phenoscope: an automated largescale phenotyping platform offering high spatial homogeneity

TECHNICAL ADVANCE/RESOURCE

Phenoscope: an automated large-scale phenotyping platformoffering high spatial homogeneity

S�ebastien Tisn�e1,2, Yann Serrand1,2, Lien Bach1,2, Elodie Gilbault1,2, Rachid Ben Ameur1,2, Herv�e Balasse1,2, Roger Voisin1,2,

David Bouchez1,2, Myl�ene Durand-Tardif1,2, Philippe Guerche1,2, Ga€el Chareyron3, J�erome Da Rugna3, Christine Camilleri1,2

and Olivier Loudet1,2,*1INRA – Institut National de la Recherche Agronomique, UMR 1318, Institut Jean-Pierre Bourgin, RD10, F–78000 Versailles,

France,2AgroParisTech, Institut Jean-Pierre Bourgin, RD10, F–78000 Versailles, France, and3ESILV – Ecole Superieure d’Ingenieurs Leonard de Vinci, Departement d’Enseignement et de Recherche Informatique et

Intelligence de l’Information, Pole Universitaire L�eonard de Vinci, 12 avenue L�eonard de Vinci, F–92916 Paris, France

Received 3 December 2012; revised 24 January 2013; accepted 25 January 2013.

*For correspondence (e-mail [email protected]).

SUMMARY

Increased phenotyping accuracy and throughput are necessary to improve our understanding of quantitative

variation and to be able to deconstruct complex traits such as those involved in growth responses to the envi-

ronment. Still, only a few facilities are known to handle individual plants of small stature for non-destructive,

real-time phenotype acquisition from plants grown in precisely adjusted and variable experimental conditions.

Here, we describe Phenoscope, a high-throughput phenotyping platform that has the unique feature of contin-

uously rotating 735 individual pots over a table. It automatically adjusts watering and is equipped with a

zenithal imaging system to monitor rosette size and expansion rate during the vegetative stage, with auto-

matic image analysis allowing manual correction. When applied to Arabidopsis thaliana, we show that rotat-

ing the pots strongly reduced micro-environmental disparity: heterogeneity in evaporation was cut by a factor

of 2.5 and the number of replicates needed to detect a specific mild genotypic effect was reduced by a factor of

3. In addition, by controlling a large proportion of the micro-environmental variance, other tangible sources of

variance become noticeable. Overall, Phenoscope makes it possible to perform large-scale experiments that

would not be possible or reproducible by hand. When applied to a typical quantitative trait loci (QTL) mapping

experiment, we show that mapping power is more limited by genetic complexity than phenotyping accuracy.

This will help to draw a more general picture as to how genetic diversity shapes phenotypic variation.

Keywords: Arabidopsis thaliana, phenotyping, high-throughput, image analysis, drought, growth, technical

advance.

INTRODUCTION

Understanding adaptation of plants to their environment is a

key issue that is addressed in many fields of study (quantita-

tive genetics, ecology, evolution, etc.) and may also poten-

tially contribute to breeding or engineering crop plants

adapted to sub-optimal conditions. Adaptation has been well

described in terms of the comparative biology of natural vari-

ants, but characterization of its underlying mechanisms is

still limited, especially for complex quantitative traits, such

as those related to plant physiology (Alonso-Blanco et al.,

2009). Arabidopsis thaliana is an interesting model because

its wide species range encompasses contrasting environ-

ments, suggesting that the numerous natural accessions

have evolved under diverse constraints. To investigate

potentially adaptive variation, natural variants may be ana-

lyzed in situ using ecological and population biology

approaches, or physiology and molecular studies may be

performed in artificial devices in which various environmen-

tal factors may be controlled to simulate diverse experimen-

tal conditions independently (Bergelson and Roux, 2010). In

Arabidopsis, although reverse genetic approaches based on

induced mutation analyses are used successfully to study

the species’ biology, they may be less efficient for revealing

© 2013 The AuthorsThe Plant Journal © 2013 Blackwell Publishing Ltd

1

The Plant Journal (2013) doi: 10.1111/tpj.12131

adaptive trajectories. Conversely, quantitative genetics is an

efficient tool for linking genetic polymorphism with pheno-

typic variation and for identifying components of the

response to environmental stresses directly from various

Arabidopsis accessions (Trontin et al., 2011).

Quantitative genetics studies often owe their success to

quantitative phenotyping rather than the qualitative or

semi-quantitative estimates that are generally used in most

classical screens (Loudet et al., 2008). Several potential

determinants of plant adaptation have been identified in

Arabidopsis, such as determination of the time to flower-

ing or germination, which are highly heritable and structur-

ing traits at the species scale (Amasino, 2010), or some

major-effect physiological changes (Baxter et al., 2010;

Poormohammad Kiani et al., 2012). However, variants

responsible for subtle quantitative variation, particularly in

response to the environment, are less well known, mainly

because these responses are highly complex, and probably

have numerous small-effect and interacting components

(Rockman, 2012). Most traits have inherent residual varia-

tion that introduces a variable error term when trying to

measure the genetic contributions. When focusing on a

specific locus (possibly with limited individual contribu-

tion), the phenotypic effect is reduced to that of the locus

at hand but the residual variation is still potentially strong.

This variation complicates (if not prevents) confirmation of

individual loci and thus identification of the underlying

molecular mechanism. To identify these determinants,

experiment size must be increased to enhance the detec-

tion power, but in most standard experimental set-ups, this

may be detrimental to the homogeneity and/or reproduc-

ibility of growing conditions. Although high-throughput

methods have been extensively developed for molecular

characterization, high-throughput macroscopic and inte-

grative phenotyping methods under controlled/laboratory

conditions require more attention. The root system archi-

tecture is of particular interest due to the complexity of its

investigation in the field, and a number of platforms have

been developed recently (Iyer-Pascuzzi et al., 2010; Clark

et al., 2013). For shoot growth study, the PHENOPSIS (Gra-

nier et al., 2006) and WIWAM (Skirycz et al., 2011) plat-

forms both represent specific set-ups for automated

phenotyping, allowing culture of approximately 200–500

Arabidopsis plants in individual pots with automatic water-

ing and imaging systems. However, even when environ-

mental conditions are strictly controlled and a standardized

protocol has been developed, some phenotypes may be dif-

ficult to reproduce due to uncontrolled sources of noise

leading to micro- and/or macro-environmental variation

(Massonnet et al., 2010). In particular, some abiotic stress

treatments are difficult to control precisely because their

severity depends on parameters that are highly heteroge-

neous across any growth room (e.g. applying mild drought

stress will strongly interact with air flow and hygrometry).

Here we present a new automated phenotyping platform

that represents a significant advance. The Phenoscope (Guer-

che et al., 2010) is an integrated device allowing simultaneous

culture of 735 individual Arabidopsis plants and high-

throughput acquisition, storage and analysis of quality phe-

notypes. Phenoscope combines the advantages of existing

automated platforms such as PHENOPSIS or WIWAM with

continuous rotation of plants during the experiment. The ben-

efits of this rotation include homogenization of the micro-

environmental conditions experienced by plants and

compensation for some of their uncontrolled effects on plant

phenotype, thus leading to enhanced statistical power to

detect significant effects. Large-scale recombinant inbred line

(RIL) experiments performed on Phenoscope produced data

with remarkable accuracy and repeatability, allowing identifi-

cation of both major- and minor-effect quantitative trait loci

(QTLs) with minimal replicates. Later validated in heteroge-

neous inbred families (HIFs), these results show how this type

of platform makes it possible to perform large-scale experi-

ments thatwould not be feasible or reproduciblemanually.

RESULTS

The Phenoscope set-up

The Phenoscope automated phenotyping platform comprises

an aluminum table (3 m 9 1 m) on a steel structure placed in

a growth room (Figure 1a) on which up to 735 plants may be

individually grown and phenotyped. On the table is a closed-

circuit track with a series of switchback turns that holds the

pots (Figure 1b). The individual pots, each designed to con-

tain a single plant, are pushed along guiding rails by coordi-

nated pusher arms that allow the robot to move each pot

sequentially across all possible positions on the table in a sin-

gle cycle (Figure S1), up to six times every 24 h. Pot move-

ments are designed so that all individual plants experience

the same external conditions on average over time, despite

any heterogeneity present in the growth room.

To maintain plants at given treatment targets, the table is

equipped with a weighing and watering station in the upper

right-hand corner (Figure 1). Pots placed on the weighing

and watering station may be weighed to control their state

and/or watered using a peristaltic pump that supplies them

with a specified nutrient solution, either at a defined volume

or at a calculated volume to reach a target weight. Non-

destructive phenotyping is performed at the imaging sta-

tion at the lower right-hand corner of the growth table (Fig-

ure 1). A digital camera takes zenithal images of plants

placed on the imaging station, labels and stores them on an

image server for further analysis. To see the robot in action,

visit http://www.inra.fr/vast/Files/PhenoFilm.avi.

Phenoscope implementation and use

A Phenoscope experiment consists of a programmed chain

of cycles in which culture of each plant is managed individ-

© 2013 The AuthorsThe Plant Journal © 2013 Blackwell Publishing Ltd, The Plant Journal, (2013), doi: 10.1111/tpj.12131

2 S�ebastien Tisn�e et al.

ually for several weeks. Each cycle comprises 735 repeti-

tions of ordered actions, namely weighing, watering, imag-

ing and pot moving, which may be activated or not,

depending on the needs of the experiment (Figure 2). Dur-

ing these actions, information is transmitted from the vari-

ous sensors to the management database and back to

record experimental data and to perform the experiment.

Various default modes may be defined and managed

through a computer interface to record initial weight infor-

mation necessary for watering management or performing

the experiment. In mode 1, all 735 empty pot weights are

recorded and stored in the management database to define

the tare of each individual grown during the experiment. In

mode 2, plugs (see Experimental Procedures) with 100%

soil water content (SWC) are placed one by one in pots

and weighed to define the reference weight used in target

weight calculations during the experiment. Mode 3 ensures

continuous movements until the beginning of mode 4,

which pilots all the actions required to run the experiment

(Figure 2 and Figure S2).

Before running mode 4, the user sets up the experiment

parameters, which are then stored in the management

database. First, the experiment duration is defined (num-

ber of days; Figure S2). Then, for each day, 2–6 cycles are

defined depending on their length (between 4 and 12 h) to

give a total of 24 h per day. The growth room and the Phe-

noscope robot are synchronized at the beginning of each

day to ensure that the Phenoscope actions are compatible

and synchronized with the growth room environment (e.g.

imaging when the light is on); each cycle is also synchro-

nized to start at the right time. Finally, the Phenoscope

actions are defined for each cycle 9 day 9 pot: the occur-

rence and type of watering (constant volume or % SWC),

the watering target (ml or %), the specification of the pump

and the activation of the camera.

Data management and analysis

Phenoscope experiments produce various types of data:

environmental and pot weight data, recorded respectively

by temperature/humidity sensors and the weighing and

watering station, are stored in the Phenoscope database

(DB), whereas images of plants recorded by the imaging

station are stored on a server for later extraction of pheno-

typic data (Figure S3). Images on the server are analyzed

using Phenospeed software, which was developed to auto-

matically and quickly process the Phenoscope images (Fig-

ure 3, Figure S4 and Methods S1). To deal effectively and

rapidly with a large number of images, most studies have

relied on basic approaches that cannot correctly determine

(a)

(b)

Figure 1. Phenoscope.

(a) Two independent Phenoscope automated phenotyping platforms placed

back to back in a 16 m2 growth room.

(b) Schematic representation of a Phenoscope table. The movements of

pots (open squares) are ensured by 39 connecting rods (green), two rakes

(red) and two pusher arms (yellow). Plants are watered on the weighing

and watering station (W&WS, blue), which comprises a weight sensor and

a device supplying the nutrient solution by a peristaltic pump. Photographs

are taken at the imaging station (IS, blue).

Figure 2. Flow chart diagram of a Phenoscope cycle.

The diagram depicts the progress through various actions (weighing, water-

ing, imaging, moving ‘one step’) according to the various modes (1–4) fol-lowed. ‘One step’ refers to Figure S1.


Phenoscope: high-throughput phenotyping 3

the rosette surface area under our experimental conditions

and throughput levels (a typical Phenoscope experiment

contains approximately 15 000 images). In addition to the

great diversity of genotypes studied, there may be highly

variable image-capture conditions: lighting, plant stage

and algae cover on the soil may be very diverse. None of

the existing state-of-the-art tools or methods (Walter et al.,

2007; Bylesjo et al., 2008; Backhaus et al., 2010; De Vylder

et al., 2012) met our needs. Thus, a dedicated image pro-

cessing algorithm was developed to address the effective-

ness and robustness of Phenoscope: Phenospeed extracts

useful information from images: the projected rosette area

(PRA, in cm2), the radius of the circle encompassing the

rosette, and the relative contributions of the red, green and

blue (RGB) components to rosette color (Figure 3c and Fig-

ure S5). Phenospeed automatically excludes rosette leaves

from neighboring plants as long as they do not overlap

with the analyzed rosette. It also automatically detects vari-

ous types of errors, especially when the rosettes reach the

edge of the image or when a leaf from a neighboring plant

overlaps with the plant of interest. Phenospeed allows the

latter case to be corrected manually, by redrawing the

edge of the actual rosette to exclude the neighboring leaf.

Otherwise, the user simply flags the associated data to

keep track of the possible mis-estimation. Finally, Pheno-

speed allows a real-time view of plant growth and pro-

duces output files containing the various phenotypic data

measured for each individual and each day of the experi-

ment. These phenotypic data, plug weight and environ-

mental data across time are stored in the Phenoscope DB,

which may be queried (Figure S6). Phenotypic data are

then curated in the Phenoscope DB to identify atypical

plants with respect to their genotype and treatment. An

interface exists to enable extraction of data for specified

experiments or to run R scripts to automatically detect out-

liers in datasets. These plants are then tagged by the user

after a manual review of plant image series, growth curves

and SWC from individual plugs across the experiment,

compared to other replicates of the same genotypes grown

under the same conditions (when available). This method

may be used to remove obvious outliers that may result

from stochastic growth defects.

Protocols for water-deficit treatments

One of the main goals and applications of Phenoscope is

to precisely control the hydric environment and allow

quantification of Arabidopsis phenotypic plasticity in

response to diverse water-deficit treatments. Performing

experiments at precise water availabilities requires control

of both soil desiccation and water supply. On Phenoscope,

the continuous rotation of individual pots is expected to

homogenize the micro-environmental conditions experi-

enced by plants which may affect the rate of evaporation.

To test this assumption, soil-drying experiments with fixed

or cycling pots were performed on Phenoscope, and water

loss was recorded 24 h after the plugs were installed (at

100% SWC). When remaining at fixed positions on the

table, pots exhibited strong spatial gradients in drying

intensities (Figure 4a), probably due to the heterogeneity

of air flow in the growth room. When the pots are cycling,

these gradients were considerably attenuated (Figure 4b).

Consequently, passive evaporation is 2.5 times more

homogeneous when pots are cycling coefficient of varia-

tion (CV = 6%) than when they remain in a fixed position

(CV = 15%; Figure S7). We have frequently encountered

this kind of gradient in other growth rooms of various

designs, and they were often even stronger depending on

how air flow is managed (data not shown).

During Phenoscope experiments, plug water loss may

be compensated for by automatic watering, which may

potentially be performed during each cycle (i.e. up to six

times a day) to maintain the user-defined level of SWC.

Watering is managed for each pot individually according

to the target SWC for each day 9 cycle. For instance, a pot

placed on the weighing and watering station is weighed,

then the peristaltic pump is instructed to deliver the

(a)

(b)

(c)

Figure 3. Arabidopsis shoot growth on Phenoscope.

(a, b) Phenoscope photographs of representative Col–0 (a) and Cvi–0 (b)

individuals at 5-day intervals (D09, D14, D19, etc.) under our conditions.

Raw photographs are shown in the top row, and the same photographs

once processed by the Phenospeed software are shown in the bottom row.

(c) Col–0 and Cvi–0 growth kinetics on Phenoscope, obtained by extracting

the projected rosette area (cm2) from the daily photographs.



amount of nutrient solution necessary to return it to its

target weight (calculated by the computer to reach the

user-defined SWC), and the pot is weighed again after

watering. According to our experience, watering twice a

day is a good compromise between maintaining a stable

SWC and the accuracy of the peristaltic pump (which must

deliver approximately 1 ml minimum per watering opera-

tion). The SWC may be held within a 0.9–1 range of the tar-

get across all cycles for various target SWC levels

(Figure 5a), whereas watering by hand (once a day) only

maintains the SWC within a 0.8–1 range of the target.

Measuring vegetative growth plasticity in response to

water availability

Typical protocols for Phenoscope drought-response experi-

ments were established to optimize plant growth and

homogeneity, treatment application and duration and to

deal with technical issues. An experiment has two distinct

phases: first, stratified seeds for each genotype are sown

on plugs placed for 8 days in the Phenoscope growth

room and constantly saturated at 100% SWC to ensure

homogeneous germination and initial seedling develop-

ment. Then, at day 8 after sowing, homogeneous seedlings

(see Experimental Procedures) are placed in a pre-deter-

mined order on the Phenoscope table (Figure S2: mode 2)

and a controlled decrease of the SWC begins. From then

on, Phenoscope is set up to adjust the SWC twice a day.

Individuals reach their target SWC and are then maintained

in this state until the end of the experiment at day 29 (Fig-

ure 5a).

As an illustration, the final projected rosette areas

reached by the Col–0 and Cvi–0 accessions under various

SWC targets are shown in Figure 5(b). Compared to the

‘control’ condition, previously defined at 60% SWC on the

basis of optimal growth of a wider range of genotypes

(Bouchabke et al., 2008), increasing SWC did not result in

any significant increase in PRA, while decreasing SWC to

40% only started to limit growth in Cvi–0. At 30% SWC, the

final PRA was significantly reduced compared to 60%

SWC, but was not further reduced at 25% SWC (Figure 5b).

Thus, moderate water-deficit conditions were defined at

30% SWC for subsequent experiments to ensure significant

responses to stress and minimal plant loss for the most

susceptible lines.

To test the effect of the Phenoscope set-up on phenotyp-

ing accuracy, two experiments were performed in parallel

(a)

(b)

Figure 4. Effect of pot cycling on evaporation heterogeneity.

Spatial distribution of water loss in free plugs over 24 h when pots are

maintained at a fixed position on the table (a) or continuously moved across

the table (b). Each square corresponds to a pot on Phenoscope. The blue-

to-red color scale indicates water loss ranging from 10 to 30% of the initial

plug weight (set at 100% soil water content).

(a)

(b)

Figure 5. Growth response of Col–0 and Cvi–0 accessions to differences in

soil water content (SWC).

(a) Evolution of SWC in plugs during the experiment. Dots represent the

SWC as defined in the experiment protocol with five target levels: 70%

(black), 60% (blue, control treatment), 40% (black), 30% (orange, mild water

deficit) and 25% (black). The observed mean plug SWC values as recorded

by the weighing station at each cycle are indicated by solid lines for control

(blue) and mild water deficit (orange) treatments.

(b) Final projected rosette area measured for Col–0 (black) and Cvi–0 (gray)

plants grown under the five treatments (five target SWC), represented rela-

tive to the value obtained at 60% SWC. For each genotype, different letters

indicate values that are significantly different from each other (Tukey test,

P < 0.05).



on Col–0 and Cvi–0 plants grown at 60 and 30% SWC, with

30 individuals for each genotype in each treatment. One

experiment was performed manually (see Experimental

Procedures) and the other experiment was performed on

Phenoscope. An analysis of variance components for the

effect of treatment (T), genotype (G) and their interaction

(G 9 T) in the whole dataset showed that the T effect was

equally highly significant in both experiments, whereas the

G and G 9 T effects were more significant in the Pheno-

scope experiment. Randomly sampling across the 30 repli-

cates showed that the significance levels (mean P value)

decreased with sample size (from 24 to four), except for

the T effect which was highly significant whatever the sam-

ple size (P < 0.001, Figure 6). In contrast, the G effect (a

mild factor, due to our choice of accessions and condi-

tions) was significant at P < 0.05 for as few as eight repli-

cates grown on Phenoscope, while at least 23 manually

performed replicates were required to reach the same sig-

nificance. The mild G 9 T effect allowed us to test the

impact of Phenoscope on a weak effect: the G 9 T effect

was only significant when 20 replicates (or more) were

grown on Phenoscope, but the interaction was undetect-

able in the manual experiment (Figure 6).

A test case: studying the growth response in the

Cvi–0 3 Col–0 RILs

Phenoscope allows definition of precise growing condi-

tions and high-throughput phenotyping, which are often

the main limitations in quantitative genetics studies, espe-

cially when working on integrative traits and interaction

with the environment. A large Cvi–0 9 Col–0 RIL popula-

tion was used as a proof of concept, first to test the accu-

racy of phenotyping on a large number of related

genotypes with great phenotypic diversity, and second to

map the genetic architecture of vegetative growth and its

plasticity with regard to water deficit. We pushed the

experimental design to its limits by using only one repli-

cate per genotype (358 RILs) 9 treatment (control versus

mild water deficit), and running the whole experiment

twice as independent biological replicates. Datasets from

the two Phenoscope experiments were very similar, and

linear models including genotype and SWC explained 78%

of the total phenotypic variance for final PRA, with the

‘experiment’ effect being non-significant (P = 0.48).

Quantitative trait loci were mapped using genotypic

mean values of final PRA and the relative expansion rate

(RER) for plants grown under each treatment (Figure 7a,

Figure S8 and Table S1). QTL mapping performed sepa-

rately on individual phenotypic values from each Pheno-

scope experiment gave very similar results and generally

led to identification of the same regions (Figure S9). Seven

QTLs were identified for the final PRA under both control

and stress conditions (chromosome 1 at approximately 6,

24 and 130 cM; chromosome 2 at approximately 45 and

80 cM; chromosome 4 at approximately 50 cM; chromo-

some 5 at approximately 105 cM), with positive allelic

effects from both accessions, and QTL models explained

approximately 60% of the phenotypic variance in each

treatment (Figure 7a and Table S1). QTL mapping for RER

showed co-localization with PRA QTLs in four regions

(chromosome 1 at approximately 130 cM; chromosome 2

at approximately 80 cM; chromosome 4 at approximately

50 cM; chromosome 5 at approximately 105 cM) (Figure 7a

and Table S1), indicating that final PRA is determined in

part by differential RER but also by early-stage growth

behavior. To validate the QTLs mapped with RILs, HIFs seg-

regating for genetic intervals covering the QTL candidate

regions on chromosome 2 (expected to include large-effect

QTLs) and chromosome 4 (expected to include small-effect

QTLs) were grown under control and stress conditions on

Phenoscope (Figure 7b and Figure S10). On chromosome

2, two distinct large-effect QTLs were validated in four HIFs

(Figure S10a,f–h), and an additional QTL was identified (Fig-

ure S10d). Moreover, HIF411 validated the interaction with

SWC of the major QTL on chromosome 2: RER was signifi-

cantly increased by Col alleles only under stress conditions

(Figure S10a). On chromosome 4, QTLs were predicted to

have weaker effects, and models were not able to separate

different loci despite potentially opposite allelic effects in

Figure 6. Influence of Phenoscope on the statistical significance of various

effects.

Comparison of the statistical significance of the genotype (‘G’), treatment

(‘T’) and interaction (‘G 9 T’) effects in experiments performed in parallel

either manually (‘M’) or using Phenoscope (‘P’). Col–0 and Cvi–0 individuals

were grown at 60 and 30% SWC (n = 30 replicates for each). For each effect

tested, the gray scale indicates the mean P value calculated from 1000 ran-

dom samples of sizes ranging from 4 to 24 individuals.



the region (Figure 7a). Three HIFs validated at least two sig-

nificant QTLs in this region, with opposite allelic effects on

final PRA and RER (Figure S10b,i,j). Antagonistic effects

segregated within HIF272: the lack of effect on final PRA

may be the result of significant positive effects of Col

alleles on RER and Cvi alleles on early growth (Figure

S10b), which may represent two distinct, but linked, loci

(although more complicated biological scenarios are possi-

ble). The positive effect on growth associated with Cvi

alleles was also independently validated in HIF182 (Figure

S10j), and HIF236 appeared to isolate the Col-positive effect

on RER (Figure S10i). At this scale, QTL mapping power

may be more limited by genetic complexity (linked and/or

interacting loci) than by phenotyping accuracy.

DISCUSSION

The current major limiting factor for identifying quantitative

variants is phenotyping. Sequencing/genotyping tech-

niques are improving constantly, and large genotypic varia-

tion datasets are now quickly released (Cao et al., 2011). In

this context, while ecological approaches to phenotyping

remain a real challenge (Bergelson and Roux, 2010), making

sense of adaptive genotypic variation requires improved,

controlled, non-destructive phenotyping.

Quantitative genetics approaches that associate geno-

typic with phenotypic variation increase constraints on

phenotyping: although genotypes (especially RILs, but

also, to a lesser extent, accessions) are partial (genetic)

replicates of each other, large population sizes are still

required (more genotypes are always better for mapping

accuracy, epistasis, etc.) and these are often not compati-

ble with precisely maintained growth conditions. Often,

increasing the experiment size through use of larger set-

ups or more independent experiments implies increasing

inherent environmental variation and the time required for

phenotype acquisition, which affects mapping power.

Beyond the initial mapping stage, fine mapping also typi-

cally requires very large population sizes to be studied in

successive steps.

For instance, QTL approaches in Arabidopsis have to

date only yielded a small number of essentially large-

effect loci that have been characterized and cloned

(Alonso-Blanco et al., 2009). Despite a few exceptions,

with sporadic cloning of medium- or small-effect QTLs

(Kroymann and Mitchell-Olds, 2005; Loudet et al., 2008),

only a few research groups are dedicated to this challeng-

ing task, especially for traits that are likely to be highly

multigenic and respond strongly to the environment, such

as leaf growth (Massonnet et al., 2010; Tisn�e et al., 2010;

Skirycz et al., 2011; Baerenfaller et al., 2012). Similarly,

only a few de novo forward genetic screens use real

quantitative phenotypes. Consequently, the loci cloned so

far are biased toward strong-effect loci, and probably

toward those that are stable and show little interaction

with the environment (requiring less control of phenotyp-

ing conditions) or with the genetic background (requiring

fewer genotypic combinations to be tested). However,

some argue that small-effect variants may be different in

nature from large-effect loci, such that the latter are not

necessarily predictive of the whole range of genetic varia-

tion important for evolution (Rockman, 2012). Nonethe-

less, this remains a very theoretical debate, and the

infinitesimal model still enjoys by-default acceptance:

there is a lack of empirical data to better address these

questions and draw a more general picture as to how

(much) and where in the pathways adaptation shapes nat-

ural variation. Moreover, the general picture (if there can

be any) may be different for traits such as flowering time

than for growth adjustment. Better access to less-limiting

and more-integrative phenotyping tools and platforms

will have a significant impact on these quantitative

approaches.

(a)

(b)

Figure 7. Genetic architecture of growth and its response to drought stress

in the Cvi–0 9 Col–0 RIL set.

(a) QTL analysis for growth-related parameters (final projected rosette area

and relative expansion rate) from the Cvi–0 9 Col–0 RILs grown under con-

trol (blue) and water deficit (orange) treatments. LOD scores from ‘Multiple

QTL Mapping’ (MQM) are plotted along the five chromosomes, with the

sign indicating the direction of the allelic effect (positive LOD scores indi-

cate a positive effect associated with the Cvi allele with respect to the Col

allele). Significance thresholds (genome-wide 5%) determined by a permu-

tation test are plotted as horizontal dotted lines.

(b) QTL effects were validated in the fixed progeny of RILs (heterogeneous

inbred family, HIF) segregating only at a residual heterozygous region indi-

cated by a rectangle along the genetic map next to the RIL name (e.g.

RIL109). Red rectangles indicate significant segregation of a growth pheno-

type in the corresponding HIF (ANOVA, P < 0.05) as detailed in Figure S10.



Studying the drought stress response is particularly

challenging as its level will be affected and interact with

virtually any parameter in the environment (Granier et al.,

2006): temperature, relative air humidity, air flow, light, soil

quality and drying, nutrient availability, etc. When a large

numbers of individuals need to be compared, experiments

in the field or in the greenhouse may potentially be homo-

geneous, but it is impossible to repeat the experiment

under the exact same conditions and/or stress intensity.

Working in growth rooms is necessary when strictly con-

trolled conditions are needed, but inherent spatial hetero-

geneity exists at all scales, precisely because the

environmental parameters are being controlled (for exam-

ple, maintaining temperature while changing the light

environment requires high air flow through the room).

Under these circumstances, we show here that continu-

ously rotating individual plants across the Phenoscope

table (Figure 1) reduces the heterogeneity perceived by the

plants and greatly improves repeatability: using Pheno-

scope reduces evaporation heterogeneity by a factor of 2.5

for hundreds of samples (Figure 4), better controls stress

intensity (always within 0.9–1 of the target SWC, compared

to 0.8–1 by hand; Figure 5) and reduces by a factor of three

the number of replicates needed to detect a specific mild

genotypic effect (Figure 6). In addition, the robot allows for

and controls the quality of complex watering conditions,

including features that are not feasible to perform manually

(e.g. several waterings per day). Moreover, compared to the

gains in a normal experimental set-up, the gains observed

here are probably conservative estimates, because our

growth room and our manually performed experiments had

already been highly streamlined over many years (Bou-

chabke et al., 2008). Furthermore, working with high num-

bers of pots is simply not possible by hand in most

laboratories, or only at the cost of even greater heterogene-

ity. Even with the number of individuals that may be han-

dled on Phenoscope (735; currently more than comparable

existing platforms for treating individual plants in a com-

pact set-up), some experimental designs may still require

replicates to be run in distinct, successive experiments (e.g.

when studying large RIL sets or performing association

mapping on relatively large accession sets, especially in

response to diverse environmental treatment conditions).

For this type of experiment, highly reproducible conditions

are desired because independent biological replicates rarely

have no effect on physiology and growth traits in such a

diverse and interacting genetic material, as witnessed previ-

ously (Loudet et al., 2003). Phenoscope was able to over-

ride this heterogeneity, as illustrated by independent QTL

mapping experiments in Cvi–0 9 Col–0 (Figure S9).

With roughly 15 000 images for each typical Phenoscope

experiment, image analysis was a challenge. Several avail-

able programs propose simple and cost-effective imaging

systems based on color filtering and adapted thresholding

or segmentation (De Vylder et al., 2012; Zhang et al.,

2012). Other systems have been developed to automate

plant imaging and information extraction (Walter et al.,

2007; Arvidsson et al., 2011). Our approach was designed

to be robust to intrinsic variations in rosette development

across the experiment or genotypes (color, shape, back-

ground soil color), and allow for a high degree of modular-

ity in imaging conditions (camera, lighting). In addition,

our system detects, takes into account and/or corrects spe-

cific incidents, e.g. when a neighboring plant extends to

within the actual plant imaging zone.

Compared to previously developed automated and inte-

grated systems designed to study shoot growth and its

response to the environment (Granier et al., 2006; Skirycz

et al., 2011), the main advantage of Phenoscope resides in

the control of environmental variation in any growth room

and at any room setting, while maximizing the number of

individuals under study. This contrasts with attempting to

homogenize a specific growth room to compare individu-

als grown across the area, as for the PHENOPSIS or WI-

WAM platforms. Automatically adjusted watering is a key

feature of these platforms and is necessary to implement

complex and precise watering regimes. A resulting limita-

tion of Phenoscope is that it requires that all plants are

moving continuously, which may have an effect on the

phenotypes expressed (although movements are very

smooth and equal for all individuals, and plants are rarely

completely undisturbed in any growth room with air flow).

Another consequence of our set-up is that the sequence of

the processed plants (for watering, imaging, etc.) remains

the same throughout the experiment, with several hours

between the time when the first and last plants of an

experiment are treated. However, if necessary, more sen-

sors could be added to the table to reduce this lag and

reduce the amount of time between data acquisition/

adjustment on the same individual.

It must be stressed that Phenoscope does not remove all

environmental variance, nor does it cancel all sources of

heterogeneity. Unexplained variation remains: one obvious

factor includes the soil used, with limited heterogeneity

among plugs but potential differences among batches (e.g.

due to the source of peat moss); another factor is variation

among individuals, especially in the initial development

steps [germination speed and vigor, which, according to

our observations and others’, probably strongly depends

on seed parameters (Zhang et al., 2012)] or later growth

defects (with some stochasticity). Removing some sources

of uncontrolled variation makes others become more

apparent than they were previously, and these need to be

taken into account or corrected (Granier et al., 2006; Mas-

sonnet et al., 2010). For instance, we are now starting to

calibrate the seeds based on their size prior to sowing on

Phenoscope, when precise comparisons are required

between related genotypes (e.g. in near-isogenic lines).



Seed size has previously been reported to explain a signifi-

cant part of the variation for shoot and root growth, espe-

cially (although not exclusively) at early stages in

development (Elwell et al., 2011).

An increase in phenotyping accuracy and throughput is

necessary to improve our understanding of the complexity

and scale of quantitative trait variation and the extent of

mutation effects on phenotypes. In QTL mapping, for

example, complex features (e.g. multiple linked loci, antag-

onistic effects, epistasis; Kroymann and Mitchell-Olds,

2005) may hinder the power to detect individual loci. Com-

bining RILs with finely phenotyped near-isogenic lines may

help to reveal more loci (Figure 7) (Koumproglou et al.,

2002; Keurentjes et al., 2007). To start to understand a sig-

nificant part of the heritability for very integrative traits

such as growth, association studies will require the pheno-

typing of large sets of accessions (Brachi et al., 2011;

Zhang et al., 2012), if not nested association mapping pop-

ulations (Bergelson and Roux, 2010).

Finally, Phenoscope lays the foundations of an auto-

mated robot that may be improved and expanded upon to

implement other stress scenarios and for non-destructive

phenotype acquisition. Although PRA represents the pho-

tosynthetically active area and has long been regarded as

a very good proxy for leaf area and biomass accumulation

(Leister et al., 1999), at least until a certain age, a better

description of the rosette may be developed by use of

extra cameras (allowing three-dimensional characteriza-

tion) and/or extra image analysis (e.g. to follow specific

leaves through time). Other complementary sensors will

soon be added to Phenoscope (i.e. thermal and fluores-

cence imaging), as well as the possibility of delivering

different watering solutions with distinct pumps to control

water and nutrients independently.

EXPERIMENTAL PROCEDURES

A detailed technical description of the Phenoscope set-up, data-base and analysis suite is available in Methods S1.

Performing a phenotyping assay

Our experiments on Phenoscope were performed using 4 cmplugs of peat moss substrate (70% blond peat, 20% perlite and10% vermiculite) wrapped in a non-woven film. Arabidopsis seedswere stratified in a 0.1% agar solution for 3 days at 4°C in the dark,and sown on plugs maintained at saturation (100% SWC) in plastictrays. Immediately after germination, seedlings were thinned oneach plug to retain only one individual, with as much synchro-nous germination as possible across plugs. For each genotype,more plugs (+30%) were sown than the number needed for theexperiment, allowing selection (essentially based on the size ofthe cotyledons to control for homogeneous early development)among 1-week-old seedlings of the same genotype or group ofgenotypes. These plugs were then installed on the robot (8 daysafter sowing).

Before starting the experiment, empty pots were individuallyweighed to define the tare. Then the selected plugs (seedlings)

were installed in the Phenoscope pots, one by one in a pre-deter-mined order to weigh the plug promptly at 100% SWC. Thisweight was used to define the target weight for each pot individu-ally, according to the watering instructions decided by the user(e.g. 60% SWC) which may differ from day to day and from pot topot. From sowing to the end of the experiment, plugs werewatered with specific nutrient solution [derived from thatdescribed previously by Loudet et al. (2003) but with 5 mM

nitrate]. The photoperiod in the growth room was set to shortdays (8 h light/16 h dark) to avoid interaction with early floweringand to optimize the study of vegetative shoot growth and theresponse to treatment during the exponential growth phase. Lightwas provided by 177 cool-white daylight fluorescent tubes to anintensity of 230 lmol m�2 sec�1. The air temperature was set to21°C during the day and 17°C at night, with a constant relativehumidity of 65%. Relative humidity and air temperature were mea-sured every 5 sec using a Rotronic HC2–S probe (www.rotronic.com). They were then averaged every 5 min and stored in thePhenoscope DB.

Manual experiments (for comparison) were performed in paral-lel in the same growth room, using the same target SWC. Trayrotation, weighing, watering, imaging and analyzing the rosettesize were performed by hand once a day, every day.

Genetic material used

The Col–0 and Cvi–0 accessions and the derived Cvi–0 9 Col–0 RILpopulation (Simon et al., 2008) have been described previouslyand were obtained from the Institut Jean-Pierre Bourgin, Ver-sailles, France (http://dbsgap.versailles.inra.fr/vnat/). Heteroge-neous inbred families (HIFs) were selected in the progeny of RILscarrying a heterozygous region at a locus of interest to validatethe relative effect of parental alleles on growth traits in an other-wise identical genetic background (Loudet et al., 2005).

Statistical analyses

All statistical analyses were performed using R software (R Devel-opment Core Team, 2007; http://www.r-project.org/). Statistical dif-ferences between accessions were tested by ANOVA, and Tukey testwas performed using the Tukey HSD() function.

For the sample size test (Figure 6), 1000 samples of 4–24 indi-viduals were randomly drawn from the complete sample of 30individuals by genotype (G) 9 treatment (T) using the sample()function. For each sample size, mean P value was obtained byANOVA, calculating the mean P value for G, T and G 9 T effect ineach of the 1000 samples.

Quantitative trait loci mapping was performed on genotypic val-ues (mean of two independent replicates) for the final PRA deter-mined at 29 days after sowing and the RER. The RER wascalculated over a time window corresponding to day 16–29 as theslope of the linear regression between the log-transformed PRAand time. QTL mapping was performed using the MQM strategyimplemented in the R/qtl package (http://www.rqtl.org/). Co-factorswere automatically backward-selected, and the scan was per-formed using the mqmscan() function for each trait and eachtreatment separately. Markers with an LOD score above a signifi-cance threshold determined by permutation tests (1000 permuta-tions, LOD = 2.42 for a significance level of 5%) were then testedtogether in an ANOVA to fit QTL models. Epistatic interactions wereinvestigated using the scantwo() function, but no significant inter-actions were observed.

An HIF strategy was chosen to validate the identified QTLs.Three replicates of three independent lines carrying fixed Col orCvi alleles in the candidate region were grown under each set of



conditions. Phenotypic data were first processed to identify poten-tial outlier plants among the nine replicates per genotype x treat-ment, i.e. plants whose individual values were outside twice theinter-quartile range. After removing outliers, G, T and G 9 Teffects on PRA and RER were then tested on data by ANOVA tovalidate the effect of the segregating region on the phenotype.

ACKNOWLEDGEMENTS

We thank P. Suply and colleagues for their help in designing andbuilding the Phenoscope tables. This work was supported byfunding from the European Commission Framework Programme7, ERC Starting Grant ‘DECODE’/ERC-2009-StG-243359 to O.L. andagence nationale de la recherche (ANR) grant ‘2Complex’/ANR-09-BLAN-0366 to O.L. This work was initially supported directly bythe Institut National de la Recherche Agronomique and its PlantGenetics Division.

SUPPORTING INFORMATION

Additional Supporting Information may be found in the online ver-sion of this article.Figure S1. Diagram of pot movements on the Phenoscope table.

Figure S2. Flow chart of various modes used for Phenoscopeexperiment management.

Figure S3. System architecture of the Phenoscope platform.

Figure S4. Processing the Phenoscope images.

Figure S5. Example of phenotypic data extraction from a Pheno-scope experiment.

Figure S6. Physical representation of the Phenoscope databasemodel.

Figure S7. Cycling the pots on Phenoscope reduces heterogeneity.

Figure S8. Phenotypic variation among Cvi–0 9 Col–0 RILs.

Figure S9. Comparing QTL analysis from two independent biologi-cal replicates.

Figure S10. Phenotypes obtained for HIFs designed to confirmQTLs located on chromosomes 2 and 4 as indicated in Figure 7(b).

Methods S1. Detailed technical description of the Phenoscope set-up.

Table S1. Parameters of the significant QTLs detected for growthtraits in the Cvi–0 9 Col–0 RILs.

REFERENCES

Alonso-Blanco, C., Aarts, M.G., Bentsink, L., Keurentjes, J.J., Reymond, M.,

Vreugdenhil, D. and Koornneef, M. (2009) What has natural variation

taught us about plant development, physiology, and adaptation? Plant

Cell, 21, 1877–1896.Amasino, R. (2010) Seasonal and developmental timing of flowering. Plant

J. 61, 1001–1013.Arvidsson, S., Perez-Rodriguez, P. and Mueller-Roeber, B. (2011) A growth

phenotyping pipeline for Arabidopsis thaliana integrating image analysis

and rosette area modeling for robust quantification of genotype effects.

New Phytol. 191, 895–907.Backhaus, A., Kuwabara, A., Bauch, M., Monk, N., Sanguinetti, G. and Fleming,

A. (2010) LEAFPROCESSOR: a new leaf phenotyping tool using contour

bending energy and shape cluster analysis.New Phytol. 187, 251–261.Baerenfaller, K., Massonnet, C., Walsh, S. et al. (2012) Systems-based analy-

sis of Arabidopsis leaf growth reveals adaptation to water deficit. Mol.

Syst. Biol. 8, 606.

Baxter, I., Brazelton, J.N., Yu, D. et al. (2010) A coastal cline in sodium accu-

mulation in Arabidopsis thaliana is driven by natural variation of the

sodium transporter AtHKT1;1. PLoS Genet. 6, e1001193.

Bergelson, J. and Roux, F. (2010) Towards identifying genes underlying ecologi-

cally relevant traits in Arabidopsis thaliana. Nat. Rev. Genet. 11, 867–879.

Bouchabke, O., Chang, F., Simon, M., Voisin, R., Pelletier, G. and Durand-

Tardif, M. (2008) Natural variation in Arabidopsis thaliana as a tool for

highlighting differential drought responses. PLoS ONE, 3, e1705.

Brachi, B., Morris, G.P. and Borevitz, J.O. (2011) Genome-wide association

studies in plants: the missing heritability is in the field. Genome Biol. 12,

232.

Bylesjo, M., Segura, V., Soolanayakanahally, R.Y., Rae, A.M., Trygg, J.,

Gustafsson, P., Jansson, S. and Street, N.R. (2008) LAMINA: a tool for

rapid quantification of leaf size and shape parameters. BMC Plant Biol.

8, 82.

Cao, J., Schneeberger, K., Ossowski, S. et al. (2011) Whole-genome

sequencing of multiple Arabidopsis thaliana populations. Nat. Genet. 43,

956–963.Clark, R.T., Famoso, A.N., Zhao, K., Shaff, J.E., Craft, E.J., Bustamante, C.D.,

McCouch, S.R., Aneshansley, D.J. and Kochian, L.V. (2013) High-through-

put two-dimensional root system phenotyping platform facilitates

genetic analysis of root growth and development. Plant, Cell Environ. 36,

454–466.De Vylder, J., Vandenbussche, F., Hu, Y., Philips, W. and Van Der Strae-

ten, D. (2012) Rosette tracker: an open source image analysis tool for

automatic quantification of genotype effects. Plant Physiol. 160, 1149–1159.

Elwell, A.L., Gronwall, D.S., Miller, N.D., Spalding, E.P. and Durham Brooks,

T.L. (2011) Separating parental environment from seed size effects on

next generation growth and development in Arabidopsis. Plant, Cell

Environ. 34, 291–301.Granier, C., Aguirrezabal, L., Chenu, K. et al. (2006) PHENOPSIS, an auto-

mated platform for reproducible phenotyping of plant responses to soil

water deficit in Arabidopsis thaliana permitted the identification of an

accession with low sensitivity to soil water deficit. New Phytol. 169, 623–635.

Guerche, P., Bouchez, D., Balasse, H. and Camilleri, C.. (2010) French Patent:

Automaton for Plant Phenotyping. Patent number PCT/FR2010/050460,

March 2010.

Iyer-Pascuzzi, A.S., Symonova, O., Mileyko, Y., Hao, Y., Belcher, H., Harer,

J., Weitz, J.S. and Benfey, P.N. (2010) Imaging and analysis platform for

automatic phenotyping and trait ranking of plant root systems. Plant

Physiol. 152, 1148–1157.Keurentjes, J.J., Bentsink, L., Alonso-Blanco, C., Hanhart, C.J., Blankestijn-

De Vries, H., Effgen, S., Vreugdenhil, D. and Koornneef, M. (2007) Devel-

opment of a near-isogenic line population of Arabidopsis thaliana and

comparison of mapping power with a recombinant inbred line popula-

tion. Genetics, 175, 891–905.Koumproglou, R., Wilkes, T.M., Townson, P., Wang, X.Y., Beynon, J., Pooni,

H.S., Newbury, H.J. and Kearsey, M.J. (2002) STAIRS: a new genetic

resource for functional genomic studies of Arabidopsis. Plant J. 31, 355–364.

Kroymann, J. and Mitchell-Olds, T. (2005) Epistasis and balanced polymor-

phism influencing complex trait variation. Nature, 435, 95–98.Leister, D., Varotto, C., Pesaresi, P., Niwergall, A. and Salamini, F. (1999)

Large-scale evaluation of plant growth in Arabidopsis thaliana by non-

invasive image analysis. Plant Physiol. Biochem. 37, 671–678.Loudet, O., Chaillou, S., Merigout, P., Talbotec, J. and Daniel-Vedele, F.

(2003) Quantitative trait loci analysis of nitrogen use efficiency in Arabid-

opsis. Plant Physiol. 131, 345–358.Loudet, O., Gaudon, V., Trubuil, A. and Daniel-Vedele, F. (2005) Quantitative

trait loci controlling root growth and architecture in Arabidopsis thaliana

confirmed by heterogeneous inbred family. Theor. Appl. Genet. 110,

742–753.Loudet, O., Michael, T.P., Burger, B.T., Le Mette, C., Mockler, T.C., Weigel,

D. and Chory, J. (2008) A zinc knuckle protein that negatively controls

morning-specific growth in Arabidopsis thaliana. Proc. Natl Acad. Sci.

USA, 105, 17193–17198.Massonnet, C., Vile, D., Fabre, J. et al. (2010) Probing the reproducibility

of leaf growth and molecular phenotypes: a comparison of three

Arabidopsis accessions cultivated in ten laboratories. Plant Physiol.

152, 2142–2157.Poormohammad Kiani, S., Trontin, C., Andreatta, M., Simon, M., Robert, T.,

Salt, D.E. and Loudet, O. (2012) Allelic heterogeneity and trade-off shape

natural variation for response to soil micronutrient. PLoS Genet. 8,

e1002814.



Rockman, M.V. (2012) The QTN program and the alleles that matter for evo-

lution: all that’s gold does not glitter. Evolution, 66, 1–17.Simon, M., Loudet, O., Durand, S., B�erard, A., Brunel, D., Sennesal, F.–.X.,

Durand-Tardif, M., Pelletier, G. and Camilleri, C. (2008) QTL mapping in

five new large RIL populations of Arabidopsis thaliana genotyped with

consensus SNP markers. Genetics, 178, 2253–2264.Skirycz, A., Vandenbroucke, K., Clauw, P. et al. (2011) Survival and growth

of Arabidopsis plants given limited water are not equal. Nat. Biotechnol.

29, 212–214.Tisn�e, S., Schmalenbach, I., Reymond, M., Dauzat, M., Pervent, M., Vile, D.

and Granier, C. (2010) Keep on growing under drought: genetic

and developmental bases of the response of rosette area using a

recombinant inbred line population. Plant, Cell Environ. 33,

1875–1887.Trontin, C., Tisn�e, S., Bach, L. and Loudet, O. (2011) What does Arabidopsis

natural variation teach us (and does not teach us) about adaptation in

plants? Curr. Opin. Plant Biol. 14, 225–231.Walter, A., Scharr, H., Gilmer, F. et al. (2007) Dynamics of seedling growth

acclimation towards altered light conditions can be quantified via

GROWSCREEN: a setup and procedure designed for rapid optical pheno-

typing of different plant species. New Phytol. 174, 447–455.Zhang, X., Hause, R.J. Jr and Borevitz, J.O. (2012) Natural genetic variation

for growth and development revealed by high-throughput phenotyping

in Arabidopsis thaliana. G3 (Bethesda), 2, 29–34.



phenoscope: an automated largescale phenotyping platform offering high spatial homogeneity

Documents