Plant defense mechanisms

Most living organisms can escape harsh environments by seeking refuge in favorable locations however, plants are immobile organisms and have to adapt to these conditions. If plants do not adapt to stressful conditions then their growth, development, yield, and seed quality may be hampered [18]. Plant stress can be categorized into two groups, biotic and abiotic. Biotic stress includes attacks on the plant by herbivores, bacteria, fungi, and other pathogens, whereas under abiotic stress the plant faces detrimental environmental conditions such as extreme temperatures, droughts, and mineral toxicity. Plants defend against such stress by activating complex networks of signaling pathways. These pathways are often activated with the help of small molecules such as Ca²⁺, reactive oxygen species, nitrogen, or phytohormones such as ethylene, jasmonic acid, abscisic acid, and salicylic acid, which serve as biological stress sensors [19]. These pathway activators often initiate a protein phosphorylation cascade to directly target defensive proteins or transcription factors to regulate the stress responsive genes [20]. Under stressed conditions, the natural metabolic homeostasis of plants is disrupted and, by activating the stress signaling pathways, plants achieve a new state of homeostasis; this process is commonly referred to as acclimation [21].

When a plant comes under drought conditions, it typically responds by implementing drought escape, avoidance, and tolerance strategies [22]. Drought escape strategies involve the plant developing high plasticity and completing its life cycle before the onset of drought, whereas under drought avoidance, the plant learns to maintain high water content in its tissues by increasing water uptake and reducing water loss [22–24]. Drought tolerant strategies are characterized by the plant developing traits such as epicuticular wax formation, osmotic adjustment, cellular elasticity, and protoplasmic resistance. These strategies allow the plant to survive in drought conditions with low tissue water content [24]. Plants do not deploy these defensive responses one at a time; instead, they implement a combination of these strategies to cope against drought [23]. Such a diverse range of defensive responses is achieved through the actions of Gene Regulatory Networks (GRNs) [24, 25]. GRNs are complex networks of genetic regulators called Transcription factors and their target genes; GRNs are directly responsible for altering the gene expression of plants when they receive environmental cues such as drought [26]. Due to these reasons, in this paper, we are interested in modeling the various GRNs, that are activated in plants in response to drought. Modeling these genetic interactions will help us establish a deep understanding of how plants deploy phenotypical defensive behavior through the actions of genes and transcription factors. Such a model will also help us identify the key regulators of drought response. The various GRNs involved in drought response in Arabidopsis are described in the following section.

Drought signaling networks

In this paper, we build a BN model from several signaling pathways involved in the drought response of Arabidopsis. Since the plant’s response to drought happens in a complex manner, it is necessary to build a comprehensive network model that can capture the multivariate and stochastic interactions taking place under drought conditions. Drought responses in plants are largely regulated by Abscisic acid (ABA) dependent and independent pathways [27]. ABA acts a sensor of drought in plants. Under drought conditions, the ABA levels increase rapidly in plants which allows them subsequently respond by closing their stomata and inducing drought responsive genes [28]. ABA regulates the expression of these genes through transcription factors in its drought signaling pathway. The basic-domain leucine zipper (bZIP) transcription factor and its subfamily of ABA-responsive element-binding protein/factor (AREB/ABF) constitute the primary transcription factors through which ABA regulates drought responsive genes [29, 30]. Under drought conditions, ABA induces AREB1(ABF2), AREB2(ABF4), ABF1, and ABF3 from this transcription factor family in the vegetative tissues of Arabidopsis [31]. ABA and another plant phytohormone Jasmonic Acid (JA) regulate the expression of the drought responsive gene RD22 in Arabidopsis via the transcription factors MYB2 and MYC2 [32, 33]. MYB2 and MYC2 act as a point of crosstalk between the ABA and JA signaling pathways. On the other hand, Dehydration-responsive element binding protein 1 (DREB1)/CBF (C-repeat binding factor) and DREB2 transcription factor families operate independently of the ABA dependent pathway to regulate the drought responsive gene RD29A. This is achieved by the actions of transcription factors DREB1A(CBF3), DREB1B(CBF1), DREB1C(CBF2), and DREB2A [33, 34]. DREB1A, DREB1B, and DREB1C are negatively regulated by a transcription factor MYB15 and positively regulated by another transcription factor, ICE1 [35–37]. While ICE1 negatively regulates MYB15, it is suppressed by transcription factors HOS1 and upregulated by transcription factor SIZ1 [38]. Among the various members of the DREB1 and DREB2 family, DREB2A and DREB1D(CBF4) play an interesting role in regulating drought response. Unlike the other DREB transcription factors discussed here, which function independently of the ABA pathway, DREB2A and DREB1D can be induced by the ABA pathway through the ABRE transcription factor family under drought conditions [33, 39, 40]. Therefore DREB2A and DREB1D serve as another point of crosstalk for both ABA dependent and independent pathways in regulating drought responsive genes. DREB2A was found to be further regulated by DRIP1. Singh et al. (2015) found that transgenic Arabidopsis overexpressing DRIP1 delayed the expression of drought responsive genes regulated by DREB2A [33]. Downstream of the DREB and ABRE transcription factors is the drought responsive gene RD29A which is heavily regulated by these transcription factors [29, 40–42].

A recent study by Li et al. (2017) identified a drought stress-activated mitogen-activated protein (MAP) kinase cascade in cotton that regulates the expression of a drought responsive transcription factor GhWRKY59. GhWRKY59 directly binds to the W-boxes of the transcription factor GhDREB2 to regulate drought response in cotton [43]. We include this ABA independent pathway in our study of the drought regulatory network in Arabidopsis, where the MAP Kinase cascade is known to converge at the transcription factor DREB2A. In building our network model, we also study the WRKY transcription factor family which is traditionally associated with defense response against pathogens. However, many studies have now shown that WRKY transcription factor is involved in the defense response against drought [44–46]. The WRKY transcription factors WRKY40, WRKY60, WRKY18 are induced by ABA to regulate the expression of RD29A [47]. WRKY18, WRKY60 are known to positively regulate the expression of RD29A, whereas WRKY40 inhibits RD29A and WRKY60 [48]. Our previous paper on modeling the WRKY transcription factor in Arabidopsis under drought further confirmed these regulatory behaviors of the WRKY transcription factor family [49]. It should be noted that there is often crosstalk between ABA dependent and other independent pathways, we noted two instances of this earlier. Another instance of the crosstalk between the JA and ABA pathways was highlighted by Mintgen et al. (2014), where WRKY60 from the ABA pathway suppresses the expression of MYB2 in the JA pathway to regulate the drought responsive gene RD22 [50]. Other than RD22, MYB2 and MYC2 also regulate the expression of another drought responsive gene ERD1 [33]. According to a study by Ollas et al. (2016), MYB2 and MYC2 regulated the expression of ERD1 through a cluster of transcription factors (ANAC019, ANAC055, and ATAF1) belonging to the NAC transcription factor family. ERD1 was found to be further regulated by the transcription factor zinc finger homeodomain 1 (ZFHD1) and the gene RD26 (ANAC072) in the ABA pathway [51]. In addition to the drought responsive genes RD29A, ERD1, and RD22, we also consider the gene RD20 in our network model. RD20 was found to be directly upregulated by the gene RD26(ANAC072) [51]. The biological interactions discussed above are summarized in Fig 1. In the next section, we create a Bayesian network model based on these signaling pathways to predict the best regulator(s) for the drought responsive genes (marked in blue in Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Drought signaling pathways in Arabidopsis.

The orange circular nodes represent elements directly regulated by ABA whereas the purple nodes represent elements regulated by JA. The two nodes colored with a mix of orange and purple represent elements regulated by both JA and ABA pathways (Crosstalk). The blue diamonds represent drought responsive reporter genes. The plain circular nodes with no colors represent the transcription factors, genes and proteins involved in the regulation of drought responsive reporter genes in an ABA independent manner. The green and red arrows represent positive and negative regulation. The arrows going into and out of ATAF1 are marked black to indicate that the nature of regulation is not known at this time.

https://doi.org/10.1371/journal.pone.0255486.g001

Bayesian network model

We observed in the previous section that plants deploy a diverse range of defense mechanisms to survive under drought conditions. These phenotypical defense responses are mediated through complex networks of signaling pathways at the genomic level. Biological signaling pathways have been successfully modeled using methods such as linear models, Boolean networks, probabilistic Boolean networks, Bayesian networks, and small molecule level models [52–57]. In order to develop a thorough understanding of these multivariate and stochastic interactions, we create a BN model of the drought signaling pathways. Unlike some modeling techniques which are solely driven by data, a BN model allows us to integrate pathway information in the form of prior knowledge along with experimental data [58]. BNs are directed acyclic graphs that represent the causal probabilistic relationships among a set of random variables and provide the conditional decomposition of the joint probability distribution of these random variables [59, 60]. Thus BNs serve as an ideal modeling paradigm to study the drought signaling pathways [58]. In this paper, our objective is to create a BN model of the drought signaling pathways outlined in Fig 1 and use this model to determine which transcription factor, protein or gene is the best regulator of drought responsive reporter genes (blue diamonds in Fig 1). The predictions made by the model can help us identify potential targets for genetic intervention techniques like CRISPR-Cas9 to create drought resistant crops.

Fig 2 represents the BN model of the signaling pathways shown in Fig 1. Every node (circle) in the network represents a gene, protein, or transcription factor in the drought signaling pathway. The black arrows or edges connecting the nodes represent the causal biological relationships we discussed in the previous section. We assume each of the nodes are binary random variables that can assume 1 for activation and 0 for inhibition. Since the nodes are random variables, associated with each of them is a parameter θ which describes the local marginal or conditional probability distribution for that node. For instance, the conditional probability parameter associated with the node representing MKK4 is given by θ_MKK4|MAP3K15. This parameter represents the activation or inhibition probability of the node representing MKK4 conditioned on the state of the node representing MAP3K15. Similarly, for the node representing the transcription factor ICE1, the local conditional probability distribution is given by θ_{ICE1|HOS1,SIZ1}. Henceforth, we will refer to local conditional or marginal probability distribution as just local probability distributions (LPD). We learn these LPDs from experimental biological data; once these LPDs are learned, the BN model is complete and can be used for carrying out inference simulations to determine the best modulator for the drought responsive genes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Bayesian network model of drought signaling pathway.

Every circular node represents a biological element in the drought signaling pathway. Every edge or black arrow represents the causal biological relationship between the nodes. Associated with every node is a θ parameter that represents the local probability distribution of the node.

https://doi.org/10.1371/journal.pone.0255486.g002

Parameter estimation in Bayesian networks

BNs consist of two major components: a directed acyclic graph (DAG) and a set of local probability distributions. The DAG can be learned from data or constructed from domain knowledge. Learning BNs from data, also known as structure learning in the literature, is an NP-Hard problem and requires us to choose a DAG from several candidate DAGs [61]. This is not very practical as we observed in in the prior sections that pathway interactions are well defined, and there can only be a single DAG representing them. Furthermore, in the context of Arabidopsis under drought, we are limited by the sizes of publicly available datasets. These datasets are not large enough to construct a reliable DAG, so we elected to create the BN model in Fig 2 using pathway information from the existing biological literature. While a DAG can be learned either using data or from domain knowledge, the local probability distributions associated with the DAG have to be estimated from experimental data. There are several ways to estimate the local probability distribution in a BN model. Typically, either a frequentist approach such as a Maximum Likelihood Estimate (MLE) or a Bayesian approach is employed. Though methods such as MLE are simple and provide a point estimate, they are only driven by data and do not take any relevant prior information into account [62]. On the other hand, a Bayesian approach provides us with the posterior distribution, which is driven by both data in the form of likelihood and relevant information in the form of a prior distribution. However, the Bayesian approach has two significant drawbacks. The first one is computing the normalizing constant or the probability of data (evidence) [63]. The normalizing constant very rarely has a closed form solution and hence can be computationally expensive to determine. The second drawback pertains to the choice of a prior distribution. Since the choice of the prior distribution is subjective and there exists no established method to select one, different choices of prior distribution will lead to different results [64]. Nonetheless, the Bayesian approach is logically rigorous and unlike frequentist approaches, once the prior distribution is established the Bayesian approach follows deductive logic. In this paper, we use a Bayesian approach to estimate the local probability distributions for the BN model outlined in Fig 2. We assumed that the nodes are binary random variables, which implies that for any node X in the BN, X = 1 (success) when the node is activated and X = 0 (failure) when the node is inhibited. Then for a single observation for any node X in the BN be modeled as a Bernoulli random variable.

Let us suppose that we have a BN model with N nodes. Then the probability with which any node X attains a state of 1 is given by θ_X. Thus if we make n (>0) independent and identically distributed observations (i.i.d) observations for each node in the BN, and if for a given node X, we observe k instances when the node attains a state of 1, then the likelihood for node X is given by: (1) (2)

P_a(X) in Eq (1) refers to the parents, if any, of node X. Since we are using a Bayesian approach to estimate the LPD of Node X, we need to select a prior distribution on the node X. Considering the computational complexity required in calculating the normalizing constant, and since the likelihood function associated with our model follows a binomial distribution by design, we assume the prior distribution on θ_X to follow a Beta distribution. Since the Beta and Binomial distributions belong to conjugate families, we know that the posterior distribution of θ_X will also follow a Beta distribution [65]. This is formulated as follows: (3) (4) where and .

In Eq (3), α_X and β_X represent the shape parameters of the Beta distribution, and in Eq (4) these parameters get updated for the posterior distribution on θ_X. We assume α_X = 1 and β_X = 1 for our calculations as the Beta(1,1) distribution corresponds to the standard uniform distribution over the interval [0, 1] [66]. Setting the prior distribution to the standard uniform distribution guarantees that we have no information regarding the prior distribution of θ_X. We chose the Beta(1,1) distribution as our prior because we do not have any domain knowledge information regarding the prior distribution of every node in the BN model. If we had such information regarding the prior distribution, they could be incorporated into this model. However, it is to be noted that choosing a different prior distribution may not allow us to reach a closed form solution for the posterior distribution on θ_X. Since the result we get in Eq (4) is a distribution and not a point estimate like what we would have obtained had we used a frequentist approach, we approximate the values for θ_X with the expected value of the posterior distribution. We do this approximation for the posterior distributions estimated at every node in the BN. This approximation for the node X has been presented in Eq (5). (5) Once these parameters are learned the BN is complete as we have both the DAG and the set of conditional probabilities. In the next section, we study the effect on drought responsive genes for intervening (activating/ inhibiting) at various nodes, then summarize our findings in the Results section.

Sampling based inference in Bayesian networks

In this section, we are interested in using the BN model in determining which nodes are the best regulators of the drought responsive reporter genes RD29A, RD20, RD22, and ERD1. Specifically, we want to study the effect on the reporter genes of intervening at the non-reporter genes. In other words, we will fix the state of every non-reporter gene node one at a time to either 0 or 1, and observe how this action (intervention) affects the LPDs for the nodes representing the drought responsive reporter genes. This kind of simulation in BNs is known as inference. Inference techniques are categorized as either exact or approximate. Exact inference techniques such as Enumeration, Variable Elimination, and Pearl’s Message Passing Algorithm are particularly efficient in polytrees or singly connected networks. One such application of exact inference was demonstrated by Vundavilli et al. to find significant nodes in the breast cancer signaling pathway [67]. Ideally, we would like to use an exact inference technique to calculate the LPDs in our BN model. However, exact inference techniques will be computationally expensive to implement as our network is multiply connected, i.e., there are at least two nodes in our BN model connected by more than one path. For instance, we can see that the nodes DREB1A and ICE1 are directly connected and are also connected through MYB15, hence making our BN model multiply connected. While exact inference algorithms work in polynomial time in polytrees, it has been shown to be NP-Hard in more generalized BNs, hence implementing them in multiply connected networks may not be practical [68]. Therefore, the size and structure of the BN govern our choice of inference techniques. This is the reason why, for determining the regulators of drought responsive reporter genes, we employ an approximate inference technique known as likelihood weighting.

Likelihood Weighting (LW) is an approximate inference technique based on stochastic simulations. Inference techniques based on stochastic simulations usually involve drawing samples from a sampling distribution, calculating an approximate posterior probability based on the samples, and then showing that the posterior probability converges to the actual probability [69]. In the context of our model, the sampling distribution will be specified by the BN in the form of LPDs. Unlike exact inference techniques, LW is generally insensitive to the network topology, however, convergence in estimating the posterior probabilities can be slow if they are close to 0 or 1 [70]. We will now describe the mathematical formulation for LW.

Consider a BN consisting of N nodes such that the DAG follows a topological ordering of {X₁,X₂,‥,X_N}. Suppose we make an observation on the node X_E in the BN, we will refer to X_E as the evidence node. Now suppose our objective is to find the effects of this observation on another node X_Q, known as the query node in the BN. Specifically, we want to estimate the posterior probability Pr(X_Q = x_q|X_E = x_e), where ‘x_q’ and ‘x_e’ are some instantiation of nodes X_Q and X_E. At this step we begin performing LW by drawing M samples from the BN for every node except for the evidence node X_E, in topological order. The generated dataset (ξ) will be a matrix with M rows and N columns, where each row represents an N-dimensional sample (datapoint) and columns represent nodes in the BN. Thus after the first iteration of the sample generation process, the datapoint will be of the form ξ⁽¹⁾ = {x₁^{(i = 1)},x₂^{(i = 1)},….,x_e^{(i = 1)},‥,x_N^{(i = 1)}}. We will repeat this process M-1 more times to obtain M such samples, thus that dataset will be of the form ξ = ξ^{{i = 1,2,‥,M}} = {x⁽ⁱ⁾₁,x⁽ⁱ⁾₂,….,x_e,‥,x⁽ⁱ⁾_N}. It should be noted that x_e, does not change across the M samples. This is because X_E is the evidence variable that has been observed and fixed. The samples for the rest of the non-evidence nodes are generated according to the LPDs associated with those nodes. For example we draw a sample x₁ for root node X₁ according to Pr(X₁). Similarly we draw a sample x₂ for the node X₂ according to Pr(X₂ | X₁ = x₁) and so on. It should be noted that all the children of node X_E have a fixed instantiation for X_E, that is x_e. We then approximate Pr(X_Q = x_q|X_E = x_e) as follows: (6)

The proof for Eq (6) is not trivial and is presented in a paper by Menon [71]. A psudo code for estimating the conditional probabilities using LW is presented in algorithm 1. We will now demonstrate LW on an example BN.

Algorithm 1: Psudo Code for likelihood weighting in Bayesian Networks

Input:

 1: BN: The Bayesian Network

 2: Q: The Query Variable, Let Q = q, that is node Q is instantiated to some value of interest q.

 3: E: The Evidence variable. Let E = e, that is node E is instantiated to some observed value e.

 4: M: Number of Samples.

Output: Probability: Estimate of P(Q = q|E = e)

 5: Initialization: X₁,X₂,‥,X_N Topological Ordering of BN

  Sampled_Data = {} {}, M by N matrix to store sampled data

  Weight = {1,…,1}, an array of size M, consisting of weights

  with values initialized to 1.

  Counts[k] = 0, where k ∈ domain of Q

 6: while iter = 1 to M do

 7:  for each node X in BN in topological order do

 8:   if X = X_i is in E then

 9:    Sampled_Data[iter][X_i] = x, where x is the value of X_i

 10:    Weight[iter] = Weight[iter] * P(X_i = x | P_a (X_i))

 11:   else

 12:    Sampled_Data[iter][X_i] = Generate random sample from P(X_i = x|P_a(X_i))

 13:   end if

 14:  end for

 15:  iter = iter+1

 16: end while

 17: k = List of row indices in Sampled_Data where Q = q

 18: Probability = Sum (Weights [k])/ Sum(Weights)

 19: return Probability

Fig 3 describes an example BN consisting of four genes A,B,C, and D. We consider the nodes representing the genes as binary random variables, which can take on the values of 1 for activation and 0 for inhibition. The LPDs for this example BN are already estimated and are presented in Fig 3. For the purpose of this example, we assume that Gene A positively regulates gene B, while it negatively regulates gene C. Gene D is upregulated by gene B, while gene C downregulates it. These effects are reflected in the LPDs for each node. Now suppose, we are interested in gene D being positively regulated, and we decide to intervene at Gene B and set it to 1. Therefore, node B = 1 serves as the evidence variable, and let us consider node D as the query variable. Then we are interested in finding the probability P(D|B = 1) using LW.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Example BN with LPDs.

Gene A positively regulates Gene B and negatively regulates Gene C. Gene B positively regulates Gene D and Gene C negatively regulates Gene D.

https://doi.org/10.1371/journal.pone.0255486.g003

In order to estimate this probability, we will need to query the BN and generate samples first. We use the topological ordering of{A,B,C,D}, another valid ordering is {A,C,B,D}. The sample generation process is described in the following steps:

1. Set the weight variable ‘W_i’ to 1. W_(i) = 1
2. The matrix Sampled_Data[iter][X_i] = is empty. This matrix will store the value of nodes A,B,C,D.
3. We start topologically at node A. Since A is not an evidence node, we sample it according to its LPD, specifically P(A). Assume this sample results in A = 1.
4. We now move on to node B. Since B is an evidence node, we do not sample it. We update, W_(i) = 1. P(B = 1|A = 1) = 1. (0.9) = 0.9.
5. We now go to node C. Since C is not an evidence node, we sample it according to its LPD, specifically P(C|A = 1). Let us assume the result of this process is C = 0.
6. We now sample node D with its LPD of P(D|B = 1,C = 0). Assume that this results in D = 1.
7. The sample generated is (A = 1,B = 1,C = 0,D = 1) with W_{(i = 1)} = 0.9. Thus Sampled_Data[1][All Columns] = [1, 1, 0, 0]
8. We repeat steps 1–7, M-1 more times to obtain a total of M samples.
9. We can then calculate P(D|B = 1) as follows:

Therefore for M = 5, if we generated sample, it would result in a 5 by 4 matrix (Sampled_Data[iter],[X_i]). Table 1 shows this matrix with an extra column for weights belonging to each sample. From the samples and weights in Table 1, we can now estimate P(D = 1|B = 1) and P(D = 0|B = 1) as follows:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Sample data from example Bayesian network.

https://doi.org/10.1371/journal.pone.0255486.t001

Experimental validation

To validate the conclusions from the Bayesian network model, we isolated Arabidopsis ataf1 (SALK_057618C) and myc2 (myc2-1, SALK_061267C; myc2-2, SALK_128938C) mutants from the Arabidopsis Biological Resource Center (ABRC) [83]. The ataf1 mutant has a T-DNA insertion in the third exon of the ATAF1 (AT1G01720) genomic DNA, both myc2 mutants have a T-DNA insertion in the exon of the MYC2 (AT1G32640) genomic DNA (Fig 8A). We germinated wild-type (WT) Col-0 and ataf1 mutant on the half-strength Murashige and Skoog (MS) medium with or without 300 mM mannitol treatment (Fig 8B). The addition of mannitol reduces water potential of growth media, which is often used to mimic drought stress (Mu et al., 2019) [84]. Although the germination rate of the ataf1 mutant was lower than WT in the medium without mannitol, the ataf1 mutant had more green cotyledon seedlings (Fig 8B) and higher green cotyledon rate (Fig 8C) than WT seedlings under 300 mM mannitol treatment. The difference became significant at nine days after germination. We also compared the green cotyledon inhibition rate of WT and ataf1 mutant on MS medium with or without mannitol. Consistently, the ataf1 mutant showed lower green cotyledon inhibition rate than WT, and the tendency became more pronounced with the increase of growth time (Fig 8D). We also germinated WT and myc2 mutants on the MS medium with or without 300 mM mannitol treatment (Fig 8E). However, there is no significant difference in the green cotyledon rate between WT and myc2 mutants with or without mannitol treatment (Fig 8F). Similarly, the green cotyledon inhibition rate between WT and myc2 mutants also did not show a significant difference (Fig 8G). Thus, our data show that the ataf1 mutant was more tolerant to the mannitol treatment, and suggests that ATAF1 plays a role in plant drought stress response. Our test conditions, such as plant growth stage, treatment, or the combination, may not be suitable to reveal the difference between WT and myc2 mutants.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Results from validation experiments.

A. The scheme of the ATAF1 and MYC2 genomic DNA and T-DNA insertion. The panel is a schematic illustration of the ATAF1 and MYC2 genomic DNA with exons (solid box), intron (lines) and 3’ untranslated region (open box). The position of T-DNA insertion of ataf1 (SALK_057618C), myc2 (SALK_061267C, SALK_128938C)was labeled. B. The ataf1 mutant is more resistant to mannitol treatment. Wild-type (WT) Col-0 and ataf1 mutant seeds were germinated on 1/2 MS medium with or without 300 mM mannitol. 30 seeds per genotype were used for each replicate. The photos were taken four-week post-germination. C. Quantification of cotyledon greening on plates corresponding to B. Seedlings with green cotyledon expansion were counted at 6–9 days post-germination. Data are shown as means ± SD (standard deviation) from three independent replicates (n = 3, *, p<0.05, Student’s t-test). D. Quantification of cotyledon greening inhibition rate on plates corresponding to B. Seedlings with green cotyledon expansion were counted at 6–9 days post-germination. Data are shown as means ± SD from three independent replicates (n = 3, *, p<0.05, Student’s t-test). E. Growth of WT and myc2 mutants on MS plates. WT and myc2 mutant seeds were germinated on 1/2 MS medium with or without 300 mM mannitol. 30 seeds per genotype were used for each replicate. The photos were taken four-week post-germination. F. Quantification of cotyledon greening on plates corresponding to E. Seedlings with green cotyledon expansion were counted at 6–9 days post-germination. Data are shown as means ± SD from three independent replicates (n = 3, no statistical significance with Student’s t-test). G. Quantification of cotyledon greening on plates corresponding to E. Seedlings with green cotyledon expansion were counted at 6–9 days post-germination. Data are shown as means ± SD from three independent replicates (n = 3, no statistical significance with Student’s t-test).

https://doi.org/10.1371/journal.pone.0255486.g008

Experimental setup

A. thaliana mutants ataf1 (SALK_057618C) and myc2 (SALK_061267C, SALK_128938C) were obtained from the Arabidopsis Biological Resource Center (ABRC). The wild-type (Col-0) and mutant plants were grown in a growth room at 23°C, 45% humidity, and 75 μE m⁻² s⁻¹ light with a 12-hr light /12-hr dark photoperiod. To detect cotyledon greening rate, 30 seeds per genotype were sterilized and germinated on half-strength Murashige and Skoog (MS) medium with or without 300 mM Mannitol treatment in each replicate. Seedlings with green cotyledon expansion were counted at 6–9 d post-germination, data are shown as means ± SD from three independent repeats (n = 3, *, p<0.05, Student’s t-test). The photos were taken four-weeks post-germination.