The amino acid and coding sequences of Tsw (A0A1C9TCM9_CAPCH) published by Kim et al. were used to search the C. chinense ‘PI159236’ scaffolds and assembled genome (Chinense.v.1.2 from the Pepper Genome Platform) for the full-length sequence of Tsw. The amino acid sequence of Tsw was also used to search the at the time most recent version of the genome of the susceptible C. annuum cv. CM334 version 1.55 for the full-length non-functional variant of tsw (Kim et al., 2017).

Tsw homologs sharing more than 30% identity and 50% coverage were retrieved from the NLRscape database (NLRscape, 2022 https://nlrscape.biochim.ro) and analyzed with NLRscape webserver tools. These were further curated by eliminating CC/NBS incomplete and motif missing sequences, and by retrieving their secondary structure propensities and LRR repeat profiles precomputed with LRRpredictor (Martin et al., 2020a) and NLRexpress (Martin et al., 2022) from NLRscape DB. Variability profiles were raised over this cleaned Tsw family set by computing Kullback–Leibler divergence of the alignment with respect to Blosum62 background distribution using Logomaker API (Tareen and Kinney, 2020) and secondary structure consensuses were generated for graphics display.

A set of ~39,469 nonredundant LRR domains (90% identity cutoff) originating from plant NLR sequences were retrieved from NLRscape (NLRscape, 2022, https://nlrscape.biochim.ro) and further processed for analysis. All sequences containing annotated LRR domains at 90% redundancy were collected using NLRscape provided online selection tools. Sequence data, domain annotations, and individual LRR repeat predictions computed with LRRpredictor (Martin et al., 2020a) were retrieved from NLRscape public datasets (NLRscape, 2022). Statistical analyses of the data were performed using in-house scripts. Figures were generated using Plotly (Plotly Technologies Inc, 2015), Matplotlib (Hunter, 2007) and ETE3 (Huerta-Cepas et al., 2016) libraries.

Due to some problems that will be clarified below, homology models were built herein by two distinct methods: (1) by using the recently developed AI-based AlphaFold (Jumper et al., 2021) and (2) by using a system adapted modeling workflow that has previously been used successfully (Slootweg et al., 2013; Kozuki et al., 2017; Zhang et al., 2019). As RAM requirements grow exponentially with the sequence length, AlphaFold failed to generate a model for the long functional C. chinense Tsw form comprising of 2,116 aa. In addition, all models of the shorter susceptible (tsw) variant of Tsw in C. annuum (1,437 aa) proposed by AlphaFold were in the open ATP active state and none in the inactive ADP resting state, which seems unrealistic. Furthermore, we found that some of the delineation of LRR repeats as well as the local 3D structure in the LRR region proposed by AlphaFold is problematic. This prompted us to develop alternate models allowing to explore a larger configurational space consistent with the local sequence propensity as well as the overall architecture of the LRR domain.

The alternate models of the Tsw-CC, -NBS and -LRR domains were built using Modeller v10.1 (Webb and Sali, 2016) primarily based on ZAR1 ADP-bound cryo-EM conformation - PDB: 6j5w (Wang et al., 2019b). The CC and NBS domains were modelled based exclusively on the ZAR1 template as it shows the highest similarity to the Tsw sequence with: CC ~15% and NBS ~30% sequence identity. By contrast the LRR domain was modelled based on the repeat delineation provided by LRRpredictor (Martin et al., 2020a) using the Joint Fragment Remote Homology Modeling (JFRHM) approach (Slootweg et al., 2013) starting from: ZAR1-LRR, FLS2-LRR (PDB: 4mn8) (Sun et al., 2013) and SCHENGEN3-LRR (PDB: 6s6q) (Okuda et al., 2020). LRRpredictor indicates that the long functional form of TSW-LRR domain has 57 LRR repeats organized as follows: a starting core region of 15 repeats followed by 7 consecutive quasi-identical blocks consisting each of 6 LRR repeats. The first 10 repeats of the core region were modelled using ZAR1 as template by preserving its curvature and all its contacts with CC-NBS. The rest of the LRR domain was modeled using the other two templates depending on the local best match. To smooth out the transition between these two regions with distinct curvatures, a ‘buffer’ region consisting of repeats 10-15 was relaxed by gradually modifying the local radius using a template mixture protocol allowing the passage from the tighter ZAR1 region, r1-r10, to the broader r15-r57 rest. The C-terminal acidic tail of LRR was not covered in the model due to the lack of 3D templates and the low confidence of the secondary structure consensus within this area. The model of the shorter susceptible variant tsw was obtained using the optimized long Tsw model, as the two sequences are highly homologous (>83.6% identity). These raw models were subjected to an optimization workflow consisting of interactive rounds of energy minimization, simulating annealing and short-term Molecular Dynamics (MD) trials followed by modelling refinements. The compliance of geometric features to known 3D structures was probed using Molprobity (Williams et al., 2018). To further optimize its stability, the model was further subjected to an equilibration stage performed in explicit TIP3P water and a 0.5 nM NaCl neutral charged environment. The system was initially minimized and gradually heated from 0 to 300K and further subjected to 5 equilibration stages of gradually relaxing restraint profiles: from stronger backbone constraints in the first stage, to no constraints in the last stage. After the equilibration, the system was subjected to a long-run MD stability trial. Simulations were performed using NAMD 3.0a8 (Phillips et al., 2005; Phillips et al., 2020) and CHARMM 16m (Huang et al., 2017) forcefield using periodic boundary conditions (PBC) and Particle Mesh Edwards (PME) electrostatics with a 12Å cutoff, 10Å switch and 14Å pairlist distances for nonbonded interactions. Constant temperature (300K) and pressure (1 atm) control was attained by using the Langevin integrator and Langevin piston. Trajectory analyses were performed using VMD (Humphrey et al., 1996) and the 3D molecular graphics displayed in the figures was generated using PyMOL (The PyMOL Molecular Graphics System, Version 2.2.3 Schrödinger, LLC).

To allow for the comparison of LRR domain lengths in plants, a set of ~39,469 LRRs from plant NLR sequences edged at 90% identity redundancy (NLRscape, 2022) was analyzed. Over three quarters of the retrieved LRRs span between 200 – 600 aa and are mainly comprised of less than 20 LRR repeats. The LRR length distribution peak is found at around 400 aa, corresponding to ~14-16 LRR repeats on average ( Figure 1A ). Both domain length and repeat number follow a unimodal distribution centered at 15 repeats and ~380 aa, with solid tails on both sides. The distribution is strongly right-sided skewed, with an elongated tail corresponding to large LRR domains. This roughly adds up to around 10% of LRR domains that are unusually long, with lengths of 1,000 to 3,400 aa that are organized in 30 or more repeats ( Figures 1A, B ). Large LRR domains are found mainly in CNL and NL receptors which display multiple examples of extremely large LRRs, of over 2,000 aa. By contrast, the RNL and TNL groups have far more regular LRRs (< 500 aa), with only few examples of more than 1,000 aa – all found in the TNL class. RNLs are quite rare, and hence far less represented in plants.

From a taxonomic point of view the vast majority of large LRRs originates from Pentapetalae, in particular from within the rosid clade: the fabids, with the bulk stemming from the Fabales order, and the malvids; Sapindales, Malvales and Myrtales orders. Interestingly, the Brasicales order in the malvids clade does not share this trend. Care should be taken in the interpretation of these results on the unequal statistical support given the uneven taxonomic distribution of NLRs in the database (shown in Figure 1C , last column).

Such large dimensions of LRR domains are puzzling as they are far larger than typical protein domains, which rarely exceed 500 aa due to stability problems (Xu and Nussinov, 1998). For instance, at 40% identity redundancy, 99.8% of the domains in the CATH database (Sillitoe et al., 2021) are shorter than 600 aa, with only 5 examples larger than 1,000 aa out of a total of 31,878, which is less than 0.02%. ( Figure 1B , CATH column). These large dimensions are favored by the loosely and modular solenoidal LRR architecture, which, in contrast to the more globular architectures, offer flexibility and elasticity to the overall fold (Kobe and Kajava, 2001). In addition, large continuous LRR horseshoes are in some cases weakened or disrupted by relatively long insertions incompatible with the LRR fold, which potentially break the LRR domain into two or more sub-regions.

The manner in which such LRR subregions would fold with respect to each other is another still open-ended question. Shorter linkers between LRR fragments generally tend to maintain the overall horseshoe shape of the domain with minor disruptions. By contrast, longer linkers that allow for rotations and translation of the downstream LRR elements might generate a wide range of overall potential conformations – from a perfect fusion of consecutive LRR fragments, with large insertion ‘isles’ as seen in some X-ray crystal structures such as the extracellular BRI1 (Hohmann et al., 2018) or PSKR1 (Wang et al., 2015b), to even a potential side-to-side LRR-LRR fragment clustering – similar to those seen in some protein-protein interactions. In assessing the LRR domain continuity, a special indicator of the local LRR stability is the embedment of the L-x-x-L-x-L pattern into a larger pattern L-x-x-(L-x-x-L-x-L)-x-x-N/C (Kajava, 1998) that provides further stability and structural reinforcement.

Large LRR domains are also difficult to explain from a functional point of view. Experimental data, from both solved 3D structures and in vivo experiments, indicate that in a large and diverse set of NLR examples interactions of the LRR with the rest of NLR domains (CC, TIR, NBS) primarily take place within the first four to eight LRR repeats (Qi et al., 2012; Slootweg et al., 2013). In ZAR1 cryo-EM structures, the LRR-CC interaction occurs within the first 4 LRR repeats, while the LRR-NB-ARC interaction is consolidated mainly within the first 8 LRR repeats – in tight contact with the ARC2 subdomain of NB-ARC (Wang et al., 2019a; Wang et al., 2019b). Sparse contacts were also shown to shape up between the further five LRR repeats and the beta-turn-beta loop structure of the ARC2 domain. All these interdomain contacts are critical for a proper functioning of these receptors, as shown by swap experiments between sections of the LRR domain of different R genes. Compelling experimental data demonstrate the importance of the compatibility between the first LRR repeats and the rest of the NLR (Qi et al., 2012; Slootweg et al., 2013; Slootweg et al., 2017; Slootweg et al., 2018). For instance, LRR swaps between potato homologs Rx1 and Gpa2, which despite their 88% sequence identity recognize unrelated pathogens, showed that successful phenotypic LRR exchanges were only obtained when preserving the original CC-NBS-first 3 LRR repeats, while the fusion of the foreign rest of the LRR domain converts the pathogen recognition (Slootweg et al., 2013; Slootweg et al., 2017; Slootweg et al., 2018).

The strong preference for 10-16 LRR repeats observed within the dataset of LRR domains ( Figure 1A ) is consistent with the existing experimental and structural data, yet it leaves a very large portion of the LRR structure unexplained in cases of extremely large LRR domains.

Starting from the Tsw sequence from C. chinense, a set of distant homologs at above 30% identity with Tsw were retrieved using the NLRscape tools (NLRscape, 2022). The set was trimmed by eliminating the highly redundant sequences by imposing a threshold of 90% identity between the identified homologs obtaining a set of 200 sequences, which we further refer to as the extended homologs group. Based on the identity matrix of the alignment, a dense group emerges, corresponding to a set of close homologs of the Capsicum Tsw which consists of ~32 nonredundant sequences, which share between 55-90% identity with Tsw on the full-length protein and above 70% on the NBS domain span. Outside this group, the remaining homologs in the extended set share less than 40% identity with Tsw on the overall sequence and below 50% on the NBS domain span, generating a neat separation.

The close homologs are all part of the Solanaceae family, specifically of the Capsicum, Solanum and Nicotiana genera, while the more distant homologs show a broader taxonomic spread primarily across various taxa orders of the malvids and rosids clades ( Supplementary File 1A ). As expected, the close homologs of Tsw displays multiple highly conserved regions, some of which are also highly conserved in the extended homologs set ( Supplementary File 1B, C ).

As specified under methods, AlphaFold automatic modelling fails for the long 2,116 aa functional version of Tsw. Additionally, the AlphaFold model does not account for the ADP resting state of the shorter susceptible 1,438 aa variant ( Figure 2 ), nor does it unambiguously match with the motif delineation proposed by LRRpredictor. This prompted us to build additional system adapted models to overcome these limitations. Such adapted models furthermore allowed us to explore an opposed limiting scenario, which in conjunction with the AlphaFold model, defines the range in which we could expect such extensive LRR domain structures to stand. While the LRR curvature proposed by AlphaFold is wide – ~60 Å in diameter with a ~25° pitch of the supercoiled solenoid – the adapted models explore a wider range of diameter (40-70 Å), pitch angles (35-95°) of the supercoil bending and optimized repeat delineation. In a conformation with a wide radius, tight pitch and only 51 repeats, such as the one proposed by the AlphaFold model, the curvature translates into ~1.5 supercoiled turns. This is compared in Figure 3 with alternate Tsw-LRR models of 57 repeats in two extreme scenarios of relaxed vs. tight radius and pitch. These suggest that pitches narrower than ~25-35° potentially imply intra-domain interactions between consecutive turns of the LRR domain. It is therefore more probable to find the real structure standing and flexing in between these two extreme limits.

The Tsw-CC domain, aa 1-129, only displays remote homology to solved CC structures. Of these, ZAR1-CC is the closest with 13-17% identity depending on the alignment method. Secondary structure predictions indicate the presence of four helical segments in +/-1 turn agreement to ZAR1. Moreover, all four helices have an amphipathic pattern consistent with a 4-helical bundle fold with the hydrophobic regions hidden inside the bundle. As in ZAR1, the Tsw-CC N-terminal region of the first helical stretch, α1, is overall hydrophobic, which might protrude the membrane in the active state similarly to that in ZAR1 ( Figure 4 ). The next two helical segments α2 and α3 have an even higher similarity to ZAR1. Tsw also contains a partial EDVID motif (aa 81- 85: “ADVAV”) known to be involved in CC-LRR interaction. However, this is not a shared feature, as the third helix (α3) shows a significant variability even in the 32 close homologs of 50-90% identity ( Figure 4 ). The absence of a conservation pressure within this region might indicate that the CC-LRR interaction pattern of Tsw family might be distinct from ZAR1.

Interestingly, the Tsw family displays highly conserved aromatic residues in positions equivalent to both the α1-α2 and the α3-α4 interconnecting loops in ZAR1. These two loops are close in distance in the 4-helical bundle structure, forming a contiguous surface exposed to the solvent. The location and degree of conservation of this patch might indicate a potential protein-protein interaction hotspot.

Lastly, α4 has the lowest homology with known templates, including ZAR1. In ZAR1 the predicted α4 is very long and was shown by cryo-EM to be unstable, as it folded differently with respect to α1, α2, and α3 in the resting and active state. While in the resting ADP state, only the first part of α4, named here α4a, closes the 4-helical bundle letting the second α4b part disordered. In the active ATP state this disordered region of α4b freezes, forming the four helical bundle and letting α4a disordered. By contrast, in Tsw the predicted α4 is significantly shorter and continues with a highly variable segment with linker-like propensities such as in the Rx1 template.

The Tsw-NBS domain (aa 158-495) shares 30% identity to ZAR1 and it preserves all the functionally significant regions: the P-loop, Walker-B, RNBS motifs A through D, the GLPL motif, and the MHD motifs ( Figure 5 ). By contrast, the N-terminal entry into the NBS diverges from the common VG motif (Wróblewski et al., 2018), which in Tsw, both in close and extended homologs groups, is of type “eeFdSR” (where lower case show the consensus amino acids with lower conservation). Not all, but only some of the intra- and inter-molecular contacts seen in ZAR1 cryo-EMs, are conserved in Tsw. As seen for other R proteins, for instance in Sw-5a and -b (De Oliveira et al., 2016), point mutations between the functional and susceptible variants generate changes in salt bridge patterns at the NBS-LRR interface such as Tsw: D420-K540 changed to N417-K537 in tsw; or Tsw: C444-E660 changed to R441-E657 in tsw.

The LRR domain of the functional C. chinense Tsw variant spans over ~1,550 residues (aa 511-2,067). LRRpredictor delineates 57 LRR repeats, with several lower probability motifs discussed in detail below. Briefly, starting from the N-terminal side, the LRR domain is structured in a core section of 15 LRR repeats (aa 511-901), followed by seven consecutive quasi-identical blocks each consisting of six LRR repeats named A-F. The last repeat, 7-F, is incomplete and followed by a highly acidic tail with high propensity for an intrinsically disordered structure (aa 2,068-2,116) ( Figure 6 ).

In Tsw, the NB-ARC-LRR linker corresponding to the region between the end of the ARC2 subdomain (aa 495) and the first proper LRR motif (aa 512 - YSHMSI) is nearly double in size to its equivalent region in ZAR1: 17 aa vs. 9 aa. In the cryo-EM structure of ZAR1, this 9 aa linker folds as an improper LRR repeat that is loosely attached to the rest of the LRR domain. In Tsw, its 17 aa equivalent has notable similarities with proper LRR repeats, both in length and secondary structure propensity. It was therefore modeled as such and referred herein as the LRR 0 repeat ( Figure 6 ).

In Tsw, the first seven LRR repeats, with which interactions with CC and NBS are expected to occur, show significant similarities with ZAR1. LRR motif predictions become problematic in the region aa 740-780. The motif probability (aa 761 - SEVVHS) is not very high here, thus this could be modeled either as a single long 44 aa repeat, or as two shorter repeats. Such a motif has a very low LRRpredictor probability. However, the propensity of this stretch for extended secondary structure and forming beta sheets is high. Hence, this was locked as a proper motif in both AlphaFold and adapted-JFRHM models.

The AlphaFold oddly models the orientation of LRR 9 in reverse, i.e. with its dorsal side oriented towards the inside of the horseshoe. None of the structurally solved LRR domains contains such an odd repeat orientation. Moreover, this orientation is inconsistent with overall NLR architecture in the resting ADP state, as such a reverse repeat would clash with the NBS domain. Therefore, most likely the local solution adopted by AlphaFold for LRR 9 is an artefact derived from fusing two templates and was corrected in the adapted-JFRHM model.

In each of the following seven blocks, LRRpredictor identifies six repeats termed herein A-F. This in contrast to AlphaFold, which merges repeats E and F in a single 53 aa strand. In our A-F delineation, all motifs on the ventral side of the horseshoe are of one of the following extended types: L-x-x-(L-x-L-x-x-L)-x-x-L-x-x/L-L-L-x/L, suggesting a structural reinforcement that was taken into consideration in the adapted models. Interestingly, all the A-F repeats display in addition a stretch of two to three consecutive hydrophobic residues on the dorsal side of the horseshoe in the 14–18 aa region of each repeat. This could potentially form a continuous hydrophobic ‘band’ over the entire structure.

The susceptible variant tsw from C. annuum lacks blocks 3–6, while the core section of block 1 and 7 follows the same LRR organization as the functional variant with only minor sequence variations (94.0% identity at nucleotide level). Apart from the missing LRR blocks, the last 23 aa at the C-terminal end – within the second half of the acidic tail – of tsw has a completely different, highly positively charged, aa sequence due to a frameshift resulting from a 1 nucleotide deletion. This makes a striking distinction between the functional vs susceptible variant, i.e. the C-terminal tail of Tsw being entirely acidic, while the tail of tsw is half acidic, half basic. Such differences could be responsible for the loss of its functionality as a resistance gene.

Interestingly, when inspecting close homologs of Tsw within the Solanaceae clade, the lengths of the LRR domains varies significantly, with most homologs not exceeding two or three repetitive blocks ( Figure 6 ). Common to all, the core region of the LRR domain shows a higher degree of conservation within the N-terminal region (LRR 1-8), which is consistent with the region’s expected involvement in NBS interaction.

Furthermore, the central C repeat of each repetitive block is highly negatively charged. This creates a noticeable repetitive pattern in 3D, and in addition makes each module slightly negatively charged. Overall, this adds up to an extreme -32 total net charge on the LRR domain span. When also considering the net charges of the CC (+6), CC-NBS linker (-7), NBS (-7), and the acidic C-terminal tail (-17), the total net charge of the overall protein becomes -57, which has to be compensated by the environment, or the folding of the protein. Tsw’s total net charge is quite extreme in comparison to other R proteins, for which it generally ranges around -10 to +10. When compensated for by length, the total net charge remains below average.

While both the AlphaFold and the JFRHM adapted models display the LRR domain as a continuous solenoid supercoil, it cannot be excluded that the continuity of the LRR domain might be disrupted or altered at the less reinforced regions of the LRR pattern due to the molecular environment or on the large biological time scale.

AlphaFold only models Tsw starting from the ZAR1 ATP active state. Moreover, the five best AlphaFold models display the CC domain in a completely different configuration and opposite orientation to that seen in ZAR1-ATP active state ( Figure 2 ). This local modeling solution adopted by AlphaFold results in a complete disruption of the interface between CC-NBS and CC-LRR as that seen in ZAR-ATP state ( Figure 2 ), suggesting that the amino acids at the predicted interface are not compatible with a CC-NBS and CC-LRR contact state.

To further investigate this incompatibility, we analyzed the corresponding CC-NBS and CC-LRR interface on the JFRHM models generated starting from the inactive ADP conformation of ZAR1. As can be seen in Figure 7 , indeed the amino acid properties at the CC-LRR interface do not match perfectly with those seen in the ZAR1 template, suggesting that the interaction between these two domains might be different in Tsw. In ZAR1, the CC-LRR interface is shared between the inactive and active conformations and has a strong electrostatic nature ( Figures 7A, B ) with the CC acidic EDVID motif area in α3 being complemented by a proportionally basic patch on the lower side of the first four LRR repeats. On the other hand, in Tsw the interface is neutral, suggesting a far weaker affinity.

By contrast, the CC-NBS interface shares a very similar electrostatic profile with that seen in ZAR1 in its resting ADP state. The NBS part of the interface consists of a negatively charged patch N_1,- (ARC1 end, ARC2 beginning) in contact with the positively charged region at the end of α3 and the outset of the α4 helix of the CC domain ( Figure 7C ).

Focusing now on the groove of formed across the NBS domain accommodating the CC-NBS linker in the JFRHM model of Tsw built starting from the ZAR1 inactive ADP template, the compatibility between the linker sequence and the NBS groove is even higher in Tsw than in ZAR1 due to a striking opposite charge matching ( Figure 7D ). In Tsw, the NBS groove is highly positively charged and the linker displays an increased net negative charge of -5 (aa 140-156), whereas the interface in ZAR1 has a more homogenous neutral charge. Together, these findings suggest that Tsw shares only partially interdomain interface profiles with ZAR1, implying that the interaction between domains might be different from the ones seen in ZAR1 during activation.