to a moderate heritability of adaptation parameters (Becker and Léon, 1988). For wide adaptation, crosses could be envisaged:

• between genotypes possessing high mean yield and the desired adaptive response (as indicated by b ≈ 1 or β ≈ 0 in joint regression analysis; β values near zero in factorial regression analysis; and genotype score on significant PC axes near zero in AMMI analysis); or

• between pairs of genotypes possessing high mean yield and specific adaptation of contrasting type (e.g. 'Orchesienne' and 'La Rocca' in Fig. 5.4 [A]).

Modelling adaptive responses can clearly highlight potential parent material in breeding for specific adaptation that contemplates - in addition to specific selection - specific germplasm produced for each subregion. Another means for identifying genetic resources on the basis of adaptive response is represented by genotype classification or ordination by pattern analysis (which conveniently takes account of both GL effects and mean yield of material - see Section 5.5). Still in the specific adaptation prospect, graphics such as Figure 5.3 can reveal genetic resources that, on the basis of large positive GL interaction effects, are likely to possess traits conferring specific adaptation to a given subregion (e.g. genotypes '18', '10' and '7' relative, respectively, to subregions C, A and D), even if no direct information on these traits is available. Genotype groups as defined by pattern analysis may also be used for this purpose.

The comparison of genetic resources of different origin in terms of adaptive responses can help locate germplasm sources of special interest in breeding for wide or specific adaptation. For example, barley landraces have proved specifically adapted to the drier and more cold-prone area of northern Syria and, therefore, have successfully been adopted as parents in specific breeding for this subregion (Ceccarelli and Grando, 1991; Ceccarelli, 1994). The adaptation of maize germplasm on a global scale clearly reflects the origin (subtropical or temperate) of the parent material (Crossa et al., 1990). The adaptive response of each lucerne landrace across artificial environments in Figure 5.4 (B) is consistent with the characteristics of the area in which it originated (Annicchiarico, 2002), while there is a clear relationship between adaptation pattern and geographic origin of the parent material for Italian lucerne varieties (Annicchiarico, 1992). These examples highlight the fact that not only the selection environment but also the origin of genetic resources has a bearing on the adaptive response of novel germplasm. Therefore, both need to be adjusted in relation to the adaptation target. The effect on the adaptation pattern of different variety types (e.g. hybrid vs. pure line - Peterson et al., 1997 - or single-cross vs. double-cross hybrid - Eberhart and Russell, 1969), may also be worthy of investigation.

Adaptive traits: preliminary indications

As anticipated (Sections 5.3 and 5.4), useful indications on adaptive traits can be obtained by the correlation of morphophysiological traits (if necessary, recorded in only a small number of test locations and averaged across observations for each entry) with the estimated adaptation parameters of genotypes (i.e. the mean yield and the relevant indicator[s] of specific response to locations, such as: b or β values in joint regression, β values in factorial regression, and genotype scores on significant GL interaction PC axes). A prerequisite for correlation analysis is the presence of significant entry variation for the morphophysiological trait in the ANOVA. In particular, characters conferring wide adaptation may be revealed by the presence of correlation with genotype mean yield and the concurrent absence of correlation with the indicator(s) of specific response to locations. On the contrary, the sole correlation with the indicator(s) of specific response suggests that specific adaptation to a given location is favoured by a specific level of the trait, whereas an intermediate trait level is beneficial for wide adaptation. Finally, characters correlated with both mean yield and the indicator(s) of specific response of the genotypes are likely either to be of major interest in a subset of locations or, again, to possess a different optimal level across locations. Whenever there is correlation with indicator(s) of specific response, additional correlation of the trait with genotype mean yield in distinct subregions may contribute to highlight the adaptive traits of local interest. In conclusion, an informative summary of results for putative adaptive traits is represented by a table that reports correlation of morphophysiological traits with

• mean yield over the region;

• the indicator(s) of specific response to locations of the genotypes; and

• genotype mean yield in individual subregions (only for traits correlated with indicator[s] of specific response).

The possible transformation of yield data prior to analysis of adaptation complicates the assessment of adaptive traits. It is recommended to use untransformed data for all estimates of genotype mean yields used for correlation with morphophysiological traits. Correlation results for the indicator(s) of specific response to locations that relates to transformed yield data should be treated with caution and, preferably, verified by correlation for responses assessed on untransformed data.

As an example, the ordination of wheat genotypes on PC 1 in Figure 5.3 is highly correlated with average values of heading date (r = 0.83) and susceptibility to frost (r = -0.76) (Annicchiarico and Perenzin, 1994). This finding, in logical agreement with the positive correlation found between site score on PC 1 and level of cold stress (as average number of frost days) on the site, suggests that:

• late heading (implying escape from cold stress) and intrinsic cold tolerance can be important traits in breeding specifically for the cold-prone subregion C;

• early flowering is preferable in breeding for the relatively warm subregion D; and

• intermediate heading associated with moderate cold tolerance is appropriate in breeding both for subregion B and for wide adaptation to the region.

The absence of correlation between any character and genotype mean yield highlights the importance of specific adaptive traits in this study. Similar investigations have revealed, for example, the crucial importance on adaptive responses of dormancy class for alfalfa in South Africa (Smith and Smith, 1992) and heading time and growth habit for barley in a Mediterranean region (van Oosterom et al., 1993). Other research has revealed the higher transpiration efficiency associated with specific adaptation to drought-prone sites of traditional cultivars relative to recent varieties of barley in Spain (Muñoz et al., 1998).

For qualitative traits, or quantitative traits of which each tested genotype possesses either of two contrasting levels, the relationship of the character with specific adaptation patterns can be:

• inferred visually, by indicating also the plant type of the genotype in ordination diagrams, such as that in Figure 5.3; or

• estimated, by averaging adaptive responses across genotypes belonging to same plant type.

Pertinent examples are provided by Simmonds (1979^[31]) for the comparison of semi-dwarf and tall wheat types by joint regression analysis, and Romagosa et al. (1993) for the comparison of barley lines isogenic for a few traits by AMMI analysis. A more complex procedure that partitions the variation for genotype response to site mean yield into contrasts for different plant types is proposed by Frensham et al. (1998).

Adaptive traits: verification and exploitation

The identification of adaptive traits by the described procedures is not conclusive in general, mainly because an adaptive trait as suggested by correlation results may actually be genetically correlated with the true, unmeasured adaptive trait. While the correlated effect may be exploited when selecting for the false adaptive trait, further experiments limited to a few sites representing different subregions and contemplating a larger set of recorded characters can be planned for verifying or integrating the information on adaptive traits of specific interest in each subregion. Such experiments are even more justified when no data on adaptive traits has been generated by the set of multi-environment trials. The investigation may be extended to molecular markers and other genetic characteristics revealed by future developments of genomics and proteomics. Managed or artificial environments may also be used for the assessment (e.g. Villegas et al., 2000), provided, however, that they can faithfully reproduce the genotype responses in the different subregions. In any case, adaptive traits indicated by low (albeit significant) correlation coefficients may be of limited importance for breeding.

The preliminary screening of novel germplasm for characters that are closely associated with the relevant indicator(s) of specific response to locations may be used for early identification of widely or specifically adapted material, thereby increasing the efficiency of selection for either adaptation target. For a given region or subregion, adaptive traits as indicated by correlations with genotype mean yield may be exploited as indirect selection criteria for yield, thereby complementing or substituting the yield information. This can be of special interest for early stages of selection that do not allow for a precise assessment of yield levels. The formal identification of a useful trait in this context proceeds according to a "black box" strategy, in which the correlated response on yield gains derived from indirect selection is estimated on a germplasm sample (in contrast with the strategy in which the response on yield gains is predicted through the comparison of isogenic lines that differ only in terms of the trait level - Jackson et al., 1996). As a prerequisite for this strategy, the germplasm sample should be sufficiently large (e.g. 25-30 genotypes at least) and representative of the genetic base. In fact, regional trials may produce valuable information in this respect only in some favourable cases (large number of tested genotypes, besides extensive observation of characters), but further experiments performed on a few crucial test sites could be planned for this purpose. For adoption as an indirect selection criterion, a candidate trait should ideally possess:

• high correlation with yield, as well as low GE interaction, across the target region or subregion;

• high broad sense heritability in individual trials (as the ratio of genotypic to genotypic plus experimental error variance components); and

• low costs for its observation.

Therefore, also the reliability of trait assessment (across locations and in single trials) is relevant. Possible curvilinear relationships between mean yield and morphophysiological traits, which cannot be detected by the linear correlation coefficient, should not be overlooked and are discussed later with respect to phenology traits. Finally, the character should also provide partly independent information when combined with other traits. An example of trait assessment integrating all of these aspects for the comparison of different direct or indirect selection scenarios in terms of predicted yield gains is provided by Annicchiarico and Pecetti (1998). In some cases, different architectures of traits may prove about as valuable, complicating the identification of useful characters (e.g. Ceccarelli et al., 1991).

The well-recognized, crucial importance as an adaptive trait of flowering and/or maturity time (Blum, 1988; Wallace et al., 1993a, 1993b) is likely to be confirmed in most investigations on seed, root, tuber or bulb crops. Information on this trait could be exploited in the early stages of selection both in a wide adaptation prospect (e.g. eliminating extremely early or late material) and in a specific prospect (e.g. excluding from further testing the subregion[s] to which the genotype is likely to be poorly adapted on the basis of its phenology). The further exploitation of this information as an indirect selection criterion may require, however, accurate definition of optimal earliness for a given region or subregion. In most cases, when plotting data of individual genotypes averaged across environments of the region or subregion, a curvilinear relation is apparent for yield as a function of flowering or maturity time. The earliness level maximizing the yield response could be estimated by solving the second-degree polynomial equation, and one or a few control varieties possessing optimal earliness could be defined (e.g. Hadjichristodoulou, 1987). In investigation of the potential interest of the phenology trait as a selection criterion, the variable could be expressed as the deviation in absolute value from optimal earliness (e.g. heading displacement in Annicchiarico and Pecetti [1998]), resulting in markedly simpler calculations. In fact, the phenology trait is frequently associated not only with stress tolerance (via the stress escape strategy that it may allow for) but also with the harvest index (via the pleiotropic effect of some of its controlling genes on the partitioning of assimilates between vegetative and reproductive development - Wallace et al., 1993a), with further implications for selection because of the well-documented importance that increased harvest index has had on the genetic improvement of seed yield potential in several crops (Donald and Hamblin, 1983). The effectiveness of selection for flowering time may be increased by highlighting the effect of its component traits (i.e. the responses to photoperiod and temperatures) in the germplasm sample evaluated in the multi-environment trials. This can be done, for example, using the freely available computer program RoDMoD (Summerfield et al., 1996). Mean yield may show a curvilinear relationship also with other morphophysiological traits (e.g. plant height - Flintham et al., 1997).

Selection for yield exclusively on the basis of useful morphophysiological traits, also termed "analytical" or "ideotype" breeding, has long since (Donald, 1968) been proposed as an alternative to "synthetical" or "empirical" breeding based on direct selection for yield, but its adoption in breeding programmes has remained limited for various reasons (Rasmusson, 1987; Loss and Siddique, 1994). Nevertheless, adaptive traits contribute to germplasm selection in most programmes, and their contribution is expected to increase (Jackson et al., 1996). The identification of these traits can also be beneficial for thoroughly exploiting the opportunities for indirect selection that may be offered by molecular markers, to which useful morphophysiological traits could be linked (Fischer, 1996). Investigating adaptive traits may be particularly useful in breeding for stressful subregions or extensive cropping systems, in which the "universal" ideotype considered of interest for seed crops by Donald and Hamblin (1983) - characterized by traits, such as a dwarf stature, an erect and determinate growth habit, few, small and erect leaves, and few or no tillers or branches - may contrast with the features of locally adapted material and prove less agronomically valid. This is suggested, for example, by Ceccarelli (1994) and by various reports in Eyzaguirre and Iwanaga (1996).

Information on crucial adaptive traits produced from multi-environment trials can also contribute to the development and the validation of crop simulation models incorporating the effect of these traits. These models can provide a powerful means for comparing different plant architectures on the basis of simulated yield responses across a target region (Hoogenboom et al., 1997; Chapman et al., 2002), further contributing to the definition of traits for selection in a wide or specific adaptation strategy.

When yield stability proves a useful breeding objective for a given region or subregion, correlation analysis (limited to data of test environments in the region or subregion) may be used for identifying morphophysiological traits associated with higher yield stability in the germplasm sample. Estimates of stability measures expressed as variances (e.g. environmental variance and Type 4 stability measures) should be used as square root values in the analysis, to approximate their distribution to the normal one. Another option is the use of yield reliability values (combining mean yield and yield stability into a unique measure of genotype merit - see Section 7.2) for correlation analysis.