Evaluation of Appropriate Reference Genes for Reverse Transcription-Quantitative PCR Studies in Different Tissues of a Desert Poplar via Comparision of Different Algorithms

Despite the unshakable status of reverse transcription-quantitative PCR in gene expression analysis, it has certain disadvantages, including that the results are highly dependent on the reference genes selected for data normalization. Since inappropriate endogenous control genes will lead to inaccurate target gene expression profiles, the validation of suitable internal reference genes is essential. Given the increasing interest in functional genes and genomics of Populus euphratica, a desert poplar showing extraordinary adaptation to salt stress, we evaluated the expression stability of ten candidate reference genes in P. euphratica roots, stems, and leaves under salt stress conditions. We used five algorithms, namely, ΔCt, NormFinder, geNorm, GrayNorm, and a rank aggregation method (RankAggreg) to identify suitable normalizers. To support the suitability of the identified reference genes and to compare the relative merits of these different algorithms, we analyzed and compared the relative expression levels of nine P. euphratica functional genes in different tissues. Our results indicate that a combination of multiple reference genes recommended by GrayNorm algorithm (e.g., a combination of Actin, EF1α, GAPDH, RP, UBQ in root) should be used instead of a single reference gene. These results are valuable for research of gene identification in different P. euphratica tissues.


Introduction
Gene expression characterization at the steady-state mRNA level is an important step in many plant biological processes. Northern blotting [1][2][3][4], is not an optimal method for gene expression quantification anymore, because it requires a large amount of RNA and is limited to detecting high-abundance transcripts. Therefore, reverse transcription-quantitative polymerase chain reaction (RT-qPCR), which can detect low-abundance mRNAs and exhibits high specificity, sensitivity, and stability [5][6][7], has become one of the most commonly used techniques [8]. However, RT-qPCR has certain disadvantages; for example, reference genes are important to correct the experimental error that occurs during RNA extraction, and cDNA preparation, but the normalized results are highly dependent on the expression stability of the reference genes selected [9]. The use of inappropriate reference genes can result in biased target gene expression profiles; thus, using suitable reference genes for data normalization is a prerequisite for any RT-qPCR study [10]. The expression of an ideal reference gene should not vary between different tissues or cells under investigation, nor in response to any experimental treatment [11]. However, extensive transcriptomic data mining and gene expression analysis studies have reported that the reliability of these endogenous controls can be influenced by various conditions [12][13][14][15][16][17]. Therefore, to obtain accurate and reliable target gene expression profiles, it is important to identify the best reference genes for use in each experimental set-up and verify these in each individual experiment before performing gene expression normalization. In common plant species, including Arabidopsis, rice, soybean, and petunia, genes associated with basic cellular processes such as those related to ubiquitin degradation, polyubiquitin, ubiquitin-conjugating enzymes, and ubiquitin ligases (collectively known as UBQ), as well elongation factors (EFs) [18][19][20][21], have been identified and used for RT-qPCR [6,12,13,19]. Nevertheless, few reference genes have been identified in forest trees, which dominate much of the terrestrial landscape on earth. Members of the genus Populus are used as a model forest species for diverse studies because of their amenability to experimental and genetic manipulation [22][23][24][25]. Studies in Populus focus mainly on improving their ability to respond to environmental stress while maintaining their high-speed growth capacity because there is strong demand for their cultivation in highly saline and arid soils in many parts of the world [26,27]. Populus euphratica Oliv., which is native to desert regions ranging from western China to North Africa, is the only arboreal species established in the world's largest shifting-sand desert, the Taklimakan Desert in China [28].
The most significant characteristic of P. euphratica is its extraordinary adaptation to salt stress [28][29][30]. This species can survive concentrations of NaCl in nutrient solution up to 450 mM while maintaining higher growth and photosynthetic rates than other poplar species [31][32][33]. Previous studies of P. euphratica have mainly focused on analyses of salt-responsive genes [26,30,34,35], miRNAs [36,37], or salt-related transcriptome sequencing [26,38,39]. In these studies, the genes up-regulated or down-regulated by salt stress are important and should be identified using validated reference genes by RT-qPCR.

RNA Quality and PCR Specificity
This study was performed in accordance with the minimum information for publication of reverse transcription-quantitative PCR experiments (MIQE)-guidelines (Table S1). P. euphratica saplings (Figure S1) were exposed to salt stress for 0, 1, 3, 6, 9 and 12 h and then leaves, stems and roots were sampled for RNA extraction ( Figure 1). The time-point "0" was used as a reference point, and sampling at further time-points represents a combination of salt-induced changes in gene expression as well as diurnal changes. Based on agarose gel electrophoresis, the intensity of the 25S rRNA band was nearly twice that of 18S rRNA band, and no genomic DNA was observed ( Figure S2). Using a spectrophotometer, the OD260/OD280 ratios of total RNA were between 1.8 and 2.0, and the OD260/OD230 ratios were greater than 1.5. Using an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA) for total RNA integrity detection, no evidence of degradation was observed. After cDNA synthesis, qPCR using the primers listed in Table 1 was performed, and to verify the specificity of these primers, the amplified products were analyzed using 2% agarose gel electrophoresis and only one band of the expected size was observed in each experiment. Meanwhile, also the presence of a single peak in the melting curve was observed ( Figure S3).   [32]; c [26] Genes used to support the suitability of the identified reference genes; d GenBank accession number; e The RT-qPCR efficiency was determined by standard curve method and then confirmed by LinReg PCR program.

Cq Value Distribution and Expression Profile of the Ten Reference Genes
The detailed information of all the genes, primers and amplicons can be seen in Table 1 and  Table S2. All RT-qPCR assays were performed on biological triplicate samples to determine the expression rates of the selected genes. Firstly, for each candidate reference gene the average Cq values of these triplicates (for each condition and tissue) were visually presented using a histogram. This allows the expression levels of the genes to be easily compared ( Figure 2 and Figure S4). One-way analysis of variance results were used to show differences in measured non-normalized expression. The Cq values ranged from 19.61 to 32.36 in root, stem, and leaf samples. UBQ showed the highest expression level irrespective of whether all tissues or just one tissue were considered. The Cq values of most of the genes ranged from 22 to 26 except TUB and UBQ. TUB showed the lowest expression level with a mean Cq value of 29.58 while UBQ overall showed the highest expression. When looking at these non-normalized data, the observed differences in measured expression within one gene under salt stress are due to both technical and biological variation. Normalization of target gene expression with reference genes will eliminate technical variation, but also erroneously impose any biological variation in reference gene expression. Therefore the expression reference genes should be minimally influenced by the salt treatment. However, these results demonstrate that the reference genes were not perfectly stable under salt stress. Indeed, considering the gene expression variation under salt stress in each tissue type, HIS in leaf (0.54 cycles), RP in stem (0.60 cycles), and UBQ in root (0.51 cycles) showed the lowest variance, while TUB in leaf (2.23 cycles), GIIα in stem (3.33 cycles), and TUB in root (2.46 cycles) showed the highest variance. This different variance in measured expression for these ten reference genes under salt stress cannot only be due to technical variation but shows that there is a biological component in reference gene variation. It is necessary to identify reference genes that have a minimal biological variation component prior to performing target gene expression studies.

Statistical and Bioinformatical Analyses of Gene Expression Stability Using ΔCt, NormFinder, geNorm, RankAggreg and GrayNorm
Given the indications for variations in expression, it was important to evaluate the stability of the ten candidate reference genes. Data from leaves, stems, or roots was analyzed separately. Many approaches exist, and to compare these, statistical and bioinformatical analyses of the data were performed according to the flowchart shown in Figure S5.

ΔCt Algorithm
In the ΔCt algorithm [42], if the ΔCt value between two genes (the Cq value of one gene minus the Cq value of a second) remains constant when analyzed in different cDNA samples, then either both genes are stably expressed among the samples or they are co-regulated. Meanwhile, if the ΔCt value fluctuates, then one or both genes are variably expressed. By introducing a third, fourth or fifth gene into the comparisons, more information is available on pairs showing less variability, and they can then be ranked or discarded [42]. Whether the distribution of ΔCt values between one gene and the other nine genes was discrete or concentrated was visually presented using a box-plot. Figure 3 and Table 2 use this ΔCt approach to reference gene selection among different tissues under salt stress.
The expression stabilities were evaluated via mean standard deviation (mSD), and the mSD values were calculated through ΔCt values between one gene and the other nine genes. The smaller mSD value of one candidate reference gene indicates more stable expression. By calculating each gene's mSD of the ΔCt values, we obtained the final rankings; the most stable gene showed the top ranking (Table 2).  Similarly, the shorter box means smaller deviation, indicating that both genes showed relatively stable expression or were co-regulated with each other.  . geNorm-based evaluation of candidate reference gene expression stability (the left three line charts) and determination of the optimal number of reference genes for normalization (the right three histograms). The lowest values of the expression stability measure M indicate the most stable expression. The pairwise variation Vn/n + 1 value is an indication of number of genes to include for normalization. Additional reference genes should be included in the analysis until the Vn/n + 1 value is below the 0.15 threshold, or until a minimum is reached.

Normfinder
Secondly, the NormFinder tool was used to identify the most suitable reference gene [43]. At least three genes and two samples per group were required for analysis using this algorithm. In our study, ten genes and six samples per group were performed. Based on amplicon efficiency, the Cq values were log-transformed and used as input [43]. Table 3 indicates that in leaves under stress, RP was the most stable gene (stability value = 0.024), while Actin was the most unstable gene (stability value = 0.164). UBQ and eIF-5A were the most stable genes in stem and root, respectively.

geNorm
For the geNorm tool, Cq values were transformed to quantities based on E −ΔCq (E = efficiency), and in each gene among different samples the highest quantities were set to 1 to obtain the input [11]. The average expression stability value, M, was calculated to rank the stability of the candidate genes. The genes with the lowest M value show the most stable expression. However, it should be noted that the strongly co-regulated genes also show low M values. The geNorm algorithm recommended an M value below a threshold of 1.5 for a gene to be considered stably expressed. However, a stricter standard using 0.5 as the threshold has also been accepted [20,44]. Based on the ranking results shown in Figure 4, almost all genes showed highly stable expression in leaf and root. According to geNorm analysis, the most optimal reference genes for RT-qPCR normalization were 60S and UBQ in leaf, and eIF-5A and UBQ in stem, GAPDH and GIIα in root.
Quantification of target gene expression relative to multiple reference genes requires calculation of the normalization factor (NF), the geometric mean of the expression of combined reference genes. geNorm software computes NF values. To determine the minimum number of reference genes to be used for normalizing the experiment, the pairwise variation (PV) of two sequential NFs (Vn/n + 1) was used to estimate the effect of introducing additional reference genes to the NF [11]. If the PV value for n genes (Vn/n + 1) is below a threshold of 0.15, n can be considered the minimum number of reference genes. Alternatively, n + 1 genes can be used when the (Vn/n + 1)-graph reaches a minimum [45]. In our study, the expression of the ten genes varied greatly in stems, such that V5/6 was used in stems. V2/3 was used in leaf and root, but it is advisable to use at least three genes to reduce the chance of selecting co-regulated genes. Thus, three reference genes should be used for normalization in leaf and root, and six in stems (n + 1 at minimum Vn/n + 1).

RankAggreg
Since the different methods showed different rankings results for each gene, we used the RankAggreg statistical method to create an aggregate order to obtain a final list of genes for each tissue. The rank positions generated using the three statistical approaches were merged, including mSD values of ΔCt, stability values of NormFinder, and M values of geNorm. The results in Figure 5 indicate that the most adequate genes tested for normalization in leaf were RP, HIS, and eIF-5A. In stem, the best three candidate genes were UBQ, eIF-5A, and HIS. In addition, eIF-5A, GAPDH and UBQ were top-ranked in root.

Figure 5.
Rank aggregation of ten genes lists using the Monte Carlo algorithm. Visual representation of rank aggregation using RankAggreg with the Monte Carlo algorithm and Spearman footrule distances. The ten candidate reference genes were ordered based on their rank position according to three stability detection methods, ΔCt, NormFinder and geNorm (gray lines). The mean rank position of each gene is shown in black, while the model computed using the Monte Carlo algorithm is indicated by a red line.

GrayNorm
GrayNorm is a recently published algorithm that finds a combination of reference genes that have the smallest possible deviation from the non-normalized data, thus are carrying the least biological variation while they are used for correcting technical variation [46]. It is based on calculating normalization factors (NFs) for each treatment group and for each possible reference gene combination. The closer the averages of the 1/NF per treatment groups are to 1.0, the less biological variation is carried and the more accurate the expression levels of GOI (genes of interest) can be calculated. There are two indices that can be employed to rank the combinations of genes, cvinter (coefficient of variation of the 1/NFs averaged per condition), and cumulative deviation over all conditions from 1.0 of the 1/NF. As suggested, cvinter may be overruled and the cumulative deviation can be used [46]. In our study "cumulative deviation from 1.0 of the 1/NF averaged per condition" was applied and the results in Table 4 and Table S3 indicate that a combination of EF1α, HIS and RP should be used in leaf, a combination of 60S, RP, eIF-5A and GAPDH (first combination of at least three reference genes) can be used in stem, and a combination of Actin, EF1α, GAPDH, RP, and UBQ should be used in root. It is worth mentioning that GrayNorm can evaluate all possible combinations of the ten reference genes by one operation while the other three algorithms can not. It is also noteworthy that the ΔCt method, Normfinder and RankAggreg yield in all tissues the same outcome.

Validation of the Stability of Selected P. euphratica Reference Genes via Differential Gene Expression Analysis of Nine Putatively Salt Responsive Genes
The relative expression levels of nine P. euphratica functional genes, PeCOBL4, PeFLA12-1, PeFLA12-2, PeFLA12-3, PeFLA12-4, PeHKT1, PeKUP3, PeNhaD1 and PeNHX2, in different tissues were analyzed by RT-qPCR. Non-normalized expression levels were compared with normalized expression levels using the reference genes selected by the various algorithms as displayed in Table 4. The non-normalized data change by using the normalization factor based on expression of the selected reference genes comprised in a normalization factor (NF). Various NFs have various effects on gene of interest expression. The smaller the difference between the normalized and non-normalized data, the more accurate the results are [46]. Figure 6 shows the induction or repression of genes by salt stress, some genes were initially induced then repressed. However, the maximum values and expression fold changes differed significantly when normalization was performed using different combinations of reference genes as proposed by the different algorithms. For example, it is known that the expression patterns of the FLA genes vary quite consistently according to the plant tissues being assessed. The mRNAs encoded by these four genes were previously found to be highly expressed in P. euphratica under salt stress [32]. As expected, steady-state mRNA levels for the FLA genes were induced by salt stress significantly ( Figure 6). When the relative expression levels of nine functional genes were normalized with the combination recommended by GrayNorm, the fold changes were smaller comparing with ΔCt, geNorm, NormFinder or RankAggreg, and the values were closer to the non-normalized data. This indicates that GrayNorm, at least in this experiment, suggests the most optimal combination of reference genes giving the highest possible accuracy.       Table 4 recommended by different algorithms. The expression levels were log2 transformed for easily comparing the up-or down-regulation, and the "1" in Y-axis and "0" in X-axis which show "1" fold change means unchanged. A to I represent nine different genes. Error bars indicate standard errors (n = 3, Bars ± SE). Columns labeled with letters "a,b,c…" indicate significant differences at p < 0.05 between different expression levels.

Discussion
Although powerful techniques, including microarrays and high-throughput measurements, have been developed to detect gene expression levels, RT-qPCR is commonly used in many laboratories. Reliable RT-qPCR data depends on suitable reference genes, which must have highly stable expression under different experimental conditions. The validation of suitable reference genes is crucial, and has been performed previously in plants [6,12,13,[18][19][20]. However, in woody plants the identification of reference genes is limited to maritime pine [47], conifers [48], Eucalyptus [49], Jatropha curcas [50] and two species of Populus, namely, P. trichocarpa × Populus deltoids [51] and Populus × euramericana [40]. To facilitate the genetic improvement of woody plants for growth in saline soils, we performed reference gene evaluation in the salt tolerant tree P. euphratica and found that none of the reference genes showed perfectly stable expression (Figures 2 and Figure S4). For example, Actin had 3.22 cycles of variation in stem; if a target gene is normalized with Actin or a 0.3 cycle variation reference gene, the final relative  quantification will have a 2 3.22−0.3 = 7.57-fold discrepancy. This may fundamentally change the target gene expression profile. Thus, candidate genes showing high-level variation should be avoided as endogenous controls. More importantly, in our pre-experiments 18S did not have significant variation, and it is often used as a candidate reference gene, such as in two Populus species for reference gene validation [40,51]. However, since the expression level of 18S was too high compared with other genes, the difference value was close to ten cycles, resulting in a 1024-fold discrepancy. To the best of our knowledge, most functional genes' expression levels are less than that of 18S, and one of the principles in selecting suitable reference genes is that the expression level of the target and reference genes should be similar. For this reason, 18S is rarely used as a reference gene, so we omitted it in our study and did not take it as an appropriate candidate reference gene. To further analyze the expression stability of candidate genes, several algorithms, including geNorm [11], NormFinder [43], BestKeeper [10], ΔCt [42], qBasePlus [52], RefFinder [53], GenEx [54], GrayNorm [46], single-factor analysis of variance, and linear regression analysis [51] have been applied. Among these algorithms, geNorm, and NormFinder have the greatest impact, and the latter one is superior. NormFinder and geNorm are free of charge. Meanwhile, geNorm has been integrated into the qBasePlus and GenEx tools, both of which are powerful RT-qPCR data analysis tools. For our data analysis, we chose the ΔCt algorithm, the two most popular tools and a new algorithm, GrayNorm, to evaluate reference gene stability in P. euphratica. The ΔCt algorithm is a primary and relatively simple approach to rank gene stability based on the SD of pair-wise differences in Ct values. This uses only Cq values for calculation, while the other three tools introduce both Cq and efficiency values. In practice, PCR efficiency is an important factor that fluctuates along with coexisting substances originating from the RNA extract and cDNA synthesis, lengths and base sequences of the primers or amplification targets, quality of the Taq enzyme, and performance of the instrument. Although these factors are not considered in the ΔCt method, the simplicity of the algorithm makes it advantageous to use. The ΔCt method can be used for the initial screening of reference genes from a small number of candidates since it requires manual calculation, while the other three methods can perform automated calculations.
When determining the stability of the ten reference genes, the five methods mentioned above yielded different ranking patterns. Each method has its own algorithm, some of which are Cq-based (ΔCt) while others are quantity-based (NormFinder, geNorm and GrayNorm). In addition, even methods with the same base algorithm showed discrepancies; for example, when transforming Cq values to quantities to obtain input, NormFinder distributes the three Cq value repeats to three groups, while geNorm uses the average Cq values of the repeats, and GrayNorm provides all the gene combinations at one time. Meanwhile, because in all RT-qPCR analysis, GOI data are divided by NF values during normalization, the algorithm of GrayNorm based on 1/NF is more visualized. The closer the average 1/NF per sample group to 1.0, the more stable of the reference gene [46]. More importantly, GrayNorm displays all the gene combinations at one time from which the optimal combination can be easily found, so if only one algorithm is required, GrayNorm is a fine choice for RT-qPCR data analysis.
In previous studies on Populus reference gene evaluation, P. trichocarpa × Populus deltoids [51] and Populus × euramericana [40], both experimental conditions are in different development stages, such as overwintered terminal vegetative buds, shoot tips including unexpanded leaves and terminal vegetative buds [51], or adventitious rooting of Populus hardwood cuttings [40], neither of them is involved in abiotic stress. In our study, we focused on the extraordinary characteristic of P. euphratica on salt stress resistance. The appropriate reference genes identified in our study were different with those in above mentioned Populus. UBQ, ACT2 and 18S were the best three candidate reference genes in P. trichocarpa × Populus deltoids, EF1α and 18S were the optimal two ones in Populus × euramericana. Meanwhile, 18S was the optimal reference gene in potato [55], however, in our study, 18S was omitted because of its too-high expression level. ACT2 (Actin) and EF1α were middle-ranked, and only UBQ was top-ranked in both P. trichocarpa × Populus deltoids and P. euphratica.
Moreover, UBQ and eIF-5A were two of the best choices for use in all tissues in our study, basically corresponding to the other species, such as rice [56], soybean [18], Oryza sativa [57], lettuce [58] and Arabidopsis [59]. This may suggest that reference genes validated in P. euphratica have commonality with other species, and these genes are conservative. Of even greater interest, HIS can be considered as a novel reference gene appropriate for RT-qPCR in P. euphratica. However, Actin, widely used as endogenous control during plant development, should be avoided in P. euphratica under salt stress.
In P. euphratica DNA microarray, GIIα have a constant expression in all experiments [60], and this indicates that it can be regarded as an ideal reference gene. However, in our study, GIIα ranks middle-level in all tissues. Its best ranks are fourth in both root and total, behind the widely used reference gene, UBQ and eIF-5A. Because the DNA microarray was focused on water deficit stress while our research conditions are salt stress, this indicates that in different experiment treatments, there are different optimal reference genes. As water deficit leads to osmotic stress while salt condition leads to both osmotic and ion stress [61], maybe GIIα is involved in P. euphratica response to ion stress, and this may have an influence on its expression stability. For GAPDH, it ranks seventhly in twenty-one candidate reference genes in Arabidopsis (GAPDH, AT1G13440) during development [13], and shows middleranks in Eucalyptus [49], berry [62], and tomato [63] during different environment conditions, even in human and mouse cells [64], and the validation results were similar with ours.
Most of the nine genes' expression levels were salt-induced as seen in Figure 6, indicating that they may participate in a salt-related mechanism in P. euphratica. The salt-resistance contribution may be somewhat reflected by the transcript levels in plants. PeNHX2, an NHX-type Na + /H + exchanger gene, reported to play a role in mediating sodium tolerance in P. euphratica [65,66], showed a much higher level (fold change) than the other eight genes in leaf, and this indicated that PeNHX2 might have a principle role in maintaining ion balance under salt conditions. The results in Figure 6 indicate that using different combinations of reference genes identified from various algorithms can influence target gene relative expression levels, and may lead to statistical significance. More importantly our results indicate that a combination of multiple reference genes recommended by GrayNorm algorithm, as shown in Table 4, should be used instead of a single reference gene. Furthermore these identified functional genes can be applied in future research in understanding of Populus salt-response signaling.

Plant Materials and Salt Stress Treatment
In this study the RT-qPCR experimental steps were performed according to the protocols [67] and MIQE guidelines [68]. Seedlings of P. euphratica were grown in the Xinjiang Uyghur Autonomous Region of China, where they are native, after which they were transferred to an open greenhouse at Beijing Forestry University in late March the following year. They were then transplanted in individual pots containing loam soil. Each pot contained three to five individuals. The potted plants were well-irrigated according to the evaporative demand, and were grown at a natural temperature, humidity level, and photoperiod in the field for three months prior to the beginning of hydroponic culture ( Figure S1). In mid-July, uniformly developed plants were washed free of soil and transferred to individual pots containing 5 L of one-quarter strength Hoagland's nutrient solution in a closed greenhouse. The temperature in the greenhouse was 20-25 °C with a 16 h photoperiod (6 AM-10 PM) and 150 μmol·m −2 ·s −1 of photosynthetically active radiation. The nutrient solutions were renewed every 24 h and the plants were raised for 20-25 days under hydroponic conditions until they showed long white primary roots. The plants were then exposed to 350 mM NaCl for 12 h. The required amounts of NaCl were added to Hoagland's nutrient solution, and plants grown hydroponically without NaCl were used as controls. Stress physiology parameters were measured at 0, 1, 3, 6, 9 and 12 h after stress treatment. The net photosynthetic rate, stomatal conductance, intercellular CO2 concentration, and transpiration rate were measured using the Li-6400 Photosynthesis System (Li-Cor Biosciences, Lincoln, NE, USA). Mature leaves from the same position, epidermis on stems, and primary roots were collected at 0, 1, 3, 6, 9 and 12 h after stress treatment. All samples were examined with three biological replicates and were immediately frozen in liquid nitrogen and stored at −80 °C.
As the factors which affect the gene expressions are salt stress and diurnal changes, we are looking at a combination of both of them. Based on the principle of single factor under material treatments, we strongly recommend the other researchers to actually include time-point controls in their own experiments, and analyze reference genes and genes of interest in treated and control samples at each time-point [46].

RNA Extraction, Quality Control, and cDNA Preparation
Stored samples were ground in liquid nitrogen to powder with a mortar and pestle. Total RNA was extracted using the cetyl trimethyl ammonium bromide (CTAB) method for plants [69]. Genomic DNA was eliminated by treating the RNA samples with DNase in the last step of the extraction process and with gDNase in the first step of the reverse transcription process. The quantity and quality of RNA were assessed with a NanoDrop2000 spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA) by detecting the OD260/OD280 and OD260/OD230 ratios, respectively. The RNA samples were also examined using 1% agarose gel electrophoresis for 15 min, and the integrities were examined using an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA) according to the manufacturer's instructions. Equal amounts of total RNA (1.8 μg) were reverse-transcribed using a TIANGEN FastQuant RT Kit (with gDNase) (Qiagen, Hilden, Germany) according to the manufacturer's protocol; 20 μL of purified cDNA samples were diluted approximately 1:10 to the same concentration in RNase-free water.

Selection of P. euphratica Candidate Reference Genes and Functional Genes, and Primer Design
GIIα has already been used as a reference gene in the gene expression analysis of P. euphratica under water deficit for four weeks [60], and was added in this study. From the transcriptomics of salt-responses [30][31][32]38,70,71] and the whole sequenced genome data of P. euphratica [26], half of the candidate reference genes can be found, and all the nine functional genes were chosen. Moreover, our results were compared with other forest species such as Eucalyptus [72] on the reference gene validation conclusions [49]. In general, half of the candidate reference genes were widely used in model plants like Arabidopsis and rice, and the others were come from above mentioned transcriptomics of P. euphratica. Ten potential housekeeping genes were selected as candidate reference genes based on the previous research [40,41]. Nine functional genes, PeCOBL4, PeFLA12-1, PeFLA12-2, PeFLA12-3 and PeFLA12-4 were selected from transcriptional data of P. euphratica [32]. PeHKT1, PeKUP3, PeNhaD1 and PeNHX2 were selected from the genome data of P. euphratica [26]. To compare the stability of the reference genes between different tissues or just in leaves, primer sequences from previous studies were used [41]. The annealing temperature of the primers was 60 ± 3 °C and the length was 18-30 bp. The amplicon was between 80 and 200 bp while the GC content ranged from 40% to 60%.

RT-qPCR Reaction Conditions
RT-qPCR was conducted in triplicate in 96-well plates with an ABI StepOne Plus instrument (ABI, Carlsbad, CA, USA). A TIAGEN SuperReal PreMix Kit (Qiagen, Hilden, Germany) with SYBR Green detection chemistry was used. The reaction volume was 20 μL containing 1 μL of diluted cDNA (corresponding to 9 ng of total RNA), 0.3 μM each primer, 10 μL of 2× PreMix Plus, and 2 μL of ROX Reference Dye. The cycling parameters were 95 °C for 15 min, followed by 45 cycles of 20 s at 95 °C and 60 s at 60 °C. Melting curves for each amplicon were then performed by heating the amplicon from 40 to 100 °C and reading at each temperature to verify the specificity of each amplification reaction. Negative controls without template were performed at the same time. The PCR products were also examined on a 2% agarose gel for 30 min to determine the amplicon size and specificity of the amplification reaction.

Determination of Reference Gene Expression Stability Using ΔCt, NormFinder, geNorm and GrayNorm
A flowchart was constructed to explain the data analysis strategy. Four calculation methods, including ΔCt [42] and three publicly available tools, NormFinder [43], geNorm [11] and GrayNorm [46] were used for expression stability validation of the ten genes in P. euphratica tissues, referred to previous study on Coffea species. [73]. The ΔCt approach was employed by comparing the relative expression of "pairs of genes" within each sample. ΔCt uses the raw Cq values as input. The mean standard deviation (SD) values of the ΔCt values (mSD) were used to rank the expression stability of the ten genes. NormFinder software (http://moma.dk/normfinder-software) is a Visual Basic Application (VBA) based on Excel for ranking the expression stability of reference genes. The lowest calculated value indicates the most stable expression. geNorm (http://medgen.ugent.be/~jvdesomp/genorm/), another VBA applet, was also used to rank the candidate reference genes based on the average expression stability values index, M. Cq values were transformed to obtain input, and the most stable gene with the lowest M value was ranked on the right. geNorm software can also be used to calculate another important index, the pairwise variation (PV), to determine the optimal number of control genes for use in normalization [11]. Download links, summaries, or detailed directions for all these methods can also be found at the website (http://www.gene-quantification.de/download.html) or for GrayNorm (https://github.com/gjbex/GrayNorm #graynorm).

Gene Stability Rankings Merged Using RankAggreg
The RankAggreg v0.4-3 (http://cran.r-project.org/web/packages/RankAggreg/) package of R program v3.0.1 (http://www.r-project.org/), which provides an easy and convenient interface to handle complex rank aggregation problems [74], was used to merge the stability measurements obtained from the three methods and establish a consensus rank of reference genes. There are two different algorithms found in RankAggreg, and according to the size of the ranking list (ten), we used the Cross-Entropy Monte Carlo algorithm (CE) to visually present the rank aggregation using a line chart. The rankings from four terms, mSD, stability, coefficient of variance (CV), and M, were used as input. The most stable two gene orders evaluated using geNorm were distinguished based on the initial M value when calculating the normalization factor (NF) value. The other bioinformatics and statistical analysis softwares used in this study including SPSS v19.0 (http://www-01.ibm.com/software/analytics/spss/), SigmaPlot v12.5 (http://www.sigmaplot.com/) and qBasePlus v2.5 (http://www.biogazelle.com/).

Expression Analysis of Nine P. euphratica Functional Genes
To explore the salt responses in the 18 P. euphratica samples and illustrate the suitability of those identified reference genes, we analyzed the relative expression levels of nine P. euphratica functional genes; namely, PeCOBL4, PeFLA12-1, PeFLA12-2, PeFLA12-3, PeFLA12-4, PeHKT1, PeKUP3, PeNhaD1 and PeNHX2, in different tissues using the optimal or worst reference genes.
PeCOBL4 gene which encodes an extracellular glycosyl-phosphatidylinositol-anchored protein, showing a distinctly different expression level in Populus × canescens (P × c) and P. euphratica (P. eu) under salt stress, and in P. euphratica the expression level is high [32]. PeFLA12-1, PeFLA12-2, PeFLA12-3 and PeFLA12-4 displayed similar expression patterns with PeCOBL4 in two Populus species [32]. In P. euphratica the Na + /K + transporter PeHKT1 is similarly expanded as tandem duplicated copies in T. salsuginea [75] and T. parvula [76], and it responds to salt stress rapidly [26]. PeKUP3 and PeNhaD1 were significantly up-regulated under salt stress when compared with salt-sensitive poplar, P. tomentosa, and both of them are involved in ion transport and homeostasis [26,65]. The transcript levels of PeNHX2 were up-regulated in roots after NaCl treatment for 6 h, and heterologous expression of PeNHX2 in yeast mutant strain R100 can improve salt tolerance [66].
The single best or worst reference gene was based on RankAggreg results, while the combination was recommended by both geNorm and RankAggreg. Cq values and amplification efficiency values were processed using the qBasePlus software [52]. For calculating the relative expression levels of the nine functional genes based on a combination of multiple reference gene, such as the six top-ranked reference genes, HIS + eIF-5A + UBQ + GIIα + GAPDH + EF1α. Firstly, we calculated out the relative quantification (RQ) of a functional gene, such as PeCOBL4, using E ΔCt method. E was the PCR efficiency of PeCOBL4 gene, 1.887, as shown in Table 1. ΔCt means the Cq values of stressed plants (1, 3, 6, 9 and 12 h in the stress stage) take away the Cq values of non-stressed plants (0 in the stress stage). Secondly, a "Normalization factor" (NF) needed to be calculated, the "NF" was calculated by geNorm or qBasePlus (geNorm is incorporated into qBasePlus now) software after log-transformed reference genes' Cq values were input, and the first column are reference genes' names and the first row are samples' names. Then dividing the "RQ" value of PeCOBL4 by the "NF" value, we get final relative expression levels. This is also the working principle of qBasePlus and geNorm according to the manuals.

Conclusions
Overall, to identify suitable reference genes, we propose using the ΔCt method to examine gene expression stability before subsequent evaluation by NormFinder, geNorm, GrayNorm, or other software programs. Our results provide an important reference gene selection guide when performing gene expression analysis in different tissues of P. euphratica.