Breeding the hyperaccumulator Noccaea caerulescens for trace metal phytoextraction: first results of a pure-line selection

Abstract To initiate the creation of phytoextraction cultivars, plants were selected from 60 populations of N. caerulescens for their high shoot biomass or Cd, Ni, and Zn concentrations. They were self-pollinated, and the selection and fixation were continued for three generations in greenhouse conditions. Selected plants showed a potential to produce 5–10 t dry matter ha−1, which is required to decontaminate soils which have been moderately contaminated with Cd. However, the high biomass genotypes could not be fixed, probably both because of their complexity and to the sensitivity of this trait to environmental conditions, and plant density in particular. The selection led to an improvement to the Cd and Zn accumulation capacities of the plants, yet caused a decrease in their Ni accumulation. This is most likely due to a decline in Ni availability in soil, rather than to a deleterious effect of inbreeding. Metal accumulation appeared to be more heritable than biomass production and fixation for the former trait should be quicker than for the latter. The accumulation capacities of the selected plants permitted offtakes representing around 25% of the soil Cd in a single cropping. This potential has to be confirmed in field conditions.


Introduction
Phytoextraction still appears to be the most relevant solution for decontaminating vast areas of cultivated soils contaminated with Cd, one of the most preoccupying contaminants of the food chain (Clemens et al. 2013). Among those species which could be used for Cd phytoextraction, the hyperaccumulator Noccaea caerulescens has the highest potential in a temperate climate (Sterckeman and Puschenreiter 2018). However, Cd phytoextraction with this species has yet to reach commercial development although many field trials have been carried out over the last 25 years (Baker et al. 1991;Brown et al. 1995;Hammer and Keller 2003;McGrath et al. 2006). The main reason is that the plant does not produce enough harvestable biomass. In the field experiments carried out with wild genotypes, many crops produced less than 1 t dry matter (DM) ha À1 , whereas a yield of 2 t DM ha À1 could be considered good and a 4 t DM ha À1 one, exceptionally high (Sterckeman and Puschenreiter 2018). Koopmans et al. (2008) estimated that a yield of 5 t DM ha À1 would be necessary to decontaminate Cd-contaminated soils (up to 10 mg Cd kg À1 ) within a reasonable time, whereas Zhao et al. (2003) calculated that a yield of 10 t DM ha À1 was advisable.
In the Woburn experiments, McGrath et al. (2000) observed yields of 7.45 t DM ha À1 and 7.80 t DM ha À1 with N. caerulescens from Whitesike and Prayon, respectively, planted at a density of 90 plants m À2 . The largest individual plants they found weighed 28 g DM (Prayon) and 31 g DM (Whitesike). At a density of 35 plants m À2 , such plants would produce around 10 t DM ha À1 . These results suggest a potential to ameliorate the biomass production of the species, for instance through pure-line selection (Gallais 2011), unless this leads to inbreeding depression. Inbreeding N. caerulescens should be possible as it has a mixed mating system, with selfing rates ranging from 0.2 to 0.5, according to Mousset et al. (2016), if not of 0.95 as observed by Riley (1956).
Until now, less has been done in relation to breeding N. caerulescens. In a U.S. patent, Li et al. (2006) claimed the invention of the Cd and Zn phytoextraction process with N. caerulescens, together with a genotype of this species, G15, considered as a subspecies and deposited under an ATCC (American Type Culture Collection) Accession Number. Genotype G15 seems to have been selected from the comparison, in 1997, of 21 genotypes of N. caerulescens cultivated on the Palmerton site (USA), contaminated by a zinc smelter (Comis 1995). G15 clearly had a higher Cd content and Cd to Zn ratio than the other genotypes, suggesting that it could derive from a calamine population from France, such as the Ganges population. However, nothing was published about the way it was obtained nor about its dry mass production or metal offtake. Apparently, G15 has never been commercialized.
Numerous N. caerulescens populations are located in Western Europe and particularly in France, where hyperaccumulating Cd genotypes were discovered (Reeves et al. 2001). Recently, we investigated the demographic history of the species in this area . This study enabled the location of new populations and the collection of seeds from numerous populations belonging to the nonmetallicolous (NM), calamine (CAL), and serpentine (SER) edaphic groups. Calamine designates any substrate highly enriched in Pb, Zn, and Cd (generally from mining or smelting areas), whereas serpentine refers to the minerals contained in the ultramafic rock on which the soil is developed. This seed collection was used to phenotype 60 different accessions of N. caerulescens in order to describe the phenotypical variability of the species and to select outstanding individuals with traits favorable to Cd, Ni, or Zn phytoextraction, i.e. high shoot biomass and/or high contents of one or two metals . Over three generations, the selected individuals were self-pollinated in order to initiate a pure-line selection. The phenotyping of the 60 initial populations showed the high natural variability of N. caerulescens' traits and its metal accumulation potential, which have been presented elsewhere . In this article, we present the evolution of growth and phenological traits, as well as of metal accumulation during the first 3 years of selection and fixation through inbreeding.

Initial plant material
The initial seeds were sampled from 60 populations, mostly located in the Western distribution zone of N. caerulescens, i.e. in Belgium, Finland, France, Germany, Luxembourg, Italy, Norway, Slovakia, Spain, Sweden, and Switzerland. There were also two populations from Slovakia, corresponding to the Eastern part of the distribution zone of the species (Koch and German 2013) (Table S1 and Figure S1 in Supporting Information). Among them were 20 CAL populations, 36 NM populations, and 4 SER populations. The seeds from most of the French populations were sampled in the field in June 2011 and 2012 (Gonneau 2014). The number of sampled plants (families) varied according to the size of the population and to the fruit maturity at the sampling date. The seeds from the other countries were provided by various partners.

Cultivation device
The cultivation device was composed of 30 polyethylene trays of 76 Â 56 Â 30 cm ( Figure S2), each filled with 140 kg of homogenized sandy-loamy soil taken from the plowed horizon of an agricultural soil (Table S2). The upper 6.4 cm soil layer of each tray, i.e. 30 kg of soil, was contaminated with Cd, Cu, Pb, Ni, and Zn in the form of salts (CdCl 2 , CuSO 4 , Ni(NO 3 ) 2 , Pb(NO 3 ) 2 , ZnSO 4 ). The Cd, Ni, and Zn target concentrations in the upper 6.4 cm were then 8.6 mg kg À1 , 196 mg kg À1 , and 993 mg kg À1 , respectively, whereas those of Cu and Pb were 40.2 mg kg À1 and 74.1 mg kg À1 , respectively. Mulching fabrics were placed at the surface of each tray, to limit weed growth and emission of contaminated dust. The fabrics were previously perforated at regular intervals to make holes (diameter 1.5 cm), to enable sowing.
The cultivation trays were installed in a greenhouse with semi-controlled growth conditions, namely an automatic watering dripper with deionized water, a natural photoperiod and target extreme temperatures of þ4 C and þ25 C, maintained by automated heating and cooling devices. Air temperature was continuously recorded with a thermograph ( Figure S3). The soil being sufficiently provided with nutrients, no fertilization was added at the beginning. More details on the cultivation device can be found in Sterckeman et al. (2017).

Plant cultivation
Three cultivations were carried out successively from each September/October (sowing) to June/July (seed harvest) of years 2012-2014, and 2014-2015, designated as 2013, 2014, and 2015, respectively. One (2013 to three seeds (2014,2015) were placed at the surface of the soil in each hole in the mulching fabric. In case of multiple seedlings, the number of plants per hole was reduced to one. In 2013, 15 plants from each population were sown in contiguous holes of a tray, two population samples being randomly affected to each tray. With 30 plants per tray, the total plant number was to be 900 and the sowing density 71 plants m À2 . However, because of insufficient germination and mortality, only 554 plants were cultivated through to harvest, with a mean density of 44 plants m À2 . As the plants appeared too crowded at this density, they were divided into two in the following cultivations. Consequently, there were 281 (22 plants m À2 ) and 426 (34 plants m À2 ) plants cultivated in 2014 and 2015, respectively, a lot of plants being lost in the second cultivation, possibly because of excess soil humidity in the first weeks of growth.
All the individuals of a given population were grouped together to avoid cross-pollination between populations. From the beginning of the flowering, the entire plants were placed in a transparent and micro-perforated plastic bag and when the first inflorescence was strong enough, it was put in a bag of the same material and the first wrapping was removed. The irrigation was adapted to the climate and the vegetation stage, in order to avoid water stress and excess drainage. The seeds from the bagged inflorescences were collected after ripening of the silicles, between mid-June and mid-July, and at the same time, the whole shoot was harvested.

Selection scheme
After phenotyping the plants originating from the 60 wild populations (F1 generation, 2013), 13 of them were selected for their capacity to accumulate Cd, Ni, or Zn and/or to produce high amounts of aerial biomass. The progenies of these 13 selected plants were cultivated during the second year (F2 generation, 2014), phenotyped and self-pollinated. From these plants, 37 were selected. Their progeny, regarded as the F3 generation (2015), was cultivated and phenotyped. Again, 37 individuals from the F3 were selected, but their progeny was not cultivated. The criteria the plants were selected for at each step are given in Table S3.

Development and growth characteristics
Selected development stages were measured : first flower buds visible (the corresponding thermal time is noted FBV), first silicle visible (FSV), first silicle dehiscent (FSD), first raceme ripe (FRR, when all the silicles of the first raceme were dehiscent). The thermal time (in degree day or C d, base temperature ¼0 C) of selected development stage was determined for each plant.
The growth variables determined on each plant were the rosette diameter (RD, cm) at FBV, the shoot dry biomass (SB, g), and the number of racemes (NR) at harvest, i.e. at the FRR stage. The shoot biomass consisted in rosette leaves, stems, bracts, and silicles.

Trace metal accumulation
Regarding trace metal accumulation, the plants were selected according to their metal content at FBV (i.e. in February-March). Indeed, it was not possible to wait for the analysis of the whole plant harvested in June or July to select the progeny to be sown in September, as the plant analysis took too long to provide results before sowing. Therefore, at FBV, 2-5 leaves of the rosette of each plant were sampled in order to determine their content in Cd, Ni, and Zn, as described in Sterckeman et al. (2017). At seed harvest, the collected shoots were also analyzed for their Cd, Ni, and Zn concentrations.

Growth and development
The mean SB values of each generation increased significantly between F1 and F2, from 5.7 g plant À1 to 16.7 g plant À1 , respectively (Table 1). However, in F3, SB dramatically decreased to 7.5 g plant À1 , which was significantly lower than the SB of F2, but still higher than that of F1. When considering the plants selected from each generation for their higher SB (SPF1 SB , SPF2 SB , and SPF3 SB ), their mean SB values were three times higher than those of F1, F2, and F3. The shoot biomass of SPF1 SB and SPF3 SB were not significantly different and were both lower than that from that of SPF2 SB . The absence of amelioration of shoot biomass over generations could indicate that the selection did not enable a fixation for this trait. This could be due to the fact that yield is generally poorly heritable (Gallais 2011). Biomass production would probably be a complex trait depending on numerous genes and it would take several inbred generations to obtain the homozygosity of favorable alleles for as All plants For a given parameter, means with different letters are significantly different according to the Kruskall-Wallis test. Note that the plants selected differed from one trait to another (Table S3). a SB of the plants selected according to metal content. many of these genes as possible. More inbred generations would, therefore, be necessary for fixation, in order to obtain plants with the desired dry mass production. However, the absence of amelioration could also be due to a dependence of the trait on the environmental conditions and a change in these (particularly plant density) over generations (Allard and Bradshaw 1964). Indeed, it can be hypothesized that the change in genome over inbred generations, in a plant with a high selfing rate cannot explain the strong and erratic variations in the individual plant biomass. Moreover, there was no correlation for SB between SPF1 (plant selected according to all criteria) and the mean values of their F2 progenies ( Figure S4). There was, however, a positive linear correlation between SB of SPF2 and the mean values of their progenies (F3) with a slope of a ¼ 0.18 (R 2 ¼ 0.56) (Figure S5), indicating a low genetic variance. This apparently absent or low genetic variance suggests that the variations in SB over generations had an important environmental origin. Indeed, the sowing density was divided by two between F1 (44 plant m À2 ) and F2 (22 plant m À2 ) and increased again in F3 (34 plant m À2 ) to an in-between level. It can be seen that SB decreased almost linearly with increasing sowing density, independently of the selection process (Figure 1a). Low sowing density might have favored the growth of the plants, reducing their competition for nutrients, water, and light. Previous studies demonstrated that plants from dense stands show lower individual biomass (Galanopoulou-Sendouka et al. 1980;Meekins and McCarthy 2002;Weiner 1990).
We also examined the evolution of traits for which there was no specific selection. The diameter of the rosette at FBV, RD, was smaller in F1 than in F2 and F3 although there was no difference in this trait between F2 and F3 ( Table 2). The trait was unrelated to sowing density as was the shoot biomass. This could be because the rosette was measured at an early stage, at which the plants were still relatively small, not contiguous and, therefore, hardly competing for resources or interacting.
The number of racemes NR, ranked as follows: F2 > F3 > F1 (Table 2). There was a linear negative correlation of NR with the sowing density (Figure 1b), suggesting that increasing space between plants favors not only their aerial biomass but also their fruit production. These results are in agreement with those of Angadi et al. (2003) who observed yield compensation in canola by producing new inflorescences at low sowing density.
The beginning of flowering, represented by the FBV stage was earlier in the F1 generation than in the F2 and F3 ( Table 2). The difference in FBV between F1 and F2 was of the order of 580 C d, i.e. an increase of 41% in F2. However, the sum of temperatures in the greenhouse, of each year was similar ( Figure S3). The F1 plants were not exposed to more cold, which could have vernalized them more than the following two generations. On the other hand, FBV decreased as the plant density increased (Figure 2).
The same trend can be observed for FSV, FSD, and FRR. Indeed, there were negative relationships between these parameters and the sowing density, whose linearity and slope increased with thermal time, from FBV to FRR ( Figure  2). The difference between F1 and F2 increased from FBV to FRR, whereas it remained almost constant between F1 and F3. F2 plants were observed to need more time to reach their development stages. Having more space enabled them to produce more shoot biomass and reproductive organs than the F1 and F3 plants and to develop more slowly.
The shoot biomass of SPF1 SB , around 18 g plant À1 corresponded to 7.9 t DM ha À1 , at a sowing density of 44 plants m À2 . A cultivar with a mean SB of SPF2 SB (48.5 g plant À1 ) would produce 10.7 t DM ha À1 (22 plants m À2 ) if expressing similar phenotypes in the field to those in the greenhouse. The mean of SPF3 SB (22.0 g plant À1 ) corresponds to 7.5 t DM ha À1 (34 plants m À2 ). These yields are much higher than those obtained in field trials with the   hyperaccumulator, which were often below 2 t DM ha À1 (Sterckeman and Puschenreiter 2018). The biomass yield potential expressed during the selection is in the range of that requested for Cd phytoextraction (Koopmans et al. 2008;Zhao et al. 2003). However, two remarks must be made about the above yield estimations in field conditions. First, the selected plants might have grown at a density lower than that used for the above calculations, which were the mean densities of the experiment. Indeed, we observed that this density varied locally, depending on the mortality in each cultivation tray. The second remark is that the shoot biomasses were obtained in climate conditions which were probably more favorable to the plant growth than those in the open-air, as it has been observed for vegetables and fruits (Khah 2012;Khoshnevisan et al. 2013). As the growth seems to be very sensitive to environmental conditions, the next step should, therefore, be an evaluation in field conditions of the yield potential of the greenhouse-selected plants.

Cadmium concentration and offtake
The Cd concentration in the rosette clearly increased over generations. It was 1.25 mmol Cd (kg DM) À1 , 5.21 mmol Cd (kg DM) À1 , and 9.24 mmol Cd (kg DM) À1 in F1, F2, and F3, respectively (Table 1). The contents in the selected plant rosettes were 3.4-6.6 times higher but similarly ranked. The sowing density seemed to have no effect on the rosette Cd content. There was a positive linear correlation between SPF1 and the means of their progenies (F2) (a ¼ 1.56, R 2 ¼ 0.90; Figure S6a) and between SPF2 and the means of their progenies (F3) (a ¼ 1.00, R 2 ¼ 0.90; Figure S6b). The latter correlation indicated a high heritability of the rosette Cd content and only a slight environmental influence. However, the former, with a slope >1 showing that the progenies accumulated systematically more Cd than their parents, could reveal an environmental effect. Indeed, the lower plant density in second year could have caused lower root competition for Cd uptake and consequently a higher rosette concentration.
The Cd concentration in shoots at harvest also increased significantly from one generation to another. However, the shoot Cd contents were much lower than those of the corresponding rosettes and increased more slowly over generations (Table 1). Indeed, the rosette-to-shoot Cd content ratio was 2.0, 5.3, and 7.5 for F1, F2, and F3, respectively. The selection indicator, i.e. the Cd content of the rosette, appeared a relatively efficient way of predicting the metal content at harvest. However, the relationships between the Cd content in rosettes and shoots presented in Figure S7 suggest that as the Cd content increases in the rosette over generations, the content in the shoot at harvest tends towards a limit, of the order of 2 mmol Cd (kg DM) À1 .
The increase in Cd content in the rosette at FBV over generations suggests a positive effect of the selection of Cd hyperaccumulating individual plants. However, the benefit of selection appeared to erode with the plant growth as shoot Cd content was lower when the plant fructified, giving the appearance of a dilution of the Cd content in the increasing shoot biomass. This dilution effect seemed to increase from F1 to F3. This is contradictory to the results of Lovy et al. (2013) who observed no difference in Cd accumulation whatever the stage of the plant. However, these authors cultivated N. caerulescens in aeroponics with no limitation in the Cd offer (i.e. with constant Cd content in nutrient solution). In this study, the accumulation of Cd in the shoot at the fructification stage seemed to tend towards a limit. Two hypotheses may explain this phenomenon. One is that the processes which lead to Cd content in the fructifying shoot at FRR are different to those that control Cd content in the rosette at FBV and that the former was not the object of selection as the latter were. The second hypothesis is that the soil Cd offer limited the plant uptake because of the depletion of the rhizosphere as the plant grew (Lin et al. 2016). In the early stages, as the plant produced numerous new roots, the soil offer was sufficient to provide them with Cd according to the plant demand, whatever its genotype. With time, the rhizosphere of the dominating old roots was depleted in Cd and the uptake was controlled by Cd diffusion in soil, which is slow and limiting (Degryse et al. 2012), as it does not correspond to the flux required to maintain constant the Cd content of the growing shoot.
To evaluate the potential for phytoextraction, the Cd offtake was calculated by multiplying the shoot Cd content at harvest (FRR) by the shoot biomass at this stage (SB), considering that the crop should be harvested when the aerial biomass peaks. The Cd offtake ranked similarly to SB: F2 > F3 > F1 and SPF2 Cd > SPF3 Cd > SPF1 Cd (Table 1). In contrast to content, the Cd offtake was negatively correlated to the sowing density ( Figure S8), as SB was. The SPF1 Cd plants with the highest uptake (around 20 mmol Cd plant À1 , Figure S9) had an offtake of 1 kg Cd ha À1 (44 plants m À2 ), i.e. about 16% of the total amount of Cd in the soil (6.4 kg Cd ha À1 ). In SPF2 Cd , plants taking up around 60 mmol Cd plant À1 would have an offtake of 1.5 kg Cd ha À1 in a single cropping (with 22 plants m À2 ), representing 23% of the soil Cd, whereas plants taking up 80 mmol Cd plant À1 would have an offtake of 2.0 kg Cd ha À1 , representing 31% of the soil Cd. Plants having the highest uptake of SPF3 Cd , i.e. around 40 mmol Cd plant À1 would also have an offtake of 1. 5 kg Cd ha À1 (34 plants m À2 ). These figures, based on performances expressed in a greenhouse, suggest that N. caerulescens could be used for Cd phytoextraction, if its potential could be fixed, homogenized, and expressed in field conditions.

Nickel concentration and offtake
The Ni concentration in the rosette decreased over generations. It was 32.3 mmol Ni (kg DM) À1 , 24.4 mmol Ni (kg DM) À1 , and 14.8 mmol Ni (kg DM) À1 in F1, F2, and F3, respectively (Table 1). The contents in the selected plant rosettes were two to three times higher but similarly ranked.
There was a positive linear correlation between SPF1 and the mean values of their progenies (F2) (a ¼ 0.86, R 2 ¼ 0.96) and between SPF2 and the mean values of their progenies (F3) (a ¼ 0.19, R 2 ¼ 0.76) ( Figure S10). The evolution of the slopes could indicate a strong decrease in the heritability of the Ni rosette content for all the plants, even for those that were not selected for the trait. However, this is difficult to explain based on genetic mechanisms. It seems more probable that the strong correlations with decreasing slopes reflect the "aging" of soil Ni, i.e. slow reactions such as diffusion in iron oxides or precipitation (Buekers et al. 2007), which would make the metal less soluble with time and less absorbable by the roots. Indeed, Buekers et al. (2007) observed a reduction of 41% in the availability of Ni added to soil (80 mg Ni kg À1 ) after 850 days. In comparison, the decrease in the shoot Ni concentration between F1 and F3, i.e. in 730 days, was about 74%.
At harvest (FRR), the shoot Ni content also decreased over generations. However, the shoot Ni contents were lower than those of the corresponding rosettes (Table 1). Indeed, the rosette-to-shoot Ni content ratio was 1.9, 2.1, and 3.2 for F1, F2, and F3, respectively. The relationships between the Ni content in rosettes and shoots ( Figure S11) were linear, with slopes corresponding to the rosette-toshoot Ni content ratio. As for Cd, these could be explained by the fact that the concentration in the shoot at FRR is controlled by physiological processes not targeted by the selection. Alternatively, these could be due to a limitation of the Ni supply from soil in the rhizosphere of the aging plants, the limitation by diffusion for Ni being smaller than that of Cd (Degryse et al. 2012), as reflected by the rosetteto-shoot concentration ratios, which were lower for Ni than for Cd.
The Ni offtake did not follow the same evolution with selection, as the 33% lower Ni shoot concentration in F2 was over-compensated by the three times higher dry mass production (Table 1). Therefore, Ni offtake of the second generation (F2 and SPF2 Ni ) was higher than that of the first one (F1 and SPF1 Ni , respectively). Ni offtake of F3 as well as of SPF3 Ni was much lower than those of the previous two generations because of the combination of the lowest Ni shoot content with a low SB.
The offtake of F1 and F2 represented 2.5 kg Ni ha À1 and 2.1 kg Ni ha À1 , respectively, i.e. about 1.7% and 1.4% of the total metal in the soil (149 kg Ni ha À1 ). Efficient plants contained ca 400 mmol Ni plant À1 (SPF1, Figure S12). Sowing them at a density of 44 plant m À2 would permit a Ni offtake of about 10 kg Ni ha À1 , i.e. 6.7% of the soil Ni content. This rate indicates that decontaminating the soil would take about 15 years to reduce content from 50 mg Ni kg À1 to ca 20 mg Ni kg À1 .

Zinc concentration and offtake
The rosette Zn content increased over generations, being 168 mmol Zn (kg DM) À1 , 194 mmol Zn (kg DM) À1 , and 215 mmol Zn (kg DM) À1 in F1, F2, and F3, respectively ( Table 1). The contents in the selected plant rosettes were nearly two times higher and similarly ranked over generations although not significantly different. This suggests a positive effect of the selection.
There was a positive linear correlation between the rosette Zn content of SPF1 and the mean contents of their progenies (F2) (a ¼ 0.95, R 2 ¼ 0.93), as well as between SPF2 and their progenies (F3) (a ¼ 0.59, R 2 ¼ 0.85) ( Figure  S13). The decrease in the slope indicated an apparent decline in the genetic variance.
At harvest (FRR), the shoot Zn content decreased over generations F1, F2, and F3. Zinc concentration in F3 was nearly two times lower than that in F1. This decrease was relatively lower than that of Ni. Moreover, the rosette Zn contents were 1.2, 1.8, and 2.9 higher than those of the corresponding shoots at harvest, for F1, F2, and F3, respectively ( Table 1). The relationships between the Zn content in rosettes and shoots ( Figure S14) were linear, with slopes corresponding to the rosette-to-shoot Zn content ratios. For the plants selected, the Zn content at FRR was almost constant over generations, with rosette-to-shoot ratios ranging from 1.3 to 2.2.
The explanation of these results is similar to those of Ni. The decline in Zn content between rosette and harvested shoot could be explained by a depletion of the metal in the rhizosphere as the plants grew. The decline in shoot Zn content at FRR over generations would be due to the loss of availability of soil Zn, as also observed by Buekers et al. (2007). The increase over generations of rosette Zn content (contrarily to Ni), could be explained by the fact that, at this stage, the rhizosphere of the young roots was not depleted in the metal, as the initial content in the soil was high enough for the positive effect of the selection to be expressed. This was not the case for Ni, as this metal was found in much lower concentrations (196 mg kg À1 , Table  S2) in the soil than Zn (993 mg kg À1 ).
The Zn offtake was more affected by SB than by the Zn concentrations. This is the reason why it was higher in F2 and SPF2 Zn than in F1 and SPF1 Zn , respectively. Offtakes of F3 and SPF3 Zn were lower than their counterparts in F1 and F2 because of the lower Zn concentration in F3 and of the much lower shoot biomass in SPF3 Zn (Table 1).
The offtake of F1 and F2 represented 21 kg Zn ha À1 and 42 kg Zn ha À1 , respectively, i.e. about 2.8% and 5.6% of the metal in the soil (750 kg Zn ha À1 ). The efficient plants contained ca 3500 mmol Zn plant À1 (SPF1 Zn , Figure S15). Growing a cultivar with such accumulation ability at a density of a 44 plants m À2 would permit the extraction of about 100 kg Zn ha À1 , i.e. 13% of the soil Zn content. This is a considerable amount of metal, whose recovery from the biomass could be useful. However, it would take about 15 years to bring the soil Zn content from 250 mg Zn kg À1 to around 35 mg Zn kg À1 , assuming a yearly extraction rate of 13% of the soil Zn content.

Conclusions
This experiment showed that the developmental, growth and reproductive traits were strongly influenced by the sowing density. Giving more space to the plants enables them to lengthen their development stages, to produce more aerial dry matter and to increase their reproductive potential.
The pure-line selection of N. caerulescens led to an improvement of the Cd and Zn accumulation capacities of the progenies but caused a decrease in their Ni accumulation. This is most likely due to a decline in Ni availability in soil than to a deleterious effect of inbreeding.
N. caerulescens has the potential to produce up to 10 t DM ha À1 , which is required to decontaminate soils moderately enriched with Cd. The Cd concentrations obtained here are consistent with this goal, as they enabled offtakes representing around 25% of the soil Cd in a single cropping. However, this potential expressed in greenhouse conditions has to be confirmed in the field, in cooler, and more contrasted climates, with more biotic and abiotic stresses.
Consequently, the breeding of N. caerulescens should be pursued in conditions close to those of the application. Ideally, it should be carried out in various pedo-climatic conditions, with stabilized soil metal availability, under realistic and, as far as possible, constant cultivation practices.
More initial individuals should be tested and the self-pollination should be pursued for six to ten generations. Moreover, it would be preferable to fix the genes of metal accumulation during the first inbred generations on numerous lines, to maintain genotypic variability on the other traits and to continue the selection for biomass production later on, after the fourth or fifth generation.