To study fungal intron evolution, we identified 1,161 orthologs among 21 fungal species and 4 outgroups (Figure 1; see Materials and methods). We aligned the amino acid sequences and mapped the corresponding intron positions onto the alignments. There were a total of 7,535 intron positions in 4.15 Megabases of conserved regions of alignment (hereafter 'conserved orthologous regions' (CORs)). Species' intron counts ranged from 0.001 introns per kilobase (kb) in CORs (in S. cerevisiae with 7 total introns) to 6.7 introns per kb (2,737 introns in humans; Figure 1). Figure 2 summarizes the average number of introns per kb of coding sequence versus median intron length. In general, major lineages are clearly separated by intron density. One exception is Ustilago maydis, a basidiomycete fungus that has many fewer introns than other members of its clade. Median intron length is inversely and significantly correlated with the average number of introns per kb (R² = 0.23, P = 1e^-4; Spearman correlation coefficient), although the trend is not significant when the hemiascomycete fungi are excluded (R² = 0.18, P = 0.06). This finding of much longer introns in the very intron-poor hemiascomycetes is intriguing, particularly in light of other peculiarities of evolution in very intron poor lineages 21. In particular, very intron-poor lineages, including hemiascomycetes (see below), have more regular 5' intronic sequences (that is, a stronger consensus sequence at the beginning of introns). Presumably, this conservation of 5' boundaries facilitates intron splicing, in which case increased intron length might be better accommodated. Comparison between other very intron-poor species and more intron-rich relatives should yield insight into the peculiarities of evolution of very intron-poor lineages. Additional data file 4 provides the summary statistics of coding sequence, intron length, and density for the sampled fungal genomes.

Patterns of intron position sharing vary across fungal species. Excluding the extremely intron-poor Hemiascomycota clade, species show between 3.7% and 38.7% species-specific intron positions, while between 32.0% and 76.5% of introns are shared with a species outside of the clade (different colors in Figure 1), and between 20.5% and 60.1% are shared with a non-fungal species. Figure 3 summarizes the pattern of species-specific and shared intron positions across the CORs. Out of 7,535 intron positions, 3,307 are species-specific positions, 1,602 of which are specific to A. thaliana. Of the 501 intron positions shared between plants and animals, from 2.76% in U. maydis to 43.2% in Phanerochaete chrysosporium (Figure 4) are shared with the various fungal species. In all, 60.7% of shared plant-animal positions are also represented in at least one fungal species.

Species within a clade share more intron positions than between clades. Another way to visualize this is using a phylogenetic tree derived from a parsimony analysis where each intron position is a binary character (Additional data file 1). We constructed a phylogenetic tree using Dollo parsimony 2223 from the intron presence absence matrix for the CORs. Dollo parsimony assumes that 0 to 1 transitions (intron gain) can occur only once across the tree for each site, and then infers a minimum number of 1 to 0 transitions (intron loss) to explain each phylogenetic pattern. Surprisingly, our species tree and parsimony tree from the intron position matrix provide nearly the same result, with two exceptions: the unresolved hemiascomycetes, which have few intron presence characters; and the position of U. maydis and S. pombe, presumably due to a high degree of intron loss in those lineages. Previous failed attempts to reconstruct phylogeny by applying parsimony analysis to intron positions experienced a similar phenomenon, with intron poor taxa artificially grouping together 19. As such, it seems possible that intron positions could be good phylogenetic characters in slowly evolving taxa, but will likely encounter problems in cases of widespread intron loss.

We next studied intron loss and gain in fungi in CORs of 1,161 genes. Four previously proposed methods showed very similar pictures, with large numbers of introns present in ancestral genomes and widespread subsequent intron number reduction along various fungal lineages (Figure 5, and tables in additional files 4 and 5). We find that the fungal ancestor was at least as intron rich as any modern fungal species and that the fungus-animal ancestor was 25% more intron-rich than any modern fungus, with at least three-quarters as many introns as modern vertebrates.

Intron number reduction has been a general feature of fungal evolution (Figure 5). We estimate that at least half of the studied fungal lineages (excluding hemiascomycetes) experienced at least 50% more losses than gains, while only between three and six experienced 50% more gains than losses (Figure 5; depending on method used, see Additional file 5). Dramatic intron reduction has occurred within each fungal clade. U. maydis' 0.21 introns per kb represent a 94% reduction in intron number relative to the basidiomycete ancestor; since the ascomycete ancestor (with at least 2.77 introns per kb), hemiascomycetes (0.01-0.07 introns per kb) species have reduced their intron number by at least 94%, S. pombe has reduced its intron number by 81% (0.52 introns per kb), and even relatively intron-rich euascomycete species (0.81-1.16 introns per kb) have undergone a 60% reduction in intron number. Interestingly, following dramatic intron number reduction in the euascomycete ancestor, intron number has remained relatively unchanged within the clade (Figure 5b), consistent with previous results 1524.

On the other hand, our results also attest to ongoing intron gain. Most species have experienced hundreds of intron gains in CORs (although many have subsequently been lost) since the fungal ancestor, and nearly every studied species is estimated to have gained more than one intron per kb since the intron ancestor. Differences in intron gain are sometimes the central determinant of modern differences in intron number. For instance, S. pombe shares as many of the 507 intron positions shared between plants and animals (most of which are likely ancestral) as most euascomycetes; euascomycete species' 50-100% more introns than S. pombe are thus primarily due not to greater retention of ancestral introns but to recent gain. Likewise, Cryptococcus neoformans retains fewer shared plant-animal introns than does Rhizopus oryzae, yet has 70% more introns, apparently due to more intron gain.

Intron evolution within hemiascomycetes provides insights into the evolution of nearly intronless lineages. The extensive loss of introns in hemiascomycetes corresponds to the position in the fungal phylogeny with a significant shift in intron structure. Intron structure in hemiascomycetes requires a six base sequence at the 5' splice site and a seven base pair site at the branching point 25. The other sampled fungi require only a limited intron splice consensus at the 5' splice site and branching point. Previous results have shown that this correspondence between greatly reduced intron number and stronger conservation of intron boundaries across eukaryotes is a general trend 21. Two explanations have been proposed. Irimia et al. 21 suggested that mutations that led to stricter sequence requirements by the spliceosome might be favored in intron-poor but not intron-rich species, in which case widespread intron loss would lead to increased strictness of splicing requirements (and thus intron boundaries). Another possibility 26 is that a shift in splicing mechanism, requiring more extensive conserved sequences at the branch point and 5' splice junction, would create a condition where introns would be more deleterious due to the additional sequence constraint necessary for splicing. In this case, increased strictness of splicing requirements (and thus intron boundaries) would drive intron loss.

Why have all of the introns then not been lost in hemiascomycete species? Some of the S. cerevisiae introns encode functional elements such as small nucleolar RNAs (snRNAs) 27 or promoter elements 28. snRNAs located in the introns of ribosomal proteins are found in orthologous loci of basidiomycetes and ascomycetes (for example, snR39 in RPL7A of S. cerevisiae), indicating their conservation since divergence from the fungal ancestor. However, only 8 of 76 snRNAs are found in the 275 nuclear introns in S. cerevisiae 9. Introns also play a role in regulation of RNA and proteins 29, perhaps through a role in recruiting factors that mediate splicing-dependent export 30. Some of the remaining introns in hemiascomycetes may also provide a necessary role as cis-regulatory containing elements or encoding factors necessary for post-transcriptional regulation, but they may also persist by chance due to low rates of loss.

On the other hand, our results show that hemiascomycete intron positions are not in general widely shared. Only one of the seven intron positions in non-Yarrowia lipolytica hemiascomycete species examined is shared with any species more distant than euascomycetes. However, six of the seven are broadly shared within the hemiascomycete lineage, suggesting either that the remaining introns are very hard to lose or that loss rates have greatly diminished within the lineage. By contrast, 14 of 23 introns present in Y. lipolytica but no other hemiascomycete are shared with a non-euascomycete, and 10 are shared with plants and/or animals; thus, widely shared introns have been preferentially lost among hemiascomycetes after the divergence with the Y. lipolytica ancestor.

Eukaryotic species vary in their numbers of introns by orders of magnitude. These differences have traditionally been attributed to alleged differences in the intensity of selection against introns across eukaryotes 3132. Additionally, it has been proposed that selection against introns could be similar, with differences in population size determining intron number 3334. Under these models, lineages with strong selection against introns (or large population size) should experience low rates of intron gain and high rates of intron loss. Lineages with weaker selection (or smaller population size) should experience more intron gain and less intron loss. Both models thus predict a strong inverse correlation between intron gain and loss rates. However, the data presented here show no clear pattern of inverse correlation (Figure 5).

These results provide an excellent opportunity to compare different previously proposed methods for reconstruction of intron evolution. There are five previously proposed methods. Dollo parsimony assumes a minimal number of changes but that once an intron is lost at a position, it is never regained 22. Roy and Gilbert's method ('RG') 1820 assumes that all intron positions shared between species are representative of retained ancestral introns, while the methods of Csűrös 16 and of Nguyen and coauthors ('NYK') 17 allow multiple intron insertions into the same site, so-called 'parallel insertion'. Carmel and coauthors' 35 method additionally allows for the possibility of heterogeneity of rates of both intron loss and gain across sites.

Previously, application of four methods (Dollo, RG, Csűrös, and NYK) to intron positions in conserved regions of 684 sets of orthologs showed very different pictures of early eukaryotic evolution. Roy and Gilbert estimated the animal-fungus and plant-animal ancestors had some three-fifths as many introns as vertebrates (among the most intron-dense known modern species) 18, while Rogozin and collaborators 19, Csűrös 16, and Nguyen and collaborators 17 all concluded that these ancestors had only half that many introns, and that higher intron densities in plants and vertebrates were due to dramatic increases in intron number. This difference has repeatedly been attributed to overestimation by the RG method 16173637, and the RG estimates have been called 'drastic' and 'generous' 2728. The rationale for this conclusion has been that if a significant number of matching intron positions represent parallel insertion, the RG method will clearly overestimate ancestral intron number.

We used all five methods to reconstruct intron evolution for the current data set. In contrast to the previous discordance, all methods now provide similar estimates for the numbers of introns in the animal-fungus ancestor. Dollo parsimony tended to be very different from the rest of the estimates for deep nodes in the tree. The Carmel and NYK methods show the most striking agreement, with less than 2% difference across all nodes except for the Opisthokont ancestor (3.3% difference). The NYK and Csűrös methods also show striking agreement, giving estimates within 2% of each other for 13 out of 18 (non-hemiascomycetes) nodes, and to within 10% for 17 out of 18. The RG method agreed with the other three methods to within 15% for all nodes except six and was not more than 30% higher than either of the other methods for any node other than the Ascomycete node. Notably, the three nodes on which RG was comparatively highest for the current data set are deep nodes near very long branches in this tree. Thus, further taxonomic sampling would likely bring even these nodes into better agreement (see below). Numbers of intron losses and gains in CORs along each branch were also estimated using all four methods. Though absolute numbers of estimated intron losses and gains along each branch varied more considerably between methods, there was a striking agreement in the relative incidence of intron loss and gain, with Csűrös (2.03 losses per gain), evolutionary reconstruction by expectation-maximization (EREM; 2.14) and NYK (2.12) nearly identical and RG only 21% higher (2.66). Notably, overall estimated numbers of gains were very similar, with only 19 more gains by RG than NYK. Results for all methods are given in Additional data files 4 and 5.

Strikingly, all four methods now estimate that the fungus-animal ancestor had at least 70% as many introns as vertebrates, 15% more than estimated by Roy and Gilbert and more than twice that previously estimated by Csűrös and NYK. Thus, it appears that the previous difference in estimated intron density in the animal-fungal ancestor was not due to overestimation by the RG method, but to a 2.5-fold underestimation by the other methods. Indeed, even the estimates of Roy and Gilbert appear to have been conservative 20.

Why should this be? Following the original authors 20, we suggest that this pattern may be due to unrecognized differences in rates of intron loss across sites. Clear differences in rates of intron loss across sites (that is, different rates of loss for introns at different positions along the same lineage) have been observed over both short 3839 and long 4041 evolutionary timescales; however, three out of four methods fail to take into account such differences in loss rate. Given the recurrent finding of differences in intron loss rates in a variety of studies, it is interesting that Carmel and coauthors' recent work did not find significant differences in rates, and that their method so closely cleaves to the findings of the other methods described here. Clearly, more study into possible differences in rates of evolution across sites, and their effects on current methods, is necessary.

We performed simulations of intron evolution that included variations in intron loss rate across sites, and reconstructed intron loss/gain evolution on each set using four of the five methods (Dollo, RG, Csűrös, EREM). We considered a four-taxa case in which taxa A and B are sisters, and taxa C and D are sisters (Additional data file 2), and in which there were 1,000 introns in CORs in the common ancestor and allowed loss rates to vary between intron positions (Figure 6). In these simulated data sets no parallel gain was allowed to occur.

There are four clear observations, each of which held over all sets of parameters. First, all methods underestimated ancestral intron density. Second, for each data set RG was closest to the real value, followed by EREM, then by Csűrös, then by Dollo parsimony. Third, the Csűrös and EREM methods consistently estimated significant numbers of parallel insertions even though none were included in the simulations - that is, both methods overestimated parallel insertions. Fourth, these trends typically increased with overall branch length. An exception to this was the lack of clear dependency of EREM on branch length.

Together, these observations suggest the following explanation for the discrepancy between previous and current estimates. In the previous data sets 19, the fungi were represented by only S. pombe and S. cerevisiae, both of which have lost the vast majority of their ancestral introns (that is, the fungal branch was very long). Under such long branch conditions, the RG method somewhat underestimated ancestral intron density, while the other methods considerably underestimated intron density and overestimated parallel insertion. In the new data set, the inclusion of fungal species that retain many more of their ancestral introns shortened the fungal branch, leading to a convergence of the four methods on better estimates (and less or no overestimation of parallel gain by NYK and Csűrös).

Indeed, the difference between NYK's estimation of the incidence of parallel gain between the present and previous data sets is striking. According to the NYK method of calculating parallel intron insertions, our data set showed very little evidence for parallel intron gain. Their method estimated 93.08 total parallel gains; thus, only 2.2 % of 4,228 shared introns were due to parallel gain. This is much less than the previous estimate that 18.5% of shared positions in the Rogozin data set were due to parallel gains. This is despite the fact that the overall number of estimated intron gains, as well as the overall number of estimated gains per kb, was higher in our data set than in the Rogozin data set. Thus, it seems that parallel gains were previously overestimated, and given the near identity of results from Csűrös method to NYK's, the same is very likely true of Csűrös' method.

This decrease in the estimated incidence of parallel gain is all the more striking given the increased number of taxa across data sets, which presumably brings with it an increased number of real gains and real parallel gains, although the implications are not entirely clear given that the species present in the current data set are not a superset of the species in the previous set. Our simulations suggest here that there will be countervailing effects of greater taxonomic sampling, with a decrease in the overestimation of parallel gains due to long-branch effects coinciding with an increase in the overall number of true parallel gains. The decrease in estimated incidence of parallel gain seen here implies that currently the former effect dominates; however, with better and better sampling the latter effect may come to dominate in future data sets. More thorough simulation studies will be necessary to more completely understand this issue.

What of other ancestral nodes of key biological interest for which the different methods gave very different estimates? The three methods' previous estimates based on the Rogozin data set also differed significantly for the fungi-animal-plant ancestor and the bilateran ancestor. In the previous data set, both ancestors were flanked by at least one very long branch, suggesting that all methods might have underestimated intron densities. The finding of intron-rich protostomes and apicomplexans would make resolution of this issue possible in the near future. This argument suggests that intron density was very high even in very early eukaryote ancestors.

Annotated genomes of many of the fungi analyzed were obtained from GenBank or directly from sequencing centers and are listed in Additional data file 3. For unannotated genomes, gene predictions were generated using a combination of ab initio and evidence based gene predictions and combined into a single composite gene call with the tool GLEAN 42. The ab initio gene prediction programs SNAP 43, AUGUSTUS 44, and Genezilla 45 were first trained on a set of genes for each genome based on alignments of conserved fungal proteins to the genome using Genewise 46 and Exonerate 47. At the start of this study, high quality annotations of Aspergillus fumigatus, Aspergillus terreus, Coprinus cinereus, Podospora anserina and Rhizopus oryzae were not available so automated annotations were generated so that these species could be included. We generated a new annotation of the v1 P. chrysosporium genome 13 as we found the previously published gene structures were not of sufficient quality based on multiple sequence alignments of the proteins with other fungal proteins. Prediction parameters derived from the closest annotated species were used with at least one round of retraining as previously described 43. Frozen versions of genome sequences, annotations in GFF format, Genome Browser 48 and Web BLAST 49 interface to the genomes, predicted coding sequences and proteins are available for download from the authors' site 50.

The predicted proteins from the 21 fungal genome annotations (Additional data file 3), were combined with the A. thaliana annotations (Feb 2005) 4 available from GenBank and the Fugu rubripes (Ensembl 30.2e, assembly 2) 51, Mus musculus (Ensembl 30.33f) 52, and H. sapiens annotations (Ensembl 30.35c) 1. The longest transcript was used for genes with multiple isoforms. The protein set was masked for low complexity sequences with pseg 53 searched in an all-against-all fashion using FASTP 54 with an expectation value cutoff of 1 × 10^-5. The output was processed with a custom Perl script to generate, for each pair of species, pairwise orthologs via best-mutual-FASTP hits. The pairwise orthologs were combined via single-linkage clustering for all sets of species into multi-way orthologs only if they formed clusters that contained exactly one protein member from each species.

The protein sequences for these orthologs were then aligned using the multiple sequence alignment program MUSCLE 55. The protein alignments were used as a guide to align the genes' coding sequences and intron positions were mapped into both the protein and coding sequence alignments using Perl language modules from BioPerl 56. The 5' and 3' ends of most genes were not alignable and many introns that occurred in these regions, in particular most of the hemiascomycete introns that tend to be within the first few codons of a gene, could not be considered in this study. Alignments of the orthologs are provided as Additional data file 8.

The alignments were evaluated for these intron positions in order to build a matrix of all intron positions. Similar to methodology in previous work 18, each observed intron column in the alignment was classified as to which species shared that intron position. Additionally, an intron position was classified as 'gapped' and removed from the final data matrix if it was within six nucleotides of a column with gaps following methodology from previous studies 19. The aligned data with intron positions inserted are available in Additional data file 7.

A random sampling of 30 of the protein alignments were used to generate a species tree by concatenating the aligned sequences and removing all gap columns from the alignment. The tree was computed and bootstrapped with MrBayes 57. The fungal species tree topology was constrained so that Stagonospora nodorum is basal to the euascomycetes for consistency with more exhaustive phylogenetic methods using larger sampling of taxa 58. Other than this constraint, the phylogenetic reconstruction was consistent with other studies that used a larger sampling of orthologous gene sequences 59.

Dollo parsimony was computed with dollop from the PHYLIP package 60 using default parameters. We generated 1,000 bootstrap replicates with seqboot and Dollo parsimony was recomputed on the replicates. The strict consensus tree was computed from these trees with consense in PHYLIP.

The resulting matrix of classified intron positions was evaluated using the RG method computed along the species tree to compute intron densities, numbers of intron gains and losses, and the fraction of introns present at different internal nodes in the tree. The NYK method was also used to construct intron loss and gain rates and densities in ancestral nodes after modification of the authors' C code. The modified RG Perl code and the NYK C code is available in Additional data file 6. The Csűrös method, implemented in intronRates.jar program 1661, was applied to the data set and allowed to find the optimal number of all-zero unobserved sites. EREM and Dollo parsimony values were computed with the EREM program 3562. The EREM values were computed under a homogenous model. The values reported in Figure 5 represent the maximum likelihood estimate from the EREM program of the numbers of predicted introns and gains and losses. Reconstructed values from all five methods are reported in the tables found in Additional data files 4 and 5. No overall comparisons between methods was made for the 'Crown' node or branches leading from this node as not all methods estimate ancestral density or rates without an outgroup.

We simulated a four-taxa case in which taxa A and B are sisters, and taxa C and D are sisters, and in which there were 1,000 introns in CORs in the common ancestor (Additional data file 2). Different introns were assigned different loss rates as given by a standard gamma distribution, with varying gamma-values. The internal branch was set to length zero (neither intron loss nor gain along the internal branch). External branch lengths were set to be of equal length, with a length chosen for each gamma value such that, on average, a given fraction (70% or 30%) of all introns present at the ancestral node were retained in each descendent taxon. We generated data sets for gamma values from 2.0 (most variation in intron loss rate) to 10.0 (least) in increments of 0.5. No insertion, parallel or otherwise, was assumed. For each set of parameters we generated expected numbers of introns with each phylogenetic distribution, and used these values, rounded to the nearest integer, as inputs for all three methods.