A physical model links structure and function in the plant immune system

Significance

Just like humans and other animals, plants need an immune system to defend themselves from disease. An important part of the plant immune system is Effector-Triggered Immunity (ETI), which can sense harmful microbes directly, by recognizing specific molecules, or indirectly, by monitoring their effect on the plant. The immune system is composed of diverse architectures governing pathogen sensing, but the functional repercussions of these structures remain unknown. Here, we develop a mathematical model of ETI that allows us to link the microscopic “parts” (genes and molecules) to the macroscopic “system” (immune sensing). We use the model to explain how ETI works and analyze how it may have been shaped by evolution.

Abstract

Effector-Triggered Immunity (ETI) is an important part of the plant immune system, allowing plants to sense and respond to harmful pathogen proteins known as “effectors.” Effectors can be sensed directly or indirectly by NLR (Nucleotide-binding Leucine-rich Repeat) proteins, many of which “guard” the plant proteins targeted by effectors. Although a few effector–target–NLR interactions have been characterized, a general understanding of how these molecular interactions give rise to a functioning immune system is lacking. Here, we present a physics-based model of ETI based on protein–protein interactions. We show that the simplest physical model consistent with the biology gives rise to a robust immune sensor and explains the empirical phenomenon of effector interference as a generic consequence of molecules competing for binding partners. Using the evolutionarily conserved ZAR1 defense gene as a model, we explain how more complex interaction networks integrate multiple pathogen signals into a single response. We then examine alternatives to a guarding architecture, including direct sensing, decoys, and blended “integrated decoy” strategies, and reveal that these sensing architectures obey functional trade-offs between their sensitivity, target protection, and proteomic cost. This allows a quantitative analysis of the trade-offs between different forms of ETI. We discuss these findings in the context of the evolutionary forces shaping the plant immune system.

Like all living things, plants must contend with diverse pathogens to survive. The plant immune system is a subject of great interest for fundamental biology but also for agriculture and food security, as experts estimate that 20 to 30% of global crop yields are lost to pests and pathogens (1). Vertebrates confront pathogen diversity with an adaptive immune system that generates diverse antibodies and optimizes them to match the environment. Plants, however, lack an adaptive immune system. Instead, they rely upon a relatively small set of genes fixed at fertilization, sensing diverse “Others” both directly and through changes to “Self.”

The plant immune system has two layers, Pattern-Triggered Immunity (PTI) and Effector-Triggered Immunity (ETI). In PTI (also known as PAMP-Triggered Immunity, where “PAMP” stands for “Pathogen-Associated Molecular Patterns”), receptors on the cell surface recognize generic pathogen molecules like bacterial flagellar proteins or fungal chitin (2). This triggers a generic response, such as closing the stomata or thickening the cell wall via callose deposition. In response to PTI, pathogens have evolved “effectors” which interfere with PTI and other important host functions (3). Effectors are delivered directly into the plant cell, for example by bacterial type-III secretion systems or fungal haustoria. There, they physically interact with plant proteins to suppress host defense. In turn, plants have evolved Effector-Triggered Immunity, where proteins called NLRs (Nucleotide-binding Leucine-rich Repeats) can either sense effectors directly or act as “guards” that sense changes in the host proteins due to the effector. This triggers a robust immune response, e.g. the Hypersensitive Response (HR), a program for localized cell death which halts the invasion of pathogens. Several reviews explain PTI and ETI in depth (4–6), as well as an emerging understanding of how the two are intertwined (7–10).

Some of the genes involved in ETI have been identified, and a few of the protein–protein interactions involved have been characterized (11–13). However, fundamental questions remain. The most basic is mechanistic: How does a robust immune sensor emerge from physical interactions between proteins? The need for a mechanistic “systems” approach has become more stark as the inventory of molecular “parts” has grown in complexity. For example, the basic “gene-for-gene” picture of a unique mapping between effectors and NLRs has been challenged by effector interference, a widespread phenomenon where one effector “interferes” with the recognition of another (12, 14). High-throughput screens of protein–protein interactions suggest an even denser network where interactions are many-to-many rather than one-to-one (11). Moreover, the “guard model” is only one example of a diversity of sensing architectures, including direct NLR–effector interactions (15), decoy proteins which have no known function other than mediating effector–NLR interactions (16), and more elaborate Integrated-Decoy structures whereby two NLRs work together, with one sensing via an integrated domain and the other performing downstream activation (13, 15). It is unknown what evolutionary pressures maintain this diversity and drive transitions between different architectures.

We find that a simple physics-based model of protein-protein interactions and positive feedback explains the function of ETI in terms of its structure. The simplest physical model consistent with the biology leads to a robust immune sensor, and effector interference is found to be a generic consequence of molecules competing for binding partners. We study how NLRs could integrate many-effector signals in more complex networks and show that the known sensing architectures obey functional trade-offs, which may explain why evolution has led to their stable coexistence. We hope this work will contribute to an integrated understanding of the plant immune system and accelerate efforts to engineer disease-resistant crops.

Results

We begin by describing a simple model for the “guard” architecture. The model has three protein types: NLRs n_i, targets t_j, and effectors e_k. In the model illustrated in Fig. 1A, ETI consists of three processes:

1. The NLR and target physically associate to form a complex with rate α₁ and dissociate with rate α₂: $n_i+t_j[{\alpha }_2]{\alpha }_1{C}_{\mathit{ij}}$,

2. The effector binds to the target with rate β₁ and disables it with rate β₂: $t_j+e_k\stackrel{{\beta }_1}{\to }{(t_j{e}_k)}^{\mathrm{b}}\stackrel{{\beta }_2}{\to }{e}_k$, and

3. The effector binds to the complex with rate γ₁ and disables the target with rate γ₂, activating the NLR: $C_{\mathit{ij}}+e_k\stackrel{{\gamma }_1}{\to }{(C_{\mathit{ij}}{e}_k)}^{\mathrm{b}}\stackrel{{\gamma }_2}{\to }{a}_i+e_k$,

Fig. 1.

A simple model of interacting proteins leads to a robust immune sensor. (A) The host “target” can be degraded by an effector (Top) or protected by an NLR (Bottom), in which case interacting with the effector leads to NLR activation and upregulation. (B) The steady-state concentrations of n₁, t₁, and a₁ as a function of effector concentration. For each component, the middle steady-states are unstable fixed points which separate the upper and lower basins of attraction. Parameters: $g_n=1$, $g_t=10$, ${\delta }_n={\delta }_t=1$, ${\alpha }_1=0.5$, ${\alpha }_2=1$, ${\beta }_1={\beta }_2=1$, ${\gamma }_1=1$, and ${\gamma }_2=5$.

where ${(xy)}^{\mathrm{b}}$ denotes an intermediate bound state and a_i is the activated form of NLR i. If we assume mass-action kinetics in a well-mixed cellular compartment, we obtain a system of 6 nonlinear ordinary differential equations with many parameters. In order to glean more biological insight, we make two simplifying assumptions. First, because molecular interactions are much faster than protein translation/degradation, we assume a separation of timescales where the bound states and complexes are in equilibrium. Thus $\frac{d}{\mathit{dt}}{C}_{\mathit{ij}}=\frac{d}{\mathit{dt}}{(t_j{e}_k)}^{\mathrm{b}}=\frac{d}{\mathit{dt}}{(C_{\mathit{ij}}{e}_k)}^{\mathrm{b}}=0$. Second, we assume that the immune sensor responds to a fixed quantity of effectors E_k. Of course, the effectors will be dynamic during a real infection, but this allows us to study how a particular immune system responds to a particular signal. With these simplifications, we obtain the following dynamics (See SI Appendix for a full derivation):

$$\begin{array}{cc}& \frac{d}{\mathit{dt}}{n}_i=g_n-{\delta }_n{n}_i-{\gamma }_1{e}_k\frac{{\alpha }_1{n}_i{t}_j}{{\alpha }_2+{\gamma }_1{e}_k}+f(a_i),\hfill \end{array}$$ [1]

$$\begin{array}{cc}& \frac{d}{\mathit{dt}}{t}_j=g_t-{\delta }_t{t}_j-{\gamma }_1{e}_k\frac{{\alpha }_1{n}_i{t}_j}{{\alpha }_2+{\gamma }_1{e}_k}-{\beta }_1{t}_j{e}_k,\hfill \end{array}$$ [2]

$$\begin{array}{cc}& \frac{d}{\mathit{dt}}{a}_i={\gamma }_1{e}_k\frac{{\alpha }_1{n}_i{t}_j}{{\alpha }_2+{\gamma }_1{e}_k}-{\delta }_n{a}_i,\hfill \end{array}$$ [3]

where g_x and δ_x denote the protein production and degradation rates for the NLR and target. The term $f(a_i)$ in Eq. 1 represents positive feedback. Immune genes are typically expressed at low levels, presumably to minimize their proteomic cost and avoid spurious immune activation (17, 18), but during infection, they rapidly accumulate to mount a robust response. Therefore, we propose a positive feedback loop whereby activated NLR a_i stimulates the expression of n_i. In what follows, we take $f(a_i)=a_i^2$, but the main results are not sensitive to the exact form of the nonlinearity (discussed below and in SI Appendix, Fig. S2). The effectors will be distributed between free molecules and those in bound states. This gives the following equation:

$$\begin{array}{cc}\hfill E_k& =e_k+{(t_j{e}_k)}^{\mathrm{b}}+{(C_{\mathit{ij}}{e}_k)}^{\mathrm{b}},\hfill \end{array}$$ [4]

$$\begin{array}{cc}& =e_k+\frac{{\beta }_1}{{\beta }_2}{t}_j{e}_k+\frac{{\gamma }_1}{{\gamma }_2}{e}_k\frac{{\alpha }_1{n}_i{t}_j}{{\alpha }_2+{\gamma }_1{e}_k}.\hfill \end{array}$$ [5]

For a given effector concentration E_k, Eq. 5 can be solved to give free e_k as a function of n_i and t_j.

Given the dynamics in Eqs. 1–3, how does the simplest network of one guard and one target respond to a single effector? Fig. 1B shows the steady states of n₁, t₁, and a₁ as a function of the effector concentration E₁. [We focus on cases where the NLR expression and degradation rates are relatively low ($g_n=1$ and ${\delta }_n=1$, respectively) compared to the target expression ($g_t=10g_n$). The results are not sensitive to the exact values of the parameters, as we show in SI Appendix, Fig. S3.] At low effector concentrations, n₁ and a₁ remain low while t₁ remains at the relatively high basal expression level. As E₁ increases, all three quantities change slowly until a critical threshold $E_1^{\ast }$ (a bifurcation in the dynamical system). Then n₁ and a₁ jump up by orders of magnitude while t₁ collapses. We identify this discontinuous transition as immune activation because we expect this dramatic change in protein expression level to be perceived by the corresponding downstream regulatory machinery. Thus the simplest possible physical model of the biology leads to an immune sensor which is robust in two senses. First, a discontinuous jump should be easier to monitor than a continuous one. Second, the bifurcation exhibits hysteresis, so the system would remain activated if E₁ were to temporarily fluctuate back below the threshold. Importantly, this switch-like activation is dependent on the nonlinear positive feedback encoded in $f(a_1)$. Linear feedback does not lead to a bifurcation, but the behavior is not sensitive to the exact form of the nonlinearity (SI Appendix, Fig. S2). The downstream signaling pathways in ETI are not fully understood, but our results suggest they may be implementing nonlinear positive feedback in NLR activation.

In a natural setting, a plant is unlikely to be exposed to a lone effector. A typical strain of Pseudomonas syringae carries roughly 30 effectors (19), and wild Arabidopsis thaliana plants have foliar microbiomes with thousands of distinct OTUs (20), including multiple known pathogens (21). How does the immune sensor described above respond to multiple effectors? In several empirical cases, the presence of effector 2 suppresses recognition of effector 1, a phenomenon called “effector interference.” For example, the P. syringae effector AvrB is recognized by the A. thaliana NLR RPM1. However, when the effector AvrRpt2 is also present, RPM1 is not activated (12). Effector interference is common — at least 33 examples have been identified (14)—and thought to be important at the ecological level. A recent survey of nearly 500 P. syringae strains found that 92% of them harbored at least one effector that elicits ETI on its own in the Col-0 accession of A. thaliana (22). This raises the question of how any pathogen manages to be successful in the face of broad-spectrum resistance, and one possibility is the use of multiple effectors to suppress recognition.

There are many potential molecular mechanisms underlying effector interference. In the case of AvrB and AvrRpt2, both interact with the same target, RIN4. AvrRpt2 acts on RIN4 by cleaving a membrane-localization domain, which prevents RPM1 from finding it at the membrane (12). We can represent this in our model as degradation of the target by e₂: $t_j+e_2\to {(t_j{e}_2)}^{\mathrm{b}}\to e_2$. Intuitively, e₂ prevents t_j from forming the complex necessary for NLR activation. Fig. 2A shows how the bifurcation diagram changes as a function of interfering effector concentration E₂. When $E_2=0$, we obtain the same steady states as Fig. 1B, with robust immune activation. As E₂ increases, the activating region shrinks ($E_2=1.1$) then disappears completely ($E_2=3$). (We note that the activated branch with high $n_i\approx 50$ persists in Fig. 2A, which we discuss below in the context of immune priming. The inactive branch with $n_i\approx 1$ always remains stable, however, so an inactive cell will remain so.)

Fig. 2.

Effector interference has a generic form regardless of the underlying mechanism. (A) The steady-state concentration of NLR n₁ as a function of E₁ for different levels of interfering effector E₂, where E₂ interferes by degrading the target. (B) A phase diagram showing the number of steady states as function of E₁ and E₂. Dashed lines with green markers show where the data in (A) lies on the phase diagram. (C) Same as (B), but for an interfering effector which slows down the binding between e₁ and the NLR–target complex C₁₁. (D) Same as (B) and (C), but for an interfering effector which binds to the complex and blocks e₁. Parameters: Same as Fig. 1, except for where γ₁ varies in (C). In (D), κ = 0.25.

Other molecular mechanisms have been implicated in effector interference. For example, HopZ3 prevents RPM1 from recognizing AvrB3 and ${\mathrm{AvrRpm}1}_{\mathrm{Psy}}$ by acetylating many of the proteins involved, including the two effectors, RIN4, and several plant kinases (RIPK, PBS1, and BIK1) (23). These posttranslational modifications function to decrease the phosphorylating activity of e₁ (i.e. AvrB3 and ${\mathrm{AvrRpm}1}_{\mathrm{Psy}}$). In other cases, e₂ binds to the complex and prevents e₁ from interacting with it (24). We can organize this complexity by constructing a “phase diagram” showing the number of steady states as a function of E₁ and E₂.

Fig. 2B shows the phase diagram for target degradation (the same mechanism as Fig. 2A). Immune activation occurs in the blue bubble around $E_1\approx 5$, where the lower branch becomes unstable and the system is forced to jump to the active state. As E₂ increases, the activating region shrinks and disappears. Fig. 2 C and D show the phase diagrams for the two other mechanisms described above (e₂ modifying the interaction rates via PTMs and e₂ protecting the complex). Despite the very different underlying mechanisms, all three phase diagrams are topologically equivalent. This explains why effector interference has a generic form regardless of the underlying mechanism. (See SI Appendix for full model descriptions.) It is not a mysterious phenomenon arising from obscure biology, but rather a natural consequence of molecules competing for binding partners.

When many effectors are present, the immune system must integrate that signal into a response. Such integration can happen at the level of gene regulation, but it can also be implemented at the level of molecular interactions if a single NLR guards multiple targets. Consider the example of ZAR1, a resistance gene common in flowering plants (25). ZAR1 offers unusually broad resistance, sensing at least six distinct effector families through a multilayer network of kinases and pseudokinases (26). Because we are interested in the question of how ZAR1 integrates multiple signals, we will study the simplified model shown in Fig. 3A. In the model, each ZAR1 molecule can form a complex with one of six kinases, which each monitor a different effector. (Full equations in SI Appendix.) The phenomenon of effector interference leads to the question of whether kinases can interfere in a similar way by competing for ZAR1 as a binding partner: When only e₁ is present, does the presence of kinase 2 to 6 affect immune activation?

Fig. 3.

ZAR1 activation depends on the dynamics of complex degradation and on effector balance. (A) A simplified model of the ZAR1 interaction network where each kinase forms a complex with ZAR1 and interacts with a single effector. (B) The critical effector concentration $E_1^{\ast }$ as a function of kinase number, for a model where complexes can degrade (orange) and a model where they are protected (blue). $E_i=0\forall i\ne 1$. For the model with complex degradation, no activation occurs for kinase number $>2$. The complex degradation rate ${\delta }_C=1$. (C) NLR bifurcation diagram where total effector concentration is the control parameter. For each set of points, E₁ is specified and the other 5 effectors are $E_j=(E_T-E_1)/5$, where E_T is the total effector concentration. The kinase expression rate $g_{\mathit{ki}}=10\forall i$. (D) The critical effector $E_1^{\ast }$ as a function of effector imbalance for different kinase expression levels. For the green curve $g_{\mathit{ki}}=2$, no activation occurs for $E_1=E_T$.

Fig. 3B shows the critical effector concentration $E_1^{\ast }$ as a function of kinase number for two models. When the complex C_i is protected from protein degradation (Fig. 3B, blue circles), complexes do not compete and the $C_1-e_1$ interaction is unaffected by $k_{i\ne 1}$. This is because at equilibrium, the fluxes of n in and out of C_i balance out, so other kinases do not deplete the reserves of n. However, when complexes can be tagged for protein degradation (Fig. 3B, orange squares), the presence of other kinases has a large effect. k₂ increases the threshold by 5×, and k₃ abolishes activation completely. This gives plants another mechanism to regulate immune response and could be tested experimentally. [We note that complex degradation does not eliminate activation in the original guard model (SI Appendix, Fig. S4)].

ZAR1 also raises an important ecological question. If the total concentration of effectors $E_T=\sum _i{e}_i$ is fixed, will ZAR1 be more sensitive to balanced effectors ($E_i=E_T/6$) or imbalanced effectors ($E_1>E_{i\ne 1})$? The answer will affect the optimal pathogen strategies and plant homeostasis. Fig. 3C shows how the bifurcation diagram shifts with effector balance for abundant kinases $g_{\mathit{ki}}=10$. As effectors become less balanced (i.e. $E_1/E_T$ goes from $1/6$ to 1), the critical effector concentration $E_T^{\ast }$ increases. A system with low $E^{\ast }$ is more sensitive, so ZAR1 is more sensitive to balanced effectors. This is because kinase expression level acts as a signal bottleneck, so more flux can go through 6 channels than 1. As a corollary, the effect becomes more pronounced as kinases become more limiting. The kinase expression level in Fig. 3C of $g_{\mathit{ki}}=10$ is very high (recall that the functional target has expression level $g_t=10$ and the NLR has $g_n=1$). Fig. 3D shows that as kinases become more limiting, the effects of imbalance become more dramatic, from a 50% shift at $g_{\mathit{ki}}=10$ to elimination of activation for $g_{\mathit{ki}}=2$ at the largest imbalance. (Note that we have varied effector concentrations, but varying the binding specificities should yield similar effects. In nature, both forms of “balance” will play a role.)

Thus far we have focused on variations of the guard model, where the NLR binds to the target and the target interacts with the effector. However, a variety of sensing architectures have been identified, which raises the question of their functional differences and the evolutionary forces driving transitions between them. For example, direct NLR–effector interactions can also play a role in ETI, particularly in the recognition of fungal pathogens (15). In other cases, the NLR–effector interactions are mediated by a “decoy” molecule which has no other known function beyond interacting with effectors (16). Fig. 4A shows schematic interaction networks for the guard, direct, and decoy architectures. Fig. 4B introduces the “Integrated Decoy” architecture, where a sensing NLR with an interacting domain works together with an activating NLR (13, 27). Typically, the pair is tightly coregulated, sharing a promoter (28). One NLR (the “activator”) can form homodimers that drive immune activation in the absence of effector. The second NLR (the “sensor”) forms heterodimers with the activator to prevent constitutive activation. When the effector interacts with the sensor’s “integrated decoy” domain, this leads to a conformational change allowing the activator’s interacting domains to bind and trigger an immune response.

Fig. 4.

Known sensing architectures obey functional, proteomic, and evolutionary trade-offs. (A) A schematic of three different sensing architectures: guard, direct, and decoy. (B) A schematic of the Integrated Decoy sensing architecture. The activating NLR can form a higher-order complex with the sensor NLR, leading to effector-sensing, or it can form a homodimer with rate f₀. Each forward reaction f_x has a corresponding back reaction b_x, omitted from the diagram for clarity. For the full model, see SI Appendix. (C) A comparison of the sensing architectures in the trade-off space of sensitivity ($E_1^{\ast }$) and protection (target after activation). (D) The critical effector $E_1^{\ast }$ as a function of the activator self-interaction rate f₀. When $f_0=0$, no activation occurs (SI Appendix, Fig. S6). Parameters: $f_1=f_2=f_3=b_1=b_2=b_3=\kappa =g_{n1}=g_{n2}={\delta }_n={\delta }_t={\beta }_1={\beta }_2=1$ and $g_t=10$.

We can use the model we have developed to compare the functional traits of the different architectures. In particular, we focus on the sensitivity (the critical effector concentration for activation, $E_1^{\ast }$) and the target protection (how much target remains after activation, which should correspond to how well ETI inhibits the pathogen’s virulence). Fig. 4C shows these two traits for the guard, direct, and decoy architectures. (For full equations and bifurcation diagrams, see SI Appendix, Fig. S5.) We hold all parameters the same and simply change the interaction network. We see that the guard architecture is relatively sensitive (i.e. low $E_1^{\ast }$) but very bad at protecting the target. On the other hand, direct recognition is less sensitive but better at protecting the target. It also suffers from the drawback of evolutionary fragility: The NLR has to evolve whenever the effector evolves, losing the elegant “sensing self” robustness of the guard architecture. However, if the direct recognition NLR is expressed at a higher level constitutively ($g_n=5$ in Fig. 4C) as suggested by several recent studies (29–31), the sensitivity trade-off is mitigated at the expense of proteomic cost. The performance of the decoy model also depends on the decoy expression level g_n. When the decoy is expressed at the same level as the target $g_n=g_t=10$, the sensitivity is low and the protection is moderate. When g_n is increased, both the sensitivity and protection improve together, but at a very high proteomic cost. Thus these three architectures are subject to functional, evolutionary, and proteomic trade-offs, suggesting the evolutionary logic underlying them.

In the integrated decoy system, most accounts implicitly assume that activation is driven by the higher-order complexes (Fig. 4B, Right branch of $f_1-f_3$) rather than activator homodimerization (Left branch, f₀), which is simply a curiosity in sensor knockout lines. However, we find that activation is dependent on activator homodimerization, and when $f_0=0$ no discontinuous immune activation occurs (SI Appendix, Fig. S6). Fig. 4D shows that when $f_0\ne 0$, however, the integrated decoy system is exquisitely sensitive. In the guard architecture, interaction parameters of $\mathcal{O}(1)$ lead to $E_1^{\ast }\approx 4$ (Fig. 1B, ${\alpha }_1=0.5$ and ${\gamma }_1=1$, in units where $g_n=1$, ${\delta }_n=1$). The integrated decoy architecture achieves this sensitivity with $f_0=0.01$, and activation persists at vanishingly small interaction rates. Thus in return for maintaining a complex genetic architecture, the plant gets a highly sensitive sensor.

Discussion

Plant Effector-Triggered Immunity encompasses a bewildering array of behaviors, genes, and molecules (32). Here, we propose a physics-based language to organize ETI into precise hypotheses and quantitative predictions. We show that the simplest physical model of the biology leads to a robust immune sensor and predict that activation requires a downstream gene regulatory network to implement nonlinear positive feedback on NLR expression levels. In the language of our model, effector interference is a natural consequence of molecules competing for binding partners, and it takes a generic form regardless of the underlying molecular mechanisms. We show how a single NLR can integrate many effector signals, and we illustrate the trade-offs between sensing architectures in terms of function, proteomic cost, and evolutionary stability.

The model raises several interesting experimental predictions and ecological implications. Little is known about how NLR expression is regulated or how NLRs activate downstream immune responses. Even in the absence of specific pathways, the model suggests looking for nonlinear positive feedback in NLR expression, which in turn would imply a discontinuous transition in NLR expression as a function of effector levels. The model could be constrained by more detailed data on protein–protein interactions and by measuring transcriptional responses at the single-cell level. The strength of the model is that it allows us to make nontrivial predictions without detailed molecular data while laying the groundwork for the next set of experiments.

The effector interference results have interesting implications of their own, suggesting that interference arises generically from shared targets rather than from evolutionary fine-tuning of effector interactions. The model implies that pathogen selection should favor the co-occurrence of effectors with the same target, allowing them to “split the signal” and avoid detection. Moreover, the persistence of the activated branch in effector interference (e.g. Fig. 2A) is suggestive of immune priming, in which a recent infection causes distant tissue to upregulate its defense genes (33). The persistence of the activated branch suggests that primed tissue may circumvent effector interference. This would ultimately be a consequence of the positive feedback where NLR activation stimulates expression: In the presence of interference, the NLR cannot activate on its own, but with a priming signal it can reach high enough levels to do so. Although this study focuses on the activation of a single NLR, the activation of multiple NLRs is an important area of future research, particularly in elucidating the delicate balance between immunity and autoimmunity (34, 35).

The ZAR1 results help explain a puzzling phenomenon. In a recent survey of almost 500 P. syringae strains, 40% were found to harbor an effector recognized by ZAR1 (22, 36). One might naively think that once a pathogen has a single ZAR1-elicitor, there is no selective pressure against acquiring more. However, only 5/150 ZAR1-eliciting strains harbored multiple ZAR1 effectors. Our model suggests an explanation: ZAR1 is more sensitive to effectors spread between kinases (Fig. 3 C and D), so there is evolutionary pressure to avoid this. Instead of splitting the signal (as in effector interference), ZAR1 integrates it.

This work helps make sense of the diverse ETI sensing architectures, showing that different strategies are useful in different contexts. Compared to the guard architecture, direct recognition is less sensitive (Fig. 4C). This makes it appropriate for opportunistic pathogens which might be present in low amounts not requiring an immune response. For example, Almario et al. found that the “core” leaf microbiome of wild A. thaliana includes the pathogenic oomycete Hpa (21), many strains of which are recognized by direct interaction of an effector with the NLR RPP4 (37). The low sensitivity of direct recognition might prevent a costly immune response when it is not necessary. In contrast, the high sensitivity of the Integrated Decoy architecture (Fig. 4D) might be appropriate for effectors with a severe virulence impact. The “sensor” NLR RRS1 has an integrated decoy domain which mimics the WRKY transcription factor domain and interacts with effectors ArvRps4 and PopP2 in the nucleus. PopP2 inhibits at least 10 TFs involved in PTI (38), and perhaps enhanced sensitivity is necessary for effectors which target transcription directly instead of the downstream proteins. These individual examples do not prove that the sensing architectures always play the same role in the field as in the model, but in the absence of statistical characterization of an unbiased sample of effectors and NLRs, they are suggestive and open the door to future research. For example, models of plant-pathogen coevolution could elucidate which sensing architectures are evolutionarily stable under various circumstances. More generally, these mechanistic insights could provide the inputs for a new generation of evolutionary models in which the interactions account for specific combinations of host and pathogen genotypes.

An important feature of the direct and decoy architectures is that they protect the effector target from degradation (Fig. 4C), potentially preserving the PTI response. This is especially relevant for targets which play a role in defense beyond the cell. For example, the Cmu1 effector secreted by Ustilago maydis (the organism responsible for maize smut) suppresses Salicylic Acid (SA) biosynthesis (39). SA is an important immune signaling molecule and is believed to be one of the mobile signals responsible for Systemic Acquired Resistance (SAR), in which an infection in one part of the plant ramps up defense genes across the organism (40). Such targets involved in SAR are promising candidates for protection via direct recognition or decoys.

In the model, the decoy architecture protects the target by “competing” with the target for effector molecules, a mechanism which was suggested in an early paper on the decoy model (16). It also provides a striking parallel to decoy receptors employed by both mammals and their viral pathogens to modulate the immune system (41). Thus our work suggests that a canonical immune regulatory mechanism extends across kingdoms of life, which would make it a very general phenomenon.

Simple models do not necessarily make quantitative predictions about particular organisms. However, they can help discover general principles governing diverse phenomena. Starting with the premise that plant ETI is implemented by physical interactions between molecules, we built a simple model which shows how the diverse recognition structures of ETI function according to a few simple rules and trade-offs. It is our hope that, beyond the specific findings presented here, this perspective based on competition for binding partners will permeate the field and help link disparate efforts in molecular biology, plant physiology, and ecology.

Materials and Methods

The differential equations for the NLR concentration n_i, target concentration t_j, and activated NLR concentration a_i (Eqs. 1–3) were analyzed numerically using Mathematica. Specifically, the fixed points were determined by setting Eqs. 1–3 equal to 0 and solving the system of equations using the “FindRoot” function. The differential equations also depend on e_k, which is itself a function of n_i, t_j, and a_i given by Eq. 5.

Details on the derivations of Eqs. 1–5 and the other sensing architectures can be found in SI Appendix. Code is available on Github at: https://github.com/BenjaminWeiner/ETI.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

There are no data underlying this work.