RPW8/HR Repeats Control NLR Activation in A. thaliana

In many plant species, conflicts between divergent elements of the immune system, especially nucleotide-binding oligomerization domain-like receptors (NLR), can lead to hybrid necrosis. Here, we report deleterious allele-specific interactions between an NLR and a non-NLR gene cluster, resulting in not one, but multiple hybrid necrosis cases in Arabidopsis thaliana. The NLR cluster is RESISTANCE TO PERONOSPORA PARASITICA 7 (RPP7), which can confer strain-specific resistance to oomycetes. The non-NLR cluster is RESISTANCE TO POWDERY MILDEW 8 (RPW8) / HOMOLOG OF RPW8 (HR), which can confer broad-spectrum resistance to both fungi and oomycetes. RPW8/HR proteins contain at the N-terminus a potential transmembrane domain, followed by a specific coiled-coil (CC) domain that is similar to a domain found in pore-forming toxins MLKL and HET-S from mammals and fungi. C-terminal to the CC domain is a variable number of 21- or 14-amino acid repeats, reminiscent of regulatory 21-amino acid repeats in fungal HET-S. The number of repeats in different RPW8/HR proteins along with the sequence of a short C-terminal tail predicts their ability to activate immunity in combination with specific RPP7 partners. Whether a larger or smaller number of repeats is more dangerous depends on the specific RPW8/HR autoimmune risk variant. Author Summary In many plant species, conflicts between divergent elements of the immune system can cause hybrids to express autoimmunity, a generally deleterious syndrome known as hybrid necrosis. We are investigating multiple hybrid necrosis cases in Arabidopsis thaliana that are caused by allele-specific interactions between different variants at two unlinked resistance (R) gene clusters, RESISTANCE TO PERONOSPORA PARASITICA 7 (RPP7) and RESISTANCE TO POWDERY MILDEW 8 (RPW8)/HOMOLOG OF RPW8 (HR). The RPP7 locus encodes intracellular nucleotide binding site-leucine rich repeat (NLR) immune receptors that can confer strain-specific resistance to oomycetes, while the RPW8/HR locus encodes atypical resistance proteins, of which some can confer broad-spectrum resistance to filamentous pathogens. There is extensive structural variation in the RPW8/HR cluster, both at the level of gene copy number and at the level of C-terminal, 21- or 14-amino acid long RPW8/HR repeats. We demonstrate that the number of RPW8/HR repeats and the short C-terminal tail correlate, in an allele-specific manner, with the severity of hybrid necrosis when these alleles are combined with RPP7 variants. We discuss these findings in light of sequence similarity between RPW8/HR and pore-forming toxins MLKL and HET-S from mammals and fungi.


Introduction
The combination of divergent parental genomes in hybrids can produce new phenotypes not seen in either parent. At one end of the spectrum is hybrid vigor, with progeny being superior to the parents, while at the other end there is hybrid weakness, with progeny being inferior to the parents, and in the most extreme cases being sterile or unable to survive.
In plants, a particularly conspicuous set of hybrid incompatibilities is associated with autoimmunity, often with substantial negative effects on hybrid fitness [1][2][3]. Studies of hybrid autoimmunity in several species, often expressed as hybrid necrosis, have revealed that the underlying genetics tends to be simple, with often only one or two major-effect loci. Where known, at least one of the causal loci encodes an immune protein, often an intracellular nucleotide binding site-leucine-rich repeat (NLR) protein [4][5][6][7][8][9][10][11][12][13]. The gene family encoding NLR immune receptors is the most variable gene family in plants, both in terms of inter-and intraspecific variation [14][15][16][17]. Many NLR proteins function as major disease resistance (R) proteins, with the extravagant variation at these loci being due to a combination of maintenance of very old alleles by long-term balancing selection and rapid evolution driven by strong diversifying selection [18][19][20]. The emergence of new variants is favored by many NLR genes being organized in tandem clusters, which can spawn new alleles as well as copy number variation by illegitimate recombination, and by the presence of leucine rich repeats in NLR genes, which can lead to expansion and contraction of coding sequences [21][22][23]. Cluster expansion has been linked to diversification and adaptation in a range of systems [24][25][26]. Several complex plant NLR loci provide excellent examples of cluster rearrangement increasing pathogen recognition specificities [19]. Substantial efforts have been devoted to decomposing the complexity of the plant immune system and interactions between its components. While many plant disease R genes are members of the NLR family, some feature different molecular architectures. One of these is RESISTANCE TO POWDERY MILDEW 8 (RPW8) in Arabidopsis thaliana, which was first identified based on an allele that confers resistance to multiple powdery mildew isolates [27]. The namesake RPW8 gene is located in a gene cluster of variable size and composition that includes multiple RPW8-like genes as well as HOMOLOG OF RPW8 (HR) genes [27][28][29]. The reference accession Col-0, which is susceptible to powdery mildew, has four HR genes, but no RPW8 gene, whereas the resistant accession Ms-0 carries RPW8. 1 and RPW8.2 along with three HR genes [27]. NLRs are distinguished by N-terminal Toll/interleukin-1 receptor (TIR) or coiled-coil (CC) domains, which, when overexpressed alone, can often activate immune signaling [30,31].
A subset of CC-NLRs (CNLs) has a diagnostic type of coiled-coil domain, termed CCR to indicate that this domain is being shared with RPW8/HR proteins. The latter have an N-terminal extension that might be a transmembrane domain as well as C-terminal repeats of unknown activity [32,33]. It has been noted that the CCR domain is similar to a portion of the animal mixed-lineage kinase domain-like (MLKL) protein that forms a multi-helix bundle [34] as well as the HeLo and HELL domains of fungi, which also form multi-helix bundles [35][36][37]. Many fungal HeLo domain proteins have a prion-forming domain that consists of C-terminal 21-amino acid repeats. This domain can form amyloids and thereby affect oligomerization and activity of these proteins [35][36][37][38][39].
We have previously reported hybrid necrosis due to incompatible alleles at the RPW8/HR locus and at the complex RECOGNITION OF PERONOSPORA PARASITICA 7 (RPP7) locus, which encodes a canonical CNL and which has alleles that provide race-specific resistance to the oomycete Hyaloperonospora arabidopsidis [54,55].
Here, we investigate in detail three independent cases of incompatible RPW8/HR and RPP7 alleles, and show that two are caused by members of the fast-evolving RPW8.1/HR4 clade. We describe how variation in the number of C-terminal repeats and the short C-terminal tail predict the degree of incompatibility between two common RPW8.1/HR4 alleles and corresponding RPP7 alleles. In a systematic intercrossing and genetic mapping program among 80 A. thaliana accessions, a series of genomic regions involved in hybrid incompatibility were identified. The underlying genes were termed DANGEROUS MIX (DM) loci. One instance, between the DM6 and DM7 regions, stood out because it is responsible for two phenotypically distinct hybrid necrosis cases (Fig 1A) [10]. Strong candidates, as previously inferred from a corresponds to the RPP7 cluster, and DM7 to the RPW8/HR cluster. We recently found an additional case of incompatibility between the DM6 and DM7 regions, with a third distinctive phenotype (Fig 1A and 2A). In addition to phenotypic differences between the three DM6-DM7 F1 hybrids, test crosses confirmed that each case was caused by different combinations of DM6 and DM7 alleles, as only certain combinations resulted in hybrid necrosis (Fig 1B). Fig Table S2). The hit in the RPW8/HR region (red arrow) stands out, but It is possible that some of the other hits that pass the significance threshold (Bonferroni correction, 5% familywise error) identify modifiers of the DM6-DM7 interaction.

6
To corroborate the evidence from mapping experiments that DM6 alleles of Mrk-0 and ICE79 were RPP7 homologs, we designed ten artificial microRNAs (amiRNAs) based on sequences from the Col-0 reference accession. AmiRNAs targeting a subclade of five RPP7 homologs that make up the second half of the RPP7 cluster in Col-0, suppressed hybrid necrosis in all three crosses, Mrk-0 x KZ10, Lerik1-3 x Fei-0 and ICE79 x Don-0 ( Fig S1 and Table S1). These rescue experiments, together with the above-mentioned test crosses, indicate that specific RPP7 homologs in Mrk-0, Lerik1-3 and ICE79 correspond to different DM6 alleles that cause hybrid necrosis in combination with specific DM7 alleles from other accessions.

A common set of RPW8/HR haplotypes affecting hybrid performances in F1 and F2 progeny
In the mentioned set of diallelic F1 crosses among 80 accessions [10], we noted that the DM6 carrier Lerik1-3 was incompatible with several other accessions, suggesting that these have DM7 (RPW8/HR) hybrid necrosis risk alleles that are similar to the one in Fei-0. Crosses with TueScha-9 and TueWa1-2 produced hybrids that looked very similar to Lerik1-3 x Fei-0 progeny, with localized spots of cell death spread across the leaf lamina along with leaf crinkling and dwarfism (Fig 1D and S2). Similar spots of cell death and leaf crinkling were observed in crosses of Lerik1-3 to ICE106 and ICE107, although these were not as dwarfed (Fig 1C,D and S2). a. If the model had a class of dead segregants that could not be counted, n was estimated to include the dead individuals for χ 2 calculation. b. In the bottom three populations, F1 phenotypes were nearly indistinguishable from normal ones and therefore both classes were combined. c. More severe than F1 hybrids with distinct DM6-DM7 phenotypes. For milder cases, the enhanced phenotypic classes were separated into two groups, with a rosette diameter of 1 cm as threshold. The rightmost numbers indicate the most severe class. d. Best-fit models using F2 segregation analyses with incompatibility alleles indicated as "A" and "B".
Hybrid necrosis often becomes more severe when the causal loci are homozygous [5,7,10,12]. To explore whether Lerik1-3 might cause milder forms of hybrid necrosis that are missed in the F1 generation, we surveyed several F2 populations involving Lerik1-3. Six segregated necrotic plants with very similar phenotypes (Fig 1D,E and S2). This makes all together for 11 incompatible accessions, which are spread over much of Eurasia ( Fig   1E). The F2 segregation ratios suggested that the DM7 allele from ICE106/ICE107 is intermediate between the Fei-0/TueWa1-2/TueScha-9 alleles and the Cdm-0/Nie-0 alleles (Table 1). Alternatively, the hybrid phenotypes might be affected by background modifiers, such that identical DM7 alleles produce a different range of phenotypes in combination with DM6 Lerik1-3 .
Because the phenotypic variation among hybrid necrosis cases involving Lerik1-3 could involve loci other than DM6 and DM7, we carried out linkage mapping with Lerik1-3 x ICE106 and Lerik1-3 x ICE107 crosses. We combined genotyping information from Lerik1-3 x ICE106 and Lerik1-3 x ICE107 F2 and F3 individuals for mapping, because the genomes of ICE106 and ICE107, which come from closeby collection sites, are very similar and because the two crosses produce very similar F1 hybrid phenotypes, suggesting that the responsible alleles are likely to be identical. We used F3 populations to better distinguish different phenotypic classes, since we did  not know the number of causal genes nor there genetic behavior. QTL analysis confirmed that the DM6 and DM7 genomic regions are linked to hybrid necrosis in these crosses (Fig 2A,B).
To narrow down the DM7 mapping interval, we took advantage of having 11 accessions that produced hybrid necrosis in combination with Lerik1-3, and 69 accessions (including Lerik1-3 itself) that did not. We performed GWAS with Lerik1-3-dependent hybrid necrosis as a binary trait [40]. The by far most strongly associated marker was immediately downstream of HR4, the last member of the RPW8/HR cluster in Col-0 (Fig 2C and Table   S2). An amiRNA matching HR4 sequences from Col-0 fully rescued both the strong necrosis in Lerik1-3 x Fei-0 and the weaker necrosis in Lerik1-3 x ICE106 (Fig 3A and Table S3). We confirmed the causality of another member of the RPW8/HR cluster in the KZ10 x Mrk-0 case with a CRISPR/Cas9-induced mutation of RPW8.1 KZ10 (Fig 3B and Fig S3).  Table S4). Both Lerik1-3 x Fei-0 and Lerik1-3 x ICE106 F1 hybrids were resistant to Hiks1, although apparently somewhat less so than the Lerik1-3 parents (Fig S4). We further tested whether RPP7-like genes from Lerik1-3 are likely to be involved in Hiks1 resistance by inoculating Hiks1 on transgenic lines carrying seven different amiRNAs against RPP7 homologs (Table S1), but none of the amiRNAs reduced Hiks1 resistance. These negative results are difficult to interpret; the amiRNAs might not efficiently knock down all RPP7 homologs in Lerik1-3 (for which the exact structure of the RPP7 cluster is unknown), Hiks1 resistance in Lerik1-3 might require other RPP loci (of which there are many in the A. thaliana genome), or Hiks1 resistance might be complex, as found in other accessions [42]. Finally, given the interaction of RPP7 with RPW8/HR in hybrid necrosis, we asked whether HR4 is required for RPP7-mediated Hiks1 resistance in Col-0. Two independent hr4 CRISPR/Cas9 knockout lines were generated in Col-0 (Fig S3), but both remained completely resistant to Hiks1 (Fig S4 and Table S4), indicating that HR4 in Col-0 is dispensable for RPP7-mediated resistance to Hiks1.

Structural variation of the RPW8/HR cluster
For reasons of convenience, we assembled the RPW8/HR cluster from TueWa1-2 instead of Fei-0; accession TueWa1-2 interacted with RPP7 Lerik1-3 in the same manner as Fei-0, the strong necrosis in Lerik1-3 x TueWa1-2 was rescued with the same amiRNA as in Lerik1-3 x Fei-0 (Table S3), and TueWa1-2 had an HR4 allele that was identical in sequence to HR4 Fei-0 . We found that the RPW8/HR cluster from TueWa1-2 had at least 13 RPW8/HR-like genes, several of which were very similar to each other (Fig 4A). For example, there were at least four copies of RPW8.3-like genes with 93 to 99.8% sequence similarity, and two identical RPW8.1 genes, named RPW8.1a, followed by distinct RPW8/HR copies.
Recapitulation experiments had identified HR4 Fei-0 (identical to HR4 TueWa1-2 and HR4 TueScha-9 ) and HR4 ICE106 as causal for hybrid necrosis (Fig 3C,D). We analyzed the phylogenetic relationship of the RPW8/HR genes in The RPW8/HR cluster of TueWa1-2 consists of RPW8/HR members from both the ancestral and the two A.
thaliana specific clades, an arrangement that has not been observed before. Using species-wide data [44], we found that accessions carrying Col-0-like HR4 alleles have simple cluster configurations, while accessions with HR4 genes resembling hybrid necrosis alleles have more complex configurations (Fig 4A). The tagging SNPs found in GWAS (Fig 4A, and Table S2) were mostly found to be associated with the complex clusters, suggesting that the tagging SNPs are linked to structural variation in the distal region of the RPW8/HR cluster ( Fig 4B).

Causality of RPW8/HR C-terminal repeats
To further narrow down the mutations that cause autoimmunity, we compared RPW8.1 KZ10     Normal Both RPW8.1 and HR4 repeats are predicted to fold into extended alpha-helices, but only RPW8.1 repeats appear to have the potential to form coiled coils [45].
The number of repeats varies in both RPW8.1 and HR4 between hybrid necrosis risk and non-risk alleles. To experimentally test the effect of repeat number variation and other polymorphisms, we generated a series of derivatives in which we altered the number of repeats and swapped different portions of the coding sequences between the RPW8.1 KZ10 risk and RPW8.1 Ms-0 non-risk alleles, and between the HR4 Fei-0 and HR4 ICE106 risk and the HR4 Col-0 non-risk alleles (Fig 5A). coding sequences of risk alleles were sufficient to induce hybrid necrosis (Fig 3C and Fig 5A, B). To simplify discussion of the chimeras, the N-terminal portion was labeled with the initial of the accession in italics ("M", "K", etc.), complete repeats were labeled with different capital letters to distinguish sequence variants ("A", "B", etc.), the partial repeat in KZ10 with a lowercase letter ("c"), and the C-terminal tails with Greek letters ("α", "β", etc.).
In RPW8.1 KZ10 , there are two complete repeats and one partial repeat, while RPW8.1 Ms-0 has only one repeat (Fig 5A). Modifying the number of repeats in RPW8.1 affected the frequency and severity of necrosis in T1 plants in a Mrk-0 background, which carries the interacting RPP7 allele, dramatically. Deletion of the first full repeat in RPW8.1 KZ10 ("K-Bcβ", with the KZ10 configuration being "K-BBcβ") substantially reduced the number of plants that died in the first three weeks of growth. The additional deletion of the partial repeat ("K-Bβ") reduced death and necrosis even further (Fig 5A). That K-Bβ still produces some necrosis, even though its repeat structure is the same as in the inactive K-Aα suggests that the polymorphism in the C-terminal tail makes some contribution to necrosis activity. It is less likely that the polymorphism in the repeats play a role, as there is only a very conservative aspartate-glutamate difference between α and β repeats.
In contrast to repeat shortening, the extension of the partial repeat ("K-BBBβ") or addition of a full repeat ("K-BBBcβ") increased the necrosis-inducing activity of RPW8.1 KZ10 , such that almost all T1 plants died without making any true leaves. However, it appears that not all repeats function equally, as removal of the partial repeat slightly increased necrosis-inducing activity ("K-BBβ"). Polymorphisms in the N-terminal non-repeat region seemed to contribute to necrosis, as swaps of the N-terminal Ms-0 fragment ("M-BBcβ" or "M-BBBβ") induced weaker phenotypes than the corresponding variants with the N-terminal fragment from KZ10. Nevertheless, we note that the normal KZ10 repeat configuration was sufficient to impart substantial necrosis-inducing activity on a chimera in which the N-terminal half was from Ms-0, which is distinguished from KZ10 by nine nonsynonymous substitutions outside the repeats.
Compared to the RPW8.1 situation, the relationship between HR4 repeat length and necrosis-inducing activity is more complex. The natural alleles suggested a negative correlation of repeat number with necrosis-inducing activity when crossed to Lerik1-3, since the non-risk HR4 allele from Col-0 has five full repeats, while weaker risk alleles such as the one from ICE106 have two, and the strong risk allele from Fei-0 has only one (Fig 5B).
Addition of a full repeat to HR4 Fei-0 ("F-RTδ", with the original Fei-0 configuration being "F-Tδ") reduced its activity to a level similar to that of HR4 ICE106 ("I-RTδ"). Deletion of a full repeat from HR4 ICE106 ("I-Tδ") modestly increased HR4 activity (Fig 5B). Together, the chimera analyses indicated that the quantitative differences between crosses of Fei-0 and ICE106 to Lerik1-3 (Fig 1 and S2) are predominantly due to variation in HR4 repeat number. This is further supported by the necrosis-inducing activity of a chimera in which the repeats in the Col-0 non-risk allele were replaced with those from HR4 Fei-0 ("C-Tδ", with the original Col-0 configuration being "C-QQRSRγ") (Fig 5B and S5). However, repeat number alone is not the only determinant of necrosis-inducing activity of HR4 in combination with RPP7 Lerik1-3 . Adding another repeat to the "F-RTδ" chimera, resulting in "F-RRTδ", increased the activity of HR4 Fei-0 again, perhaps suggesting that there is an optimal length for HR4 to interact with RPP7 Lerik1-3 .
Taken together, the swap experiments led us to conclude that naturally occurring variation in the configuration of RPW8/HR repeats play a major role in quantitatively modulating the severity of autoimmune phenotypes when these RPW8/HR variants are combined with RPP7 alleles from Mrk-0 and Lerik1-3. At least in the case of HR4, we could show directly that the short C-terminal tail also affects the hybrid phenotype, while for RPW8.1 this seems likely as well, given that the repeats between different alleles differ less from each other than the tails.  (Fig 6A) were predicted to be as frequent as the long non-risk variants (Fig 6A, B and Table S5), whereas for RPW8.1, only seven accessions were predicted to have the long RPW8.1 KZ10 -type risk variant (Fig 6A and Table S6).
To confirm the short read-based length predictions, RPW8.1 was PCR amplified from 28 accessions and HR4 from 113 accessions (Fig 6A-D and Table S5 and S6). This not only confirmed that the Illumina predictions were accurate, but also revealed new variants with different arrangements of HR4 repeats, although none were as short as HR4 Fei-0 or HR4 ICE106 (Fig 6A, B). The short necrosis-risk HR4 variants are found across much of the global range of A. thaliana (Fig 6C), whereas the much rarer necrosis-risk RPW8.1 KZ10 -like variant was exclusive to Central Asia. We also observed that sequences of the two short HR4 types were more conserved than the longer ones, with each short type belonging to a single haplotype, while the long necrosis-risk HR4 alleles belonged to multiple haplotypes (Fig 6D) Table S5). The extensive information on RPW8.1/HR4 haplotypes allowed us to use test crosses to determine whether interaction with either RPP7 Mrk-0 or RPP7 Lerik1-3 is predictable from sequence, specifically from repeat number (Fig   6E, F). As expected, accessions with the longest, Type 1, RPW8.1 KZ10 -like alleles (Fig 6E, pink) produced necrotic hybrid progeny when crossed to Mrk-0, whereas accessions carrying the two shorter Type 2 and 3 alleles did not (Fig 6E and Table S7). The situation was similar for HR4; all but two of the tested accessions with the shortest HR4 Fei-0 -like alleles (Fig 6F, red) produced strongly necrotic progeny when crossed to Lerik1-3, while accessions carrying the second shortest HR4 ICE106 -like alleles (Fig 6F and Table S8) produced more mildly affected progeny. Hybrid progeny of Lerik1-3 and accessions carrying other HR4 alleles did not show any signs of necrosis (Fig 6F). Necrosis was correlated with reduction in overall size of plants, which in turn correlated with RPW8.1/HR4 repeat length (Fig 6F and Table S9). Finally, HR4 Fei-0 -like alleles in two accessions caused a mild phenotype similar to HR4 ICE106 , suggesting the presence of genetic modifiers that partially suppress autoimmune symptoms.

Discussion
The RPW8/HR cluster is remarkably variable in terms of copy number, reminiscent of many multi-gene clusters carrying NLR-type R genes [16]. While the first three genes in the cluster, HR1, HR2 and HR3, are generally well conserved, there is tremendous variation in the number of the other genes in the cluster, including RPW8.1/HR4.
Nevertheless, that the HR4 hybrid necrosis-risk allele is not rare and widely distributed, accounting for half of all HR4 carriers (Fig 6B, C), suggests that it might provide adaptive benefits, as postulated before for ACD6 hybrid necrosis-risk alleles [12]. effectors. In such a scenario, at least some RPP7 proteins might act as guards for RPW8/HR proteins and sense their modification by pathogen effectors [16,52].
Can we conclude from the MLKL homology that RPW8 and HR proteins form similar pores as MLKL?
Unfortunately, this is not immediately obvious, as a different mechanism has been suggested for fungal proteins with HeLo and HELL domains [35][36][37]. For MLKL, it has been suggested that the multi-helix bundle directly inserts into the membrane, whereas for the fungal protein, it has been proposed that the multi-helix bundle regulates the ability of an N-terminal transmembrane domain to insert into the membrane. An N-terminal transmembrane domain has been predicted for RPW8 [27], but although RPW8 proteins can be membrane associated [53,54], the insertion of this domain into the membrane has not been directly demonstrated.
We have shown that differences in protein structure, rather than expression patterns or levels, are key to the genetic interaction between RPW8/HR and RPP7. While we do not know whether the proteins interact directly, allele-specific genetic interactions are often an indicator of direct interaction between the gene products [55]. Moreover, reminiscent of RPW8/HR and RPP7 interaction, the activity of the fungal HeLo domain protein HET-S is regulated by an NLR protein [38].
In conclusion, we have described in detail an intriguing case of hybrid necrosis in A. thaliana, where three different pairs of alleles at a conventional complex NLR resistance gene cluster, RPP7, and alleles at another complex, but non-NLR resistance gene cluster, RPW8/HR, interact to trigger autoimmunity in the absence of pathogens. Our findings suggest that within the immune system, conflict does not occur randomly, but that certain pairs of loci are more likely to misbehave than others. Finally, that genes of the RPW8/HR cluster can confer broad-spectrum disease resistance, while at least one RPP7 member can confer race-specific resistance, provides yet another link between different arms of the plant immune system [56].
The binary trait with accession information was submitted to the easyGWAS platform [40], using the FaSTLMM algorithm. A -log10(p-value) was calculated for every SNP along the five A. thaliana chromosomes.

RPW8.1/HR4 length prediction
Short reads from the 1001 Genomes project (http://1001genomes.org) were mapped using SHORE [58] with 5 mismatches allowed per read. Sequences of the RPW8/HR clusters from Col-0 and KZ10 were provided as references and the covered region for RPW8.1 KZ10 and HR4 Col-0 was retrieved.

RPW8.1/HR4 sequence analysis
Overlapping fragments covering the HR4/RPW8.1 genomic region were PCR amplified from different A. thaliana accessions (oligonucleotides in Table S11). Fragments were cloned and Sanger sequenced. A maximum-likelihood tree of coding portions of exons and introns was computed using RaxML [59] and visualized with Figtree.

Population genetic analysis
The geographical distribution of the 113 accessions carrying different HR4 alleles was plotted using R (version 0.99.903). Packages maps, mapdata, mapplots and scales were used. A haplotype network was built using a cDNA alignment of 113 HR4 alleles from different accessions. The R packages used were ape (dist.dna function) and pegas (haploNet function).

Histology
Cotyledons from 18 day-old seedlings were collected and 1 ml of lactophenol Trypan Blue solution (20 mg Trypan Blue, 10 g phenol, 10 ml lactic acid, 10 ml glycerol and 10 ml water) diluted 1 : 2 in 96% ethanol was added for 1 hour at 70°C. Trypan Blue was removed, followed by the addition of 1 ml 2.5g/ml chloral hydrate and an overnight incubation. The following day, the de-stained cotyledons were transferred to 50% glycerol and mounted on slides.