Dataset

We used a collection of genome assemblies that included 82 haploid isolates of Sordariales fungi and two outgroups representing 27 genera (S1 Table). Assembly size ranged from 28.3Mb to 57.7Mb, with an average size of 39.3Mb (standard deviation [s.d.]: 6.0Mb). L50 ranged from two to 704 contigs (mean: 83.8; s.d.: 128.9) and N50 from 14.1Kb to 11.2Mb (mean: 2.29Mb; s.d.: 2.82Mb; S1 Table). The completeness of assemblies, as estimated using Busco [47], ranged between 77.3 and 99.2% (mean: 96.1%; s.d.: 3.7%), and the number of predicted genes varied from 7,066–14,970 (mean: 10,582; s.d.: 1,788; S1 Table). Together, these observations indicate that our dataset is of sufficient quality to probe into the evolution of a family of genes such as NLRs.

Species tree inference

To place our analyses in an evolutionary context, we constructed a species tree using two different species tree inference methods, based on single-copy orthologs. Orthology analysis of the 82 ingroup genomes and two outgroup genomes identified 77,480 orthogroups, of which 2367 orthogroups were single-copy orthogroups that were present in >80% of the genomes included in the dataset (referred to as SCO80), and 1030 orthogroups were single-copy orthogroups present in all genomes (referred to as SCCO). The species tree inferred using the gene tree summary method implemented in Astral-III [48] based on SCO80 was supported by high confidence values (Fig 2A). Local Posterior Probability (localPP) was maximal for all nodes, except the node including Podospora bellae-mahoneyi and P. pseudocomata (localPP: 0.92), and the node carrying the three N. tetrasperma lineages (localPP: 0.68 and 0.99). The species tree recovered the main Sordariales families as monophyletic groups, and confirmed the polyphyly of Podospora, Cercophora, and Chaetomium. The total-evidence species tree inferred using a maximum-likelihood approach based on the concatenation of SCCO sequences was mostly congruent with the Astral-III species tree (S1 Fig). Incongruence between Astral and total-evidence species trees was limited to the nodes with the lowest bootstrap support in the latter. All subsequent analyses were carried out using the Astral-III phylogram as the species tree, after having inferred branch lengths using the SCCO sequences (Fig 2A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. NLR richness across Sordariales.

(A) Sordariales species tree built with Astral-III and 2367 gene trees; branch length was estimated using RAxML-ng and 1030 single-copy orthologs present in all taxa. In panels A, D-G, I, and J, colors represent Sordariales families, as defined by Wang et al. [49]. (B) Frequency of the different types of NB domains in individual NLRomes. (C) Number of NLR clusters identified in genomes with an N50 greater than 1Mb. (D) Number of NLR genes per genome. (E-F-G) Relationships between the number of NLR genes and the GC content, the RIP affected area, and the repeat content. (H-I-J) Relationships between the number of NLR clusters and the bimodal GC distribution, the number of NLR genes, and the RIP-affected area. For continuous parameters, a linear regression is presented, and the adjusted R² is reported. For the bimodal GC distribution, a correlation index r was calculated using a threshold model.

https://doi.org/10.1371/journal.pgen.1011739.g002

High variability in NLR content across Sordariales

To investigate whether the extensive variability in NLR content observed at the scale of Dikarya is also observed at the more restricted scale of Sordariales, and to investigate if there are differences between families or within families of Sordariales, we used a combination of Pfam-A and Sordariales-specific HMM profiles which identified 4152 genes with an NB domain and canonical C-terminal repeats (Ankyrin, TPR, WD40, or HEAT; Fig 3). The methodology employed to predict NLR genes is summarized in Fig 3 and detailed in the Methods section. HMM profiles were generated from alignments of domains identified using Pfam-A, with the aim of improving detection sensibility in Sordariales.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Bioinformatic identification of NLRs in 84 Sordariales genomes, with counts of the number of candidate genes and orthogroups identified.

https://doi.org/10.1371/journal.pgen.1011739.g003

The 4152 genes were clustered into 144 orthogroups, of which 97 orthogroups with ≤5 candidate NLRs were visually inspected and corrected for gene prediction errors, and 47 orthogroups with >5 candidate NLRs were subjected to additional annotation steps. We generated HMM profiles from NB domains for all 47 orthogroups. These profiles were used to detect genes with NB domains initially missed during domain prediction. Among genes with an NB annotation, we only kept genes longer than 1kb and we aligned the remaining 3,758 NLR candidates against the 84 genome assemblies using Exonerate to (i) correct truncated NLRs, (ii) identify NLRs missed by gene prediction, and (iii) identify NLRs with non-canonical structures, either due to the presence of non-canonical domains or the lack of any domain at the C-terminus, N-terminus, or both. The final set of NLRs included 4769 NLR sequences: 4613 within the 82 Sordariales genomes (S2 Table) and 153 within the two outgroups. Orthology analysis of the 4613 candidate NLR sequences identified 485 NLR orthogroups: 91 in Chaetomiaceae, 229 in Lasiosphaeriaceae, 119 in Podosporaceae, and 46 in Sordariaceae (S3 and S4 Tables). The majority of NLRs (3,858 – 84%) clustered in 191 large orthogroups (39%) comprising six or more NLRs. The remaining NLRs (688 – 15%) were grouped into 227 (47%) orthogroups containing two to five NLRs, and 67 NLRs (1%) were singletons. In most orthogroups, NLRs were unevenly distributed among the different species, with only 14 orthogroups (3%) including sequences from all species of the corresponding Sordariales family: 10 in Podosporaceae, four in Lasiosphaeriaceae, one in Sordariaceae, and none in Chaetomiaceae. Most of the large orthogroups (153 – 80%) and many of the small orthogroups (64 – 28%) were species-specific and thus included paralogous sequences.

The average number of NLRs per Sordariales genome was 56 (s.d.: 42.8), and ranged from 10 NLRs in Crassicarpon thermophilum and Trichocladium antarcticum, to 162 NLRs in Lasiosphaeria ovina (Fig 2D and S1 Table). At the family level, the average number of NLRs was 60 in Chaetomiaceae (s.d.: 34.7; coefficient of variation [CV]: 0.58 + /- 0.19), 122 in Lasiosphaeriaceae (s.d.: 25.8; CV: 0.21 + /- 0.11), 30 in Sordariaceae (s.d.: 26.4; CV: 0.89 + /- 0.17), and 71 in Podosporaceae (s.d.: 21.6; CV: 0.31 + /- 0.37). A Kruskal-Wallis test indicated a significant family effect on the number of NLRs, and post-hoc Wilcoxon-Mann-Whitney tests revealed statistically significant differences between all pairs of families, except Chaetomiaceae and Podosporaceae. At the family scale, we observed extensive variation in NLR content within Sordariaceae, with fire-associated Neurospora taxa presenting smaller repertoires than their soil- or dung-associated relatives. Using phylogenetic generalized least squares to account for knowledge of phylogenetic relationships, we found no significant association between the number of NLRs and assembly quality statistics (N50, L50, number of contigs, total length, or Busco score; S5 Table), indicating that the observed variation in NLR content is primarily biological in origin.

No strong relationship between mechanisms regulating genomic repeat elements and NLR content

The variation in NLR content between Sordariales families suggests historical duplication events in certain lineages despite defense mechanisms against genomic repeat elements. To examine the relationship between the number of NLRs and the stringency of mechanisms involved in regulating genomic repeat elements, such as repeat-induced point mutations (RIP) [38], we used phylogenetic generalized least squares [50] and a threshold model [51]. RIP is a mechanism of fungal genome defense that induces C-to-T mutations in repetitive DNA to prevent their expression. The number of NLRs was positively significantly correlated with the GC content (adjusted R²: 0.07; Fig 2E) and negatively significantly correlated with the repeat content (adjusted R²: 0.06; Fig 2G) and the percentage of RIP-affected regions (adjusted R²: 0.12; Fig 2F), but correlations were relatively weak (S5 Table). Transposon invasion followed by RIP can lead to genomic compartmentalization of AT-rich regions and a bimodal distribution of GC content [52]. We found that differences in the number of NLRs between genomes with and without bimodal distributions of GC content were not statistically significant. Together, these analyses indicated that mechanisms of defense against genomic repeat elements had no strong impact on the number of NLRs.

NLRs form clusters in NLR-rich species

The lack of a clear correlation between NLR content and RIP stringency should not be seen as evidence that NLRs evade defense mechanisms, but rather as a sign that they have diversified through processes other than tandem gene duplication. To better understand how NLR repertoires diversified in the Sordariales, we studied the presence of NLR clusters. In plants, NLRs often form clusters, which may represent a means for generating new functional diversity through unequal crossovers (leading to gene loss and the formation of new genes), as well as gene conversion events between partially orthologous genes [25,53,54]. These changes in sequence and genomic organization may be facilitated by the repetitive structure of the C-terminal region of NLRs.

Before determining whether the NLRs form clusters, we had to define what a cluster is in the Sordariales. Plant NLR clusters are usually defined as sets of NLRs separated by less than 200 kb and by a limited number of other genes, usually up to eight [35]. However, this definition may not be suitable for fungal genomes, which are generally more gene-dense than plant genomes. We therefore sought to determine the optimal genomic size and number of NLRs to define clusters in Sordariales. We focused on the 37 Sordariales genomes with an N50 higher than 1Mb, in which 2349 NLRs were identified. We then compared the number of clustered genes for NLRs and 1000 sets of randomly-selected genes while varying the maximum distance and the number of non-NLR genes between NLRs of the same cluster. For each genome, the size of each of the 1000 sets of randomly-selected genes was equal to the number of NLRs. We observed that for genomes in which NLR genes formed more clusters than randomly-selected genes, the difference in the number of clusters was maximized for cluster length varying between 20–45 kb and with a maximum of eight non-NLR genes between two NLRs. Based on these simulations, we defined a Sordariales NLR cluster as a group of two or more NLRs, with two consecutive NLRs being separated by less than 40 kb and less than eight non-NLR genes.

Using these criteria, we identified that 602 (26%) of the 2349 NLRs formed 272 clusters in 26 genomes (Figs 2C and 4 and S1 and S6 Tables). No NLR clusters were identified in the remaining eleven genomes, corresponding to the seven conidiating species of Neurospora (N. crassa clade A, N. crassa clade B, N. hispaniola, N. tetrasperma L9, N. tetrasperma L6, N. sitophila clade 3, N. discreta PS4B) and four Chaetomiaceae species (Thermothelomyces heterothallica, Thermothielavioides terrestris, Thermothelomyces thermophila, T. antarcticum). We then examined the cluster size, the variability of domain assortments within clusters, and estimates of the stringency of defense against genomic repeats, to gain insight into the mechanisms underlying cluster formation. Among the 272 clusters, 164 consisted of NLRs sharing the same type of NB domain, and only 16 clusters consisted of paralogous NLR copies. The number of NLRs per cluster was relatively limited, with 225 clusters of two NLRs, and 47 clusters of three to five NLRs. The numbers of NLR genes and NLR clusters were highly and significantly positively correlated (adjusted R²: 0.86; Fig 2I). No strong associations existed between the number of NLR clusters and assembly quality. Among assembly quality statistics (N50, L50, number of contigs, total length, or Busco score), only the Busco score showed a positive, significant but relatively weak correlation with the number of NLR clusters (adjusted R²: 0.12; S5 Table). GC and repeat content did not correlate with the number of NLR clusters. Both the RIP bimodal distribution (mean correlation r: -0.45; 95% confidence interval not overlapping zero; Fig 2H) and the percentage of RIP-affected regions (adjusted R²: 0.14; Fig 2J and S5 Table) showed a negative significant correlation with the number of NLR clusters. These results, particularly the correlation between the number of NLRs and the number of clusters, suggest that organizing into clusters may enable NLR repertoires to diversify. RIP prevents tandem gene duplications, but processes such as unequal crossing over may allow the formation of clusters without strict duplications.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of NLRs along Sordariales genome.

(A) Distribution of non-clustered NLRs (blue) and clustered NLRs (red) along the 26 genomes with an N50 larger than 1Mb. All scaffolds have been concatenated to simplify the representation. Genomes are organized from top to bottom in decreasing number of clustered NLRs. (B) Distribution of the seven non-clustered NLRs (blue) and the 11 clustered NLRs (red) along the scaffold 10 of Podospora didyma.

https://doi.org/10.1371/journal.pgen.1011739.g004

Another important feature of plant NLRs whose existence we sought to examine in Sordariales was their frequent organization as head-to-head pairs. Paired NLRs in plants follow a “sensor-helper” model in which the sensor NLR binds the exogenous ligand while the helper NLR triggers the immune response. The NLR sensor often includes an integrated decoy [55]. Among the 225 clusters of two NLRs, 13 were organized as head-to-head NLRs, identified in 11 genomes (S7 Table). Most head-to-head NLRs (19/26) had a complete tripartite structure, and only one paired NLR contained a non-canonical domain (Lipocalin_5; S7 Table). These results indicate that head-to-head NLRs are rare in Sordariales and that Sordariales NLRs have not converged evolutionarily towards the same type of head-to-head organization as plant NLRs.

Sordariales display two types of NACHT NLR proteins

As a minority of 55 NLRs presented an NB domain (Fig 2B) that could not be identified as NACHT or NB-ARC, we set out to clarify their identity. Out of 4613 Sordariales NLRs, 3350 (73%) presented a NACHT domain, 1141 (25%) an NB-ARC, and 55 NLRs were identified as both NACHT and NB-ARC, with close E-values in analyses with HMMER. In the NLR-specific orthology analysis, all 55 NLRs clustered with an additional seven NACHT NLRs, one NB-ARC NLR, and seven NLRs without identified NB domain, forming four family-specific orthogroups. To clarify the identity of the NB domain present in these 70 NLRs, we built an alignment and examined the Walker B motif that is expected to distinguish NB-ARC and NACHT domains [23]. We observed that the first acidic residue (aspartic acid) of the Walker B was not directly followed by another acidic residue, as expected for the NB-ARC, but by a glycine. We also found that another acidic residue (aspartic acid) was present three positions downstream of the first one. These features fitted the typical NACHT consensus motif hhhhD[GAS]hDE described in Leipe et al. [23].

To ascertain why our HMM profiles failed to identify a minority of NLRs with a typical NACHT Walker B as NACHT domains, we further examined their similarity with other NACHT and NB-ARC domains, using four sequences of these NACHT domains, 325 sequences of other NACHT domains, and 85 sequences of NB-ARC domains. All sequences represented different orthogroups. We split NB sequences into two fragments (from the N-terminal end to the start of the Walker B motif, and from the start of the Walker B to the C-terminal end; Fig 5A). For each fragment, we computed distances between sequences using alignment-free methods of sequence comparison [56,57], and visualized distance matrices using principal coordinates analysis (PCoA). In PCoAs based on the first and second fragment (Figs 5E-H and S2), NB-ARC and bona fide NACHT fragments were clearly distinguished along the first coordinate. In the PCoA based on the distance between C-terminal sequences (i.e., second fragment; Figs 5F, 5H and S2), the four unclassified NB sequences clustered with other NACHT sequences, apart from the NB-ARC sequences. In the PCoA based on the distance between N-terminal sequences (i.e., first fragment; Figs 5E, 5G and S2), the four unclassified NB sequences were intermediate between NACHT and NB-ARC sequences. Thus, these distance-based analyses indicate that the four unclassified NB sequences, while matching the NACHT definition based on the Walker B motif, present dissimilarities with the other NACHT sequences identified in Sordariales NLRs.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Two types of fungal NACHT NLRs: the N-NACHT and the T-NACHT.

(A) Domain structure of an NLR, showing the two fragments considered in remaining panels, and the relative position of the HETHS. (B) Cladogram of 122 fungal species representative of the whole fungal kingdom adapted from James et al. [58]. (C) Red dots indicate genomes in which N-NACHT NLRs were identified. (D) Absolute numbers of N-NACHT, NB-ARC, T-NACHT NLRs, and NLRs with an undetermined NB domain. NLRs were defined as genes containing an NB and a canonical C-terminal domain. (E-H) Principal coordinate analyses of matrices of alignment-free genetic distances of the first and second NB fragments of NACHT and NB-ARC NLRs as defined in (A). Normalized Google Distance (G-H) is based on 3-mers, while the Lempel-Ziv Complexity distance (E-F) relies on information theory. (I-T) Alignment of the Walker A, Walker B, and HETHS for 20 N-NACHT NLRs of Sordariales, 20 N-NACHT NLRs of non-Sordariales fungi (10 from Ascomycota and 10 from Entomortierella parvispora), 20 T-NACHT NLRs of Sordariales, and 20 NB-ARC NLRs of Sordariales.

https://doi.org/10.1371/journal.pgen.1011739.g005

To identify the genomic features that differentiate the four unclassified NB sequences from the other NACHT sequences, we examined the HElical THird domain of STAND proteins (HETHS), also named winged helix domain or ARC2 domain, which is a helical bundle located at the C-terminus of the core P-loop domain of STAND proteins (Fig 5A). Leipe et al. [23] proposed a subdivision of NACHT genes into two families: the NAIP-like and TLP1-like families, partly based on the HETHS domain. The NAIP-like family is defined as having a consensus motif FhHxxhQE[YF]hxA in the HETHS domain (note that we corrected a typo in the motif reported in the original publication). Unlike other NACHT sequences, our four representatives of divergent NACHT sequences presented the consensus FLHLT[FY]QEYFAA at the HETHS domain (Fig 5K), which placed them in the NAIP-like family of NACHT domains, while other NACHT sequences represented the TLP1-like family (Fig 5Q). In the following, the two domains are referred to as N-NACHT for the NAIP-like form (which is rare) and T-NACHT for the TLP1-like family (which is more frequent).

To re-evaluate the distribution of NB domains considering two types of NACHT domains, we constructed HMM profiles based on the alignments of N-NACHT on the one hand, and T-NACHT on the other hand. Among the 4,613 Sordariales NLRs, T-NACHT was the most common NB domain (3388 NLRs, 73.4%), followed by NB-ARC (1150 NLRs, 24.9%), and N-NACHT (70 NLRs, 1.5%). The NB domain could not be identified for five NLRs (0.1%), which do not cluster with any NLR with annotated NB in orthology analyses. These five genes were still considered NLRs because they passed our cutoff on identity levels in similarity analyses with Exonerate. All genomes harbored both T-NACHT and NB-ARC NLRs, but only 33 of the 82 genomes had N-NACHT NLRs, with one to five N-NACHT NLR genes per genome. Together, these analyses show that using the diagnostic motifs of the different types of NB domains allows the building of HMM profiles that detect NB domains with increased sensitivity.

Origin of N-NACHT NLRs in Sordariales

To investigate the origin of N-NACHT NLRs in Sordariales, we characterized their distribution throughout the Fungal Kingdom. We assembled a dataset of 122 proteomes representative of the whole Fungal Kingdom (Fig 5B and S8 Table). Using both Pfam-A and HMM profiles, we predicted NLRs within these proteomes, which we defined as genes presenting both an NB domain and canonical C-terminal repeats. We identified a total of 3993 NLRs in 77 of the proteomes (Fig 5D and S9 Table), with NLR numbers ranging from one in species of Chytridiomycota and Mucoromycota to 315 in Trichoglossum hirsutum (Ascomycota: Helotiales). NLRs were missing in Saccharomyces and Schizosaccharomycota species, in the earliest branching lineages of Basidiomycota, and in some of the early branching lineages of fungi (Zoopagomycota, Blastocladiomycota, and Rozellomycota). Among the 3993 fungal NLRs, we identified 2722 T-NACHT NLRs, 1180 NB-ARC NLRs, 44 N-NACHT NLRs, and 47 NLRs with an undetermined NB (Fig 5D and S9 Table). N-NACHT NLRs were identified only in seven species of Ascomycota and one species of Mucoromycota (Entomortierella parvispora, Mortierellales) (Fig 5C and S9 Table). The number of N-NACHT NLRs ranged from one to seven in the seven Ascomycota species, but reached 26 in Entomortierella parvispora. We conclude from these analyzes that N-NACHTs are not an innovation specific to Sordariales, but that they have been acquired or lost several times during fungal evolution.

NLR domain assortments in Sordariales represent a subset of the larger Dikarya

At the scale of Dikarya, NLRs exhibit high diversity of N-terminal and C-terminal domains, and a wide variety of combinatorial domain assortments [2,16]. We aimed to investigate whether this pattern persisted on a smaller evolutionary scale, within Sordariales, or if previous measurements of NLR repertoire variability at the Dikarya scale had been underestimated due to using a limited number of genomes per phylum. We annotated the three types of NB domain (N-NACHT, T-NACHT and NB-ARC), N-terminal and C-terminal domains (Fig 6A) in Sordariales using Pfam-A or Sordariales-specific HMM profiles. We next assigned C-terminal repeats to four categories of canonical (i.e., identified as main components of NLR architectures) NLR repeats (TPR, Ankyrin repeats, WD40 repeats, and HEAT) for 3104 NLRs (67%), and N-terminal repeats to eight canonical N-terminal domains (Goodbye-like, HeLo-like, HET-like, Patatin, PNP_UDP_1, RelA_SpoT, sesB-like, or amyloid) for 3324 NLRs (72%) (Fig 6A). The most frequent N-terminal domain across Sordariales were sesB-like, HeLo-like, Goodbye-like, the most frequent NB domain was T-NACHT, and the most frequent C-terminal repeats were Ankyrin, TPR, and WD40 (Fig 6A). The most frequent NB/C-terminal domain associations were NB-ARC/TPR for NB-ARC NLRs, T-NACHT/Ankyrin or T-NACHT/WD40 for T-NACHT NLRs, and N-NACHT/HEAT for N-NACHT NLRs (Fig 6B). These most frequent NB/C-terminal domain assortments represented 2990 (96%) of the 3104 NLRs exhibiting a canonical C-terminal domain. Among the NLRs with a canonical N-terminal domain, most NLRs with a HeLo-like domain (820 – 97%), a HET-like domain (179 – 92%), a Goodbye-like domain (729 – 90%), or a RelA_SpoT domain (14 – 82%) were T-NACHT NLRs. NLRs with a Patatin domain were almost exclusively of the NB-ARC type (146 – 95%). PNP_UDP_1 and amyloid domains were associated with T-NACHT (170 – 54% and 68–44%, respectively) or NB-ARC (145 – 46% and 87–56%, respectively). The sesB-like domain was the only canonical N-terminal domain that was associated with all three types of NB domains: T-NACHT (624 – 65%), NB-ARC (287 – 30%), and N-NACHT (53 – 5%) (Fig 6B), and the only canonical N-terminal domain identified in N-NACHT NLRs.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Architectural variability of NLRs in Sordariales.

(A) Distribution of NLRs based on their canonical N-terminal, NB, and C-terminal annotations. (B) Associations between canonical N-terminal, NB, and C-terminal domains. Category ‘undetermined’ represents NLRs without any domain predicted. Category ‘missing’ represents truncated NLRs (<20 aa in N-terminus and <50 aa in C-terminus). Category ’mixed’ represents NLRs with several canonical domains. Category ‘non-canonical’ represents NLRs without a canonical domain but with a non-canonical domain.

https://doi.org/10.1371/journal.pgen.1011739.g006

NLRs with similar architectures were generally grouped into the same orthogroups. Most large orthogroups (≥6 NLRs) included a single type of C-terminal domain (166 – 87%) or a single type of N-terminal domain (120 – 63%). Furthermore, for 8% and 17% of large orthogroups, minor C-terminal or N-terminal types, respectively, were represented in only one or two sequences. The NLR architectures in the largest orthogroups reflected the most frequent NB/C-terminal and NB/N-terminal domain associations. For instance, the largest NB-ARC orthogroups included NLRs with TPR repeats, while the largest T-NACHT orthogroups included NLRs with either Ankyrin or WD40 repeats.

Only a subset of all possible domain combinations was observed in Sordariales. Out of the total possible 24 (8 × 3) N-terminal/NB and 12 (3 × 4) NB/C-terminal combinations, only 17 and 7 were observed, respectively. Only 33 of the 96 tripartite domain architectures were identified (S10 Table). Hence, not all tripartite combinations were observed in Sordariales, but the realized proportion of the 96 possible tripartite domain assortments was only marginally lower in Sordariales (39.3%) than previously reported in Dikarya (44.4%) [2,16].

Sordariales NLRs encompass a wide variety of non-canonical domains

In addition to canonical domains, additional non-canonical domains can be found in fungal NLRs. In plants NLRs, non-canonical domains are generally rare, but they may play a crucial role in the immune response by acting as integrated decoys, i.e., as domains targeted by pathogens that were duplicated and fused to NLRs to act as baits to trigger the defense response [59]. We examined the presence and putative function of non-canonical domains in the Sordariales to determine if they could correspond to bona fide integrated decoys or if, on the contrary, they are domains that should be added to the list of canonical domains.

We identified 185 non-canonical domains in 378 Sordariales NLRs (Fig 7A and S11 Table). Non-canonical domains were identified either at the N-terminal or C-terminal regions, either on their own (indicated in Fig 6B), connected with a canonical domain, or connected with another non-canonical domain. Most of these non-canonical domains (170–92%) were found solely in one to three NLRs. The 185 domains were identified in 327 T-NACHT NLRs (10%), 50 NB-ARC NLRs (4%), and one N-NACHT NLR.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Fungal NLRs include non-canonical domains.

(A) Number of NLRs with non-canonical domain annotations, with surface area proportional to the number of NLRs. Names and numbers of NLRs are specified only for the most frequent domains. (B) Examples of three NLR architectures that include a zinc finger domain or a SPRY domain in C-terminal position. The hexagons repeated in tandem in NLR bombom-1_g5425 reflect that the zinc finger motif was repeated three times in the C-terminal position. In the two other NLRs, the zinc finger and SPRY motifs were found only once. These three NLRs are characteristic of architectures containing zinc finger and SPRY domains. (C) Architecture of NLR chafim-1_g9256 of Chaetomium fimeti. This NLR includes a PKinase domain in C-terminal position, which could act as a sensor to detect pathogen-derived effectors.

https://doi.org/10.1371/journal.pgen.1011739.g007

The three most abundant types of non-canonical domains were binding domains of types zinc-finger, SPRY and PKinase (Fig 7A). We identified zinc-finger domains in 57 NLRs, in most cases in the C-terminal position (52 NLRs), and in five cases as the only N-terminal domain. Thirty-four NLRs contained C2H2 zinc-finger type, while 23 NLRs contained treble clef zinc-finger type (17 ZZ, three C1, two FYVE and one TRASH) [60]. Only 16 zinc-finger domains, all of treble clef type, were located next to canonical C-terminal repeats (Fig 7B). A SPRY domain was found in 46 NLRs, always found in the C-terminal position and always associated with Ankyrin repeats (Fig 7B). These analyses indicate that the most frequent non-canonical domains were binding domains, and consistent with their binding activity, they were predominantly located in the C-terminal position.

PKinase domains were the third most common non-canonical domains (12 NLRs; Fig 7A), and nine Pkinase domains represented good candidates for being integrated decoys. PKinase domains were found at the C-terminus of nine T-NACHT NLRs in two orthogroups exclusive to Chaetomiaceae and Podosporaceae. Aside from the PKinase domain, all nine NLRs had a complete canonical tripartite architecture (Fig 7C). We built a gene tree for all NB domain sequences from both orthogroups (S3 Fig). The genealogy of the NB domain revealed that all nine PKinase-including NLRs were monophyletic and thus that all nine PKinase domains may have originated from a single integration event. To test the hypothesis of a single integration event, we compared PKinase sequences of the nine NLRs with all Sordariales proteomes. We observed high similarity between the NLR PKinases and other PKinase domains in single-copy core genes identified as CHK-2 proteins. We calculated BLOSUM62 distance index between all pairs of PKinase domains, both from NLRs and CHK-2 proteins (S12 Table). BLOSUM62 distance was low among CHK-2 sequences (between 0 and 0.29) and NLR sequences (between 0.04 and 0.46), as well as between CHK-2 and NLR sequences (between 0.32 and 0.45). The similarity between PKinase sequences in NLRs and CHK-2 proteins supported the hypothesis that their PKinase domain was integrated from a CHK-2 gene and maintained as a putative integrated decoy.

Extensive variation in NLR repertoire size among Sordariales species

Variation in the number of NLRs in Sordariales was extensive. The variation in the number of NLRs could be as much as sixteen times amongst different species of the order, and between two to thirteen times within a family. The variation in repertoire size was also shown at the broader Dikarya scale, ranging from zero NLRs in Saccharomyces sp. to 602 NLRs in Fibularhizoctonia sp. [2,4]. The variation observed in Sordariaceae, the family of Neurospora relatives, however, allows us to gain insight into the ultimate (i.e., eco-evolutionary) factors underlying NLR diversification. Indeed, dung- or soil-associated Neurospora species harbored, on average, over three times as many NLRs as conidiating, wild fire-associated species of Neurospora (N. crassa, N. discreta, N. perkinsii, N. metzenbergii, N. intermedia, N. hispaniola, N. tetrasperma, and N. sitophila). Differences in repertoire size in Neurospora and hypothetically in the broader Sordariales could be explained by variation in the diversity of mycoparasites on and in plants or soil, assuming NLRs are significant players in fungal innate immunity. To investigate this hypothesis further, a better understanding of the life cycle of Sordariales is required.

NLRs can be organized in clusters in NLR-rich species

Our data also allowed us to gain insight into the proximal (i.e., genomic) factors underlying variation in NLR repertoire size. We discovered that the relationship between NLR content and the stringency of defense mechanisms against genomic repeat elements (e.g., RIP) was weak, indicating that NLRs have diversified by means other than tandem gene duplication. The organization of Sordariales NLRs into clusters, in particular, could be one of the proximal factors explaining the size of the repertoires. Having proposed a definition of NLR clusters that consider the higher gene density of fungi compared to plants, we searched for such clusters in the 37 Sordariales genomes with the longest assembled scaffolds. Most Sordariales genomes presented clusters of NLRs, and the numbers of NLR clusters and NLR genes were highly positively correlated, while the number of NLR clusters was weakly correlated with the strength of RIP. In species that exhibited NLR clusters, approximately 25% (maximum 46%) of NLRs were clustered, although clusters were relatively small, with less than five NLR genes. This trend is in contrast to what is observed in plants where often more than 50% of NLRs are clustered, and these clusters can contain over 10 NLR genes [35,61]. Regardless, clusters could contribute to diversification by promoting unequal crossing over and gene conversion among partially homologous genes, favored by the repetitive nature of the C-terminus of NLRs.

Improved domain annotation in Sordariales uncovers three types of nucleotide-binding domains

Using a combination of Pfam-A and Sordariales-specific HMM profiles, we were able to substantially improve the annotation of nucleotide-binding domains compared to previous studies [2,4], with only 0.4% of NB domains remaining undetermined. By examining the Helical Third section of the nucleotide-binding domains [23], we were able to demonstrate that, similarly to animals, fungi have NACHT domains of both NAIP-like and TLP1-like types, which are referred to as N- and T-NACHT, respectively. Unlike animals, where the rare form of NACHT (N-NACHT) is almost exclusively associated with LRRS, it is exclusively associated with HEAT repeats in fungi. Recently, it was shown that the N-NACHT/HEAT architecture is also present in Rickettsiales [17], an order of obligate intracellular proteobacterial endosymbionts of animals, suggesting a possible origin through horizontal transfer. To explain the patchiness of the distribution of this architecture, the hypothesis of a single ancestral horizontal transfer from Proteobacteria to Opisthokonta, followed by multiple loss events, is more parsimonious than independent transfers in fungi and animals. Further comparative studies with a high level of phylogenetic density should help clarify the origin of the N-NACHT/HEAT architecture.

Variability of NLR domain assortments in Sordariales

The repertoire of domain assortments identified in the Sordariales represents a subset of what is observed in Dikarya [2,4] but with only a small part of tripartite combinations in Dikarya not found in Sordariales. None of the major N-terminal, NB, or C-terminal domains detected in the Dikarya were absent from the Sordariales. Eight types of N-terminal domains were identified in Sordariales, and among the previously reported N-terminal domains [2,4], only minor N-terminal domains such as C2 and Peptidase_S8 were rare or absent in Sordariales. Identifying the NAIP-like type of NACHT also allowed us to demonstrate that HEAT repeats represent a major type of C-terminal domain in fungal NLRs, not merely a sub-type of TPR repeats. These results indicate that previous measurements of NLR repertoire variability at the Dikarya scale were not underestimated due to the limited number of species included per sub-phylum. Our findings also indicate that the assortment of domains in Sordariales is ancient and has mostly been preserved as inherited from their Dikarya ancestor.

Sordariales NLRs harbor putative integrated decoys

We systematically explored non-canonical domains in Sordariales NLRs, which allowed us to identify 185 different non-canonical domains, present in 8% of NLRs. The two most commonly found non-canonical domains were zinc-finger and SPRY, which are domains generally described as involved in binding other proteins, amino acids, or molecules [62–64]. These domains were almost exclusively found in the C-terminal position. Zinc fingers were identified in some plant NLRs, but their function remains unknown so far. Some authors state that they could be integrated decoys [39], while others argue for a nucleic acid-binding property [65] or another undetermined property [66]. In humans, the B30.2 domain, which consists of PRY and SPRY domains, is the binding part of TRIM27 in the NOD2-TRIM27 interaction [67]. In teleost, the NLR-B30.2 subgroup comprises NLR proteins with the B30.2 domain as a constitutive part of their architecture [68,69]. In these NLRs, the B30.2 domain likely functions as the pathogen recognition module and is up-regulated during viral infections [33,68]. Given their binding properties and frequency, we hypothesized that the zinc-finger and SPRY domains may act as sensor domains within Sordariales NLRs, independently or combined with canonical domains. Thus, these domains may constitute two additional canonical C-terminal domains of fungal NLRs.

Among the non-canonical domains, the PKinase domain, which is the third most frequent, represents a potential integrated domain. PKinase domain proteins are known targets of microbial effectors in plants or animals [39,70,71]. There are many examples of PKinase proteins guarded by NLRs [72], and PKinase domains are often found as integrated domains in plants NLRs [39,55,73]. In fungal NLRs, the PKinase domain has been considered canonical in previous studies [2,4], despite being rare. In Sordariales, we identified a PKinase domain in 12 NLRs only. In nine of the 12 NLRs, the PKinase domain was in C-terminal position. These PKinase sequences showed high similarity with a PKinase domain found in a non-NLR single-copy core orthogroup, which includes the PRD-4 gene in N. crassa. PRD-4 is a serine/threonine kinase protein, the ortholog of human checkpoint kinase 2 (CHK2) [74]. CHK2/PRD-4 is involved in multiple pathways in mammals and N. crassa, including DNA damage response and circadian cycles [74–77]. In humans, CHK2 can also interact with viral proteins [78]. These observations suggest that the PKinase domain integrated into Sordariales NLRs might be a decoy for viruses or microbial effectors targeting CHK2/PRD-4.

Concluding remarks

Our study aimed to enhance the understanding of NLR domain assortments in Dikarya in general, and Sordariales in particular. We found that fungi have NACHT domains of both NAIP-like and TLP1-like types, similar to animals. We identified NLR clusters in most taxa, similar to plants, and we provided evidence of an association between the number of NLR and the number of NLR clusters. Thus, our results indicate that fungal NLRs present characteristics shared with fungi’s sister group, the animals, and characteristics reflecting evolutionary convergence with plant NLRs. These findings enhance our understanding of the evolutionary history of fungal NLRs and offer new possibilities for comparative immunology.

Dataset

Methods used for genome assembly, gene prediction, and Busco score were previously described in Hensen et al. [79]. Our dataset included 84 isolates, with two outgroups, eight Podosporaceae, 22 Chaetomiaceae, 39 Sordariaceae, and 13 Lasiosphaeriaceae (sensu Wang et al. [49]) (S1 Table). We chose the definition of Lasiosphaeriaceae from Wang et al. [49], instead of the one presented by Hensen et al. [79], because this former definition is monophyletic and because we had too few (from zero to four) genomes for each of the families represented in Lasiosphaeriaceae sensu lato to make meaningful comparisons with the three other families. Each morphological species was represented by one isolate, except Neurospora discreta, N. crassa, N. tetrasperma and N. sitophila, for which we included one isolate per subspecific category (“category” is a group [80], lineage [81], clade [82], or phylogenetic species [80,83,84]). Orthology among predicted genes was determined using Orthofinder 2.3.3 using MSA for gene tree inference and Diamond 0.9.22 for similarity searches. Genome assemblies, gene models, HMM profiles, orthogroups and NLR sequences are available from Zenodo (https://doi.org/10.5281/zenodo.14204294).

Species tree inference

We used two different approaches to infer a species tree of the 84 isolated. The first approach relied on Astral-III (5.7.5). Only single-copy orthogroups with less than 20% species missing (SCO80) were included in the analysis. We used the 2378 SCO80 and not simply the 1030 single-copy core orthogroups (SCCO) because Astral is non-sensitive to gene tree incompleteness (i.e., to missing genes), while reducing the number of gene trees based on missing data may reduce the phylogenetic signal [85]. SCO80 sequences were aligned at the codon level (i.e., keeping sequences in the coding reading frame) with TranslatorX v1.1 [86] using MAFFT as the aligner and default parameters for Gblocks v0.91b [87]. Sequences with more than 60% gaps or missing data were removed [88], sequences were realigned using TranslatorX, and this was repeated until no sequence had more than 60% gaps or missing data. Only alignments over 100 bp were retained, and the final dataset included 2367 SCO80. Gene trees were inferred in RAxML-ng 0.9.0 using the GTR + G model with 50 starting trees (25 maximum parsimony trees and 25 random trees). Branch support was estimated in the same analysis by calculating 1000 bootstrap trees and applying the MRE-based bootstopping test. We detected outliers using TreeShrink 1.3.7 with the following parameters: -m per-species -b 20 (a sequence is considered an outlier only if removing it reduces the tree size by more than 20%) -q 0.02 (a sequence is considered an outlier only if its potential tree-size reduction value is among the 2% highest for the individual on the dataset). TreeShrink detected outliers in 391 gene trees (17%). Outlier sequences were removed from their corresponding SCO80, re-processed with TranslatorX and submitted to gene tree inference using RAxML-ng. Genes trees were re-rooted using nw_reroot (with default parameters) from Newick Utilities [89]. Low-support branches (bootstrap value <10) were collapsed using nw_ed from Newick utilities (option ‘i & b<10’). Finally, all 2367 re-rooted and collapsed gene trees were used for species tree inference with Astral-III 5.7.5, with default parameters. Branch lengths for the Astral species tree were estimated using the concatenated sequences, aligned at the codon level, of all 1030 SCCO, using RAxML-ng 0.9.0 with the following options: --evaluate --brlen scaled --opt-model on --opt-branches on.

The second approach inferred a maximum-likelihood genealogy from the concatenation of sequences, aligned at the codon level, of all 1030 SCCO. The concatenated sequence represented 1,530,669 bp. The genealogy was inferred in RAxML-ng 0.9.0 using the GTR + G model with 160 starting trees (80 maximum parsimony trees and 80 random trees). Branch confidence was estimated in the same analysis by calculating 100 bootstrap trees.

NLR identification

Orthology analysis

Our initial orthology analysis with Orthofinder identified 47 NLR orthogroups from whole predicted proteomes. To summarize previous Methods sections, we then (1) fixed gene prediction errors, (2) removed non-NLR sequences, (3) added new NLR candidate sequences, (4) split orthogroups based on NB domain type. We then conducted a new orthology analysis by segregating NLRs according to their NB type.

Identification of NLR clusters

To determine whether certain NLRs formed clusters in Sordariales genomes, we first defined what is a fungal NLR cluster. NLR clusters in plants consist of at least two NLRs within a 200 kb window, separated by no more than eight non-NLR genes [35]. However, since fungal genomes are denser in genes than plant genomes, we tested various combinations of two criteria: the maximum distance between two NLRs (0, 0.5, 1, 1.5, 2, 2.5, 3, 3.5, 4, 4.5, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45, 50, 100, 150 or 200 kb) and the maximum number of non-NLR genes between two NLRs (0, 1, 2, 3, 4, 5, 6, 7 or 8). For each of the 37 genomes with an N50 higher than 1Mb, and for each combination of maximum inter-NLR distance and maximum number of inserted non-NLR, we determined the number of NLR genes identified as clustered and assessed the probability this number could be obtained by chance only. To do so, we randomly selected X genes, X being the number of NLRs in a given assembly, and counted how many would be identified as clustered based on the assumed maximum inter-NLR distance and maximum number of inserted non-NLR. We repeated this procedure 1,000 times. We finally compared the number of clustered NLRs to the mean number of clustered random genes. We selected parameter values 40 kb for the maximum distance between NLRs and eight for the maximum number of non-NLR genes as this maximized, for 80% of the assemblies, the difference between the number of clustered NLRs and the mean number of clustered genes.

Association between NLRs and other variables

Phylogenetic generalized least squares (PGLS) analysis is a method to investigate the relationship between variables while accounting for the shared evolutionary history among species. We used phylogenetic generalized least squares analysis to examine the relationships between genome features, the number of NLR genes, and the number of NLR clusters. Genome features were the GC content, the repeat content, the RIP affected area, N50, L50, the number of contigs, the total length of the assembly, and the Busco score. GC-, RIP-, and repeat-related variables were obtained from Hensen et al. [79]. PGLS analyses were conducted using the caper and geiger packages in R.

We also compared the number of NLR genes and the number of NLR clusters to the bimodal distribution of GC content, also obtained from Hensen et al. [79]. Since the bimodal distribution of GC content is a binary parameter, we used the threshold model proposed by Felsenstein [92] implemented in the treshBayes function from the phytools package in R [51]. For each analysis, we ran 2x10⁶ generations, sampling every 1,000 generations, we removed half of the chain as burn-in, and we determined the highest posterior density interval of 95% probability.

Functional annotation of NLRs

Having identified 4613 NLR genes in Sordariales, we functionally annotated them using built-in Pfam-A (version 33.1) HMM profiles, our Sordariales-specific HMM profiles (see previous Methods sections), and HMM profiles from reference [16], referred to as Wojciechowski profiles. HMMER 3.2.1 was run on each of the 82 sets of Sordariales NLRs (hmmsearch -E 1e-4) with all three HMM profiles.

To annotate NB domains, we only used Sordariales-specific HMM profiles, as Wojciechowski profiles do not cover NB domains. Three situations were encountered: (1) for 60 sequences, no NB domain was predicted; (2) for 3909 sequences, one and only one NB domain type was identified; (3) for 644 sequences, two or three different, overlapping, NB domains were predicted. For the 644 sequences with multiple overlapping NB domains predicted, we retained the domain for which the lowest E-value (the “best” annotation) was lower than the squared E-values of other competing predictions, or we considered that the NB domain was present but undetermined.

Functional annotation of other C-terminal and N-terminal domains was performed using the same approach as with the NB domain, using built-in Pfam-A, Sordariales-specific and Wojciechowski HMM profiles. If different overlapping domains were predicted with different HMM profiles, we prioritized predictions with Wojciechowski profiles over predictions with Sordariales-specific profiles, and prediction with Sordariales-specific profiles over predictions with Pfam-A. If different overlapping domains were predicted with the same HMM set of profiles, we retained the domain for which the lowest E-value (the “best” annotation) was lower than the squared E-values of other competing predictions, otherwise all predictions were kept.

We considered canonical the C-terminal domains in categories Ankyrin repeats, TPR, WD40 repeats, or HEAT repeats. We considered canonical the N-terminal domains in categories HeLo-like, sesB-like, Goodbye-like, HET-like, PNP_UDP_1, RelA_SpoT and Patatin. We merged categories HeLo-like domains and sesA, as well as NAD1 and Goodbye because, as noticed in Dyrka et al. [2], we observed extensive overlap between the sesA and HeLo-like annotated sets and the NAD1 and Goodbye annotated sets. We also grouped all amyloid motif domains under the “amyloid” category. All the canonical and amyloid N-terminal domains are listed in S15 Table. All other domains were categorized as non-canonical domains.

HMM profiles for Sordariales T-NACHT, NB-ARC, and N-NACHT domains

Having discovered that the fungal NACHT domain could be subdivided into two sub-types (T-NACHT and N-NACHT; see Results), we built improved versions of our Sordariales-specific HMM profiles. For each type of NB, NLR sequences were aligned using muscle 5.1 (default settings), and only the NB domain was kept. The resulting 2258 T-NACHT, 839 NB-ARC, and 45 N-NACHT sequences were re-aligned using muscle 5.1 (-super5). The alignments were used to build HMM profiles with HMMER 3.2.1 (hmmbuild --amino --informat afa).

NB alignment-free sequences comparison

We relied on alignment-free approaches to compare T-NACHT, NB-ARC, and N-NACHT sequences. We randomly sampled one NB domain per NLR orthogroup. The resulting 325 T-NACHT, 85 NB-ARC, and four N-NACHT were split into two fragments, between the Walker A and Walker B motifs (Fig 3A). Alignment-free distances were computed using the Alfree package [57]. Principal Coordinates Analysis (PCoA) was performed using the cmdscale function in R.

Sequence comparisons between PKinase domains in NLRs and non-NLRs

We identified a set of nine NLRs in Podosporaceae and Chaetomiaceae that all included a C-terminal PKinase domain. To investigate the evolutionary origin of these NLR sequences and test the hypothesis of a single integration of the PKinase domain, we inferred the genealogy of the NB domain of all 40 sequences belonging to the two corresponding NLR orthogroups (11 Podosporaceae sequences, 29 Chaetomiaceae sequences). NB sequences were aligned with MUSCLE 3.8.31, and their genealogy was inferred using RAxML-ng 0.9.0 with the following commands: --all --model GTR + G --bs-trees 100 --tree pars [80],rand [80]. In parallel, PKinase sequences were used for similarity searches using blastp (blast 2.9.0; blastp -max_target_seqs 2000 -threshold 300) against sequences of all Sordariales proteomes. Having identified high similarity between PKinase sequences of the nine NLRs and other PKinase sequences in non-NLR genes belonging to orthogroup OG0001798, we compared PKinase sequences from NLRs and non-NLRs. We aligned sequences in Jalview 2.11.2.5 using MUSCLE 3.8.31 (default settings), and we calculated the BLOSUM62 distance index between pairwise sequences using the DistanceCalculator function from the Python package Bio.Phylo.TreeConstruction.

Identification of N-NACHT NLRs in Fungi

To investigate the distribution of NLRs with an N-NACHT NB domain across the fungal kingdom, we assembled a dataset of 122 freely available proteomes representative of the whole fungal kingdom, except Sordariales, and we selected one proteome per fungal order reported in James et al. [58]. Proteomes were obtained from the NCBI (National Center for Biotechnology Information), the JGI (Joint genome Institute) or the supplementary materials from published articles (S8 Table). The dataset included two Rozellomycota, three Blastocladiomycota, ten Chytridiomycota, one Olpidiales, six Zoopagomycota, eight Mucoromycota, 39 Basidiomycota and 53 Ascomycota. We used Pfam-A and Sordariales-specific HMM profiles in HMMER 3.2.1 (hmmsearch -E 1e-4) to identify NLRs as genes harboring a least a NB domain and a canonical C-terminal domain. Leucine-Rich Repeats were included in C canonical domain for this analysis (Pfam-A HMM profiles: LRR19-TM, LRRC37, LRRC37AB_C, LRRCT, LRRFIP, LRRNT, LRRNT_2, LRR_adjacent, LRR_RI_capping, NEL, TTSSLRR, CENP-F_leu_zip, LRV, LURAP, and LRR_k with k in [1–6,8–12]).