Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 Apr 30;122(2):e70169. doi: 10.1111/tpj.70169

In silico prediction method for plant Nucleotide‐binding leucine‐rich repeat‐ and pathogen effector interactions

Alicia Fick 1,2, Jacobus Lukas Marthinus Fick 3, Velushka Swart 1,2, Noëlani van den Berg 1,2,✉
PMCID: PMC12042882  PMID: 40304719

SUMMARY

Plant Nucleotide‐binding leucine‐rich repeat (NLR) proteins play a crucial role in effector recognition and activation of Effector triggered immunity following pathogen infection. Genome sequencing advancements have led to the identification of a myriad of NLRs in numerous agriculturally important plant species. However, deciphering which NLRs recognize specific pathogen effectors remains challenging. Predicting NLR–effector interactions in silico will provide a more targeted approach for experimental validation, critical for elucidating function, and advancing our understanding of NLR‐triggered immunity. In this study, NLR–effector protein complex structures were predicted using AlphaFold2‐Multimer for all experimentally validated NLR–effector interactions reported in literature. Binding affinities‐ and energies were predicted using 97 machine learning models from Area‐Affinity. We show that AlphaFold2‐Multimer predicted structures have acceptable accuracy and can be used to investigate NLR–effector interactions in silico. Binding affinities for 58 NLR–effector complexes ranged between −8.5 and −10.6 log(K), and binding energies between −11.8 and −14.4 kcal/mol−1, depending on the Area‐Affinity model used. For 2427 “forced” NLR–effector complexes, these estimates showed larger variability, enabling identification of novel NLR–effector interactions with 99% accuracy using an Ensemble machine learning model. The narrow range of binding energies‐ and affinities for “true” interactions suggest a specific change in Gibbs free energy, and thus conformational change, is required for NLR activation. This is the first study to provide a method for predicting NLR–effector interactions, applicable to all pathosystems. Finally, the NLR–Effector Interaction Classification (NEIC) resource can streamline research efforts by identifying NLRs important for plant–pathogen resistance, advancing our understanding of plant immunity.

Keywords: Nucleotide‐binding leucine‐rich repeat, effector, effector triggered immunity, plant–pathogen interactions, NLR–effector interactions, technical advance

Significance Statement

We show that binding energy and affinity values may be used to predict novel plant–pathogen NLR–effector interactions. Thus, this is the first study to provide a machine learning‐based tool for predicting NLR–effector interactions with high accuracy, which can be used to advance our understanding of the biochemical mechanisms regulating NLR activation.

INTRODUCTION

During plant–pathogen interactions, both organisms produce a wide array of proteins that significantly influence the outcome of the infection (Ammari et al., 2016; Li et al., 2022; Naidoo et al., 2018). Plant Nucleotide‐binding leucine‐rich repeat (NLR) proteins recognize pathogen effectors secreted into the cytoplasm of host cells, triggering immune responses in the plant (Contreras, Lüdke, et al., 2023; Dodds et al., 2024; Kourelis & van der Hoorn, 2018). However, when effectors evade recognition by NLRs, they can disrupt the host's immune signaling pathways, leading to successful pathogen infection (Zhang et al., 2022). Given their critical role in plant defense, NLRs are the focus of many plant–pathogen interaction studies. Understanding NLR–effector interactions is crucial for advancing crop breeding programs and NLR engineering (Contreras, Lüdke, et al., 2023; de Araújo et al., 2019).

The growing availability of plant genome and RNA‐sequencing data has significantly contributed to the identification of numerous NLR genes, particularly in important agricultural crops (Li et al., 2023; Liu et al., 2024; Santos et al., 2022; Yu et al., 2023). Additionally, research has successfully modified NLR proteins to either resurrect protein function or to expand on the number of recognized effectors, which may ultimately enhance pathogen resistance levels (Contreras, Pai, et al., 2023; Maidment et al., 2023; Tamborski et al., 2023). However, identifying specific NLRs that recognize particular pathogen effectors and activate immune responses remains challenging due to the complexity of plant–pathogen interactions and the multitude of potential NLR–effector pairings (Barragan & Weigel, 2021; Yang et al., 2019; Zheng et al., 2021). Experimental validation methods, such as yeast two‐hybrid systems or co‐immunoprecipitation assays, are technically demanding, time‐consuming, and expensive, limiting the number of interactions that can be studied in research settings (Zheng et al., 2021).

Studying NLR–effector interactions and identifying functionally important NLRs can be streamlined by focusing on NLRs that directly bind effectors, function as singletons, and activate immune responses (Adachi & Kamoun, 2022; Contreras, Lüdke, et al., 2023). Some NLRs operate in NLR networks, where sensor NLRs bind to effectors and activate helper NLRs to trigger immune responses (Contreras, Lüdke, et al., 2023), while others guard effector‐targeted proteins, without directly binding to the effectors themselves (Jones & Dangl, 2006). Focusing on NLRs that directly bind to effectors simplifies the prediction of unknown NLR–effector interactions. Of 419 experimentally validated NLRs, 67 NLRs directly recognize 93 effectors and do not form part of NLR networks (Brabham et al., 2024; Kourelis et al., 2021; Redkar et al., 2023).

Most data on NLRs are from the Asterid group, which exhibits highly diversified sensor and helper NLR families (Adachi & Kamoun, 2022; Wu et al., 2017). In contrast, studies on NLRs in other plant groups such as Rosids, including NLRs from Arabidopsis thaliana, Vitis vinifera (grape), and Manihot esculenta (cassava), reveal a higher proportion of singleton NLRs, with no network‐related NLRs detected in these genomes (Wu et al., 2017). This suggests that more singleton NLRs are likely to be found in species whose genomic sequences are not yet available. Prigozhin and Krasileva (2021) demonstrated that direct‐recognition NLRs display higher amino acid (AA) diversity per AA site as measured by Shannon entropy scores. These higher scores are particularly observed in the leucine‐rich repeat domains of NLRs (NLRLRR), which bind effectors and govern recognition specificity. Therefore, Shannon entropy scores may be useful in predicting singleton NLRs with direct effector‐recognition capabilities (Prigozhin & Krasileva, 2021). Once potential direct‐recognition NLRs are identified, predictions can be made regarding the specific effectors each NLR recognizes.

Various programs for predicting plant–pathogen protein–protein interactions are available, as reviewed by Li and Zhang (2016) and Yang et al. (2019) and Lei et al. (2024) (AraPathogen2.0). These tools use a range of approaches, including protein domain analysis (Deng et al., 2002), sequence similarity (Matthews et al., 2001), gene co‐expression, and protein structure to predict interactions (Zhang et al., 2012). Machine learning (ML) programs have significantly advanced predictions of plant–pathogen protein–protein interactions, but they are primarily trained on data from A. thaliana and Oryza sativa (rice) (Karan et al., 2023; Lei et al., 2024; Yang et al., 2019; Zheng et al., 2021). However, due to the specificity of NLR effector‐recognition being influenced by single‐nucleotide mutations in both NLRs and effectors, current programs are inadequate for predicting NLR–effector interactions in plant species without available protein–protein interaction data (Dodds et al., 2001; Ortiz et al., 2022; Segretin et al., 2014; Tamborski et al., 2023). The scarcity of “true” NLR–effector interaction data further complicates accurate predictions, making it difficult to identify novel NLR–effector interactions (Sharma, 2022).

In this study, we aimed to develop an in silico method for predicting NLR–effector interactions. These predictions are expected to accelerate our understanding of plant–pathogen interactions and enable a more targeted approach to investigate both NLR‐ and pathogen effector functionality (Jones et al., 2024). As more data on NLR–effector pairs become available, dedicated ML programs can be developed to analyze plant–pathogen protein–protein interactions with greater accuracy. We demonstrate that NLR–effector protein structures can be predicted with acceptable accuracy using AlphaFold2‐Multimer, showing strong comparability to experimentally validated structures, thus making them suitable for studying NLR–effector interactions. The predicted structures were then used to calculate binding affinities (BA) and binding energies (BE) through multiple ML models, which were combined to train an Ensemble learning model. Significant differences in BA‐ and BE‐values were observed between “true” NLR–effector partners and nonfunctional (“forced”) NLR–effector partners that do not activate plant immune responses. These differences enabled the prediction of NLR–effector interactions using the trained Ensemble model. Our findings enhance the understanding of NLR–effector interactions, NLR functionality, and effector triggered immunity (ETI) activation.

RESULTS AND DISCUSSION

NLRLRR –effector interaction prediction using a protein structure‐based approach

Although ML programs have been developed to predict plant protein–protein interactions using a structure‐based approach, these tools do not allow users to train new models on specific datasets for particular protein families (Dodds et al., 2001; Ortiz et al., 2022; Segretin et al., 2014; Tamborski et al., 2023). Moreover, these programs have not been specifically designed to predict NLRLRR–effector interactions. Therefore, our objective was to develop a novel method tailored to predict NLRLRR–effector interactions by focusing on comparisons between “true” and “forced” NLRLRR–effector protein complex data.

First, we evaluated the ability of AlphaFold2‐Multimer to accurately predict the structures of NLRLRR–effector complexes and determine a suitable AlphaFold (AF) confidence score threshold for reliable predictions (Evans et al., 2022). This assessment involved comparing three cryogenic‐electron microscopy (cryo‐EM) structures of known NLRLRR–effector complexes (Sr35‐AvrSr35, RPP1‐ATR1, and Roq1‐XopQ) with 90 predicted complex structures that exhibited a range of AF confidence scores (Ma et al., 2020; Martin et al., 2020; Zhao et al., 2022). This approach allowed us to establish a reliable confidence score for subsequent analyses of NLRLRR–effector interactions.

A strong correlation was observed between AF confidence scores and DockQ scores (R = 9.3, P < 0.01), as well as TM‐scores (R = 0.85, P < 0.01), indicating that AF confidence scores are a reliable indicator of NLRLRR–effector complex prediction accuracy (Figure 1). All but one of the predicted NLRLRR–effector complexes with AF confidence scores >0.42 exhibited DockQ scores >0.23 and TM‐scores >0.6, suggesting that these predicted complexes closely resemble cryo‐EM structures. Additionally, CAPRI‐ranked complexes of acceptable, medium, and high quality were found to have AF confidence scores ≥0.4, further supporting the reliability of AF confidence scores in assessing the accuracy of predicted structures. These strong correlations suggest that AF confidence scores can serve as a reliable metric for evaluating the accuracy of predicted NLRLRR–effector complexes. Based on these findings, an AF confidence score ≥0.42 was selected as the appropriate cut‐off value for identifying predicted NLRLRR–effector structures with acceptable accuracy. Previous studies have also demonstrated that structures with AF confidence scores between 0.4 and 0.6 are generally accurate, with AF confidence scores often used as proxies for estimating protein–peptide interactions (Bret et al., 2024; Homma et al., 2024; Yin et al., 2023; Yin & Pierce, 2024). Notably, structures predicted using a template‐free approach, without the incorporation of PDB100 templates, exhibited the lowest AF confidence scores. These findings suggest that for reliable and accurate prediction of NLR–effector complexes, the use of PDB100 templates is strongly recommended. To further validate the use of AF confidence scores >0.42 as an indicator of predictive accuracy, we compared the interacting AAs between the predicted NLRLRR–effector complexes and cryo‐EM structures. This comparison focused on AA rotation and their spatial location within the protein, providing additional evidence that AF confidence scores are a meaningful indicator of the structural fidelity of predicted interactions.

Figure 1.

Figure 1

Open in a new tab

Evaluation of the model accuracy based on the correlation between AlphaFold (AF) confidence scores, DockQ scores, and TM‐scores.

The distribution of AF confidence scores (calculated as 0.2 × pTM + 0.8 × ipTM) for 90 predicted NLRLRR–effector structures is shown in relation to (a) DockQ scores and (b) TM‐scores. Structures for Sr35‐AvrSr35, RPP1‐ATR1, and Roq1‐XopQ complexes were predicted using different protocols in AlphaFold2‐Multimer v.3 and compared against their native cryo‐EM structures. The models were also ranked according to the stringent protein–protein complex evaluation criteria set by the CAPRI community. Dotted lines indicate the thresholds for accurately predicted structures: DockQ scores >0.23 and TM‐scores >0.6. An AF confidence score of ≥0.42 was identified as the cutoff for acceptable prediction accuracy of both the complex and complex interface (PDB structure IDs: 7XVG, 7CRB, and 7JLU). The correlation between AF confidence scores and structural accuracy was measured using Pearson's correlation coefficient.

(c) Three examples of Sr35‐AvrSr35 predicted structures are shown with varying AF confidence scores, compared to the cryo‐EM Sr35‐AvrSr35 structure. Predicted structures are color‐coded based on their CAPRI rankings, with the cryo‐EM structure shown in gray. This comparison highlights the relationship between AF confidence scores and the accuracy of predicted structures relative to experimental data.

As expected, a strong negative correlation was observed between AF confidence scores and both i‐RMSDbb (deviation in predicted interface backbone atoms) scores (R = −0.89, P < 0.01) and i‐RMSDsc (deviation in predicted interface sidechain atoms) scores (R = −0.91, P < 0.1) (Figure 2). Using an RMSD threshold of <3 Å to indicate acceptable accuracy and comparable placement of the predicted atoms relative to cryo‐EM structures, 33 predicted structures were identified as highly accurate (Manandhar et al., 2022; Ramírez & Caballero, 2018). These structures had an average AF confidence score of 0.84, underscoring the importance of a high AF confidence score when analyzing specific interacting AAs between NLRsLRR and effectors in silico. Further comparisons of hydrogen bond locations and interacting AA positions between predicted and cryo‐EM NLRLRR–effector complexes revealed significant inaccuracies in the predicted interacting AAs for complexes with AF confidence scores <0.42 (Figure 3a–c). These inaccuracies were particularly pronounced for effector proteins, with greater variability observed in the specific interacting AAs when comparing predicted structures (with different AF confidence scores) to their corresponding cryo‐EM structures. This is unsurprising, as effectors often interact within the concave region of the LRR domain, and their positioning during interaction can vary considerably (Ma et al., 2020; Martin et al., 2020; Zhao et al., 2022). While the NLR's interacting AAs are generally restricted to the concave side of the LRR domain, effectors can adopt a wider range of orientations and positions within the complex, contributing to the observed variability in predicted interactions.

Figure 2.

Figure 2

Open in a new tab

Evaluation of the model accuracy based on the correlation between AF confidence scores and interface RMSD scores.

Correlation between AF confidence scores and interface root mean squared difference (RMSD) scores for (a) interface backbone (i‐RMSDbb) and (b) interface sidechain (i‐RMSDsc) residues. The RMSD scores were calculated by comparing 90 predicted NLRLRR–effector complex structures (30 Sr35‐AvrSr35 complexes, 30 RPP1‐ATR1 complexes, and 30 Roq1‐XopQ complexes) to their corresponding cryogenic‐electron microscopy (cryo‐EM) structures. The predictions were made using AlphaFold2‐Multimer, following different protocols. AF confidence scores were calculated using pTM‐ and ipTM values (0.2 × pTM + 0.8 × ipTM). Pearson's correlation coefficient values are indicated above each graph.

Panels (c, d) provide visual examples for i‐RMSDbb and i‐RMSDsc scores, respectively, using Sr35‐AvrSr35 protein complexes as examples. In these images, interacting atoms between the NLRLRR and effector proteins from the cryo‐EM structure are shown in black, while interacting atoms from predicted structures with varying AF confidence scores are color‐coded. Higher i‐RMSDbb and i‐RMSDsc scores indicate greater structural deviations between the predicted complexes and the cryo‐EM reference structures (PDB structure IDs: 7XVG, 7CRB, and 7JLU).

Figure 3.

Figure 3

Open in a new tab

Location of amino acids (AAs) involved in the formation of hydrogen bonds during NLRLRR–effector interactions.

Cryogenic‐electron microscopy (cryo‐EM) structures (native) and three representative AlphaFold2‐Multimer predicted structures for (a) Sr35‐AvrSr35, (b) RPP1‐ATR1, and (c) Roq1‐XopQ are shown. Each protein structure from an NLRLRR–effector pair is displayed separately, with the NLRLRR structure positioned above the effector structure. AAs that form hydrogen bonds during the interaction are depicted in black. For each predicted complex, the AF confidence scores (calculated as 0.2 × pTM + 0.8 × ipTM) are displayed, along with DockQ scores, TM‐scores, and RMSD scores for both i‐RMSDbb and i‐RMSDsc residues. These values were obtained by comparing the predicted structures to the corresponding native cryo‐EM structures.

Panels (d, e) show the percentage of shared AAs forming hydrogen bonds in predicted structures compared to their native counterparts for both NLRLRR and effector proteins, across varying AF confidence scores. The percentage is calculated relative to the total number of hydrogen bonds within the respective predicted NLRLRR–effector complex. For predicted structures, AAs involved in hydrogen bond formation that are not shared with the native structure are indicated according to how far they are from the native AA location. These non‐shared AAs are color‐coded as 1 (red), 2 (turquoise), 3 (purple), 4 (blue), or 5 (pink) AAs away from the original bond location in the native structure, either in the N‐ or C‐terminus direction. A significant positive correlation (R = 0.62; P < 0.01) was observed between the number of shared bonds for NLRLRR structures in predicted versus native complexes. However, no significant correlation was observed for bonds formed between predicted and native effector proteins (R = 0.11; P > 0.05). This indicates that while NLRLRR structures with high AF confidence scores reliably predict hydrogen bonding interactions, the same cannot be said for effector proteins, where the variability in predicted bond locations is higher.

It is worth noting that, on average, 80% of the interacting AAs are shared between predicted NLRLRR structures with AF confidence scores >0.6 and their corresponding cryo‐EM structures. Additionally, less than 20% of predicted interacting AAs are located within 1–5 residues of the interacting AAs in the related cryo‐EM structure (Figure 3d). As shown in Figure 3(a–c), hydrogen bonds (depicted in black) in similar regions of the predicted NLRLRR domains when compared to their cryo‐EM counterparts. These findings suggest that NLRLRR–effector complexes with AF confidence scores >0.6 can be reliably used to investigate which AAs are likely to form bonds with the recognized effector protein.

In contrast, for predicted structures with AF confidence scores <0.6, the percentage of shared interacting AAs decreases significantly to 53% (R = 0.62; P < 0.01). Additionally, more interacting AAs tend to be located further from their interacting AAs in cryo‐EM structures. While these models with lower AF confidence scores may still be informative for in silico investigations of interacting AAs, the accuracy is more limited. Importantly, this trend was not observed for interactions relative to the effector protein itself. There was no significant difference in the percentage of shared bonds between predicted and cryo‐EM structures when analyzing AAs interacting with the effector (R = 0.11; P > 0.05) (Figure 3e). This indicated that the predicted bonds between AAs in the effector protein cannot reliably predict which AAs have a higher probability of contributing to effector recognition by an NLRLRR, highlighting a limitation in using predicted models to infer effector‐specific interactions.

Comparisons between “true” and “forced” NLRLRR –effector interactions

Since NLRLRR–effector structures with AF confidence scores >0.42 have demonstrated acceptable accuracy, we compared structures of “true” and “forced” NLRLRR–effector interactions to identify any notable differences between these two classes. In total, 6022 structures were predicted, including 93 “true” NLRLRR–effector interactions. Neither model nor interface accuracy (plDDT, pTM, and ipTM values) showed significant differences between “true” and “forced” interaction complexes (Figure 4). This is in contrast with a previous study by Martin (2024), which suggested that high ipTM values could distinguish between “true” and “forced” interactions. Our findings indicate that these metrics primarily reflect the confidence level in the predictive accuracy and the interface location between NLRLRR‐ and effector structures (Evans et al., 2022; O'Reilly et al., 2023), rather than distinguishing “true” or “force” interactions. Therefore, these metrics cannot reliably predict whether an unknown NLRLRR–effector is “true” or “forced.” Our findings align with a study on interactions between human genome maintenance proteins, where AFM confidence metrics were also insufficient for identifying “true” interactions (Schmid & Walter, 2025).

Figure 4.

Figure 4

Open in a new tab

Confidence metrics and interaction interface comparisons for “true” and “forced” NLRLRR–effector structures predicted by AlphaFold2‐Multimer.

No significant differences were observed between 93 “true” and 5930 “forced” NLRLRR–effector interactions for (a) pLDDT, (b) pTM, (c) ipTM, or (d) the number of predicted contacts. Statistical significance was assessed using Wilcoxon rank‐sum tests for box‐and‐whisker plots. Correlations in the scatterplot were evaluated using Pearson's correlation coefficient (R = 0.45; P > 0.05), indicating no statistically significant relationships.

The number of predicted contacts between NLRLRR and effector proteins also did not differ significantly when comparing “true” and “forced” interactions (Figure 4d). Previous studies have indicated that the number of contacts correlates with BA, suggesting “true” NLRLRR–effector interactions might have more contacts (Raucci et al., 2018; Vangone & Bonvin, 2015). However, the specific AAs involved in forming these contacts also influence BA (Yi et al., 2024). Thus, while the total number of predicted contacts is similar between “true” and “forced” NLRLRR–effector complexes, differences in AAs involved in the interaction could explain variations in BA. These findings reinforce the conclusion that neither AlphaFold2‐Multimer predicted structures, nor the associated confidence metrics (plDDT, pTM, or ipTM values), nor visual inspection of predicted structures can be used to reliably predict novel NLRLRR–effector interactions.

Binding affinity‐ and binding energy‐values for “true” and “forced” NLRLRR –effector complexes

Given the lack of identifiable differences in predicted structures between “true” and “forced” NLRLRR–effector interactions, we proceeded to investigate the interaction strength by predicting binding affinity (BA)‐ and binding energy (BE)‐values. After excluding structures with AF confidence scores <0.42, 58 “true” and 2427 “forced” NLRLRR–effector complex structures were retained for further analysis. The BE and BA values for each complex were predicted using the Area‐Affinity web tool, with the predicted complex structures as input (Yang et al., 2023). As expected, there was considerable variation in both BE and BA values among the NLRLRR–effector complexes when analyzed across the 97 models used by Area‐Affinity (Figures 5 and 6, respectively). BE values ranged from −0.049 kcal/mol−1 (protein–protein Generated nonlinear model [linear fitting] 8) to −40.283 kcal/mol−1 (protein–protein Linear model 10), while BA values ranged from −0.036 log(K) to −29.535 log(K) for the same models. However, no significant difference in BE and BA values were observed between “true” and “forced” NLRLRR–effector complexes within any specific model. Interestingly, most models consistently predicted higher BE and BA values for some “forced” NLRLRR–effector complexes, suggesting stronger interactions between proteins that do not naturally interact. For instance, the average predicted BE and BA values across all models for the “true” interaction of Rpi‐blb2LRR‐Avrblb2 were −12.681 kcal mol−1 and −9.344 log(K), respectively. In contrast the “forced” interaction of Rpi‐blb1LRR‐Avrblb2 yielded BE and BA values of −14.728 kcal/mol−1 and −10.798 log(K), respectively. The lower BE and BA values suggest that Avrblb2 binds more tightly to Rpi‐blb1LRR, although this interaction does not result in the activation of host immune responses (Champouret et al., 2009; Oh et al., 2014).

Figure 5.

Figure 5

Open in a new tab

Predicted binding energies for “true” and “forced” NLRLRR–effector interactions.

(a) Binding energies were predicted for protein complexes using 60 protein–protein machine learning models, and (b) 37 antibody–protein machine learning models, both from the Area‐Affinity platform. NLRLRR–effector complex structures were predicted using AlphaFold2‐Multimer, and only complexes with AF confidence scores (calculated as 0.2 × pTM + 0.8 × ipTM) > 0.42 were included in the binding energy analysis. “True” interactions are depicted by filled box‐and‐whisker plots, while “forced” interactions are represented by unfilled plots.

Figure 6.

Figure 6

Open in a new tab

Predicted binding affinities for “true” and “forced” NLRLRR–effector interactions.

(a) Binding affinities were predicted for protein complexes using 60 protein–protein machine learning models, and (b) 37 antibody–protein machine learning models, both from the Area‐Affinity platform. NLRLRR–effector complex structures were predicted using AlphaFold2‐Multimer, with only complexes having AF confidence scores (calculated as 0.2 × pTM + 0.8 × ipTM) > 0.42 included in the binding affinity analysis. “True” interactions are represented by filled box‐and‐whisker plots, while “forced” interactions are shown using unfilled plots.

These results suggest that only NLRLRR–effector interactions with certain BE and BA values are likely to trigger NLR activation and consequently immune response activation (McBride et al., 2022). This raises an interesting question: why do BE and BA values show less variability for “true” NLRLRR–effector interactions compared to “forced” interactions? BA and BE values are influenced by several factors, including the number of hydrogen bonds between interacting proteins and the size of the interaction interface (Kastritis & Bonvin, 2013; Klebe & Böhm, 1997). Given the diversity of “true” NLRLRR–effector interactions, particularly in terms of protein size, we expected larger variations in BE and BA values. We hypothesize that the lower variability in BE and BA values of “true” interactions may be due to coevolutionary processes, as proposed by the Red Queen dynamic hypothesis (Lighten et al., 2017; Van Valen, 1973). According to this model, NLRLRR domains and effectors are engaged in a continuous coevolutionary arms race: NLRs evolve to recognize effectors, while effectors evolve to evade recognition (Rabajante et al., 2016; Råberg, 2023; Singh et al., 2018). This dynamic may lead to an equilibrium in BA between NLRs and effectors, where a conserved conformational change within NLRs is required for NLR activation. As a result, “true” NLRLRR–effector interactions may exhibit more consistent BE and BA values, reflecting a finely tuned balance shaped by this evolutionary pressure.

These results suggest that while effector recognition specificity is largely governed by the LRR domain, the NB‐ARC domain may ultimately be responsible for determining NLR activation (Locci & Parker, 2024). This aligns with previous research that reported comparable Shannon entropy scores for regions within both the NB‐ARC‐ and LRR domains, indicating similar mutation rates (Prigozhin & Krasileva, 2021). This suggests that NB‐ARC domains may play a role in effector binding, which could explain why effectors with stronger BA to an “incorrect” LRR domain fail to activate the corresponding NLR. This supports a hypothesis proposed by Tamborski and Krasileva (2020) regarding NLR activation mechanisms. The authors suggested that both the “switch model” and the “equilibrium‐based switch model” could coexist within a system. According to the “switch model,” NLR activation occurs only when an effector binds to an ADP‐bound NLR, triggering an intramolecular signal that induces a conformational change, leading to an ADP‐ATP switch and subsequent activation of defense signaling (Bernoux et al., 2016; Maekawa et al., 2011; Tameling et al., 2002; Williams et al., 2011). In contrast, the “equilibrium‐based switch model” suggests that NLRs continuously cycle between ADP‐ and ATP‐bound states, with effectors capable of binding to either state. Together, these models highlight the central role of the NB‐ARC domain in regulating NLR activity and suggest that its function may be modulated by the overall auto‐inhibitory strength of other domains (such as CC/TIR or NB domain) or by NLRLRR–effector recognition.

Implementation of the Ensemble learning model

After training and testing 777 different Ensemble models, the highest accuracy of 99.03% was achieved for an Ensemble model using the AdaBoost method. Results from the confusion matrix showed 20 “forced” NLR–effector interactions classified as being “true” (false‐positive), and three “true” interactions as “forced” (false‐negative). This model was trained with 202 maximum splits, 61 learners, a learning rate of 0.08 sec, and the ReliefF algorithm for selecting the 21 most important features. These features (BA or BE values from various Area‐Affinity ML models) were primarily predicted by Area‐Affinity Nonlinear models, which displayed the greatest variability in BA or BE values for this particular model. Optimal hyperparameter tuning resulted in a minimum classification error rate after just four iterations, further demonstrating the model's high accuracy. This model can classify NLRLRR–effector interactions as either “true” or “forced” with very high accuracy. The trained model has been implemented as the NLRLRR–Effector Interaction Classification (NEIC) application, which is now publicly available. Users can classify unknown interactions by following the steps outlined in Figure 7, or as explained on the associated GitHub site (https://github.com/koosfick/NEIC1.0). However, caution is advised when investigating NLRLRR‐bacterial effector interactions. Since only four predicted NLRLRR‐bacterial effector complexes had acceptable accuracy (AF confidence score >0.42), the model was trained using only these four complexes. Effector genes from different pathogenic species exhibit considerable variation in mutation rates, sizes, and amino acid composition, suggesting that separating data based on effector type (fungal, oomycete, or bacterial) in future studies, may further improve the Ensemble model's accuracy (Erijman et al., 2014; Gómez‐Pérez & Kemen, 2021; Kemen et al., 2011; Yi et al., 2024). As more NLRLRR–effector interaction data become available, future studies could focus on developing interaction–prediction models tailored to specific effector types, which would enhance both the precision and utility of ML models in this field.

Figure 7.

Figure 7

Open in a new tab

Step‐by‐step guide for studying NLRLRR–effector interactions using NLRLRR–Effector Interaction Classifier (NEIC).

Begin by obtaining NLR leucine‐rich repeat (LRR) domain sequences and effector sequences (with signal peptide‐sequences removed). These sequences can be used for protein complex prediction. NLRLRR–effector complex structures can be predicted using AlphaFold2‐Multimer. Only include complex structures with AF confidence scores >0.42 (AF confidence scores = 0.2 × pTM + 0.8 × ipTM) for further analysis. Next, predict binding affinity (BA) and binding energy (BE) values using the Area‐Affinity platform. Ensure the predicted data are transformed into the correct input format using the Data_Transformation sheet. Once formatted correctly, the NEIC can be used to predict which interactions are most likely to occur.

NEIC was tested on Rpi‐chc1.2‐ and rpi‐tub1.3 complexes interacting with two variations of PexRD31, following manual mutation of NLRLRR sequences (Monino‐Lopez et al., 2021). The results from NEIC mostly aligned with the experimental findings from Monino‐Lopez et al. (2021), which were tested using co‐infiltration in N. benthamiana, achieving 75% accuracy (Figure S1). Four mutated RB::C2 NLRs (RB::C2_14–19; RB::C2_19‐A; RB::C2_19‐B; RB::C2_18) were experimentally shown to activate localized cell death following recognition of PexRD31‐B and PexRD31‐C, while four other mutants (RB::C2_19‐C; RB::C2_17+18; RB::C2_16; and RB::C2_16+19) failed to induce cell death when co‐expressed with PexRD31‐B and PexRD31‐C (Monino‐Lopez et al., 2021). NEIC correctly classified 12 out of 16 interactions as “true” or “forced,” indicating that NEIC can be used to investigate the impact of single AA mutations on NLRLRR–effector interactions, though with slightly reduced accuracy. Despite this, NEIC's ability to correctly classify most interactions suggests it can still provide a valuable tool for guiding experimental validation of NLR–effector interactions. In future studies, all possible NLRLRR variations could be generated through in silico mutation, allowing for the identification of crop varieties with enhanced resistance. This research addresses a crucial aspect of understanding the plant immune system, as emphasized by Jones et al. (2024).

CONCLUSION

Despite the inherent limitations of predictions, such as the reliance on AF‐predicted structures, the program developed in this study offers valuable insights into the mechanisms of NLR–effector interactions. These insights can help generate hypotheses that streamline experimental procedures, ultimately accelerating molecular efforts to understand plant immune responses, particularly in non‐model organisms. Our findings demonstrate that NLRLRR–effector structures with AF confidence scores >0.42 show acceptable comparability with experimentally validated structures, making them useful for investigating NLRLRR–effector interactions. Additionally, we found that BEs and BAs for “true” NLRLRR–effector interactions tend to fall within a specific range, which can assist in identifying novel interactions. Importantly, all programs used in this study are freely available online, enabling the prediction of new NLR–effector interactions without requiring access to high‐performance computing resources. This provides a more efficient and targeted approach for studying these interactions in planta. As more data on NLRLRR–effector interactions become available, specialized AI models can be developed to further improve prediction accuracy beyond the capabilities of current tools. Notably, this is the first program specifically designed for predicting NLRLRR–effector interactions with high accuracy, making a significant step forward in plant immune system research.

EXPERIMENTAL PROCEDURES

NLR and effector sequences

NLR and effector protein sequences, from 93 known interactions, previously identified by Kourelis et al. (2021), were downloaded from the Uniprot online database (https://www.uniprot.org). NLR sequences included those from barley (Hordeum vulgare subsp. vulgare), bread wheat (Triticum aestivum), crab apple (Malus × robusta), tomato (Solanum lycopersicum), muskmelon (Cucumis melo), Arabidopsis (A. thaliana), Indica rice (O. sativa sub sp. indica), Japonica rice (O. sativa sub sp. japonica), flax (Linum usitatissimum), potato (Solanum tuberosum, Solanum bulbocastanum, Solanum × edinense, Solanum venturi, Solanum chacoense), soybean (Glycine max), tobacco (Nicotiana benthamiana), cultivated einkorn wheat (Triticum monococcum sub sp. monococcum), rye (Secale cereale), wild chili pepper (Capsicum chacoense), and maize (Zea mays). Additionally, NLR sequences from S. chacoense (Rpi‐chc1.2_543–5; QZA82918.1) and S. tuberosum (rpi‐tub1.3_RH89‐039‐16; QZA82920.1) were also retrieved from Uniprot. Specific amino acid residues in these sequences were manually altered based on experimentally validated mutations reported by Monino‐Lopez et al. (2021). In total, eight mutated versions of these NLRLRR structures were used in the study.

Pathogen effector sequences included oomycete effectors (Hyaloperonospora arabidopsidis; Phytophthora infestans; and Phytophthora sojae), fungal effectors (Blumeria graminis f. sp. hordei; B. graminis f. sp. secalis; B. graminis f. sp. triticale; B. graminis f. sp. tritici; Fusarium oxysporum f. sp. lycopersici; F. oxysporum f. sp. melonis; Magnaporthe oryzae; Melampsora lini; Parastagonospora nodorum; and Pyrenophora tritici‐repentis) and bacterial effectors (Erwinia amylovora; Pseudomonas syringae; Ralstonia pseudosolanacearum; Xanthomonas campestris pv. vesicatoria; Xanthomonas oryzae pv. oryzae; and Burkholderia andropogonis). Additionally, two P. infestans PexRD31 sequences (XP_002897637 and XP_002897638) were also used for investigating interactions with Rpi‐chc1.2 and rpi‐tub1.3 (Monino‐Lopez et al., 2021).

Leucine‐rich repeat domains of NLRs (NLRLRR) and effector signal domains were identified using Homologous Superfamily results from the online InterPro web tool with default parameters. For protein prediction, only the NLRLRR domain sequences were used, as they are the main determinants for effector recognition specificity in singleton NLRs (Kourelis et al., 2021; Locci & Parker, 2024). For effector sequences, signal domains were removed before further analysis. The complete dataset used is available in Data S1.

NLRLRR –effector protein complex prediction

NLRLRR–effector protein complex structures were predicted using AlphaFold2‐Multimer v.3 in Google Colab (Accessed September 2023–July 2024), employing PDB100 templates and unpaired multi‐sequence alignments (MSA) (Evans et al., 2022). Complex structures for all known interacting NLRLRR–effector partners (93 complexes) were predicted, along with “forced” NLRLRR–effector complexes (5930 complexes). For each complex, five structures were generated, and the one with the highest combination of pLLDT (per‐residue model confidence score), pTM (predicted template modeling score), ipTM (interface pTM), and AF confidence score (0.8 × ipTM + 0.2 × pTM) was selected for further analysis (Evans et al., 2022). All these scores range from 0 to 1, representing low and high predicted model accuracy, respectively. An AF confidence score >0.5 suggests acceptable model accuracy; however, accurate protein structures with lower AF confidence scores have also been reported (Bret et al., 2024; Johansson‐Åkhe & Wallner, 2022; Si & Yan, 2024; Yin & Pierce, 2024). As such, it is necessary to determine a suitable AF confidence score cutoff value to reliably identify accurate NLRLRR–effector complexes.

Assessment of NLRLRR –effector protein complex prediction accuracy

The accuracy of the predicted structures was evaluated by comparing them with available cryogenic‐electron microscopy (cryo‐EM) structures. NLR dimers and NLRs with integrated domains were excluded from these comparisons. The cryo‐EM structures used for this analysis included Sr35‐AvrSr35 (7XVG), RPP1‐ATR1 (7CRB), and Roq1‐XopQ (7JLU), all obtained from RCSB Protein Data Bank (https://www.rcsb.org; Ma et al., 2020; Martin et al., 2020; Zhao et al., 2022). For each of these three NLR–effector complexes, structures were predicted both with and without the use of PDB100 templates, employing paired, unpaired, and unpaired‐paired MSA options to deliberately introduce variability and errors in the prediction process. This approach enabled the comparison of structures with different AF confidence scores to the corresponding cryo‐EM structures. Five structures were generated for each prediction scenario, yielding a total of 90 structures for analysis.

The predicted structures were compared to their corresponding cryo‐EM structures using Dockground CAPRI‐Q to obtain DockQ scores, TM‐scores, Critical Assessment of Predicted Interactions (CAPRI) rankings, and root mean squared difference (RMSD) values for both backbone (i‐RMSDbb) and sidechain (i‐RMSDsc) interface residues (Basu & Wallner, 2016; Collins et al., 2022; Lensink et al., 2017; Lensink & Wodak, 2010; Zhang & Skolnick, 2005). DockQ scores >0.23 and TM‐scores >0.6 were set as cutoff values to identify accurately predicted structures. Additionally, CAPRI criteria for protein–protein complex models were applied to rank the models based on their structural deviation from the cryo‐EM structures (Lensink et al., 2017; Lensink & Wodak, 2010). The CAPRI accuracy categories include high accuracy (fraction of native contacts [fnat] ≥ 0.5, Ligand Root mean square [L‐RMS] ≤ 1, and Interface Root mean square [I‐RMS] ≤ 1), medium accuracy (fnat ≥ 0.5, L‐RMS > 1 and I‐RMS > 1, or 0.3 ≤ fnat ≤ 0.5, L‐RMS ≤ 5 or I‐RMS ≤ 2), acceptable accuracy (fnat ≥ 0.3, L‐RMS > 5 and I‐RMS > 2, or 0.1 ≤ fnat ≤ 0.3, and L‐RMS ≤ 10, or I‐RMS ≤ 4), and incorrect predictions (fnat < 1, or L‐RMS > 10 and I‐RMS > 4). After analyzing the results from DockQ, TM‐scores, and CAPRI rankings, an AF confidence score cutoff was determined to identify NLRLRR–effector complexes with acceptable accuracy in the predicted structures.

The NLRLRR–effector predicted complexes were visualized using ChimeraX v. 1.5. Interacting amino acids (AA) between the NLRLRR and effector were identified using the “contacts” command, considering any AA within 8 Å to be an interaction (UCSF ChimeraX; Meng et al., 2023; Pettersen et al., 2021). Hydrogen bonds were identified using the “hbonds” command with a distance tolerance of 2.5 Å. The residue numbers involved in hydrogen bond formation were extracted and compared between the predicted‐ and corresponding cryo‐EM structures. In cases where the interacting residue numbers differed between the two structures, the distance from the original interacting residue was measured in both the N‐ and C‐terminal directions to assess positional shifts.

Prediction of binding affinities and energies for “true” and “forced” NLR–effector partners

The Area‐Affinity online tool (accessed December 2023–July 2024) was used to predict BA and BE values (also referred to as Gibbs free energy) for all predicted NLRLRR–effector complex structures with an AF confidence score >0.42. Predictions were based on complex structures provided in .pdb format, using both the protein–protein and antibody–protein prediction programs within Area‐Affinity (Yang et al., 2023). These programs utilize a total of 97 ML models, including linear models, various nonlinear models (such as random forest and neural network), constructed nonlinear models, generated nonlinear models, and mixed models, for BA and BE predictions. Results from all 97 models were aggregated to compare BA and BE values between “true” and “forced” NLRLRR–effector complex structures with AF confidence scores >0.42. Area‐Affinity was selected for its accuracy, which is either superior to or comparable with other predictive tools like PRODIGY and Lisa (Raucci et al., 2018; Xue et al., 2016; Yang et al., 2022, 2023). Given the limited availability of experimentally validated data for NLR–effector BEs and BAs, a combination of both protein–protein and antibody–protein models was used to enhance the reliability of the BE and BA predictions.

Training of ensemble learning model

An Ensemble classification model was trained and tested for classifying NLRLRR–effector interactions as being “true” or “forced,” based on the predicted BA and BE values from the combined 194 Area‐Affinity models. The Ensemble model was chosen due to its ability to combine multiple weak decision tree learners into a robust classification algorithm, thereby improving predictive performance (Dong et al., 2020; Hastie et al., 2009; Mienye & Sun, 2022). The model was implemented using the MATLAB Classification Toolbox v.24.1.0.2603908 (https://www.mathworks.com) and trained on the predicted BA and BE values for “true” and “forced” NLRLRR–effector complex structures with AF confidence scores >0.42. To validate the model's accuracy and mitigate overfitting, a 10‐fold cross‐validation scheme was employed. For model testing, 25% of the data were set aside as a hold‐out sample to ensure the model was tested on unseen data after training and validation. Given the limited number of “true” interactions available for classifier training, which could negatively impact model accuracy, the misclassification cost was adjusted. Predicting a “true” interaction as “forced” was assigned a cost of 20, while predicting a “forced” interaction as “true” was assigned a cost of 10. This weighting helped to mitigate the potential imbalance in the dataset and reflect greater importance of accurately identifying “true” interactions.

To optimize hyperparameters during model training, the Optimizable Ensemble Classifier was employed. This method automatically adjusts various hyperparameters throughout the training process to identify the best combination for optimal performance. The hyperparameters explored included Ensemble methods (Bag, GentleBoost, LogitBoost, AdaBoost, and RUSBoost), number of learners (ranging from 10 to 500), learning rate (0.001 to 1), maximum number of splits (1 to 966), and number of predictors to sample (1 to 50). This approach ensured that the Ensemble hyperparameters were fine‐tuned to achieve the most accurate model. Additionally, different feature ranking algorithms (None, Chi2, ReliefF, anova, and Kruskal–Wallis) were tested along with varying the number of features retained (1–194) in each model to further enhance the accuracy of the trained models. In total, 777 Ensemble models were trained and evaluated based on criteria such as overall accuracy, minimum classification accuracy, and accuracy when classifying interactions from mutated NLRLRR data (unseen during training and validation) obtained from Monino‐Lopez et al. (2021). The most accurate model was then exported as the NLRLRR–effector interaction classification application (NEIC) available at https://github.com/koosfick/NEIC1.0. The data transformation sheet (Data_Transformation.xlsm) used for transforming BA and BE data before analyzation and classification using NEIC can be found using the same resource site. This application, developed using MATLAB App Designer and MATLAB Compiler, integrated the trained model with the ReliefF algorithm to classify NLRLRR–effector interactions from new input data.

The experimental procedure, including the workflow for training and testing the ensemble classification model, hyperparameter optimization, and feature selection, is outlined in Figure 8.

Figure 8.

Figure 8

Open in a new tab

A visual representation of the experimental procedure followed in this study.

Protein sequences for NLRLRR and effectors were retrieved from UniProt, with leucine‐rich repeat domains (LRRs) and effector sequences prepared for protein prediction using AlphaFold2‐Multimer. Effector sequences were processed by removing signal domain‐encoding regions prior to structure prediction. The accuracy of the predicted structures and the selection of an acceptable AF confidence score cutoff value were evaluated by comparing predictions with available cryogenic‐electron microscopy (cryo‐EM) structures. After establishing an AF confidence score cutoff value, binding affinity (BA) and binding energy (BE) values were predicted for all reliable NLRLRR–effector complexes using Area‐Affinity. These predicted BA and BE values were then used to train an Ensemble learning model designed to classify NLRLRR–effector interactions as either “true” or “forced.” The trained model provides a robust framework for accurately identifying these interactions based on predicted binding metrics.

Statistical analysis and graphics

All statistical analyses were performed using RStudio v. 1.4.1106 (RStudio Team, 2020), and graphs were produced using the online Plotly web server (https://chart‐studio.plotly.com).

CONFLICT OF INTEREST

The authors declare no conflicts of interest.

Supporting information

Figure S1. AlphaFold‐Multimer predicted structures of Rpi‐chc1.2 mutants interacting with PexRD12‐B and PexRD31‐C effectors. For each protein complex, AlphaFold (AF) confidence scores, experimental interaction data, and NEIC classifications are provided. Experimental data were obtained from Monino‐Lopez et al. (2021), with interactions that activate hypersensitive responses indicated as “True” and those not associated with hypersensitive responses as “N‐I.” AF confidence scores were calculated as 0.2 × pTM + 0.8 × ipTM.

TPJ-122-0-s001.pdf (13.6MB, pdf)

Data S1. Identity information of all plant NLR and pathogen effectors used for predicting NLR–effector complex structures.

TPJ-122-0-s002.xlsx (15.2KB, xlsx)

ACKNOWLEDGMENTS

The authors extend their gratitude to Prof. Warren du Plessis for his invaluable insights and discussions on classification models and their implementation. We also thank Mr. Daniel Opperman for his expertise on the methods used for NLRLRR–effector complex predictions and his constructive feedback on the manuscript.

DATA AVAILABILITY STATEMENT

The data that supports the findings of this study are available in the supplementary material of this article.

REFERENCES

  1. Adachi, H. & Kamoun, S. (2022) NLR receptor networks in plants. Essays in Biochemistry, 66(5), 541–549. Available from: 10.1042/EBC20210075 [DOI] [PubMed] [Google Scholar]
  2. Ammari, M.G. , Gresham, C.R. , McCarthy, F.M. & Nanduri, B. (2016) HPIDB 2.0: a curated database for host‐pathogen interactions. Database, 2016, baw103. Available from: 10.1093/database/baw103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Barragan, A.C. & Weigel, D. (2021) Plant NLR diversity: the known unknowns of pan‐NLRomes. Plant Cell, 33(4), 814–831. Available from: 10.1093/plcell/koaa002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Basu, S. & Wallner, B. (2016) DockQ: a quality measure for protein‐protein docking models. PLoS One, 11(8), e0161879. Available from: 10.1371/journal.pone.0161879 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bernoux, M. , Burdett, H. , Williams, S.J. , Zhang, X. , Chen, C. , Newell, K. et al. (2016) Comparative analysis of the flax immune receptors L6 and L7 suggests an equilibrium‐based switch activation model. The Plant Cell, 28(1), 146–159. Available from: 10.1105/tpc.15.00303 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Brabham, H.J. , Gómez De La Cruz, D. , Were, V. , Shimizu, M. , Saitoh, H. , Hernández‐Pinzón, I. et al. (2024) Barley MLA3 recognizes the host‐specificity effector Pwl2 from Magnaporthe oryzae . The Plant Cell, 36(2), 447–470. Available from: 10.1093/plcell/koad266 [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Bret, H. , Gao, J. , Zea, D.J. , Andreani, J. & Guerois, R. (2024) From interaction networks to interfaces, scanning intrinsically disordered regions using AlphaFold2. Nature Communications, 15(1), 597. Available from: 10.1038/s41467-023-44288-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Champouret, N. , Bouwmeester, K. , Rietman, H. , van der Lee, T. , Maliepaard, C. , Heupink, A. et al. (2009) Phytophthora infestans isolates lacking class I ipiO variants are virulent on Rpi‐blb1 potato. Molecular Plant‐Microbe Interactions, 22(12), 1535–1545. Available from: 10.1094/mpmi-22-12-1535 [DOI] [PubMed] [Google Scholar]
  9. Collins, K.W. , Copeland, M.M. , Kotthoff, I. , Singh, A. , Kundrotas, P.J. & Vakser, I.A. (2022) Dockground resource for protein recognition studies. Protein Science, 31(12), e4481. Available from: 10.1002/pro.4481 [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Contreras, M.P. , Lüdke, D. , Pai, H. , Toghani, A. & Kamoun, S. (2023) NLR receptors in plant immunity: making sense of the alphabet soup. EMBO Reports, 24(10), e57495. Available from: 10.15252/embr.202357495 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Contreras, M.P. , Pai, H. , Selvaraj, M. , Toghani, A. , Lawson, D.M. , Tumtas, Y. et al. (2023) Resurrection of plant disease resistance proteins via helper NLR bioengineering. Science Advances, 9(18), eadg3861. Available from: 10.1126/sciadv.adg3861 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. de Araújo, A.C. , Fonseca, F.C.D.A. , Cotta, M.G. , Alves, G.S.C. & Miller, R.N.G. (2019) Plant NLR receptor proteins and their potential in the development of durable genetic resistance to biotic stresses. Biotechnology Research and Innovation, 3, 80–94. Available from: 10.1016/j.biori.2020.01.002 [DOI] [Google Scholar]
  13. Deng, M. , Mehta, S. , Sun, F. & Chen, T. (2002) Inferring domain‐domain interactions from protein‐protein interactions. In: Proceedings of the Sixth Annual International Conference on Computational Biology. Washington, DC, USA: Association for Computing Machinery, pp. 117–126. [Google Scholar]
  14. Dodds, P.N. , Chen, J. & Outram, M.A. (2024) Pathogen perception and signaling in plant immunity. The Plant Cell, 36(5), koae020. Available from: 10.1093/plcell/koae020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Dodds, P.N. , Lawrence, G.J. & Ellis, J.G. (2001) Six amino acid changes confined to the leucine‐rich repeat β‐strand/β‐turn motif determine the difference between the P and P2 rust resistance specificities in flax. The Plant Cell, 13(1), 163–178. Available from: 10.1105/tpc.13.1.163 [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Dong, X. , Yu, Z. , Cao, W. , Shi, Y. & Ma, Q. (2020) A survey on ensemble learning. Frontiers of Computer Science, 14(2), 241–258. Available from: 10.1007/s11704-019-8208-z [DOI] [Google Scholar]
  17. Erijman, A. , Rosenthal, E. & Shifman, J.M. (2014) How structure defines affinity in protein‐protein interactions. PLoS One, 9(10), e110085. Available from: 10.1371/journal.pone.0110085 [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Evans, R. , O'Neill, M. , Pritzel, A. , Antropova, N. , Senior, A.W. , Green, T. et al. (2022) Protein complex prediction with AlphaFold‐multimer. bioRxiv. 2021.2010.2004.463034. Available from: 10.1101/2021.10.04.463034 [DOI]
  19. Gómez‐Pérez, D. & Kemen, E. (2021) Predicting lifestyle from positive selection data and genome properties in oomycetes. Pathogens, 10(7), 807. Available from: 10.3390/pathogens10070807 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Hastie, T. , Tibshirani, R. & Friedman, J. (2009) Ensemble learning. In: Hastie, T. , Tibshirani, R. & Friedman, J. (Eds.) The elements of statistical learning: data mining, inference, and prediction. New York, NY: Springer New York, pp. 605–624. [Google Scholar]
  21. Homma, F. , Lyu, J. & van der Hoorn, R.A.L. (2024) Using AlphaFold multimer to discover interkingdom protein‐protein interactions. The Plant Journal, 120(1), 19–28. Available from: 10.1111/tpj.16969 [DOI] [PubMed] [Google Scholar]
  22. Johansson‐Åkhe, I. & Wallner, B. (2022) Improving peptide‐protein docking with AlphaFold‐multimer using forced sampling. Frontiers in Bioinformatics, 2, 959160. Available from: 10.3389/fbinf.2022.959160 [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Jones, J.D.G. & Dangl, J.L. (2006) The plant immune system. Nature, 444(7117), 323–329. Available from: 10.1038/nature05286 [DOI] [PubMed] [Google Scholar]
  24. Jones, J.D.G. , Staskawicz, B.J. & Dangl, J.L. (2024) The plant immune system: from discovery to deployment. Cell, 187(9), 2095–2116. Available from: 10.1016/j.cell.2024.03.045 [DOI] [PubMed] [Google Scholar]
  25. Karan, B. , Mahapatra, S. , Sahu, S.S. , Pandey, D.M. & Chakravarty, S. (2023) Computational models for prediction of protein‐protein interaction in rice and Magnaporthe grisea . Frontiers in Plant Science, 13, 1046209. Available from: 10.3389/fpls.2022.1046209 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Kastritis, P.L. & Bonvin, A.M.J.J. (2013) On the binding affinity of macromolecular interactions: daring to ask why proteins interact. Journal of the Royal Society Interface, 10(79), 20120835. Available from: 10.1098/rsif.2012.0835 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Kemen, E. , Gardiner, A. , Schultz‐Larsen, T. , Kemen, A.C. , Balmuth, A.L. , Robert‐Seilaniantz, A. et al. (2011) Gene gain and loss during evolution of obligate parasitism in the white rust pathogen of Arabidopsis thaliana . PLoS Biology, 9(7), e1001094. Available from: 10.1371/journal.pbio.1001094 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Klebe, G. & Böhm, H.J. (1997) Energetic and entropic factors determining binding affinity in protein‐ligand complexes. Journal of Receptors and Signal Transduction, 17(1–3), 459–473. Available from: 10.3109/10799899709036621 [DOI] [PubMed] [Google Scholar]
  29. Kourelis, J. & van der Hoorn, R.A.L. (2018) Defended to the nines: 25 years of resistance gene cloning identifies nine mechanisms for R protein function. The Plant Cell, 30(2), 285–299. Available from: 10.1105/tpc.17.00579 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Kourelis, J. , Sakai, T. , Adachi, H. & Kamoun, S. (2021) RefPlantNLR is a comprehensive collection of experimentally validated plant disease resistance proteins from the NLR family. PLoS Biology, 19(10), e3001124. Available from: 10.1371/journal.pbio.3001124 [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Lei, C. , Zhou, K. , Zheng, J. , Zhao, M. , Huang, Y. , He, H. et al. (2024) AraPathogen2.0: an improved prediction of plant‐pathogen protein‐protein interactions empowered by the natural language processing technique. Journal of Proteome Research, 23(1), 494–499. Available from: 10.1021/acs.jproteome.3c00364 [DOI] [PubMed] [Google Scholar]
  32. Lensink, M.F. & Wodak, S.J. (2010) Docking and scoring protein interactions: CAPRI 2009. Proteins: Structure, Function, and Bioinformatics, 78(15), 3073–3084. Available from: 10.1002/prot.22818 [DOI] [PubMed] [Google Scholar]
  33. Lensink, M.F. , Velankar, S. & Wodak, S.J. (2017) Modeling protein‐protein and protein‐peptide complexes: CAPRI 6th edition. Proteins: Structure, Function, and Bioinformatics, 85(3), 359–377. Available from: 10.1002/prot.25215 [DOI] [PubMed] [Google Scholar]
  34. Li, H. & Zhang, Z. (2016) Systems understanding of plant‐pathogen interactions through genome‐wide protein‐protein interaction networks. Frontiers of Agricultural Science and Engineering, 3, 102–112. Available from: 10.15302/J-FASE-2016100 [DOI] [Google Scholar]
  35. Li, H. , Hu, R. , Fan, Z. , Chen, Q. , Jiang, Y. , Huang, W. et al. (2022) Dual RNA sequencing reveals the genome‐wide expression profiles during the compatible and incompatible interactions between Solanum tuberosum and Phytophthora infestans . Frontiers in Plant Science, 13, 817199. Available from: 10.3389/fpls.2022.817199 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Li, X. , Ma, L. , Wang, Y. , Ye, C. , Guo, C. , Li, Y. et al. (2023) PlantNLRatlas: a comprehensive dataset of full‐ and partial‐length NLR resistance genes across 100 chromosome‐level plant genomes. Frontiers in Plant Science, 14, 1178069. Available from: 10.3389/fpls.2023.1178069 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Lighten, J. , Papadopulos, A.S.T. , Mohammed, R.S. , Ward, B.J. , Paterson, I. , Baillie, L. et al. (2017) Evolutionary genetics of immunological supertypes reveals two faces of the red queen. Nature Communications, 8(1), 1294. Available from: 10.1038/s41467-017-01183-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Liu, G. , Fang, Y. , Liu, X. , Jiang, J. , Ding, G. , Wang, Y. et al. (2024) Genome‐wide association study and haplotype analysis reveal novel candidate genes for resistance to powdery mildew in soybean. Frontiers in Plant Science, 15, 1369650. Available from: 10.3389/fpls.2024.1369650 [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Locci, F. & Parker, J.E. (2024) Plant NLR immunity activation and execution: a biochemical perspective. Open biology, 14(1), 230387. Available from: 10.1098/rsob.230387 [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Ma, S. , Lapin, D. , Liu, L. , Sun, Y. , Song, W. , Zhang, X. et al. (2020) Direct pathogen‐induced assembly of an NLR immune receptor complex to form a holoenzyme. Science, 370(6521), eabe3069. Available from: 10.1126/science.abe3069 [DOI] [PubMed] [Google Scholar]
  41. Maekawa, T. , Cheng, W. , Spiridon, L.N. , Töller, A. , Lukasik, E. , Saijo, Y. et al. (2011) Coiled‐coil domain‐dependent homodimerization of intracellular barley immune receptors defines a minimal functional module for triggering cell death. Cell Host & Microbe, 9(3), 187–199. Available from: 10.1016/j.chom.2011.02.008 [DOI] [PubMed] [Google Scholar]
  42. Maidment, J.H.R. , Shimizu, M. , Bentham, A.R. , Vera, S. , Franceschetti, M. , Longya, A. et al. (2023) Effector target‐guided engineering of an integrated domain expands the disease resistance profile of a rice NLR immune receptor. eLife, 12(12), e81123. Available from: 10.7554/eLife.81123 [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Manandhar, S. , Sankhe, R. , Priya, K. , Hari, G. , Kumar B, H. , Mehta, C.H. et al. (2022) Molecular dynamics and structure‐based virtual screening and identification of natural compounds as Wnt signaling modulators: possible therapeutics for Alzheimer's disease. Molecular Diversity, 26(5), 2793–2811. Available from: 10.1007/s11030-022-10395-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Martin, J. (2024) AlphaFold2 predicts whether proteins interact amidst confounding structural compatibility. Journal of Chemical Information and Modeling, 64(5), 1473–1480. Available from: 10.1021/acs.jcim.3c01805 [DOI] [PubMed] [Google Scholar]
  45. Martin, R. , Qi, T. , Zhang, H. , Liu, F. , King, M. , Toth, C. et al. (2020) Structure of the activated ROQ1 resistosome directly recognizing the pathogen effector XopQ. Science, 370(6521), eabd9993. Available from: 10.1126/science.abd9993 [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Matthews, L.R. , Vaglio, P. , Reboul, J. , Ge, H. , Davis, B.P. , Garrels, J. et al. (2001) Identification of potential interaction networks using sequence‐based searches for conserved protein‐protein interactions or “interologs”. Genome Research, 11(12), 2120–2126. Available from: 10.1101/gr.205301 [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. McBride, J.M. , Eckmann, J.P. & Tlusty, T. (2022) General theory of specific binding: insights from a genetic‐mechano‐chemical protein model. Molecular Biology and Evolution, 39(11), msac217. Available from: 10.1093/molbev/msac217 [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Meng, E.C. , Goddard, T.D. , Pettersen, E.F. , Couch, G.S. , Pearson, Z.J. , Morris, J.H. et al. (2023) UCSF ChimeraX: tools for structure building and analysis. Protein Science 32(11), e4792. Available from: 10.1002/pro.4792 [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Mienye, I.D. & Sun, Y. (2022) A survey of ensemble learning: concepts, algorithms, applications, and prospects. IEEE Access, 10, 99129–99149. Available from: 10.1109/ACCESS.2022.3207287 [DOI] [Google Scholar]
  50. Monino‐Lopez, D. , Nijenhuis, M. , Kodde, L. , Kamoun, S. , Salehian, H. , Schentsnyi, K. et al. (2021) Allelic variants of the NLR protein Rpi‐chc1 differentially recognize members of the Phytophthora infestans PexRD12/31 effector superfamily through the leucine‐rich repeat domain. The Plant Journal, 107(1), 182–197. Available from: 10.1111/tpj.15284 [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Naidoo, S. , Visser, E.A. , Zwart, L. , Toit, Y.D. , Bhadauria, V. & Shuey, L.S. (2018) Dual RNA‐sequencing to elucidate the plant‐pathogen duel. Current Issues in Molecular Biology, 27, 127–142. Available from: 10.21775/cimb.027.127 [DOI] [PubMed] [Google Scholar]
  52. O'Reilly, F.J. , Graziadei, A. , Forbrig, C. , Bremenkamp, R. , Charles, K. , Lenz, S. et al. (2023) Protein complexes in cells by AI‐assisted structural proteomics. Molecular Systems Biology, 19(4), e11544. Available from: 10.15252/msb.202311544 [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Oh, S.K. , Kim, H. & Choi, D. (2014) Rpi‐blb2‐mediated late blight resistance in Nicotiana benthamiana requires SGT1 and salicylic acid‐mediated signaling but not RAR1 or HSP90. FEBS Letters, 588(7), 1109–1115. Available from: 10.1016/j.febslet.2014.02.028 [DOI] [PubMed] [Google Scholar]
  54. Ortiz, D. , Chen, J. , Outram, M.A. , Saur, I.M.L. , Upadhyaya, N.M. , Mago, R. et al. (2022) The stem rust effector protein AvrSr50 escapes Sr50 recognition by a substitution in a single surface‐exposed residue. New Phytologist, 234(2), 592–606. Available from: 10.1111/nph.18011 [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Pettersen, E.F. , Goddard, T.D. , Huang, C.C. , Meng, E.C. , Couch, G.S. , Croll, T.I. et al. (2021) UCSF ChimeraX: structure visualization for researchers, educators, and developers. Protein Science, 30(1), 70–82. Available from: 10.1002/pro.3943 [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Prigozhin, D.M. & Krasileva, K.V. (2021) Analysis of intraspecies diversity reveals a subset of highly variable plant immune receptors and predicts their binding sites. The Plant Cell, 33(4), 998–1015. Available from: 10.1093/plcell/koab013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Rabajante, J.F. , Tubay, J.M. , Ito, H. , Uehara, T. , Kakishima, S. , Morita, S. et al. (2016) Host‐parasite red queen dynamics with phase‐locked rare genotypes. Science Advances, 2(3), e1501548. Available from: 10.1126/sciadv.1501548 [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Råberg, L. (2023) Human and pathogen genotype‐by‐genotype interactions in the light of coevolution theory. PLoS Genetics, 19(4), e1010685. Available from: 10.1371/journal.pgen.1010685 [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Ramírez, D. & Caballero, J. (2018) Is it reliable to take the molecular docking top scoring position as the best solution without considering available structural data? Molecules, 23(5), 1038. Available from: 10.3390/molecules23051038 [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Raucci, R. , Laine, E. & Carbone, A. (2018) Local interaction signal analysis predicts protein‐protein binding affinity. Structure, 26(6), 905–915 904. Available from: 10.1016/j.str.2018.04.006 [DOI] [PubMed] [Google Scholar]
  61. Redkar, A. , Cevik, V. , Bailey, K. , Zhao, H. , Kim, D.S. , Zou, Z. et al. (2023) The Arabidopsis WRR4A and WRR4B paralogous NLR proteins both confer recognition of multiple Albugo candida effectors. New Phytologist, 237(2), 532–547. Available from: 10.1111/nph.18378 [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. RStudio Team . (2020) RStudio: integrated development for R. Boston, USA: RStudio, PBC. Available from: http://www.rstudio.com/ [Accessed April 2024]. [Google Scholar]
  63. Santos, M.D.L. , Resende, M.L.V.D. , Alves, G.S.C. , Huguet‐Tapia, J.C. , Resende, M.F.R.D.J. & Brawner, J.T. (2022) Genome‐wide identification, characterization, and comparative analysis of NLR resistance genes in Coffea spp. Frontiers in Plant Science, 13, 868581. Available from: 10.3389/fpls.2022.868581 [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Schmid, E.W. & Walter, J.C. (2025) Predictomes, a classifier‐curated database of AlphaFold‐modeled protein‐protein interactions. Molecular Cell, 85(6), 1216–1232. Available from: 10.1016/j.molcel.2025.01.034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Segretin, M.E. , Pais, M. , Franceschetti, M. , Chaparro‐Garcia, A. , Bos, J.I.B. , Banfield, M.J. et al. (2014) Single amino acid mutations in the potato immune receptor R3a expand response to phytophthora effectors. Molecular Plant‐Microbe Interactions, 27(7), 624–637. Available from: 10.1094/MPMI-02-14-0040-R [DOI] [PubMed] [Google Scholar]
  66. Sharma, V. (2022) A study on data scaling methods for machine learning. Journal for Global Academic and Scientific Research, 1(1), 31–42. Available from: 10.55938/ijgasr.v1i1.4 [DOI] [Google Scholar]
  67. Si, Y. & Yan, C. (2024) Protein language model‐embedded geometric graphs power inter‐protein contact prediction. eLife, 12, RP92184. Available from: 10.7554/eLife.92184 [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Singh, P.K. , Ray, S. , Thakur, S. , Rathour, R. , Sharma, V. & Sharma, T.R. (2018) Co‐evolutionary interactions between host resistance and pathogen avirulence genes in rice‐Magnaporthe oryzae pathosystem. Fungal Genetics and Biology, 115, 9–19. Available from: 10.1016/j.fgb.2018.04.005 [DOI] [PubMed] [Google Scholar]
  69. Tamborski, J. & Krasileva, K.V. (2020) Evolution of plant NLRs: from natural history to precise modifications. Annual Review of Plant Biology, 71, 355–378. Available from: 10.1146/annurev-arplant-081519-035901 [DOI] [PubMed] [Google Scholar]
  70. Tamborski, J. , Seong, K. , Liu, F. , Staskawicz, B.J. & Krasileva, K.V. (2023) Altering specificity and autoactivity of plant immune receptors Sr33 and Sr50 via a rational engineering approach. Molecular Plant‐Microbe Interactions, 36(7), 434–446. Available from: 10.1094/MPMI-07-22-0154-R [DOI] [PMC free article] [PubMed] [Google Scholar]
  71. Tameling, W.I.L. , Elzinga, S.D.J. , Darmin, P.S. , Vossen, J.H. , Takken, F.L.W. , Haring, M.A. et al. (2002) The tomato R gene products I‐2 and Mi‐1 are functional ATP binding proteins with ATPase activity. The Plant Cell, 14(11), 2929–2939. Available from: 10.1105/tpc.005793 [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Van Valen, L. (1973) A new evolutionary law. Evolutionary Theory, 1, 1–30. [Google Scholar]
  73. Vangone, A. & Bonvin, A.M. (2015) Contacts‐based prediction of binding affinity in protein‐protein complexes. eLife, 4, e07454. Available from: 10.7554/eLife.07454 [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Williams, S.J. , Sornaraj, P. , deCourcy‐Ireland, E. , Menz, R.I. , Kobe, B. , Ellis, J.G. et al. (2011) An autoactive mutant of the M flax rust resistance protein has a preference for binding ATP, whereas wild‐type M protein binds ADP. Molecular Plant‐Microbe Interactions, 24(8), 897–906. Available from: 10.1094/MPMI-03-11-0052 [DOI] [PubMed] [Google Scholar]
  75. Wu, C.H. , Abd‐El‐Haliem, A. , Bozkurt, T.O. , Belhaj, K. , Terauchi, R. , Vossen, J.H. et al. (2017) NLR network mediates immunity to diverse plant pathogens. Proceedings of the National Academy of Sciences of the United States of America, 114(30), 8113–8118. Available from: 10.1073/pnas.1702041114 [DOI] [PMC free article] [PubMed] [Google Scholar]
  76. Xue, L.C. , Rodrigues, J.P. , Kastritis, P.L. , Bonvin, A.M. & Vangone, A. (2016) PRODIGY: a web server for predicting the binding affinity of protein‐protein complexes. Bioinformatics, 32(23), 3676–3678. Available from: 10.1093/bioinformatics/btw514 [DOI] [PubMed] [Google Scholar]
  77. Yang, S. , Li, H. , He, H. , Zhou, Y. & Zhang, Z. (2019) Critical assessment and performance improvement of plant‐pathogen protein‐protein interaction prediction methods. Briefings in Bioinformatics, 20(1), 274–287. Available from: 10.1093/bib/bbx123 [DOI] [PubMed] [Google Scholar]
  78. Yang, Y.X. , Huang, J.Y. , Wang, P. & Zhu, B.T. (2023) Area‐affinity: a web server for machine learning‐based prediction of protein‐protein and antibody‐protein antigen binding affinities. Journal of Chemical Information and Modeling, 63(11), 3230–3237. Available from: 10.1021/acs.jcim.2c01499 [DOI] [PMC free article] [PubMed] [Google Scholar]
  79. Yang, Y.X. , Wang, P. & Zhu, B.T. (2022) Importance of interface and surface areas in protein‐protein binding affinity prediction: a machine learning analysis based on linear regression and artificial neural network. Biophysical Chemistry, 283, 106762. Available from: 10.1016/j.bpc.2022.106762 [DOI] [PubMed] [Google Scholar]
  80. Yi, C.H. , Taylor, M.L. , Ziebarth, J. & Wang, Y. (2024) Predictive models and impact of interfacial contacts and amino acids on protein‐protein binding affinity. ACS Omega, 9(3), 3454–3468. Available from: 10.1021/acsomega.3c06996 [DOI] [PMC free article] [PubMed] [Google Scholar]
  81. Yin, R. & Pierce, B.G. (2024) Evaluation of AlphaFold antibody‐antigen modeling with implications for improving predictive accuracy. Protein Science, 33(1), e4865. Available from: 10.1002/pro.4865 [DOI] [PMC free article] [PubMed] [Google Scholar]
  82. Yin, R. , Ribeiro‐Filho, H.V. , Lin, V. , Gowthaman, R. , Cheung, M. & Pierce, B.G. (2023) TCRmodel2: high‐resolution modeling of T cell receptor recognition using deep learning. Nucleic Acids Research, 51(W1), W569–W576. Available from: 10.1093/nar/gkad356 [DOI] [PMC free article] [PubMed] [Google Scholar]
  83. Yu, H. , Wu, X. , Liang, J. , Han, Z. , Xiao, Y. , Du, H. et al. (2023) Genome‐wide identification of nucleotide‐binding domain leucine‐rich repeat (NLR) genes and their association with green peach aphid (Myzus persicae) resistance in peach. BMC Plant Biology, 23(1), 513. Available from: 10.1186/s12870-023-04474-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  84. Zhang, Q.C. , Petrey, D. , Deng, L. , Qiang, L. , Shi, Y. , Thu, C.A. et al. (2012) Structure‐based prediction of protein‐protein interactions on a genome‐wide scale. Nature, 490(7421), 556–560. Available from: 10.1038/nature11503 [DOI] [PMC free article] [PubMed] [Google Scholar]
  85. Zhang, S. , Li, C. , Si, J. , Han, Z. & Chen, D. (2022) Action mechanisms of effectors in plant‐pathogen interaction. International Journal of Molecular Sciences, 23(12), 6758. Available from: 10.3390/ijms23126758 [DOI] [PMC free article] [PubMed] [Google Scholar]
  86. Zhang, Y. & Skolnick, J. (2005) TM‐align: a protein structure alignment algorithm based on the TM‐score. Nucleic Acids Research, 33(7), 2302–2309. Available from: 10.1093/nar/gki524 [DOI] [PMC free article] [PubMed] [Google Scholar]
  87. Zhao, Y.B. , Liu, M.X. , Chen, T.T. , Ma, X. , Li, Z.K. , Zheng, Z. et al. (2022) Pathogen effector AvrSr35 triggers Sr35 resistosome assembly via a direct recognition mechanism. Science Advances, 8(36), eabq5108. Available from: 10.1126/sciadv.abq5108 [DOI] [PMC free article] [PubMed] [Google Scholar]
  88. Zheng, C. , Liu, Y. , Sun, F. , Zhao, L. & Zhang, L. (2021) Predicting protein‐protein interactions between rice and blast fungus using structure‐based approaches. Frontiers in Plant Science, 12, 690124. Available from: 10.3389/fpls.2021.690124 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Figure S1. AlphaFold‐Multimer predicted structures of Rpi‐chc1.2 mutants interacting with PexRD12‐B and PexRD31‐C effectors. For each protein complex, AlphaFold (AF) confidence scores, experimental interaction data, and NEIC classifications are provided. Experimental data were obtained from Monino‐Lopez et al. (2021), with interactions that activate hypersensitive responses indicated as “True” and those not associated with hypersensitive responses as “N‐I.” AF confidence scores were calculated as 0.2 × pTM + 0.8 × ipTM.

TPJ-122-0-s001.pdf (13.6MB, pdf)

Data S1. Identity information of all plant NLR and pathogen effectors used for predicting NLR–effector complex structures.

TPJ-122-0-s002.xlsx (15.2KB, xlsx)

Data Availability Statement

The data that supports the findings of this study are available in the supplementary material of this article.

Articles from The Plant Journal are provided here courtesy of Wiley

RESOURCES