The exapted TEs (ETEs) analyzed in this study were selected from the study conducted by Hoen and Bureau (2015). Reference genes used in the assays were pre-selected from the literature as already having a phenotype under similar or identical phenotyping conditions. The availability of T-DNA insertion mutant alleles for each ETE and reference gene was identified through The Arabidopsis Information Resource (TAIR; Stanford University, Stanford, CA USA) genome browser (version 10) (Figure S1). Alleles were selected according to the likelihood that they would disrupt the ETE or reference gene coding potential (Table S1). If available or applicable, ETEs, or reference genes with previously described T-DNA insertion mutants were used.

All seeds (ETE mutants and reference lines) were obtained directly from the Arabidopsis Biological Resource Center (Ohio State University, Columbus, OH USA) or through TAIR portal. Wild-type A. thaliana ecotype Col-0 seeds were originally obtained from Lehle Seeds (Round Rock, TX USA). Bulked seeds were dried for 10 days after harvest and stored at 4°C. Seeds were vapor-sterilized in a desiccator jar as describe in Bent (2000).

Oligonucleotide primers were designed using the SIGnAL T-DNA primer design tool (http://signal.salk.edu/tdnaprimers.2.html) using default parameters, except Ext5 = 400 was used. Primers were ordered en masse (Integrated DNA Technologies, IA USA) at a 25-nmole scale in a 96-well format. Three primers were used in each PCR with two target-specific T-DNA flanking primers and a primer that anneals to the left-border of the T-DNA (Table S1).

Experimentally validated homozygous lines were used in the study. Seeds were either grown on square Petri plates containing 40 ml of ½ strength MS (Murashige and Skoog) media, 2.5 mM MES, 1% sucrose, and 0.8% agar or were sown individually into cells (3 cm²) of a plug tray (8 × 12) containing soil (2 parts Pro-Mix: 1 part vermiculite: 1 part perlite; Premier Tech Horticulture, QC Canada) pre-moistened with a fertilizer solution (20N: 20P: 20K). After stratification for 3 days in the dark at 4°C, plates or trays were transferred to a growth chamber (Conviron, MB Canada; model A1000) with constant 16 h light (120 μmole quanta/m²/s): 8 h dark at 22°C and 70% humidity. A leaf from 3-week-old seedlings (diameter <5 mm) was harvested and genomic DNA extracted using a modified NaOH-based protocol (Klimyuk et al., 1993).

PCR was performed in a 25 μl total volume consisting of 1 μl of genomic DNA template (40–110 ng/μl), 1X PCR buffer, 2 mM MgCl₂, 0.1 mM dNTPs, 0.5 U of native taq polymerase (Invitrogen, CA USA), and 0.4 μM of each primer (see above). Reactions were performed on 0.2 ml 96-well thermal cyclers (Applied Biosystems, ON Canada; model Veriti) using the program of 94°C for 1 min, 40 cycles: 94°C for 1 min, 60°C for 1 min, and 72°C for 2 min, 72°C final extension for 10 min, and a 4°C hold. Reaction products were run on 1.2% agarose gels and visualized using a gel documentation station (Bio-Rad, CA USA; model EZ System). Homozygous lines were identified by comparing the observed vs. predicted band sizes. Reactions with wild-type and no DNA were used as controls. Lines that did not yield clear homozygous individual using three primers (see above) were re-genotyped using two separate reactions, one with only the gene specific primer pair and the other the T-DNA and one gene-specific primer pair. Lines that still did not yield any clear homozygous individuals were replanted and re-genotyped. Validated unambiguous homozygous seedlings were repotted or transferred into 4 cm square pots, staked, and allowed to grow to seed set.

Plants were grown at the Phytotron located at McGill University, Department of Biology (Montreal, Quebec, Canada; http://www.biology.mcgill.ca/Phytotron).

For the salt, arsenic and phosphate assays, image acquisition and analyses were performed at the McGill Plant Phenomics Platform (MP3; Montreal, Quebec, Canada; http://mp3.biol.mcgill.ca) using a customized version of the LemnaTec scanalyzer High-Throughput Screening (MP3-HTS) (LemnaTec GmbH, Wuerselen, Germany), where up to 60 plates per run can be analyzed. In this study, images were only taken with the visible light camera piA2400-17gc (VIS) and the fluorescent light camera scA1400-17gc (FLUO). For the freezing assay, image acquisition was done using a Nikon D3100 camera with AF-Micro Nikkor 60 mm 1:2.8G ED mounted on a static support. Image files were stored on a Dell R910 server (MP3 server) with 256 GB of RAM and two MD1200 storage devices (72 TB).

A custom image analysis algorithm was developed using java 1.7.0-45 (http://www.java.com) and ImageJ library v1.48 (http://imagej.nih.gov/ij/) (Schneider et al., 2012) for salt, arsenic, phosphate and freezing. The statistical analysis script was written in R v3.0.2 (Team, 2013) (http://www.r-project.org/) for all the assays. A modified version of the R script published in Camargo et al. (2014) was used to generate the figures Figure 2, Figures S2–S5. PostgreSQL v 9.3.1.was used to build the database. The analyses were implemented on the MP3 server. P-values were adjusted for multiple comparisons using the method of Benjamini & Hochberg of false discovery rate control (Team, 2013). Only lines having an FDR-adjusted P-value lower than 0.05 were considered to have a stress-related phenotype.

The salt assay was adapted from the methods described by Zhang et al. (2011). Three reference salt sensitive genes were used in the assays: SOS2 (At5g35410; SALK_016683) and SOS3 (At5g24270; SALK_137171, CS859749) as reported by Shi et al. (2000), and DDF1 (At1g12610; SALK_127759, SALK_114390) as described by Magome et al. (2008). Seeds (ETE mutant lines, wild-type, and reference lines) were sown on square Petri plates with grid (Simport, Canada) containing MS medium buffered at pH 5.6 with 1% v/w sucrose, 2.5 mM 2-(N-morpholino) ethane sulfonic acid (MES), and 0.8% agar. After 3 days stratification at 4°C in darkness, the plates were placed in a growth chamber (Conviron model A1000, Canada) programmed for a 16 h light (120 μmole quanta/m²/s), 8 h dark photocycle (22°C, 70% humidity). To induce salt stress, 4-day old seedlings were transferred to MS medium supplemented with 150 mM NaCl (Zhang et al., 2011), a concentration corresponding approximately to the LD50 for wild-type. Assays were performed using 36 seeds per plate for each of the three independent biological replicates. For the standard growth condition, lines were grown for 12 days on MS medium without NaCl.

Image acquisition was performed using the VIS and FLUO cameras. Visible light RGB images were taken for each Petri dish using a “one frame one plate” setup with top light illumination. Pixels having saturation (S) of the HSB color space higher than 49 were retained (74 for the standard growth condition). A grayscale transformation was then done keeping only the pixels with values between 0 and 140 and from 150 to 255. The RGB-values of the selected pixels were tagged as foreground to identify the objects using an adapted version of the “combined contour tracing and region labeling” algorithm proposed by Burger and Burge (2008). Objects having an area >40 px and an Euclidean distance lower than 130 to the center of the grid cell were selected. Only the largest object in each grid cell was kept. As a result, each seedling was represented by one object. The hue channel of the HSB color space was divided in 16 categories where each object pixel was classified (Berger et al., 2012). Fluorescent light images were acquired using the same frame configuration as visible light. Intensity images were converted to HSB color space. Pixels with S higher than 20 were tagged as foreground. After identifying the objects using the Burger and Burge adapted algorithm, objects having an area >20 were selected (Burger and Burge, 2008). Only the largest object within each grid cell was kept as representation of the seedling. The same color classification was done as above. An Euclidean distance matrix of the objects was calculated with the color classes of the visible light camera joined to the classes of the fluorescent camera using as input for a hierarchical cluster analysis. The resulting hierarchical cluster tree was divided into two groups. The number of seedlings belonging to each group was counted and tabulated per line. The most representative lines of each group were analyzed to determine the characteristic of each group. To detect the candidate lines, a Fisher's exact test was used on 2 × 2 contingency tables built for each of the lines having the target line and the wild-type as columns and the clustered groups as rows.

The phosphate assay was based on the methods described by Misson et al. (2004). Three reference phosphate responsive genes were used in the assays: LPR1 (At1g23010; lpr1-2, SALK_102704) and LPR2 (At1g71040; SALK_138670) as reported by Svistoonoff et al. (2007), and IPK1 (At5g42810; SALK_065337) as first characterized by Stevenson-Paulik et al. (2005). Seeds sown on square Petri plates (Simport, Canada) containing 1/10 strength MS medium (0.15 mM MgSO₄, 2.1 mM NH₄NO₃, 1.9 mM KNO₃, 0.5 mM NaH₂PO₄, 0.3 mM CaCl₂, 0.5 μM KI, 10 μM Fe, 10 μM H₃BO₃, 10 μM MnSO₄, 3 μM ZnSO₄, 0.01 μM CuSO₄, 0.01 μM CoCl₂, 0.1 μM Na₂MoO₄, 2 μM EDTA, 0.03 μM thiamine, 0.24 μM pyridoxine, 0.4 μM nicotinic acid, 55 μM inositol, and 14 mM MES; pH 5.8), 0.5% sucrose, and 0.8% agar (Murashige and Skoog, 1962). In the Pi limiting medium, a concentration of 5 μM NaH₂PO₄ was added instead of 0.5 mM NaCl was used to replace the equivalent amount of sodium provided by NaH₂PO₄. After 3 days at 4°C in the dark, plates were placed in a vertical position in a growth chamber (Conviron model A1000, Canada) programmed for a 14 h light (22°C), 10 h (18°C) dark photocycle (70% humidity, 120 μmole quanta/m²/s) and grown for 9 days, after which images were acquired. Six seeds per plate were used for each of two independent biological replicates, for a total of 12 plants per line. The standard growth condition consists of growing seedlings as described above on the 1/10 strength MS medium with 5 μM NaH₂PO₄.

Image acquisition was performed using the VIS camera. A setting of one frame for each Petri dish and bottom light illumination was used. Images were transferred to the MP3 server for further analysis. Every image was converted to a grayscale image and the mean auto local threshold algorithm was applied to make a draft separation of the background. The grid embossed on the Petri dish was digitally removed from the image to not confound downstream analysis. Horizontal lines >50 px and vertical lines >240 px were removed. After eroding 1 px size and removing top and bottom border lines from 50 to 260 and 1,698 to 1,798 Y coordinates, the resultant image was re-converted to RGB using the original image as template. A list of objects was created using the same algorithm as used for the salt assay and only the objects having an area >109 px, a circularity lower than 0.70, and an Euclidean distance lower than 145 to the center of the grid cell were retained, representing the root portion of the plant. For the shoot part of the plant, only the pixels having the bright (B) of the HSB color space higher than 23 and lower than 138 were kept from the original images. A list of object was created and only the objects having an area >10 px, a circularity rate >0.099 and an Euclidean distance to the center of the grid cell lower than 130 were considered to belong to the shoot part of the seedlings. The shoot and root objects of each plant were merged into one object using the location of the shoot as a start point and the vertical lines of the grid on the plate as reference. The results provided a list of objects where each of them represents one plant. A one-way ANOVA on the major axis calculated from the digital plant was used between each line and wild-type. The major axis is defined as the axis where a physical body requires less effort to rotate. It extends from the centroid to the widest part of the object (Burger and Burge, 2008).

The freezing stress assay was performed as previously described (Xin and Browse, 1998). The protocol without acclimation and with a freezing temperature of −8°C proved to be a good trade-off both in terms of resolution (discriminating survivability) and stress effect (number of affected seedlings per plate) for high throughput. Two lines for the freezing sensitive reference gene SFR2 (At3g06510; SALK 106253C, and SALK 000226) were used in the assays, as described by Moellering et al. (2010). Seeds were sown on 0.9% agar round plates (100 × 15 mm) with Gamborg's B5 basal salt medium (Gamborg et al., 1968) (PhytoTechnology Laboratories, USA) and 1.5% sucrose. After 2 days of stratification at 4°C in the dark, plates were transferred to a growth chamber (Conviron model A1000, USA) set at 22°C and constant 24 h light (120 μmole quanta/m²/s and 70% humidity) for 10 days. Then, plates were sat on ice in a freezing chamber (Conviron model PGW36, Canada) for 16 h at −1°C in the dark. Ice chips (Hoshizaki ice machine, model F-450 BAB, USA) were then sprinkled directly on seedlings, after which the temperature of the freezing chamber was decreased at a rate of 1°C per hour until it reached −8°C and then held for a duration of 2 h. After the freezing treatment, plates were thawed at 4°C in the dark for 12 h. Then, seedlings were recovered at 22°C under constant 24 h light for 48 h. Afterwards, image acquisition was done as described below. Three independent biological replicates were performed using 60 seeds per plate each time. For the standard growth condition, lines were grown on Gamborg's B5 basal salt medium for 10 days as described above.

For the image analysis method, pixels with blue (B) of RGB color space lower than 127 and (red(R) − blue(B))/(red(R) + blue(B)) (Kawashima and Nakatani, 1998) >0.20 were considered as foreground. The identification of the objects from the foreground was performed using the same algorithm as in the salt assay. Only the objects having an area between 10,000 and 40,000 px were considered as representations of seedlings. The hue channel of the HSB color space was divided in 32 categories where each object pixel was classified (Berger et al., 2012) and the categories were re-converted to RGB. The objects were then divided into two groups according to the amount of leaf damage represented by the yellowish that was calculated from the number of pixels in the color classes 4,12,20,28,36 (hue channel of the HSB color spectrum) over the total number of pixels. The number of seedlings belonging to each group (less or more of 50% of damage) were counted and tabulated per line. The statistical analysis is the same as in the salt assay.

The arsenic stress assay was based on a protocol described by Lee et al. (2003). The arsenic resistant gene ARS5 (At5g42790, ars5-2, CS440215) was used as a reference gene, as described by Sung et al. (2009). Seeds were sown on minimal medium containing 2.5 mM H₃PO₄, 5 mM KNO₃, 2 mM MgSO₄, 1 mM Ca(NO₃)₂, 1 mM MES, 0.5% sucrose, and 0.8% agar (pH 5.7). Macronutrients were replaced with 750 μM potassium arsenate (Sigma-Aldrich Life Science. USA), a concentration determined to be the mean lethal dose (LD50) for wild-type seedlings. Plates were stratified for 2 days at 4°C in the dark and then transferred to a growth chamber (Conviron model A1000, Canada) set at 22°C, constant light (120 μmole quanta/m²/s) and 70% humidity for 5 days after which images were acquired. Three independent biological replicates were performed, where 36 seeds per line were used in each replicate. For the standard growth condition, lines were grown for 5 days on minimal medium without potassium arsenate.

Image acquisition was performed using the VIS and FLUO cameras. The Petri dish covers were not removed during image acquisition to avoid contamination. A setting of one frame one plate for each Petri dish and bottom light illumination was used for the VIS camera. Pixels having red (R) lower than 254, green (G) lower than 246, and blue (B) lower than 254 of the RGB color space were retained. Identification of the objects from the foreground was done using the same algorithm as in the salt assay (see above). Objects having an area >10 px, a circularity ratio >0.099 and an Euclidean distance lower than 130 to the center of the grid cell were selected. Only the largest object within each grid cell was kept. As a result, each seedling was represented by one object. The fluorescent light images were acquired using the same frame configuration as the visible light images. The intensity images were converted to HSB color space. Pixels with S higher than 50 were tagged as foreground. After identifying the objects using the Burger and Burge adapted algorithm (Burger and Burge, 2008), objects having an area >20 were selected. Only the largest object within each grid cell was kept as representation of the seedling. The color classification, clustering, statistical test and filtering used were as described in the salt assay.

The phenotypic results of the mutant lines for each assay are available on our phenomics database at http://mp3.biol.mcgill.ca/dteassays.

As compelling as the features gleaned by computational methods are in arguing a functional contribution for ETEs, it is nevertheless prudent to confirm genotype-phenotype relationships by experimentation. We exploited the availability of T-DNA large insertion mutant populations for A. thaliana (Alonso et al., 2003) to establish one to four homozygous putative knockout lines per candidate ETE, which we used for phenotyping in trait assays (Figure S1, Table S1). The confirmation of the homozygosity of the mutant alleles was achieved using PCR-based genotyping (Table S1). In addition to wild-type A. thaliana (Col-0) and T-DNA mutant lines from reference genes known to affect each trait, we assayed 103 independent T-DNA mutant lines representing 50 of the 67 novel and previously characterized ETEs (Table S2). The remaining novel ETEs could not be assayed due to the limited availability or absence of available T-DNA mutant lines within regions likely to significantly alter the function of the gene. Among these 50 genes, about half (24) were known ETEs with the remainder (26) being novel (Hoen and Bureau, 2015).

We looked at the response of the ETE T-DNA mutant lines grown under standard laboratory growth conditions, and then subjected to four abiotic stress conditions, namely phosphate limitation, freezing temperature, and exposure to semi-lethal concentrations of either arsenic or salt. To make this study more amenable to high-throughput, we designed plated-based assays for each condition (standard and abiotic stress) to assess the A. thaliana ETE T-DNA mutant lines for phenotypes at the seedling stage, and acquired images of each plate using an automated phenomics platform. We also developed a custom phenotypic data analysis pipeline using the acquired images as input for each assay (Figure 1) (see section Materials and Methods). The use of a phenomics platform increased the reproducibility of the measurements, such that different measurements like leaf size, color, and root length, were more comparable (Fahlgren et al., 2015; Vello et al., 2015; Flood et al., 2016) The pipeline of a phenomics analysis requires three steps, namely: image analysis, data mining, and statistical analysis (Figure 1A). Image analysis has two major components: first, segmentation, where the image is partitioned into sets of pixels to select those representing the plant (i.e., the digital plant), and second, the calculation of the morpho-colorimetric features (e.g., projected leaf area), based on the digital plant. We developed specialized segmentation algorithms for each experimental condition (e.g., salinity, freezing), accounting for differences in medium composition, experimental requirements, and growth patterns in each protocol. For example, while for most protocols the plastic plate lids were removed to capture plate images, for the arsenic experiment they were not removed because of safety reasons and, as a result, back light rather than a top light was used to avoid reflection off the lids. Another example is the type of information of interest for each condition: for the phosphate limitation assay the pixels representing the roots were our main interest, whereas for the salinity assay it was the shoots (see section Materials and Methods).

We then used data mining techniques to identify proxies for the main phenotypic traits that were previously identified in studies of phenotypic responses in A. thaliana under the different stress conditions used in this study (Xin and Browse, 1998; Lee et al., 2003; Misson et al., 2004; Verslues et al., 2006; Zhang et al., 2011). Data mining techniques do not require human intervention to evaluate the phenotype and are amenable to high-throughput protocols. For the arsenic and salt assays, a cluster of color classification has been shown to be optimal to assess germination and survivability, respectively (Berger et al., 2012; Vello et al., 2015). In the former case, a seed is represented by a brown color whereas a seedling's predominant color is green. In the latter case, the seedlings that are alive retain some green areas (Figure 1). The data mining results for the arsenic and salt assays were validated by a blind manual scoring of 216 seedlings of random T-DNA lines, and showed over 90% agreement (Figure S2). For the freezing assay, a color classification of yellowish tones was used as a proxy for leaf damage. Finally, no color classification was required for the phosphate limitation assay since the root elongation was the main phenotypic trait (Figure 1). The statistical approaches that were used to analyze the data mining results are described in the following section.

In addition to the 103 ETE T-DNA mutant lines, we also included “reference lines” corresponding to T-DNA mutant lines of well-characterized genes associated with one of the four stresses (Table S2). In total, more than 36,500 seedlings, representing 119 lines, were screened and analyzed (Table S2).

Of the 50 ETEs included in our study, 46 had at least 1 T-DNA mutant line with a significant phenotype under one of the four stress conditions using the main phenotypic trait analysis. To guide our analysis of these TE-derived genes, we created a flowchart to group ETEs into categories (Figure 4). For example, ETEs for which one or more T-DNA mutant lines displayed a phenotype in only one growth condition were labeled as “Category 1,” whereas ETEs having one or more T-DNA mutant lines displaying phenotypes in more than one condition were labeled as “Category 2.” To strengthen the evidence that the phenotypes we detected for the T-DNA mutant lines were linked to the mutated ETEs, we labeled as “Category 3” the ETEs for which only one T-DNA mutant allele was available, or where a mutant allele failed to show a phenotype due to a suboptimal location of the gene-specific T-DNA insertion. ETEs for which two or more T-DNA mutant alleles were available and found significant under a given condition were labeled as “Category 4” (Figure 4). Finally, to single out ETEs that had a detected mutant phenotype principally caused by the exposure to abiotic stresses, we identified ETEs for which more than one T-DNA mutant line was found significant in the same stress assay (if not, counted in “Category 5”). This filtering approach led to a subset of 26 ETEs (“Category 6”) (Figure 4). This stringent subset represents experimentally well-supported ETEs derived from different TE superfamilies, almost half of which (12/26) are previously reported ETEs, while the remainder (14/26) are novel (Hoen and Bureau, 2015).

We found four T-DNA mutant lines that differed significantly from wild-type under standard growth conditions, which correspond to four different ETEs including the known ETEs FRS7 (line 49; P = 0.006) and MUG2 (line 102; P = 0.005) (Table S2; Figure 3). Nonetheless, these two genes still figure in the count number of “Category 6” for ETEs responding to abiotic stress because they had two T-DNA mutant lines showing a strong phenotype under one or more stresses, with P-values differing by an order of magnitude of 3 (FRS7) and up to 19 (MUG2) under a given stress compared to standard growth conditions (Table S2). Also, we previously showed, using the same T-DNA mutant lines as in this study, that MUG1 and MUG2 single mutants show subtle but significant phenotypes under standard growth conditions [37]. However, the phenotypes were detected at a later developmental stage than in this study, and therefore could explain why mutant lines for MUG1 did not appear to be significantly different from wild-type under standard growth conditions (Figure 3).

We were interested in singling out ETEs that appeared to have a strong and more specific response to abiotic stress. We considered a response to be strong and specific if a gene had: (1) at least two T-DNA lines found significant in the same stress assay, and (2) an FDR corrected P-value lower than 0.01 (see section Materials and Methods). Applying these filters resulted in a list of 25 ETEs, with 24 of them belonging to “Category 6” and one to “Category 1” (Table 2). While almost all the ETEs in “Category 6” (24/26) met both filtering criteria for one abiotic stress condition, half of them (12/24) met the P-value filtering of 0.01 for more than one stress condition (Table 2). This suggests that while some ETEs may play a role related to a specific type of stress, others may be responsive to stress conditions in general.

From the ETEs listed in Table 2, we have detected previously unreported phenotypic responses under abiotic stress conditions for a number of well-characterized ETEs. Among the known ETEs, we found that all four members of the MUGA gene family have a salt tolerance phenotype and, in addition, MUG1 (At3g04605) and MUG2 (At2g30640) have a mutant phenotype in the arsenic assay. Moreover, the related FHY3 ETE (At3g22170) also has a salt tolerance phenotype, whereas FAR1-related sequences (FRS) genes have a phosphate limiting phenotype. FRS1 is the only ETE for which we found a strong mutant phenotype under three abiotic stress conditions (i.e., salt, arsenic, and phosphate) (Table 2). With respect to the FAR1/FHY3 gene family, it is worth mentioning that even though the ETE FAR-RED IMPAIRED RESPONSE (FAR1; At4g15090) did exhibit a significant mutant phenotype in three assays (i.e., salt, arsenic, and phosphate), it was not included in the filtered list as only one previously uncharacterized T-DNA line was available when we conducted this study (Table 3). A number of other ETEs, belonging to “Category 1” and “Category 3” also had one T-DNA mutant line available for this study, but based on their phenotype, it would be worth conducting subsequent experiments with additional T-DNA mutant lines for these genes (Table 3). The ETE gene families MUGA, MUGB, and FAR1/FHY3 are all derived from Mutator-like elements (MULEs) DNA transposons (Yu et al., 2000; Lisch et al., 2001; Cowan et al., 2005). Of the 26 novel ETEs tested in this study, 13 passed the final triage, excluding four where only one T-DNA mutant line was available (Tables 2, 3). Although the majority of them are derived from hAT and Pif-Harbinger DNA transposons, we also identified one ETE derived from a Ty1/copia retrotransposon (At1g21260). Nine of these novel ETEs present in the final triage have strong phenotypes in more than one assay (Table 2). Taken together, the identified and previously reported phenotypes suggest that many of the triaged ETEs are pleiotropic (Ouyang et al., 2011). Stress-related phenotype among members of the same ETE family were also detected, as in the case of MUGA (salt) and FAR1/FHY3 (phosphate), which is also observed for TE-mediated stress induced gene expression (Makarevitch et al., 2015).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.