Introduction
Organic farming often leads to specific environmental conditions, which are more stressful for the plants since the use of chemical inputs is not permitted 1 . While modern varieties fit the needs of conventional farming in industrialized countries (homogeneity, adaptation to mechanization and high input management), they often are not adapted to agronomic practices that decrease the use of inputs and fossil energy. In addition, dimensions of integrated systems such as the need for biomass for animal feed, competitive ability with weeds, efficient uptake and utilization of nitrogen and end-use quality for traditional foods are often not taken into account in conventional breeding programsReference Wolfe, Baresel, Desclaux, Goldringer, Hoad, Kovacs, Loschenberger, Miedaner, Østergard and Lammerts van Bueren 2 – Reference Dawson, Murphy and Jones 5 . This often leads farmers in North Africa, Latin America or Asia to cultivate landraces or historic varieties instead of modern cultivarsReference Ceccarelli and Grando 6 , Reference Smith, Castillo and Gomez 7 , since landraces may be adapted to heterogeneous environments and specific objectivesReference Jarvis, Hodgkin, Sthapit, Fadda and Lopez-Noriega 8 .
To succeed in developing new varieties adapted to these kinds of heterogeneous environments, participatory plant breeding (PPB) has been implemented in several casesReference Ceccarelli, Grando, Bailey, Amri, El-Felah, Nassif, Rezgui and Yahyaoui 9 , Reference Patto, Moreira, Almeida, Satovic and Pego 10 . PPB aims at developing varieties adapted to farmers’ needs in contrasted low input environments (such as organic management in Europe), while maintaining genetic diversity. It is based on (i) accounting for genotype×environment×management interactions through decentralized selection; (ii) collaboration between researchers, breeders, farmers and other stakeholders; and (iii) the development and use of appropriate genetic diversity for breeding.
Efficient breeding in stressful environmental conditions will require that environmental complexity is taken into accountReference Desclaux, Nolot, Chiffoleau, Leclerc and Gozé 11 . Decentralized selection in many sites is needed in order to conduct direct selection in the target environmentReference Ceccarelli, Grando, Bailey, Amri, El-Felah, Nassif, Rezgui and Yahyaoui 9 , which has proved more efficient than indirect selection from favorable to stressful environmentsReference Wolfe, Baresel, Desclaux, Goldringer, Hoad, Kovacs, Loschenberger, Miedaner, Østergard and Lammerts van Bueren 2 . In addition, participation of farmers is required to benefit from their experience and expertise in varietal evaluation in their particular environmentReference Ceccarelli, Grando and Carena 12 , and to implement selection in a way that will respond to their specific needs. This participation also empowers farmers and leads to more autonomy with respect to varietal choices and the promotion of farmers’ rightsReference Morris and Bellon 3 , Reference Halewood, Deupmann, Sthapit, Vernoy and Ceccarelli 13 .This approach has already proved to be efficient in developing countriesReference Smith, Castillo and Gomez 7 , Reference Ceccarelli, Grando and Carena 12 , Reference Vom Brocke, Trouche, Weltzien, Barro-Kondombo, Gozé and Chantereau 14 – Reference Bachmann, Cruzada and Wright 18 but has started only recently in Europe, with new programs showing promising resultsReference Patto, Moreira, Almeida, Satovic and Pego 10 , Reference Osman, Almekinders, Struik and van Bueren 19 – Reference Dawson, Rivière, Berthellot, Mercier, De Kochko, Galic, Pin, Serpolay, Thomas, Giuliano and Goldringer 23 .
In France, the demand of organic farmers for adapted varieties first resulted in the cultivation of landraces and historic varietiesReference Thomas, Dawson, Goldringer and Bonneuil 24 . In 2003, a group of organic farmers who wanted to conserve agricultural biodiversity and enhance their seed autonomy founded the association Réseau Semences Paysannes (RSP, the Farmers’ seed network). The RSP is a network of farmers’ associations that conserve, multiply and exchange landraces, old varieties and other farmers’ varieties 25 . On-farm management has been shown to be an effective method to conserve agricultural biodiversity, complementary to in situ conservation on research stations and ex situ conservation in gene banksReference Enjalbert, Dawson, Paillard, Rhoné, Rousselle, Thomas and Goldringer 26 , Reference Thomas, Demeulenaere, Dawson, Khan, Galic, Jouanne-Pin, Remoue, Bonneuil and Goldringer 27 . On-farm management is a key activity for the conservation of genetic resources, as underlined by the 2004 International Treaty on Plant Genetic Resources for Food and Agriculture (ITPGRFA) under the control of the FAO. Nevertheless, an in situ conservation system solely based on landraces and old varieties may not fit all farmers’ needs. Farmers from the RSP have become interested in the development of new varieties that both conserve crop biodiversity and are adapted to the current organic farming practices 25 . Because in the context of industrialized agriculture and an institutionalized seed supply system, the classical inter-generational transmission of knowledge has disappeared, PPB development presents unique challenges both at the technical (genetic, agronomic and analytical) and at the organizational levelReference Thomas, Dawson, Goldringer and Bonneuil 24 , Reference Bonneuil and Demeulenaere 28 .
Several strategies can be applied to start a new breeding scheme, such as crossing, mixing or selecting in the available landraces or locally adapted populationsReference Goldringer, Enjalbert, David, Paillard, Pham, Brabant, Cooper, Spillane and Hodgkin 29 . Here, we chose to create new populations through manual crosses, thus generating a broad range of new allelic combinations. The farmer–baker who initiated the project, Jean-François Berthellot (J.F.B.), chose the parents according to their baking and milling quality, their history and geographical area of cultivation and their agronomic behavior, based on the knowledge he has acquired by growing them on his farm for the past 5 years, and through collaborations with other farmers from the RSP. The goal of making crosses was to combine the bread-making quality and agronomic resilience of landraces with historic varieties (first half of the 20th century) and a few more recent varieties, which are more resistant to lodging. Before the populations derived from these crosses were distributed to a large network of farms, mainly in France, an experiment was performed to assess the impact of farmer's mass selection in an early generation.
From conversations with farmers, it is clear that while they are looking for certain characteristics in their varieties, they are also looking to maintain more phenotypic diversity within varieties than is normally present in modern varieties. They often mention that one of the benefits of more heterogeneous population-varieties is their increased stability over years, due to the within-variety heterogeneity that buffers environmental fluctuations. Within-field varietal diversity has been shown to increase the functionality, resilience and stability of agricultural ecosystemsReference Newton, Akar, Baresel, Bebeli, Bettencourt, Bladenopoulos, Czembor, Fasoula, Katsiotis, Koutis, Koutsika-Sotiriou, Kovacs, Larsson, Pinheiro de Carvalho, Rubiales, Russell, dos Santos and Vaz Patto 30 , Reference Hajjar, Jarvis and Gemmill-Herren 31 . Genetically diverse varieties may combine quantitative and qualitative resistance, thus providing more durable disease resistanceReference Wolfe 32 – Reference Zhu, Chen, Fan, Wang, Li, Chen, Fan, Yang, Hu, Leung, Mew, Teng, Wang and Mundt 34 . Phenotypic variability within varieties has also been found to be associated with an increase in associated biodiversityReference Chateil, Goldringer, Tarallo, Kerbiriou, Le Viol, Ponge, Salmon, Gachet and Porcher 35 , Reference Tooker and Frank 36 . Within-field varietal diversity may not only contribute to yield stability but also to the stability of quality, as shown in wheat where varietal mixtures increase uptake efficiency of nitrogenReference Mille, Fraj, Monod and De Vallavieille-Pope 37 . Finally, the on-farm management of such diverse varieties contributes to the in situ conservation of genetic resources for plant breedingReference Jarvis, Hodgkin, Sthapit, Fadda and Lopez-Noriega 8 , Reference Patto, Moreira, Almeida, Satovic and Pego 10 , Reference Enjalbert, Dawson, Paillard, Rhoné, Rousselle, Thomas and Goldringer 26 , Reference Louette and Smale 38 – Reference Pujol, David and McKey 40 .
While for an autogamous plant like wheat, most of the selection is often done among varieties that are nearly pure lines, at a more advanced generation, and not within segregating populationsReference Ceccarelli and Grando 6 , Reference Ceccarelli, Grando, Bailey, Amri, El-Felah, Nassif, Rezgui and Yahyaoui 9 , Reference Joshi and Witcombe 41 , Reference Ceccarelli, Grando, Tutwiler, Baha, Martini, Salahieh, Goodchild and Michael 42 , this study assessed the diversity created in the program and the response to farmers’ mass selection in early generations after crossing (F2) in terms of trait means and genetic variance within and among families. Collaboration with farmers (participation) at all stages of the research study was critical to reach these objectives.
Materials and Methods
Context and experimental design
On the initiative of J.F.B., a farmer active in the RSP, a PPB project was started in 2005 with researchers from INRA Le Moulon. The project was extensively discussed with all farmers, RSP coordinators and researchers. Ninety crosses were made on J.F.B.'s farm between different historic wheat varieties, landraces and modern varieties created for organic agriculture. These landraces, historic varieties and modern varieties had been cultivated on his farm for at least 5 years. Most of these populations came from the national seed banks at INRA Clermont Ferrand, France; and from Switzerland (varieties created by Peter Kuntz) and Germany (varieties created by Bertold Heyden).
The first (F1) and second (F2) generations of progeny of the 90 different families (one family is derived from each cross, numbered 1–90) have been grown on his farm in 2006–2007 and 2007–2008. Selections were made of individual spikes in a sub-set of the F2 families. Seed of these spikes was bulked for each family and 35 of these families were evaluated in the F3 generation with their corresponding unselected bulk at INRA Le Moulon (Gif sur Yvette, France) in 2008–2009. Three families had two different selected versions, so the total number of selected populations was 38. When two selections were made within a family, this was indicated by a letter following the number (14a and 14b, 34a and 34b, 42a and 42b). There were three complete block replicates of the 35 families, with paired rows of each version (selected or bulk) for each family. Rows were 1.20 m long with 20 seeds sown per plot. Paired versions within families were randomized in each block but the selected and the bulk families of each pair were always grown side by side. This maximized the power to detect differences between the two versions. In addition, Renan, a pure line cultivar frequently used in organic agriculture in France was used as a check variety and as a point of comparison for the farmers. Renan was included twice in each replicate of the experiment.
Measurements
Qualitative observations and quantitative measures were taken on the main tiller of each of five plants for each version of each family in each replicate (i.e., a total of 15 plants for each version of each family). The traits measured and their abbreviations are listed in Table 1. These traits were chosen by farmers and also based on phenotypic descriptors used for variety registration. After field measurements were made on each plant, the spike was cut and individually bagged. At the technical facilities at INRA Le Moulon, measurements were taken on all spikes collected. Grain from each replication of each version within families was analyzed for technology traits at INRA Clermont Ferrand using near infrared spectroscopy (Foss NIRSystem 6500), using whole grain. Although this prediction is less precise than on whole grains flour, it was chosen because it is not destructive and this allowed us to replant the seeds. The correlation between estimated and true value using this method in the calibration sample was (Table 1): protein (0.86), hardness (0.77), test weight (TW; 0.80), mixing time integral (MTI; 0.7) and dough strength (W; 0.75). This was assessed on modern cultivars (G. Branlard, personal communication).
Traits measured. The traits analyzed are in bold.
If the end of the spike was broken, the existing spike length was used for calculations of spike density, and the SL itself was treated as missing data for the analysis. If only the last spikelet was missing, 5 mm was added to the spike length measurement.
Statistical analysis
First, we tested for an overall version effect using an ANOVA model with all effects fixed:
$$Y_{ijkl} = {\rm \mu} + {\rm family}_i + {\rm rep}_j + {\rm version}_k + {\rm \varepsilon} _{ijkl} $$
Where Y ijkl , is the phenotypic value measured for plant l of version k of family i in replication j, μ is the general mean, family i is the effect of family i, rep j is the effect of repetition j, version k is the effect of the version (k=selected or non-selected) and ε ijkl is the random error term.
Model (2) was used to test for specific selection effect dependent on the family. This was similar to Model 1 but the version effect (selected or non-selected) was nested within family (version(family) ik ).
The statistical analysis was implemented in SAS v 9.2 proc GLM 43 . Two-sided tests between the least-square (LS) means for selected and non-selected versions within each family were made with Tukey's multiple comparison procedure in SAS. This was done using the SLICE function on the LS Means for version (family), which tests for an effect of version within each family. To visualize the response to selection for multiple traits and families, the log(p-value) of the two-sided test described above were recorded in a matrix of traits×families (18×38). We used a log transformation in order to weigh the significant changes. When the means of the selected versions were higher than the bulk, then log(p-value) were multiplied by −1 (so that the values were positive), else by +1 (so that the values remained negative). Ward's clustering procedure was used on the data in this matrix to group each version of the different families by the similarities in their responses to selection. This clustering and the resulting dendrogram and heatmap visualization were done with the heatmap function in R 44 .
Repeatability is a measure of the proportion of phenotypic variation due to genetic causes. It was estimated using the genetic and residual variance estimated for each version (selected or bulk) from the ANOVA mixed model:
$${Y_{ijk} = {\rm \mu} + {\rm family}_i + {\rm rep}_j + R_{ijk}} $$
Where Y ijkl is the phenotypic value measured for plant k of family i in repetition j, μ is the general mean, family i is the family i (random effect), rep j is the repetition j (fixed effect) and R ijk is the error. Mixed models were used to estimate variance components using the function VarCorr in the R package lme4 44 . Thus the repeatability was estimated as:
$${r = \displaystyle{{{\mathop{\rm var}} (G)} \over {{\mathop{\rm var}} (G) + \displaystyle{{{\mathop{\rm var}} (R)} \over 3}}}}$$
Where var(G) was the estimated among-family genetic variance, and var(R) the estimated residual variance. As the families are progenies of F2 spikes, the error R ijk of the model includes both the within-family genetic variance plus the environmental variance.
We estimated the average within-family genetic variance as the difference between the error variance (var(R) from model (3)) and the residual variance of the check variety Renan. As the check is a pure line, i.e., genetically homogeneous, the variation observed was only due to the environment. More information on the calculation of within-family variance is provided in supplemental material. For the data where there was no individual data but only plot mean values, only the replication effect was used in the model.
Correlations among traits before and after selection were also calculated, methods and results are given in the supplementary information.
Results
Trait variation
Plant height (PH) was rather high in both the selected and bulk versions with a mean of 134.5 and 132.7 cm, respectively. This was also the case for last-leaf-to-spike distance (LLSD) with means of 28.6 and 27.7 cm, respectively, and for spike length (SL) with means of 12.8 and 12.7 cm, respectively (Table 2). Such high values are typical of landraces and historic varieties. In general, the range of variation of the trait values was quite large, indicating the high level of diversity generated in these populations.
Summary for each trait for the data for selected version.
Summary for each trait for the data for bulk version.
In the models (1) and (2), the family effect was significant for all traits (Tables 3a and 3b), while the effect of replication was significant only for protein, W and MTI, which might be due to the fact that the grain composition traits are highly dependent on the environmentReference Bassett, Allan and Rubenthaler 45 , Reference Herndl, White, Graeff and Claupein 46 .
Results of the ANOVA (model 1) with DF(family)=37, DF(version)=1, DF(rep)=2.
F values are presented. Stars indicate the significance of the F test; *: 0.05>P-value>0.01; **: 0.01>P-value>0.001; ***: P-value<0.001.
Results of the ANOVA (model 2) with DF(family)=37, DF(version in family)=38, DF(rep)=2.
F values are presented. Stars indicate the significance of the F test; *: 0.05>P-value>0.01; **: 0.01>P-value>0.001; ***: P-value<0.001.
Response to selection for trait means
Seven traits changed significantly after selection with a common trend over all families among which thousand kernal weight (TKW) and protein showed highly significant changes (respectively positive and negative). Six other traits changed significantly but the direction of the response depended on the family (Fig. 1). For 24 families out of 38, significant differences between bulk and selected versions were found for at least one trait.
Evolution between selected (grey) and bulk (black) versions for four traits: (a) grain weight per spike (GW_Spike) in grams, (b) plant height (PH) in cm, (c) thousand kernel weight (TKW) in grams and (d) protein concentration (Protein) in %. Stars represent significant differences between the means: *, 0.05<P-value <0.01; **, 0.01<P-value<0.001; ***, P-value<0.001. tem=control (Renan).
The overall version effect was significant at P<0.01 for TKW and protein (Table 3a), and significant at P<0.05 for PH, earliness, GW_Spike, TW, W and MTI. This indicated a unidirectional response to selection for these traits: although the response was not always significant for the different families, we found a trend towards an increase for TKW, GW_Spike and PH and towards a decrease for protein (Fig. 1). The effect of the version within family was significant for seven traits out of 16: PH, LLSD, SL, total number of spikelets (SpTot), TKW, sterility and earliness (Table 3b) indicating that for LLSD, SL, SpTot and sterility the response was dependent on the family.
While there was a significant global version effect for PH (Table 3a), the magnitude of the selection response was specific to each family (Table 3b). PH significantly increased from 120 to 150 cm for family 64; from 118 to 140 cm for family 80; from 70 to 110 cm for family 34a, and from 100 to 115 cm for family 34b, whereas PH decreased significantly from 130 to 125 cm for family 24 (Fig. 1b).
As can be seen in Figs 1a, 1c and 2, selection for GW_Spike, KN_Spikelet, TKW, earliness and TW always increased trait values when response was significant. Moreover, TKW increased for 26 families over the 38, of which four cases were significant (Fig. 1c). In contrast, hardness, W and MTI selection always decreased trait values when the response was significant. Finally, PH, LLSD, sterility, awns, SpTot, color, density, protein, and SL changed in both directions. It is interesting to point out that protein tended to decrease after selection for 30 families over the 38 selected, but only two families decreased significantly. For one family (family 60), protein increased significantly after selection. There were 14 families that did not respond to selection for any of the traits measured, 11 families that responded to selection for only one trait, six families that responded to selection for two traits, six families for three to eight traits and one family for 12 traits.
Change in the phenotypic mean between bulk and selected versions of F3 families for several traits. If the mean decreases: -, 0.05>P-value>0.01; - -, 0.01>P-value>0.001; - - -, P-value<0.001. If the mean increases: +, 0.05>P-value>0.01; ++, 0.01>P-value>0.001; +++, P-value<0.001. Earliness=days to flowering, so an increase means that a family flowers later. Color that increases means darker spike. Awns that increases means more awns.
For family 34, the two selections (‘a’ and ‘b’), led to different responses: eight significant changes for ‘b’ and two significant changes for ‘a’ that were in the same direction as for ‘b’ (for PH and LLSD). For family 42, the two selections (‘a’ and ‘b’), led to similar response patterns. For family 14, neither of the selections led to many changes from the bulk version. The dendrogram shows that variables linked to technological properties (Protein, MTI, W) had similar response patterns (Fig. 2). LLSD and PH also responded similarly to each other, as did TKW and TW.
Response to selection at the variance level
Repeatability ranged from 0.13 for GW_Spike within the bulk versions to 0.81 for PH within the bulk versions, with higher values for morphological traits such as PH, LLSD, SL, earliness and SpTot and for grain composition traits (Table 4). This indicated good control of environmental variation and the ability of the experimental design to discriminate among families both before and after selection. Note that the lowest value was for the trait that is the closest to grain yield (GW_Spike).
Differences in repeatability, genetic variance among F3 and genetic variance within F3 for each version: bulk and selected for each trait;
$\sigma = \sqrt {(variance)} $
.
GW_spike, TKW, protein and MTI had greater repeatability in the selected version compared to the bulk (increases from +32.2 to +139.4%). Repeatability for SL, TW, W, PH, KN_Spike, SpTot, earliness and density did not change much in the selected version (changes from −4.8 to +10.5%). Repeatability for hardness, sterility, KN_Spikelet and LLSD was lower in the selected version (decreases from −9.2 to −28.3%) (Table 4).
Selection increased the among-family genetic variance for GW_Spike, TKW and MTI (increases from +25.4 to +66.0%) while it was reduced in a more limited proportion for LLSD, KN_Spikelet, earliness and hardness (decreases from −11.4 to −22.7%). Little change was observed for the other traits (Table 4). Selection increased the average within-family genetic variance for sterility (+32.7%) while a marked decrease was observed for GW_Spike, TKW, protein, TW and MTI (between −44.7 and −13.3%) (Table 4).
Discussion
In this study, we analyzed F2 derived F3 families from 35 crosses among a wide range of landraces, historic and more recent varieties and selected by an organic farmer. First we comment on the diversity created by the crosses. Then, we discuss the response to farmer's selection within early generation families (mean and variance). We provide specific examples of directional selection, correlations among selected traits, and the influence of parental varieties on the response to selection in the supplementary information. Finally, we discuss how these results can be integrated in the ongoing PPB program.
Creation of diversity for selection
In conventional selection programs, breeders usually seek to decrease the within-family genetic variance in order to obtain uniform lines. However, in this project, one of the goals was to maintain within-family variance and thus maintain the genetic potential for continuing on-farm selection. Overall, after selection, this variation remained high. This is positive as it will allow the farmers to continue selecting within populations.
Both before and after selection, differences among families were highly significant for all traits (Tables 3a and 3b). This was consistent with the objective of the crosses to generate a large range of diversity in order to increase the chance of developing populations that might adapt to contrasting environmental conditions. Traits where uniformity is of importance for standardized production may also not be as critical in situations where production and value-added processing occurs on-farm or for a small artisanal market. For the specific example of baking force W, see supplementary information.
Because this experiment was done in a common garden with all populations grown in a single environment, the observed phenotypic diversity is likely due to the genetic diversity found in our panel of populations. The populations are in the F3 generation derived from crosses among very diverse parents. Assessing the relative contribution of among- and within-family variability is of importance to develop appropriate selection procedures in PPB.
In general, genetic among-family variation was high for most traits (Table 4), leading to high repeatability, except for some characteristics of the spike such as KN_spike and GW_spike. This is consistent with classical findings in quantitative genetics for such complex traits. When selecting, selection among families can be the first stepReference Ceccarelli, Grando, Bailey, Amri, El-Felah, Nassif, Rezgui and Yahyaoui 9 , Reference McElhinny, Peralta, Mazón, Danial, Thiele and Lindhout 17 , Reference Ghaouti, Vogt-Kaute and Link 21 , Reference Ceccarelli, Grando, Tutwiler, Baha, Martini, Salahieh, Goodchild and Michael 42 . With diverse and distinguishable families, farmers can find populations that better suit their specific environments and practices.
The within-family genetic variance was also quite high for many traits (Table 4). This is expected as we are dealing with segregating populations. After selection among families, selection within families can be the second step in PPB. Within-family diversity can be used in an evolutionary plant breeding approachReference Murphy, Lammer, Lyon, Carter and Jones 47 where populations are mainly submitted to natural selection. To increase selection efficiency, farmers may apply mass selection to guide the evolution of the population towards phenotypes of interest for themReference Smith, Castillo and Gomez 7 , Reference Dawson, Rivière, Berthellot, Mercier, De Kochko, Galic, Pin, Serpolay, Thomas, Giuliano and Goldringer 23 , Reference Joshi and Witcombe 41 , Reference Gyawali, Sunwar, Subedi, Tripathi, Joshi and Witcombe 48 , Reference Virk, Chakraborty, Ghosh, Prasad and Witcombe 49 . Researchers and farmers conducting selection have to be careful to monitor competition among individuals within the population (for light and uptake of nutrients in the soil) when using either evolutionary breeding or mass selection, as competition ability might be negatively correlated with some traits desired by farmers (quality and yield)Reference Wolfe, Baresel, Desclaux, Goldringer, Hoad, Kovacs, Loschenberger, Miedaner, Østergard and Lammerts van Bueren 2 , Reference Goldringer, Enjalbert, David, Paillard, Pham, Brabant, Cooper, Spillane and Hodgkin 29 .
Assessing the impact of farmers' mass selection
A significant positive response was found for morphological and phenological traits such as PH, LLSD, earliness or SL. Within the RSP, farmers are often looking for plants that are tall but resistant to lodging because they have noticed a positive relationship between PH and the maintenance of grain filling under stress which impacts yield and grain quality. From farmers' observations, taller plants also are more competitive with weeds, and farmers can use the extra straw for livestock or for soil fertility management. The measure of LLSD was proposed by the farmers, because they found that a greater distance between the spikes and the foliage may prevent leaf diseases from jumping to spikes. They also observed an improvement in grain filling and grain maturation with longer LLSD if leaves die because of disease or abiotic stress, as the stem resources can be used to continue grain filling. This is corroborated in the scientific literatureReference Blum 50 .
Traits such as TKW or GW_Spike, which are more difficult to assess visually in the field, also increased after selection in families where there were significant differences. Family 80 was very heterogeneous and responded to 12 traits out of 18 measured. This can be related to the parents used (see supplementary information). Ceccarelli et al.Reference Ceccarelli, Grando, Tutwiler, Baha, Martini, Salahieh, Goodchild and Michael 42 also showed the ability of farmers to select superior populations when working with early generations.
In discussions among farmers, they said that they may have particular selection criteria but always adapt to the specific populations they observe and take a more holistic approach to selection. As already noticed by Ghaouti et al.Reference Ghaouti, Vogt-Kaute and Link 21 , farmer selection is integrative, it does not favor individual traits but instead overall plant vigor, field productivity and quality. In our study, traits such as PH, SL, SpTot, density and color are indicators of such vigor for the farmer, leading to homogenization and uniform selection. However, if there are plants that look interesting but do not fit the ‘type’, the farmers will select them anyway, leading to more heterogeneous samples.
Selection in the F2 may be early for an efficient response to selection as segregation is not complete (i.e., (1/2)2 heterozygotes expected), and the covariance between F2 plants and F3 plants from the same family is expected to be moderate. Moreover, as there is a high level of heterozygosity, differences are difficult to assess. Most of the selection in PPB programs described in the literature are carried out within more advanced generations, for example F3, F4, F5 in barley and beanReference Ceccarelli and Grando 6 , Reference Almekinders, Centeno, Torrez, Olivera, Suarez and Carrasco 15 , Reference Ceccarelli, Grando, Tutwiler, Baha, Martini, Salahieh, Goodchild and Michael 42 ; F5 for riceReference Joshi and Witcombe 41 ; and F3 and F4 in sorghumReference Vom Brocke, Trouche, Weltzien, Barro-Kondombo, Gozé and Chantereau 14 . In this study, the farmer's objective was to save time and subject the populations to the conditions of the target environment as soon as possible. An original aspect of this experiment is that all crosses were performed on-farm and the early generations were also cultivated on the same farm by the farmer who initiated the project. In most reported cases of PPB, crosses were made in the research station and the early generations were also grown at the research stationReference Ceccarelli, Grando, Bailey, Amri, El-Felah, Nassif, Rezgui and Yahyaoui 9 , Reference Vom Brocke, Trouche, Weltzien, Barro-Kondombo, Gozé and Chantereau 14 , Reference Almekinders, Centeno, Torrez, Olivera, Suarez and Carrasco 15 , Reference Ceccarelli, Grando, Tutwiler, Baha, Martini, Salahieh, Goodchild and Michael 42 . Here, the farmer applied mass selection without being influenced by the researchers. This approach differs from other programs where researchers train farmersReference Gyawali, Sunwar, Subedi, Tripathi, Joshi and Witcombe 48 or apply selection before the farmersReference Vom Brocke, Trouche, Weltzien, Barro-Kondombo, Gozé and Chantereau 14 , Reference Virk, Chakraborty, Ghosh, Prasad and Witcombe 49 . The objective here was to assess specifically the effect of the farmer's selection, which is based on his unique knowledge of his farming system. The efficiency of farmers’ selection has been demonstrated in other cases on rice, barley or quinoa but it was most often a screening among families at a later generationReference Ceccarelli and Grando 6 , Reference McElhinny, Peralta, Mazón, Danial, Thiele and Lindhout 17 , Reference Joshi and Witcombe 41 .
The farmer whose selections were studied is not representative of all farmers in the RSP. He has extensive experience growing and observing population-varieties and may be described as an ‘expert farmer’. The objective of the study was to characterize the response to mass selection by an expert farmer in order to assess the potential of, and limits to, this approach for motivated farmers involved in PPB programs. These results show that mass selection within families can be effective for some traits even at early generations, and that genetic diversity is maintained within families for future selection. These results will contribute to improve farmers' understanding of the impact of their selection and thus may affect their future selection.
Conclusion
The objective of the farmer in this study was to select improved populations, based on farmers' criteria, while maintaining the potential for future selection within populations. Based on our results this goal seems to have been achieved. This study has helped lay out the basis for the implementation of a PPB program by creating new populations with broad diversity which can then be distributed for selection according to farmers' criteria. The understanding and the analysis of the results were possible because of the interaction between farmers and the research team. This collaboration is the basis of the program and has led to a better knowledge of farmer variety management and its impact on genetic diversity.
Selection on-farm is new for farmers in FranceReference Bonneuil and Demeulenaere 28 . More time and exchanges of knowledge are needed for farmers to regain the knowledge and skills of selectionReference Storup and De Kochko 51 . These newly created wheat families were sent to farmers all over France and are now managed by around 25 farmers under different environments and practices. These farmers are collaborating in the PPB program which is the basis of the methodology. Several farmers have started mass selection within the populations. The next step will be to study the evolution of these newly created families under farmers' selection and management practices as well as evolutionary pressure in diverse environments in terms of molecular and phenotypic diversity and on-farm agronomic and quality traits.
Acknowledgements
We thank Jérôme Enjalbert, Mathieu Thomas for useful discussion. We thank Gerard Branlard from INRA Clermont Ferrand for his time when doing NIRS analysis and his help with understanding quality measurements. P. Rivière was funded by a grant from DIM ASTREA (Ile de France region). J. Dawson was funded by an INRA postdoctoral research fellowship and during 5 months by a postdoctoral fellowship from the European Community's Seventh Framework Programme (FP7/2007–2013) under the Grant Agreement n245058-Solibam (Strategies for Organic and Low-input Integrated Breeding and Management). The research leading to these results has received funding from the above mentioned FP7 Solibam programme and from the Specific Targeted Research Project of the European Commission 6th Framework Program Priority 8.1 SSP: Opportunities for farm seed conservation, breeding and production Proposal/Contract no.: SSP-CT-2006-04434.
Supplementary material
To view supplementary material for this article, please visit http://dx.doi.org/S1742170513000343.
