The NRC network shows limited gene expansion in Asterales compared to Solanales

The presence of the NRC network across Asterids prompted us to investigate the extent of its gene copy number in Asterales, a Campanulid order, compared to the well-studied Solanales order from Lamiids (Fig 1A) [30]. For this purpose, we extracted a total number of 21,232 non-redundant NLR sequences from a dataset of de-novo annotated 40 Solanales and 29 Asterales genomes with harmonized annotations using NLRtracker, representing a diverse set of species and genera from the two orders (S1 and S2 Data) [9,73]. Based on phylogenetics analysis, we then extracted the monophyletic NRC network comprising 7,020 sequences, including the reference NRC-H and NRC-S from RefPlantNLR (S1 Fig and S3 Data) [9]. In Asterales and Solanales, there are on average 300 and 338.6 non-redundant NLRs per genome, respectively. However, NRC-type NLRs on average only comprise 6.6% of the NLRome in Asterales genomes compared to 50.3% in Solanales (S2 Fig).

Compared to Solanales, Asterales show much less expansion of NRC-S and NRC-H genes (Figs 1B and S2). To further understand the evolutionary patterns of diversification in Asterales we performed phylogenetic analysis on the 592 NRC-H and NRC-S NLRs, 88 and 504, respectively. We divided the 88 Asterales NRC-H into two phylogroups. One group, containing 27 sequences phylogenetically grouped with NRC0, while the remaining 61 sequences formed a separate phylogroup and grouped with known Solanales NRC helpers, such as NRC2 and NRC4, referred to as Ast-NRCs hereafter (Fig 1B) [30]. All 504 Asterales NRC-S formed a well-supported phylogroup next to the Solanales SD-type sensors, but none were grouped with the Rx-type Solanales sensors (Fig 1B). Although Asterales NRC-S belong to the SD-type class, they lack the Solanaceous domain characteristic at their N-termini (Fig 1B and S3 Data) [37]. These observations suggest notable differences in the evolutionary pace and expansion of the NRC network between Solanales and Asterales orders.

In common lettuce, the 17 NRC sequences (comprising 2 helpers and 15 sensors) represent only 5% of the total NLRome, which consists of 359 sequences. In contrast, TIR-NLRs dominate the NLRome with 206 sequences, making up over 57%. The CC-NLR, CC_R-NLR, and CC_G10-NLR classes contain 97, 9, and 42 sequences, respectively (S3 Fig). This indicates a disproportionate expansion of different NLR classes, with a limited expansion of NRC-H and NRC-S in lettuce.

Asterales NRC networks exhibit lineage-specific expansion of NRC-H and NRC-S

Given the observed differences in the expansion of NRC network between Solanales and Asterales, we sought to investigate the Asterales NRC network gene copy number and genomic structure in more detail, focusing on lettuce. We analyzed 40 NRC-H and NRC-S sequences from common lettuce (Lactuca sativa) and two wild lettuce species (Lactuca saligna and Lactuca virosa) (S4 Data). Through phylogenetic analysis, we divided the Lactuca NRC sequences into four phylogroups: one helper phylogroup containing sequences that grouped with either NRC0 or tomato NRC2 (referred to as Ast-NRC1 in each species hereafter), and sensor phylogroups 1, 2, and 3 (referred to as sensor group 1,2, and 3 hereafters.) (Fig 2A). These phylogenetic groups were defined by well-supported branches (Figs 2A and S5). Based on the previous categorization of NLR families in lettuce, NRC sequences are distributed across four RGC groups: (1) the NRC helper group as RGC7, (2) sensor group 1 as RGC26, (3) sensor group 2 as RGC27, and (4) sensor group 3 as RGC9 (S3 Fig) [65,66].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Unlike LsatNRC0, Ast-LsatNRC1 is genetically linked to NRC-S in lettuce.

(A) Phylogenetic tree of NB-ARC domains from 40 NRC sequences of three Lactuca species (Lactuca sativa, Lactuca saligna, and Lactuca virosa) constructed using IQ-TREE 2 [76]. Lactuca NRC gene family is divided into an NRC-H and three well-supported NRC-S phylogenetic orthogropus. Numbers on tree nodes indicate bootstrap support values. AtZAR1 and NbZAR1 were used as outgroup sequences. SlNRC0 and SlNRC2 were used as reference sequences. At: Arabidopsis thaliana; Nb: Nicotiana benthamiana; Sl: Solanum lycopersicum; Ast: Asterales. (B) Physical map of Lactuca sativa NRC sequences on the chromosomes. While Ast-LsatNRC1 is genetically linked to group 3 NRC-S, LsatNRC0 is located on Chromosome 3, not linked to any sensor NLRs. Previously described Major Resistance Clusters (MRCs) are shown for NRCs [66,67]. Lsat: Lactuca sativa.

https://doi.org/10.1371/journal.pgen.1012245.g002

NRC-H and NRC-S are often found in genetic clusters like NRC0 and the Solanaceae NRC6 [36,44]. Therefore, we examined the physical location of NRC sequences in the lettuce genomes. Unlike previous observations in various Asterid species, lettuce NRC0 (LsatNRC0; Helixer_NC_056625.2_001598.1), on chromosome 3, was not located near any sensor sequences (Fig 2B) [44]. In contrast, Ast-LsatNRC1 (Helixer_NC_056630.2_002654.1) was found to be in a physical cluster (<100kb) with sensors from groups 2 and 3, all residing on chromosome 8. Notably, all of group 3 sensors are clustered together on this chromosome. This genomic cluster is part of the previously reported MRC8c (Fig 2B) [64–66]. The remaining lettuce NRC-S sequences, including those from group 1, are scattered across chromosomes 4, 8, and 9, as part of MRCs 4, 8a, and 9a, respectively (Fig 2B) [64–66]. Only LsatNRC0 is not part of any of the reported MRCs (Fig 2B).

Due to a lack of chromosome-level assemblies for the wild lettuce species, we could not determine if the NRC-H and NRC-S in these genomes are genetically clustered together. However, our genome analysis revealed Ast-LvirNRC1 (Helixer_CAKMRJ010003334.1_001958.1) positioned near group 2 and 3 sensors on contig2395 (Figs 2A and S4). We further examined five other species across Asterales to investigate whether NRC-S are in genetic clusters with the NRC helpers as a general trend (S5 Fig). We classified the NRC sequences in these species based on their phylogenetic relationship to the Lactuca NRC-S and NRC-H. In Cichorium intybus, CiNRC0 (Helixer_CM042016.1_001204.1) is located within 500K base pair distance of a group 1 NRC sensor on chromosome 8 (Helixer_CM042016.1_001201.1), whereas Ast-CiNRC1a (Helixer_CM042010.1_003195.1) and Ast-CiNRC1b (Helixer_CM042010.1_003196.1) are clustered with sensors from group 2 and 3 on chromosome 2 (S6 Fig and S6 Data). In Cynara cardunculus, CcNRC0 (Helixer_NC_037542.1_000084.1) is in close proximity with a group 1 sensor (Helixer_NC_037542.1_001617.1) on chromosome 15, and Ast-CcNRC1 is in a genetic cluster with four group 1 sensors on contig NW_020200580.1 (S7 Fig and S6 Data). In Helianthus annuus, Chrysanthemum lavandulifolium and Codonopsis lanceolata, none of the NRC-H are clustered with any NRC-S (S6 and S7 Figs and S6 Data). These findings suggest variations in the physical clustering of NRC-H and NRC-S within the Lactuca genus and between different Asterales species.

LsatNRC0 and Ast-LsatNRC1 are required for cell death signaling of distinct NRC-S phylogenetic groups

To investigate the functional connections of the lettuce NRC network and to test the phylogeny-based classification of these NLRs into NRC-S and NRC-H, we used Agrobacterium-mediated transient co-expression assays (Agroinfiltration) in an nrc2/3/4 CRISPR knockout line of Nicotiana benthamiana [30]. We agroinfiltrated autoactive mutants of the putative NRC-S and NRC-H to test their functionality, as the cognate effectors for these sensors remain unidentified. These mutants were generated by introducing a D-to-V mutation within the MHD motif of the NLRs, located near the C-terminus of the NB-ARC domain [15]. We first tested whether transiently expressed NRC-S or NRC-H alone could induce cell death in N. benthamiana. While none of the NRC-S triggered cell death on their own, the autoactive mutants of the two lettuce NRC-H displayed a macroscopic hypersensitive cell death response (HR) (S8A Fig).

We then co-expressed the NRC-S and NRC-H in pairs to assess their potential functional connections and the dependence of lettuce NRC-S on the two NRC helpers. Among the 16 tested NRC-S, 9 displayed visible HR as autoactive mutants when co-expressed with one or both wild-type NRCs. While the cell death phenotype of functional sensors was consistent across two replicates, protein accumulation was not assessed for sensor–helper pairs that did not trigger cell death. Three NRC-S induced cell death through only LsatNRC0, while four responded solely through Ast-LsatNRC1. Notably, two sensors were functionally connected with both LsatNRC0 and Ast-LsatNRC1 (Figs 3 and S8B and S7 Data). Specifically, all LsatNRC0-dependent sensors belonged to group 1, while Ast-LsatNRC1-dependent sensors resided in group 3. NRC-S that functioned with both helpers are group 2 members (Fig 3A). Furthermore, we assessed if lettuce NRC-H and NRC-S possessed the MADA motif required for cell-death activity of previously characterized NRC-H. While both NRC helpers possessed MADA motifs, none of the 15 tested sensors got a hit for MADA HMM profile (Fig 3 and S8 Data) [15]. This was consistent with the cell death assays showing NRC helpers but not NRC sensors are able to cause cell death on their own [15,77].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. LsatNRC0 and Ast-LsatNRC1 mediate the cell death signaling of distinct NRC-S phylogenetic groups.

(A) Co-agroinfiltration cell death assays of NRC-H and NRC-S. 14 NRC-S expressed as autoactive mutants (D-to-V mutation in the MHD motif) with two wild-type NRC-H. Sensor-helper pairs agroinfiltrated together causing cell death are outlined with red squares. While group 1 and group 3 sensors can only signal through LsatNRC0 and Ast-LsatNRC1, respectively, group 2 sensors can signal through both. Sensor-helper pairs with no visible cell-death were not assessed in terms of protein accumulation. Lsat: Lactuca sativa; Ast: Asterales. (B) Lettuce NRC network model. The 9 validated NRC-S can cause cell death through LsatNRC0 and Ast-LsatNRC1 in a partially redundant manner.

https://doi.org/10.1371/journal.pgen.1012245.g003

Although we showed functional connection between NRC-H and NRC-S in common lettuce, this remains to be experimentally determined in other species including those from Lactuca virosa, Lactuca saligna, and other Asterales species. However, based on phylogenetic relationships and absence of both NRC0 and group 1 sensors, we assumed the NRC network in Lactuca species follow a similar specificity profile following phylogenetic grouping. Taken together these observations indicate the presence of a functional NLR network with partially redundant signaling pathways in lettuce (Fig 3B).

In contrast to NRC helpers, NRC sensors fail to form resistosome-like oligomers when modeled with AlphaFold 3

Recently, we demonstrated that AlphaFold 3 [78] can differentiate between helper and sensor NLRs within pairs by predicting resistosome-like structures for helpers but not for sensors [23]. Building on these findings, we applied a similar approach to the lettuce NRC network to investigate potential differences in the in-silico oligomerization of NRC-H and NRC-S. Using AlphaFold 3, we modeled five replicates of the two NRC-H and 16 NRC-S sequences from lettuce as hexamers in the presence of 50 oleic acid molecules, simulating the plasma membrane environment. To evaluate these models, we extracted and analyzed key metrics, including pLDDT (per-residue measure of local confidence), pTM (predicted template modeling score for overall structure), ipTM (interface-specific pTM), and per-chain pTM (S8 Data). AlphaFold 3 successfully modeled resistosome-like oligomers for both NRC-H sequences, though it was unable to generate a complete N-terminal funnel in any of the LsatNRC0 predicted structures (Fig 4). This was consistent with our previous observations when modeling NRC0 oligomers [19]. In contrast, none of the NRC-S from the three sensor groups formed resistosome-like structures (Fig 4). Across all replicates, NRC-S models consistently showed low confidence with high predicted aligned errors (Figs 4 and S9). Structurally, NRC-S displayed lower metrics compared to NRC helpers, with pTM and ipTM values below 0.5 and per-chain pTM values below 0.7 (Figs 4 and S9). Conversely, helpers had consistently higher metrics, with pTM and ipTM exceeding 0.5 in all replicates and an average per-chain pTM of 0.75 (Figs 4 and S9). These results suggest that AlphaFold 3 can be used to distinguish between NRC-S and NRC-H and their capacity to form higher-order resistosome-like oligomers.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. In contrast to NRC-H, NRC-S failed to form resistosome-like structures when modeled with AlphaFold 3.

Structural modeling with AlphaFold 3 was used to predict the ability of the tested lettuce NRC-H and NRC-S to form resistosome-like hexameric oligomers. Modeling included 50 oleic acid molecules to approximate the plasma membrane. Key structural metrics, including pLDDT (per-residue measure of local confidence), pTM (predicted template modeling score for overall structure), ipTM (interface-specific pTM), and per-chain pTM, are plotted for each protein (n = 5). Example predicted structures from the fourth replicate are displayed in top and side views, colored with the pLDDT color scale indicating local confidence levels. Lsat: Lactuca sativa; Ast: Asterales.

https://doi.org/10.1371/journal.pgen.1012245.g004

NRC-S dependent on Ast-LsatNRC1 are under stronger positive selection compared to LsatNRC0-dependent sensors

To examine the selection pressures that affected mutations in the NRC helper and NRC-S phylogenetic groups, we conducted positive selection tests on Lactuca NRC sequences using CodeML in PAML v4.10.7 software [79]. These tests compare evolutionary models that allow a proportion of sites to evolve under positive selection (ω = dN/dS > 1) against null models that do not. A significantly better fit of the positive selection model indicates that diversifying selection has driven amino acid changes at specific positions. Individual sites under positive selection are then identified using Bayes Empirical Bayes (BEB) analysis [80]. For each NRC group, we compared the M8 model allowing positive selection against its null counterpart, M8a, using likelihood ratio tests (see Methods; S10 Data).

For NRC-H, the M8a vs M8 comparison did not reach our 1% significance threshold (p = 0.037), and BEB analysis did not identify any sites at either posterior probability threshold. Similarly, although the M8a vs M8 comparison was significant for sensor group 2 (p = 0.003), BEB analysis did not identify any sites at either posterior probability threshold (S10 Data). A significant LRT unaccompanied by BEB-supported sites indicates that the improved fit of M8 is not underpinned by identifiable positively selected codons and is most parsimoniously interpreted as a false positive of the M8a vs M8 test, which is known to be sensitive to alignment noise and gene conversion. In contrast, both sensor groups 1 and 3 showed evidence of positive selection (S10 Data). BEB analysis under the M8 model identified five sites under positive selection in group 1 (posterior probability > 0.95), three of which localized to the CC domain when mapped onto an AlphaFold 3-predicted structure (Helixer_NC_056630.2_000504.1; Fig 5B). Sensor group 3 showed the strongest signatures, with 48 sites under positive selection (posterior probability > 0.95), 33 of which exceeded 0.99. The majority of these sites mapped to the LRR domain (Helixer_NC_056630.2_001736.1), localized to its concave face, a region implicated in effector binding [51,70] (Fig 5C).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Ast-LsatNRC1-dependent sensors are under more positive selection compared to NRC0-dependent sensors and NRC helpers.

(A) The y-axis represents the posterior probability that each site belongs to the positively selected site class (ω > 1), as estimated by the Bayes Empirical Bayes (BEB) method under the M8 model (beta + ω) from the CodeML package in PAML [79]. Sites with posterior probability (P) > 0.95 and > 0.99 are indicated by circles and triangles, respectively. No sites under positive selection were detected in NRC-H or group 2 sensors. In contrast, group 1 and 3 NRC-S showed several sites under positive selection, with most in group 3 localized to the LRR domain. (B) Sites under positive selection (red) mapped onto predicted structures of representative NRC-S from groups 1 and 3, generated using AlphaFold 3. (C) Zoomed-in view of the LRR domain from a representative group 3 NRC-S (Helixer_NC_056630.2_001736.1), highlighting sites under positive selection (red). Most sites under positive selection are located at the concave side of the LRR domain. CC: Coiled-Coil; NBD: Nucleotide-binding domain; HD1: Helix domain 1; WHD: Winged-helix domain; LRR: Leucine-rich repeat.

https://doi.org/10.1371/journal.pgen.1012245.g005

To formally test whether NRC subclades differed in average selective pressure, we used PAML’s branch model 2 (model = 2, NSsites = 0), assigning each phylogroup (helpers, sensor groups 1, 2, and 3) its own ω parameter and comparing it to the one-ratio model (M0) by likelihood ratio test. The four-ratio model fit significantly better than M0 (2·ΔlnL = 305.66, df = 4, P = 6.5 × 10⁻⁶⁵), with ω = 0.12 for the helpers, 0.35 and 0.39 for sensor groups 1 and 2, and 0.58 for sensor group 3 (S10 Fig and S10 Data). The helper clade therefore shows the lowest average ω, group 3 the highest, and groups 1 and 2 intermediate values, consistent with the M8 BEB results placing positive selection most prominently in group 3. Although the four subclades differ in sequence number and tree depth, their total tree dS lengths from the M0 fit are broadly comparable (1.51–2.87; S10 Data), with group 3’s larger size (24 sequences) offset by its shallower per-branch divergence (≈0.06 substitutions/branch versus 0.17–0.25 in the other subclades). To visualise branch-level heterogeneity, the free-ratio model (M1) is shown as S11 Fig; per-branch ω estimates are presented descriptively rather than as a formal test, and branches with ω ≥ 10 are flagged in the figure caption as artefacts of low synonymous substitution rates rather than meaningful estimates.

Next, we estimated the pairwise nonsynonymous (d_N) and synonymous (d_S) substitution rates and d_N/d_S ratios (ω) for members of each NRC-H and NRC-S group using the Nei and Gojobori (1986) method (S12 Fig and S11 Data) [81,82]. Analyses were conducted on full-length sequences and individual domains (CC, NB-ARC, and LRR), where ω > 1 implied presence of positive selection. Among full-length and domain-based comparisons, NRC-H had lower average ω compared to all NRC-S groups (S13 Fig). While sensor group 3 on average had overall higher ω than the rest of NRC-H and NRC-S, the CC domain in group 1 sensors had ω > 1 values, indicative of positive selection (S13 Fig). Overall, the pairwise d_N/d_S ratios of NRC-H and NRC-S phylogroups and selection analysis pointed to NRC-S dependent on Ast-LsatNRC1 being under stronger positive selection compared to NRC-S dependent on LsatNRC0.

Extraction of NRC network sequences

A total number of 29,261 NLRs were compiled from the NLRtracker output of 39 Solanales and 29 Asterales de-novo annotated high-quality genome assemblies from NCBI (https://www.ncbi.nlm.nih.gov/datasets/) and a S. melongena chromosome-level genome assembly from GWH genome warehouse datasets [87] (S1 Data) [73,88]. Based on the simple domain annotation from NLRtracker we only kept NLRs with “CNL”, “CNLO”, “CN”, “OCNL”, “CONL”, “NL”, “NLO”, “ONL”, “BCNL”, “BNL”, “BCN”, “BCCNL”, “BNLO”, “BOCNL”, “RNL”, “TN”, “TNL”, “TNLO”, “TNLJ” domain architectures. This led to keeping 25,042 NLRs. The remaining NLRs were then deduplicated using an R script (remove redundancy based on 100% sequence similarity) to keep 24,518 non-redundant sequences. Finally, based on NLRtracker domain output, any entry with NB-ARC domain shorter than 250 amino acids and longer than 400 amino acids were removed. This led to a final 21,232 sequences that was used for the phylogenetics analysis.

To construct a phylogenetic tree of the extracted NLRs, the NB-ARC domain sequences were aligned with RefPlantNLR NB-ARCs using FAMSA v2.2.2 [default options] [89]. The alignment was then used to construct a phylogenetic tree using FastTree v2.1.11 [-lg] [9,75]. The NRC network along with CcRPP1 and its orthologs, including 7,040 sequences, was extracted based on the presence of reference sequences from a well-supported branch containing NRC-H and NRC-S. The NB-ARC domain of the extracted sequences were aligned using MAFFT v7.525 [90]. The alignment was used to construct a new phylogenetic tree in the same way as before. The resulting tree was rooted at the CcRPP1 phylogroup.

Phylogenetics and gene distance analyses

Plant taxonomy trees and taxa divergence times were obtained from TimeTree 5 database (timetree.org) [74]. 122 NRC sequences including helpers and sensors from Lactuca sativa, Lactuca saligna, Lactuca virosa, Cichorium intybus, Helianthus annuus, Chrysanthemum lavandulifolium, Cynara cardunculus, and Codonopsis lanceolata present in the dataset were extracted based on phylogeny. The NB-ARC sequences of extracted proteins were aligned with AtZAR1, NbZAR1, NbNRC2, SlNRC0, as references, using MAFFT v7.525 [--anysymbol] [90]. FastTree v2.1.11 [-lg] was used to generate the tree of NRC network in the mentioned species [75]. We then used IQtree v2.3.0 [-B 1000 -m MFP] to generate a phylogenetic tree of 40 sequences in the NRC network in the Lactuca genus [76,91]. The resulting tree was divided to helpers, sensor group 1, sensor group 2, and sensor group 3 based on branch lengths and reference sequences.

To obtain the physical location of the genes on the chromosome, first the GFF output from Helixer was simplified using GffRead [default options] and then imported into R using rtracklayer package [92,93]. The physical distances between genes was calculated using a custom R script and gene coordinates were exported to be visualized on chromosomes using MG2C online tool [94]. The output was then visualized manually.

Linking previous lettuce NLRome classifications to the new annotation

Lettuce NLRome IDs from Christopoulou, et al., 2015, and Lettuce V8 genome assembly from Phytozome 13 were extracted [64,95,96]. We then used BLASTP as part of BLAST+ v2.16.0 to identify corresponding gene identifiers in the Phytozome 13 annotation to the new Helixer annotation [97]. The BLAST output was then processed in R and the gene coordinates, major resistance loci information, and RGC groups were retrieved for the new Helixer annotation. The corresponding genes between the Phytozome 13 and Helixer annotations were identified based on BLAST hits with >98% sequence identity and bit scores above 1500 (S12 Data).

The same approach was used to retrieve corresponding sequences from the RefSeq reference annotation of Lsat_Salinas_v11 (GCF_002870075.4) [98] (S12 Data).

Agroinfiltration and cell death assays

The NRC sequences used in the cell death assays were originally retrieved from the Phytozome 13 annotation. These sequences were cross-checked against the Helixer annotation, with two sequences absent from the latter. All tested sequences are listed in S8 Fig.

To test the functionality of lettuce NRCs, annotated genes were cloned using the Golden Gate Modular Cloning (MoClo) kit [99] and the MoClo Plant Parts kit [100]. In brief, coding sequences were synthesized by GENEWIZ as Golden Gate Level 0 pICH41155 modules and subsequently transferred to the binary vector pICH86988 through BsaI digestion (Weber et al., 2011). Native BsaI sites in the original sequences were domesticated as needed. In addition to wild-type clones, a D-to-V mutation within the MHD motif was introduced to generate autoactive versions [15] of the lettuce NRCs. Verified clones were transformed into A. tumefaciens strain GV3101 pMP90 for ectopic expression in N. benthamiana.

For cell death assays, wild-type and nrc2/3/4 N. benthamiana plants were grown in a controlled environment growth chamber at 22–25°C with a 16-hour light period and 45%/65% relative humidity during the light/dark cycle. Light intensity was approximately 200 µmol/m²/s. Four- to five-week-old plants were used for agroinfiltration, following the methods described by Bos et al. (2006) [101]. The final OD₆₀₀ of all Agrobacterium suspensions was adjusted to 0.5 in infiltration buffer (10 mM MES, 10 mM MgCl₂, and 150 µM acetosyringone, pH 5.6). This concentration was used for both expressing clones individually and co-expressing NRC-H and NRC-S of interest. The cell death phenotype was imaged 4–5 days post-agroinfiltration.

Sequence polymorphism and selection analyses

Nucleotide sequence of the NRC sequences were obtained using a custom python script from a compiled CDS file of the annotated genomes. Codon-based nucleotide alignments were generated for each NRC-H and NRC-S phylogroup generated using MACSE v2.07 [-prog alignSequences -seq] [102]. Positive selection tests were conducted on resulting alignments using CodeML in PAML v4.10.7 [79]. For each NRC phylogroup (helpers, sensors C1, C2, and C3), two site models were evaluated: M8a (beta + ω = 1) and M8 (beta + ω). Likelihood ratio tests (LRT; 2ΔL = 2(lnL_alternative – lnL_null)) were performed comparing M8a vs M8 (df = 1, χ² critical value = 6.63 at 1% significance). Sites under positive selection were identified using the Bayes Empirical Bayes (BEB) method [80] under the M8 model, with posterior probabilities > 0.95 (and the more stringent > 0.99) used to indicate strong support.

To compare average selective pressure across NRC subclades, we ran PAML branch model 2 in CodeML (PAML v4.10.7; model = 2, NSsites = 0, CodonFreq = 2) [79]. assigning each of the four subclades (helper, sensor groups 1, 2, and 3) its own ω parameter via $1–$4 labels placed at each subclade’s most recent common ancestor in the rooted tree. This four-ratio model was tested against the one-ratio model (M0; model = 0) by likelihood ratio test (df = 4). Per-subclade tree dS lengths were obtained by summing the M0 branch-level dS estimates over each subclade’s stem and internal branches. The free-ratio model (M1; model = 1) was also run for exploratory visualisation of branch-level heterogeneity (S10 Fig) and is not used for hypothesis testing; in this model, branches with ω ≥ 10 reflect low synonymous substitution rates (dS ≈ 0) and are flagged as ambiguous rather than indicative of positive selection.

Pairwise nonsynonymous (d_N) and synonymous (d_S) substitution rates were estimated for full-length sequences and individual domains (CC, NB-ARC, and LRR) in each phylogroup using the Nei and Gojobori (1986) method [81,82]. These rates were extracted from the “2NG.dN” and “2NG.dS” outputs of CodeML using a custom R script. All figures were generated with the ggplot2 package in R [103].

Protein structure prediction and mapping Shannon entropy to protein structures

AlphaFold 3 [Seed = 1, Ligand = ADP] was used to predict the monomeric structures of representative sequences from each phylogroup (S9 Data) [78]. For oligomeric structures, AlphaFold 3 with [Ligand = OLA x 50] was used to model sequences (S9 Data). First model for each sequence was modeled with Seed = 1. Four other replicates were modeled for each sequence with random seeds. Model confidence and metadata were processed and plotted with R. All scripts available at [https://github.com/amiralito/lettuce_salad].