Fifteen new and two existing tetillid cox 1 sequences from four genera were considered in our analysis: Cinachyrella (9 specimens/6 species), Craniella (five specimens/4 species) Paratetilla (one specimen) and Tetilla (two specimens/two species). To avoid bias we only considered sequences that included all the insertion site positions of both the previously discovered and newly discovered tetillid introns. However, since we did not amplify complete cox 1 sequences, introns present at the beginning (the first 108 - 156 bp are missing, depending on the sequence) or at the end of the gene (the last 615 - 1074 bp) would not be detected by our study. The complete cox 1 sequence of the sponge A. queenslandica was used as reference for numbering positions in the alignment. The use of reference sequence allowed a comparison of insertion sites between cox 1 CDSs of different lengths, as well as between partial and complete sequences. Among the 15 new sequences only five were found to possess a longer cox 1 sequence, most probably indicating the presence of introns. Only the Cinachyrella alloclada sample was found to have an 1140 bp long insertion after position 723. This insertion is similar in length and insertion site to those previously found for the intron of C. levantinensis and the second intron of P. onkodes (introns 723, Figure 1). All other insertions are shorter and inserted at different positions of the alignment. Two Cinachyrella sp. samples from Zanzibar as well as the Cinachyrella sp. 3743 sample from Australia (which probably belongs to the same species), have a 985 bp long insertion that is inserted after position 714 (introns 714, Figure 1). This insertion site is similar to that of the first intron of P. onkodes. Finally, the Tetilla radiata sample has a 948 bp long insertion that is inserted after position 870 (intron 870, Figure 1). Interestingly, the introns present in the cnidarians Palythoa sp. and Savalia savaglia are also inserted at this position. A BLASTX analysis 20 of the inserted sequences allowed us to identify a single LAGLIDADG ORF in each insertion.

All inserted sequences were found to fold into the canonical group I intron secondary structure, confirming that the insertions are indeed introns. However, not all introns had the same secondary structure. Three different secondary structures, corresponding to introns 714, 723 and 870, were reconstructed from the five sequences (Figure 2). Introns 723 are represented in sponges by the introns of C. levantinensis, C. alloclada and the second intron of P. onkodes. They are characterized by an absence of P2 as well as large P6 (69-96 bp) and large P9 (113-116 bp) regions (Figure 2b, 12 ). Introns 714 were identified in the three Cinachyrella sp. samples from Zanzibar and Australia. The first intron of P. onkodes was also identified as an intron 714. These introns are characterized by an absence of P2 and P9 as well as the presence of a large loop (> 10 bp) in their P1 structure (Figure 2a). It should be noted that the Cinachyrella sp. and P. onkodes intron 714 structures differ in their P5 region, which is longer in Cinachyrella sp. Finally, an intron 870 was only found in the T. radiata sequence (Figure 2c). This last intron is characterized by the presence of a P2 and a longer P5 than other sponge introns. In all introns most of the LAGLIDAGD ORF is located in the loop associated with the paired region P8. Each LAGLIDADG ORF was found to contain at least one UGA codon. This codon corresponds to a stop codon in the standard genetic code but to tryptophan in the "mold, protozoan, and coelenterate" mitochondrial code (i.e., the demosponge mitochondrial code), refuting the possibility of a nuclear origin for the introns. It is worth noting that most mitochondrial genetic codes could code for a LAGLIDADG sequence. It is thus not possible to determine the origin of the intron based on a genetic code. High sequence similarity exists among the introns inserted at the same position. The percentage of identity ranges from 85 to 99% between two sponge introns inserted at the same position, and from 43 to 53% between introns with different insertion sites.

To identify the evolutionary relationships of the LAGLIDADG sequences encoded within the three intron forms, as well as their putative origin, a phylogenetic tree was reconstructed based on the LAGLIDADG protein sequences encoded within sponge introns together with similar sequences identified by a BLASTP search 21 (see Methods section). The endonuclease ORF encoded within the intron and the intron section that takes part in creating the intron's secondary structure can have different histories 1 . In this analysis, the section of the intron involved in creating its secondary structure was excluded. Consequently, the phylogenetic result allowed us to establish whether the LAGLIDADG sequences encoded within introns of similar structure and insertion sites are closely related. The resulting tree was unrooted (Figure 3). The first intron of P. onkodes was not included in this analysis since it lacked an ORF. The LAGLIDADG sequences of sponges belonged to three unrelated clades corresponding to the different insertion sites and secondary structures, which suggest that the LAGLIDADG ORFs were transmitted with their introns. Interestingly, not only the clade comprising the ORFs of introns 723 (cf. 9 12 13 ) but also the clade comprising the ORFs of introns 870 included cnidarian representatives (i.e., the corals Palythoa sp. and S. savaglia 11 , which had an intron inserted at the same position as T. radiata). The clade comprising the ORFs of introns 714, in contrast, was not found to be related to any cnidarian sequence. Another unrelated animal clade included only cnidarian sequences. The LAGLIDADG sequences of this clade originated from introns inserted at position 888. Goddard et al. 10 studied the co-phylogeny of the LAGLIDADGs encoded within introns 888 and their host.

None of the animal LAGLIDADG sequences were found to be closely related (i.e. similarity below 65%) to any fungi or plant sequence.

To identify whether closely related introns were hosted by closely related sponges, a phylogenetic tree was reconstructed based on cox 1 sequences (Figure 4) using the maximum likelihood (ML) criterion under the TrN + Γ + I model and Bayesian analyses under the CAT + Γ4 site-heterogeneous model 22 . The monophyly of Tetillidae was recovered with maximum support (Bayesian posterior probability: PP = 1.0; and ML bootstrap percentage: BP = 100). In contrast, most of the represented tetillid genera were found to be paraphyletic. More specifically, the representative of Paratetilla was found to be closely related to Cinachyrella species . Similarly, the sequences of Tetilla leptoderma and Craniella sp. 3878 were found to be more closely related to each other than to other representatives of their respective genera. These observations indicate the need for a closer examination of the Tetillidae phylogeny, using additional markers and a larger taxon sampling. Surprisingly, cox 1 sequences possessing introns did not always cluster together. While the three Cinachyrella spp. specimens possessing introns 714 formed a monophyletic clade, C. alloclada and C. levantinensis, which possessed introns 723, belong to distant lineages .

In order to examine whether the sponge introns were transmitted vertically, we checked for co-evolution between cox 1 coding sequences (CDSs) and intron sequences (the sequences included both the LAGLIDADG and the non-coding regions involved in the intron secondary structure) nesting within the cox 1 genes. Such methods are usually used to explore co-speciation between two organisms (e.g., host - parasite relationship). In the case of cox 1 sequences and their introns it is indeed possible to consider the introns as parasites of the cox 1 genes. Since in Tetillidae we have three unrelated "parasitic" introns (714, 723, 870), the history of each intron should be considered separately. Because only a few species were found to possess an intron 714 or an intron 870, our co-speciation analyses were only based on introns 723. The cox 1 and intron 723 sequences of 20 corals and three sponges possessing such introns were considered in these analyses.

Following Goddard et al. 10 , we first conducted tests of competing evolutionary hypotheses. More specifically, using reciprocal approximate unbiased (AU) tests we evaluated, for each data set separately whether the likelihood scores obtained under the intron ML topology (Figure 5a) were significantly different from the likelihood scores obtained under the cox 1 ML topology gene (Figure 5b). Both the cox 1 and the intron data sets reject the H₀ hypothesis that the two genes could share the same topology (p < 0.001).

Because reciprocal AU tests are more adapted when node support is high 10 we also applied a non-parametric version of Huelsenbeck and Bull's likelihood ratio test for detecting conflicting signals 10 23 . Like the AU tests, the LRT test assumes an identical phylogeny for the cox 1 and LAGLIDADG genes. However, in this test both data sets are not considered separately. The test computes the difference between the likelihood scores obtained when both data sets have different topologies ('two trees model') and the likelihood scores obtained when both data sets have the same topology ('one tree model'). This difference is then compared to a null distribution generated by non-parametric bootstrapping (see Methods). Unlike the AU tests, the likelihood ratio test does not reject the H₀ hypothesis that the one-tree hypothesis (both genes have the same topology) is favored over the two-tree model (each gene has a different topology), although marginal significance is observed (p = 0.072; out of the 500 replications performed 36 were found to have a smaller log-likelihood ratio statistic than the original data set, Additional file 2).

The fact that the reciprocal AU tests reject the hypothesis of co-phylogeny and that the LRT test is marginally significant indicates the presence of at least one incongruent node between the LAGLIDADG and cox 1 trees. These results support the hypothesis of horizontal gene transfer of introns. However they do not exclude the possibility that in some clades a vertical transmission is the most likely hypothesis. We therefore also conducted a global test of co-evolution, as well as a test on each host parasite link, using the program ParaFit 24 . In this approach, matrix permutations are used to compare cox 1 and intron patristic distances. The analysis is thus unconstrained by the phylogenetic tree. Unlike the AU and likelihood ratio tests, the null hypothesis in ParaFit is that the host and parasite phylogenies are randomly associated, and thus that both data sets assume a different topology. Hence, a significant p-value indicates the existence of at least one congruent host-parasite link between the LAGLIDADG and cox 1 trees. In such a case, the tests on each host parasite link allow us to identify the co-speciating host-parasite pairs. This test revealed that the global co-speciation parameter, was significant (p < 0.001), albeit low (ParaFitGlobal = 0.0047). Only 3 out of 23 pairwise co-speciation links examined were significant, and again the link values were low (0.0019 < ParaFitLink1 < 0.002, 0.001 < p < 0.006). These three significant links represented the relationships between each of the sponges sampled and their intron sequences. When these three taxa were removed from the analysis, the global test indicated no co-speciation (ParaFitGlobal = 0.00001, p > 0.05). The ParaFit result thus suggests that intron 723 could have been vertically transferred in tetillid and homoscleromorph sponges. However, this result is most probably an artifact due to the fact that only three sponges, belonging to different distant classes, were considered in this analysis. In support of such an idea, it is worth noting that the two tetillid LAGLIDADG proteins are closer to the cnidarian sequences than to the homoscleromorph sequence (PP = 0.98, BP = 80, Figure 3). The latter result contradicts the current view that sponges are either monophyletic 25 or that homoscleromorphs are closer to cnidarians 26 . It seems thus more appropriate to explain the distribution of intron 723 in homoscleromorphs and tetillids by at least two independent transfer events. However, we cannot exclude the possibility that this intron has been vertically transmitted in tetillids.

The cox 1 gene of 15 tetillid sponges was amplified. The origin of the samples is indicated in Additional file 3. The DNA of each sample was isolated using a modified PVP protocol 42 . The cox 1 sequences were amplified using two step nested PCRs. The conditions of PCR amplifications were: 94°C for 2 min; 35 cycles at 94°C for 50 sec, 50°C for 50 sec, 72°C for 4 min; and a final elongation at 72°C for 10 min. Different primer pairs were used depending on the species considered. The list of primer used and their sequences are indicated in Additional file 4. Amplified fragments were directly sequenced on an ABI PRISM 3100 (Applied Biosystems) genetic analyzer. The sequences were submitted to GenBank under accession numbers HM032738-HM032752.

Intron insertion-sites were determined manually, based on nucleotide and amino-acid alignments of cox 1 sequences. The core structure of the cox 1 introns was determined with Citron 43 . Peripheral hairpin structures were predicted with Mfold 44 45 using default setting. Finally, the graphic visualizations of the secondary structures were generated with RnaViz 46 . ORFs were sought for within each intron sequence. For each ORF identified, a BLASTX search 20 was conducted to determine the protein family it belonged to, if any.

Following Rot et al. 12 , a LAGLIDADG dataset was constructed using BLASTP searches. Each of the LAGLIDADG sequences found in sponge introns was used as query. This dataset included mostly fungi, as well as plants, cnidarians and sponges. GenBank accession numbers are indicated in Figure 3. The protein sequence alignment was conducted with the L-INS-i algorithm under the JTT-200 substitution matrix, as implemented in Mafft version 6 47 . Due to the high variability of the LAGLIDADG sequences, the use of Gblocks 48 or Soap 49 to remove poorly aligned region of the alignment, resulted in matrices less than 100 amino-acids (aa) long. The LAGLIDADG data set was therefore corrected manually. Sections of the alignment with more than one third of missing data were removed. Additionally, the 5' region of the LAGLIDADG ORF which takes part in the folding of the intron (i.e., the first 63 aa positions of the LAGLIDADG alignment that correspond to the largest such region) were removed as well. The final LAGLIDADG data set included 217 characters; 3 of which were constant (Additional file 5). All of the 214 variable characters were phylogenetically informative. The reconstruction of the tree was conducted with RAxML 7.0.4 50 with 100 bootstrap repeats, under the CAT + Γ + I, and with Phylobayes 3.2 51 under the CAT model. The Phylobayes analysis included two chains with 18,100 cycles (3,400,000 generations) while 4,500 cycles were discarded as burnin. The maxdiff value of this run was 0.069.

The cox 1 data set included the 15 DNA sequences obtained as well as two tetillid sequences available in GenBank. Since Astrophorida has been shown to be the sister clade of Spirophorida (e.g. 38 39 ), four astrophorid sequences were used as outgroup. Accession numbers are indicated in Figure 4. A codon alignment of the cox 1 sequences was obtained using the online version of Pal2Nal 52 . The underlying protein reference data set was aligned using MAFFT version 6 47 with the L-INS-i algorithm. The program Gblocks 48 was then used to exclude regions that were poorly aligned. The nucleotides downstream to the intron insertion site (i.e., 18 nucleotides after each insertion) were also removed since co-conversion of cox 1 exonic sequences can occur after intron insertion 53 . The cox 1 data set used in the phylogenetic analysis included 927 characters; 687 of them were constant and 303 had missing data. Among the 240 variable sites 190 were phylogenetically informative (Additional file 6). Phylogenetic reconstruction was conducted using both the maximum likelihood (ML) and the Bayesian approaches. The ML tree was reconstructed with PAUP* 4 54 under the TrN + Γ + I model of sequence evolution and using the tree bisection reconnection (TBR) branch-swapping algorithm and 100 random sequence addition starting trees. The TrN + Γ + I model was found to be the best fitting ML model using Modeltest 3.7 55 . Branch supports were estimated based on 100 bootstrap repetitions. A Bayesian tree was reconstructed with Phylobayes 3.2c 51 under the GTR CAT model of sequence evolution. The analysis included two chains with a total run length of 13,000 cycles (490,000 generations) while 3200 cycles were discarded as burnin. The maxdiff value of this run was 0.038.

Three different introns were found in sponges (see Results section). However, co-speciation tests, between intron sequences and their cox 1 host sequence, could only be performed for intron 723. Other introns did not include enough representatives, except intron 888, which was already studied by Goddard et al. 10 .

A total of 20 species of corals and three species of sponges possessing intron 723 were considered. The nucleotide data sets of each gene were aligned with MAFFT version 6 47 with the L-INS-i algorithm. After manual removal of ambiguously aligned positions the cox 1 and intron data sets were respectively 630 and 1078 bp long (Additional files 7, 8. The program PAUP* 4 54 was used to reconstruct the evolutionary relationships based on cox 1 and intron sequences, and to obtain the corresponding matrices of patristic distances. The two ML phylogenetic trees were obtained using the TBR branch-swapping algorithm under the best model of sequence evolution identified by Modeltest 3.7 55 .

Three co-speciation tests were conducted. Following Goddard et al. 10 the first approach was to compare the ML topology of both genes under each data set. Reciprocal approximate unbiased tests as implemented in Consel 0.1i 56 were applied. The site-wise log-likelihoods were obtained using PAUP* 4 54 . The second approach was to perform a non parametric version of Huelsenbeck and Bull's likelihood ratio test 23 . This test compares the log-likelihood L₀ obtained under the hypothesis that the cox 1 gene and the intron share a single tree (each marker being allowed to evolve under different model and branch-length parameters), with the log-likelihood L₁ obtained under the hypothesis that each data set evolved independently, resulting in different trees and sets of parameters. Computations were performed with PAUP* under the GTR + Γ + I model for both gene. The log-likelihood ratio statistic (d) is defined as:

d = 2 ( ln L 0 − ln L 1 )

The null distribution of d was obtained using non-parametric bootstrap. Five hundred bootstrap files were generated separately for the cox 1 and the intron data sets using Mesquite 2.72 57 . Each cox 1 bootstrap file was concatenated with one LAGLIDADG bootstrap file. The cox 1 and intron sequences were then randomized within each concatenated file using a Perl script designed for this purpose. Three ML trees were constructed for each data set. The first was based on the first 630 positions of the concatenated and randomized matrix, the second on the last 1078 positions and the third on the total matrix length (1708 bp). The tree reconstructions were performed with PAUP* using rounds of heuristic searches starting with a neighbor-joining (NJ) tree and using tree bisection-reconnection (TBR) branch-swapping. The initial model parameter values were those estimated by Modeltest. After a first round of heuristic search the parameters were estimated on the resulting tree and then used for the subsequent round of heuristic search. The process was repeated until all parameters converged. Ln L₀ was computed by assuming that the first 630 bp and the last 1078 bp shared the ML tree obtained using the whole data set (albeit each partition was allowed to evolve under different model and branch-length parameters). Ln L₁ was computed by assuming that the first 630 bp and the last 1078 bp had different trees and different models of evolution. The third approach applied the program ParaFit 24 to perform a global test of co-speciation (ParaFitGlobal) as well as local tests for each cox 1 - intron link. Following the recommendation of the ParaFit manual, the only parameter considered was ParaFitLink1. This parameter is indeed more adapted when the global test is significant but the local tests show a mixed trend. The principle coordinates (PCOs) of the patristic matrices were calculated with DistPCoA 58 using the Lingoes correction.