We analyzed the Csl genes in the 17 genomes based on BLAST 39 and HMMER 40 searches (see Methods for details). We found that all predicted Csl proteins contain either the Pfam Cellulose_synt domain (PF03552, 898 aa long) or the glycosyltransferase family 2 (GT2) domain (PF00535, 149 aa long) but not both. For example, all CslA and CslC proteins in Arabidopsis have the GT2 domain but not the Cellulose_synt domain, while all other Arabidopsis Csl proteins contain only the Cellulose_synt domain, which is consistent with a previous finding that CslA and CslC have a different origin than the other Csl families 22. We then grouped all proteins with the GT2 domain into a set denoted as the GT2 dataset, and those with the Cellulose_synt domain into the Cellulose_synt dataset. These two datasets do not share any common proteins.

By querying the Pfam Cellulose_synt domain, we identified two cyanobacterial proteins from the fully sequenced bacterial genomes. These two proteins were originally identified by Nobles et al. in 2001 23, and here are used to root the eukaryotic Cellulose_synt phylogeny.

To characterize the identified Csl genes from the 17 genomes, we built phylogenetic trees for each of the two aforementioned datasets, based on the multiple sequence alignments of both the full length proteins and the conserved Pfam domains (see Methods for details). Figures 1A and 1B show the un-rooted maximum likelihood (ML) trees for the Cellulose_synt dataset and the GT2 dataset, respectively. The number and the species information of the genes included in the phylogeny are given in Table 2. Neighbor joining (NJ) trees for the two datasets were also constructed and given in the Additional file 1 [see Additional file 1].

The number of proteins in each Csl family, calculated based on their groupings as shown in Figure 1, is given in Table 2. No Csl genes were detected in the three stramenopile algae (two diatoms: Thalassiosira pseudonana, Phaeodactylum tricornutum and one brown tide alga: Aureococcus anophagefferens) and the red alga Cyanidioschyzon merolae. Only one single-copy gene, which is most closely related to CslA/C, was identified in each of the six green algae. Overall, 77 proteins from land plants have the GT2 Pfam domain, which are assigned to either the CslA or the CslC family (Figure 1b). Six additional proteins from green algae also have the GT2 Pfam domain and are the most homologous to both the land plant CslA and CslC genes. The Cellulose_synt Pfam domain is found in 211 proteins, which are assigned to the other Csl families (Figure 1a). Based on their taxonomic distribution and the phylogenies shown in Figure 1, we have divided the 294 identified Csl proteins into three categories. The first category consists of CslA and CslC, the most conserved Csl families in our analyses. These families are found in all seven land plants and all six green algae sampled in this study. The second category contains three phylogenetically related families, CesA, CslD and CslF (Figure 1a). Among these families, CslF is specific to grasses while CesA and CslD are present across all the sampled land plants, including the two lower plants (P. patens of bryophytes and S. moellendorffii of lycophytes in Table 1). The last category includes the remaining Csl families, which appear to be confined to the five sampled seed plants (Table 2). Such a restricted distribution suggests a likely late origin of these Csl families and possible functional roles specific to seed plants.

Figure 1 shows all the nine previously known Csl families 14, namely CesA and CslA/B/C/D/E/F/G/H, each forming a well-supported group. A new Csl family was discovered very recently in cereals and named CslJ 2021. Here we found this family not only in sorghum but also in poplar and grape, suggesting that it is another seed plant-specific Csl family in addition to the CslE family. The sorghum sequence (Sb03g047220) was found in both this study and previous papers 2021. According to the phylogeny (Figure 1a; [see Additional file 1]), CslJ is closely related to CslG. We have examined if the CslJ family contains functional genes, knowing that neither Arabidopsis nor rice contains this family (Table 2). The following evidence suggests that genes of this family are not likely to be pseudogenes. First, our Ka/Ks analysis (see Methods for details) shows that members of this family are under negative selection (median of Ka/Ks = 0.23, Figure 2). Second, all members of this family (Table 3) have EST sequences in the NCBI EST database, indicating that these genes are expressed. And third, members of this gene family have also been found in unfinished genomes such as barley, wheat and maize 2021.

Six green algal Csl genes (each green alga contains a single-copy of this gene; see Table 2) form a well-supported group distinct from the land plant CslA and CslC genes (Figure 1b). Inasmuch this green algal Csl group possibly represents the most homologous genes of the common ancestor of CslA and CslC in all the land plants.

The ratio between Ka and Ks (see Methods for details) has been widely used to measure the selection pressure on proteins 41. Generally, a lower selection pressure indicates a higher evolutionary rate. Using Ka, Ks and the Ka/Ks ratio as proxies, people have found significant correlations between protein evolutionary rates and numerous features derivable from genome sequences, such as the number of protein-protein interaction partners, gene expression levels, the essentiality of a gene and the number of paralogs of a gene 424344 (and papers cited therein). In particular, genes that originated relatively recently through gene duplications or by other mechanisms usually evolve more rapidly than the more ancient genes 45. In order to compare the relative evolutionary rates among the Csl families that may have come into being through duplications 46, we have calculated the Ka/Ks ratios (see Methods for details) and conducted rigorous statistical analyses on the computational results.

The comparisons of the distribution of the Ka/Ks ratios across different Csl families are shown in Figure 2a–d, from which it is clear that the narrowly distributed Csl families (CslB/H/E/G/J) tend to have higher Ka/Ks ratios than those widely distributed Csl families (CslA/C/D and CesA). This observation is statistically supported by pair-wise Wilcoxon tests (Table 4a–d). Interestingly, this observation remains to be true when we compared the genes of the Csl families across different groupings of the plant genomes, namely across (a) all the seven land plants (Figure 2a), (b) the five seed plants (excluding moss and spike moss from (a); Figure 2b), (c) the three dicot plants (Figure 2c), (d) the two monocot plants (Figure 2d), and (e) each of the seven plants (data not shown). Specifically, we found that the dicot-specific CslB and CslG families have evolved the most rapidly, followed by the monocot-specific CslH, and then by the seed plant-specific CslE and CslJ. In addition, the monocot-specific CslF has evolved significantly more rapidly than CslC and CesA, but not than CslA and CslD (Table 4d). Overall, we found that the CslB/H/E/G/J families have evolved more rapidly than the other Csl families, which lends further support for the hypothesis that these families might have diversified relatively recently to acquire new functions specific to the seed plants.

We have compared the Csl genes across different plant genomes by inspecting the phylogeny of each family and analyzing the gene structures. We present here a detailed comparative analysis of the less studied and narrowly distributed CslB/H/E/G/J families. The analyses of the other Csl families are given in the Additional file 2 [see Additional file 2].

Our analyses did not detect any CesA and Csl genes in the three stramenopile algal genomes nor in the red algal genome, but this does not necessarily mean that no stramenopiles or red algae contain these families. In fact, a CesA gene has been identified very recently in an yet sequenced red alga Porphyra yezoensis 13. Actually, the cellulose synthases, which assemble into the so-called terminal complex (TC), have been found across all classes of organisms. TCs can be morphologically classified into rosette TCs or linear TCs 474849, for each of which the component CesAs have rather different domain structures 48. While the rosette TCs have been extensively studied and found in all seed plants, the linear TCs may be the most ancient, given their wider distribution across bacteria, fungi, animals and many classes of algae including stramenopiles 4850.

Since the focus of this study is on CesAs that form rosette TCs, we have adopted a rather conservative filtration procedure (see Methods for details), which may have excluded some GT2 proteins that are possibly forming linear TCs. For instance, one GT2 gene from P. patens and seven GT2 genes from S. moellendorfii were removed by our filtration procedure (data not shown) but have been found to be similar to the CesA genes of cyanobacteria and the red alga Porphyra yezoensis 32.

Despite all the previous studies about the distribution of CesAs, very little has been done to identify the other Csl families across different organisms. So what is known about these families is fragmented at the best. For example, a sequence fragment from Chlamydomonas was reported to have diverged early from the CslA and CslC families 22. In our study, we found that all the six sampled chlorophyte green algae have a single-copy gene that is the most homologous to the land plant CslA and CslC families. Our phylogenetic analysis indicates that the CslA and the CslC families might have evolved via a gene duplication event that occurred uniquely in land plants after they split from green algae (Figure 1b).

CslAs and CslCs have been characterized to encode mannan synthases 1516 and xyloglucan synthases 19, respectively. It is tempting to speculate that the mannan synthesis might represent the ancestral function for the single-copy CslA/C-like genes in green algae, since mannan is present in both charophyte algae and chlorophyte algae while xyloglucan is absent from these green algae 5152.

Recent studies have shown that the plant CesA genes and some other Csl families are likely of cyanobacterial origin 2223, possibly as a result of intracellular gene transfer from plastids (or cyanobacterial endosymbionts). In a Csl gene phylogeny built by Nobles and Brown, CslG was suggested to be the first Csl family evolved in plants, followed by CslE, CslB, CesA and CslD 22. However, our search of fully sequenced plant and algal genomes has shown that CslG is among the most narrowly distributed Csl families, and is absent from our sampled algae and lower plants (Table 2). Intuitively, such a distribution pattern suggests that CslG might not be the earliest plant Csl family.

When constructing the phylogeny of Csl families (Figure 6), we included two cyanobacterial CesA protein sequences (YP_322086.1 from Anabaena variabilis ATCC 29413 and NP_487797.1 from Nostoc sp. PCC 7120) as the out-group of the plant CesA and Csl genes. These sequences are the only two hits (see the first section of Results) in our hmmsearch with the Pfam Cellulose_synt domain against all fully sequenced prokaryotic genomes (with E-value cutoff < 1.0). These two cyanobacterial sequences are phylogenetically more closely related to the land plant rosette-TC-forming CesAs than to the other bacterial linear-TC-forming CesAs 22, and thus have been proposed as the progenitor of all the rosette-TC-forming CesAs and Csl genes. The rooted Csl phylogeny (Figure 6) suggests a major early split between CesA/CslD/F and the other Csl families (at node I). Clearly neither CslG nor any of the other Csl families could be considered as the earliest in plants, because the cyanobacterial CesAs are not placed closer to any particular plant Csl family than to the others. Instead, the Csl phylogeny suggests that, after the establishment of plastids (or cyanobacterial endosymbionts) in the ancestral plant, the cyanobacterial derived CesA gene might have undergone several rounds of sequence and function divergence (see nodes I to IV in Figure 6). The evolutionary relationship of CslB/H to the other families is less clear from our phylogeny since the grouping of CslB/H with CslE/G/J has only a modest support (61% bootstrap value).

The finding that CslB/H/E/G/J split earlier than CslD/F (I vs. III) is surprising, because CslD is found in lower plants (moss and spike moss) whereas CslB/H/E/G/J are not. One plausible explanation is that some ancestral genes of the CslB/H/E/G/J families might have been present in the most recent common ancestor of the land plants, but were lost in moss and spike moss later. To test this hypothesis, we performed a TFASTY 53 search of the Csl proteins against the P. patens genomic sequences and the S. moellendorffii genomic sequences (Table 2), respectively, after masking out all predicted genes. Interestingly we found that some short sequence fragments are more similar to CslB/E/G than to the other Csl families, suggesting the possibility that they are the remnants of CslB/H/E/G/J. For example, the P. patens scaffold_21 region from position 2584724 to position 2584398 (in base pair) is 28% identical (51% similar) to the poplar CslB protein eugene3.00141027 from position 7 to position 109 (in amino acid), whose Cellulose_synt pfam domain is from position 9 to position 747; and the S. moellendorffii scaffold_577 region from position 192 to position 741 is 27% identical (52% similar) to the rice CslE protein Os09g30120.1 from position 533 to position 708, whose Cellulose_synt pfam domain is from position 15 to position 737. Similar searches against green algal genomic DNAs also found some homologous regions to the land plant CslB/H/E/G/J genes.

Although the common ancestor of the CslB/H/E/G/J families may have split from the other Csl families during the early evolution of plants, their diversification might have occurred recently. This is supported by the narrow distribution of CslB and CslG in dicots and CslH in monocots. Additionally, we have shown that all these five families have evolved rapidly (Figure 2). Paralogous genes of these families typically form a monophyletic group in our phylogenies and have highly similar gene structures (Figure 3, 4, 5). These findings suggest that recent intragenomic duplications have played a major role in the rapid sequence and functional diversification of these families 46. A simple TBLASTN 39 search of the Csl proteins against the masked plant genomic DNA sequences (with all annotated gene models masked) found many homologous DNA fragments, indicating that they are likely pseudogenic relics after gene or genome duplications. A detailed analysis of these DNA fragments could possibly lead to a deeper understanding about the evolution of these Csl families.

We downloaded the genome, proteome, and gene prediction and annotation data for the 17 genomes from various sources (Table 1). Specifically, the Arabidopsis data were from The Arabidopsis Information Resource (TAIR version 7.0) 54, rice data from The Institute for Genomic Research (TIGR version 5.0) 55, grape data from 56, red algae data from 57, and data for all the others are from Joint Genome Institute (JGI) as of Dec. 2007. The Arabidopsis and rice proteome data include alternatively splicing variants, while all the other proteomes do not. We included these splicing variants of the two organisms in our analyses, but counted them as one gene in our statistics in Table 2. For instance, AtCslA3 has three known splicing variants (AT1G23480.1, AT1G23480.2 and AT1G23480.3); all these three proteins were included in our phylogenetic analyses, although we count them as one single gene in Table 2. We specified these variants in our gene structure figures such as Figures 3, 4, 5. In addition, 597 fully sequenced prokaryotic genomes were downloaded from 58 as of Dec. 2007.

We downloaded known Csl sequences from the Cell Wall Navigator database 5960, including all known Arabidopsis and rice Csl genes as well as sequences from other species in UniProt 61. We took this data set as the initial query to search against the annotated protein sequences of the 17 genomes.

There are two Pfam 40 domain models for the Csl proteins: PF03552 (Cellulose_synt) and PF00535 (Glycos_transf_2 or GT2), both of which were searched in our analyses. In addition, we searched bacterial cellulose catalytic domain model PF03170 (BcsB) but did not find any significant hits (with E-value cutoff < 1.0) in plants and algae. We ran hmmsearch against proteins of the 17 genomes by querying these HMM models in the ls mode (global with respect to query domain and local with respect to hit protein; see details in the manual of HMMER package). We also performed hmmsearch against the 597 prokaryotic genomes.

We have processed the search hits obtained from the above BLAST and HMMER searches in order to build an accurate Csl gene catalog for each genome:

Two datasets were prepared for our phylogenetic analysis: protein sequences that contain the PF00535 (GT2) domain and those that have the PF03552 (Cellulose_synt) domain. Multiple protein sequence alignments (MSAs) were performed on both the full length regions and the conserved Pfam domains for the two datasets. MAFFT 62 was used in these alignments using two highly accurate methods: L-INS-i and E-INS-i. L-INS-i is considered to be the most accurate MSA method 6364, and E-INS-i performs well on sequences with large unalignable regions (see manual of MAFFT). The resulting MSAs were manually edited to remove gaps and ambiguously aligned regions. We have also inspected the MSAs for the presence of the DXD and D, D, D, QXXRW motifs that are characteristic of possessive β-glycosyltransferases 65. The original MSAs, the edited MSAs and the resulting phylogenetic trees are all available in the Additional file 4 [see Additional file 4].

The ProtTest v1.4 package 66 was run on the computed MSAs to select the best-fit models for phylogenetic analyses. We found the combination of JTT+I+G+F models to be the best one for our phylogeny reconstruction. The maximum likelihood (ML) trees were built using PhyML 67, while neighbor-joining (NJ) trees were built using MEGA4 68 considering the above models. Specifically, PhyML analyses were conducted using the JTT model, 100 replicates of bootstraps, an estimated proportion of the invariable sites (I), four rate categories, an estimated gamma distribution parameter (G), and optimized starting BIONJ tree. MEGA analyses were conducted using the JTT substitution model, 500 replicates of bootstrap, pair-wise detection of gaps or missing data, gamma distributed rate among sites and the gamma parameter set at 1.0 (G).

The evolutionary rate of proteins can be estimated by calculating the evolutionary distances that are often measured by calculating the ratio of the number of nonsynonymous substitutions per nonsynonymous site (Ka) and the number of synonymous substitutions per synonymous site (Ks) 43. For each Csl family, in order to obtain the longest possible alignment, we firstly built an initial multiple sequence alignment on the full length Csl protein sequences, and then manually examined the alignment to remove fragmental sequences that introduced long gaps into the alignment (e.g. for those less well annotated genomes, the predicted protein sequences are often of low quality and fragmented). We then rebuilt the multiple sequence alignment on the remaining sequences and reconstructed an un-rooted ML tree.

We transformed the amino acid sequence alignment into codon sequence alignment by using pal2nal 69. The coding sequences were obtained from the downloaded genome data. The maximum likelihood estimation of Ks, Ka and Ka/Ks values for each gene family was conducted by running codeml in PAML 70, using the above tree and the codon alignment as the input. The computation of Ka/Ks ratios for a group of genes based on their phylogeny is conducted under the assumption that each gene evolves at an independent rate. We used this model to compute the Ka/Ks ratio for each gene of each Csl family. Analyses were also performed using the conserved domain regions; the result as shown in Figure 2 remains unchanged (data not shown).

The gene structure information was parsed from the GFF file downloaded along with the genome data, and was used as the input for the graphic display at the Gene Structure Display Server of Peking University 71.