Raw Roche 454 GS-FLX Titanium reads were quality and adapter trimmed and size selected to yield 681,722 cleaned reads with a mean length of 372.6 bp and 254 Mbp of total sequence data (Table 1, Figure 1A). Reads were first assembled in MIRA v3.0rc4 55 and the resulting assembly was passed through a second assembly step in CAP3 56 to join additional contigs (Table 2). The resulting final assembly consisted of 56,256 unique sequences (i.e. retained singletons plus primary and secondary contigs, hereafter referred to as unigenes; Additional file 1). Unigenes had a mean length of 547.2 bp and summed to a total assembly length of 30.79 Mbp (Table 1, Figure 1B). The average read-depth coverage for the final unigene assembly was 7.0× (Table 2). The distribution of unigene coverage was highly left-skewed toward low coverage with an extremely long tail (maximum coverage was 2,078×; Figure 1C, D). The steep decline in read-depth coverage suggests that cDNA normalization was effective and is typical for a normalized library 15 .

Because information about the actual size and composition of the transcriptome is often unknown, we utilized a simulation-based tool, ESTcalc 57 , to estimate the expected depth and breadth of transcriptome coverage for this data set. The model for transcriptome coverage backing ESTcalc was parameterized using the well-characterized Arabidopsis thaliana transcriptome and several "next-generation" sequencing runs using normalized and non-normalized cDNA libraries 57 . Using the results from these simulations (retrieved using ESTcalc), our dataset is expected to cover 87% of the nucleotide positions in the transcriptome (Table 3), with every gene represented by at least one read (i.e. percent of genes tagged).

Additionally, 70% of the genes are predicted to be sequenced to 90% of their length. Consistent with these estimates, we were able to identify 333 of 357 (93.3%) Arabidopsis genes that are conserved as single copy genes across all Eukaryotes (i.e. ultra-conserved orthologs; UCOs 58 ). Similarly, we detected 754 of 959 (78.6%) shared single copy tribes from Arabidopsis thaliana, Populus trichocarpa, Vitis vinifera, and Oryza sativa in our classification of unigenes in the PlantTribes database 59 60 ). These two gene sets (UCOs and shared single copy tribes) represent a highly conserved subset of genes expected to be present in eukaryotic and plant genomes, respectively, and can be used as a proxy for gene detection and sampling breadth. As a final measure of gene detection in this data set, we utilized a bootstrapped data resampling approach using the distribution of reads in our final assembly (see methods section) to generate a unigene accumulation curve which plots the number of unigenes detected as a function of sequencing effort (Figure 2). Using this method, we estimate that on average 90%, 95% and 99% of the unigenes were tagged after approximately 378,683; 455,145; and 543,727 reads were sampled (Figure 2). On average, it took 59 reads to detect each of the last ten unigenes.

To identify potential contaminant sequences in the sample or sequencing library, we examined the taxonomic distribution of blastx hits for each unigene searched in the NCBI nr database. We examined both the taxonomic classification of the best hit as well as the lowest common ancestor (LCA) assignment for each unigene using MEGAN v.3.7.2 61 . 34,740 unigenes had a positive a blast hit, of which only 1.8% had a best hit to an organism outside of the green plants and 1.1% received an LCA assigned taxon which is not within, or a super set of land plants (Table 4). We also examined the unigene set for potential genomic DNA contamination by screening unigenes for blastn hits to the complete chloroplast genome sequence of Pteridium aquilinum (HM535629 62 ). None of the chloroplast sequences identified in the transcriptome were longer than 3.5 kb or contained more than five adjacent genes (most spanned only a single gene) and thus can reasonably be considered putative transcripts 63 64 . That we did not detect any long fragments of chloroplast DNA in the transcriptome assembly gives us confidence that our DNase treatment during RNA extraction was effective and the resulting cDNA library used in sequencing is free of contaminant genomic DNA.

Unigenes were annotated with gene ontology (GO) terms, enzyme codes, and conserved protein domain functions using the Blast2GO suite 65 66 67 . Unigenes were first interrogated against the NCBI nr protein database using a blastx e-value threshold of 1e-10, keeping the top 10 high scoring alignments, resulting in 34,740 unigenes (61.8%) with positive blast hits. The best blastx hits for these unigenes corresponded to 22,596 unique protein accessions in the nr database. The longest open reading frame (uncorrected six-frame translations automated in Blast2GO) from 29,357 unigenes (52.2%) had positive matches to conserved protein domains using InterProScan (IPS) searches implemented in Blast2GO. These results (nr blastx and IPS) were used to assign 87,137 GO terms to 25,999 unigenes (Additional file 2). These GO terms were used to map 11,243 enzyme codes to 8,993 unigenes. Enzyme codes were then used then to retrieve and color 144 KEGG pathway maps. To assess whether the frequency of functional categories present in the Pteridium transcriptome differ significantly from the suite of gene functions present in other plants, we compared the GO terms assigned to Pteridium unigenes with the GO classification for all genes in the Arabidopsis thalliana genome (TAIR9 GO annotation downloaded on 13 September 2010) using a two-tailed FDR-corrected Fisher's exact test. When using the full GO classification, none of the GO terms in Pteridium are significantly enriched or underrepresented relative to the full GO classification for Arabidopsis (FDR-corrected alpha = 0.05). To examine broad-level classification of gene functions in the bracken transcriptome, we mapped GO terms to the GO-slim vocabulary (Figure 3) and repeated the Fisher's exact test. 42 GO-slim categories were overrepresented and 88 categories were underrepresented in the Pteridium transcriptome relative to the Arabidopsis thalliana GO-slim annotation (FDR-corrected alpha = 0.05; Additional file 3).

Unigenes were classified into 6,987 tribe (inflation level 3) and 9,395 orthogroup MCL clusters (Additional file 4) in the PlantTribes gene family database on the basis of the best blastx hit to the inferred protein models of ten complete plant genomes included in an updated version of the PlantTribes database ( 60 and CWD, unpublished). To evaluate the level of gene overlap between the Pteridium gametophyte transcriptome and other land plants, we examined overlap in both PlantTribes orthogroup cluster membership and blastx hits for predicted proteins in Physcomitrella patens, Selaginella moellendorffii, and Arabidopsis thaliana (Figure 4). Among genes in the Arabidopsis genome with positive blastx hits with Pteridium unigenes, we examined for the presence of "gametophyte genes" previously identified in the literature. Honys and Twell 68 used microarrays to identify 1,355 genes specifically expressed in haploid male gametophyte tissues in Arabidopsis, that is, genes consistently expressed in at least one of four male gametophyte developmental stages and absent in six sporophytic tissue gene expression profiles. Similarly, Yu et al. 69 and Wuest et al. 70 identified 911 genes (combined) that were significantly over-expressed in female gametophytic cells relative to sporophytic tissues. In total, we identified 1,156 known Arabidopsis gametophyte genes that produced significant alignments with Pteridium unigenes in our blastx search (Figure 5).

A total of 2,679 perfect di-, tri-, tetra-, and pentanucleotide simple sequence repeats (SSRs) longer than 9, 8, 6, and 5 repeats, respectively, were identified in 2,285 unigenes (Additional file 5) using msatCommander 71 . Sufficient flanking sequences existed to design high quality primers for 548 potentially amplifiable SSR loci. PCR primers were chosen using Primer3 72 as implemented in msatCommander 71 (Additional file 6). Since this RNA was extracted from gametophytes derived from spores collected from a single diploid sporophyte, we are unable to determine the level of variation present at these SSR loci in natural populations.

To identify and classify expressed repeat sequences, we screened the Pteridium unigenes with RepeatMasker, using RepBase sequences belonging to land plants (Embryophyta). In total, 416 retrotransposons were identified, representing 0.17% of the total unigene sequence length (Table 5). Additionally, 269 DNA transposons were identified, representing 0.07% of the total sequence length (Table 5).

Because contaminant adapter/primer sequences, polyA/T repeats, and low complexity end sequences can substantially compromise de novo assembly and can be difficult to completely remove (KM Dlugosch, personal communication), we aggressively filtered and trimmed the reads beyond the default instrument-level processing routines at the cost of sequence information loss (approximately 11.6 Mbp were removed, representing 4.4% of the filter-passed data).

Considering the sheer quantity and depth of sequencing produced by next-generation sequencing platforms, we deemed this an acceptable level of loss to improve accuracy in the assembly. We also used a two-step assembly strategy to minimize redundancy in our final unigene sequence set. We adopted this approach because MIRA is able to handle the large number of reads produced by 454 sequencing and utilizes a multi-pass strategy to identify and correct sequencing and assembly errors to produce a highly accurate assembly, but is sensitive to uneven sequencing depth of coverage and allelic diversity, resulting in a high number of redundant contigs. CAP3 is a proven and efficient DNA sequence assembler that can be used to join highly similar overlapping sequences, but is unable to handle the large number of reads produced by new high-throughput sequencing platforms. By combining these two assembly tools, we were able to join contigs and singletons that failed to assemble in MIRA to reduce sequence-level redundancy in our final unigene sequence set.

In examining the taxonomic distribution of nr blastx hits for the unigenes, we identified only a small proportion of sequences with best blast hits or LCA assignments outside of the green plants. When we examine these hits in greater detail, we find that many of them only align to short conserved domains, are hypothetical proteins of unknown function from model organisms, or are genes which are conserved across broad taxonomic levels, such as cytochrome P450, alpha-tubulin and dynein proteins. Additionally, because no other fern genomes have been sequenced, some of these sequences may represent novel fern genes. Thus, the evidence indicates that there is very little heterospecific sequence contamination in these data.

While the simulations that underlay ESTcalc are based on the well characterized Arabidopsis thaliana floral transcriptome (approximately 18,000 genes with transcripts averaging 1,500 bp long) and assume perfect cDNA normalization and sequence assembly, Wall et al. 57 show that their results were highly predictive for empirical datasets from diverse eukaryotic species and tissues, making their simulations useful as a null model for predicting transcriptome coverage in other organisms. The predictions for transcriptome coverage produced by ESTcalc are largely consistent with that observed in our two-step assembly. However, the larger assembly size, greater number of unigenes, and shorter unigene lengths observed in our data set relative to the ESTcalc prediction may be explained by imperfect cDNA normalization or inefficient de novo assembly. Additionally, it is also becoming evident that with increased transcriptome sequencing throughput, it is possible to capture a richer, more nuanced, picture of transcriptome complexity (e. g. partially processed transcripts and alternative splice forms) 73 74 75 . This increased information content, however, presents significant challenges for de novo assembly and often results in a fragmented or partially redundant assembly 76 77 . Also consistent with the ESTcalc estimate that we have tagged all of the transcripts present in this sample, our unigene accumulation curve shows that the rate of new unigene detection for this cDNA library has declined to the point that additional sequencing is unlikely to detect new genes, but may however serve to condense and join non-overlapping contigs in our assembly. Similar approaches to evaluate sufficient sampling in transcriptome projects have been used by other researchers when other information about the transcriptome is absent 15 78 .

The GO functional categories represented in the Pteridium gametophyte transcriptome are not significantly different from the suite of functional categories present in the full Arabidopsis genome GO annotation. Most of the unigenes annotated with a cellular component are localized to plastids or mitochondria, but a large number of them are also targeted for ribosomes or the plasma membrane (Figure 3). The molecular function of unigenes is heavily dominated by binding nucleic acids or proteins and metabolic activity, including hydrolase and kinase activity (Figure 3). The biological processes represented include all of the major cellular processes from transport and cellular organization to transcription, translation, and metabolism (Figure 3).

Visual examination of annotated/colored KEGG maps (not shown) indicates that we have captured all of the genes required for glycolysis, the citrate cycle, and plant hormone biosynthesis including gibberellin, abscisic acid, strigolactone, cytokinin, brassinosteroid, and auxin. We also detected Enzyme code signatures for most of the genes involved in nucleic and amino acid metabolism and chlorophyll biosynthesis.

The PlantTribes database contains an objective MCL cluster-based classification system for plant genes and gene families 60 79 80 . By identifying similar sequences in this classification system, we assigned unigene sequences with putative gene family identities. The most abundant of these gene families present in the unigene set was the pentatricopeptide repeat protein (PPR) family, with over 600 unigenes classified as PPR proteins. We were also able to identify 65 unigenes classified in the MADS-box transcription factor family. Using this classification, gene sequences from Pteridium can be extracted for gene families of interest for use in studies of gene family evolution or phylogenomics. The overlap in orthogroup membership and blast hits for proteins in Arabidopsis thaliana, Selaginella moellendorffii, and Physcomitrella patens is similar (Figure 4A/B), but some striking differences can be observed. In both the PlantTribes and blast-based Venn diagrams, most of the unigenes which were identified in Arabidopsis, Selaginella, or Physcomitrella are also shared across all three species. In the PlantTribes classification, most of the genes are shared with Arabidopsis (21,649 unigenes), the species in this comparison that shares the most recent common ancestor with Pteridium, while slightly fewer and approximately equal gene representation is shared with Selaginella and Physcomitrella (19,649 and 19,485 unigenes, respectively). This is in contrast to the blast-based examination of gene set overlap in which Arabidopsis again has the greatest number of unigenes with hits (23,148 unigenes), but Physcomitrella has hits with 6,122 more unigenes than Selaginella (Figure 4). At first this seems counterintuitive because Pteridium shares a more recent common ancestor with Selaginella than with Physcomitrella. This pattern may be explained by the expression of "gametophyte genes" in Pteridium that are conserved with genes in the Physcomitrella genome, however little is known about the expression and evolution of genes between sporophyte and gametophyte stages. Both Physcomitrella and Pteridium have maintained a homosporous life cycle with a large independent gametophyte stage. These life history differences may also play a role on the selective pressures and/or constraints influencing gene evolution and more work is needed to address these hypotheses.

In our examination of Arabidopsis gametophyte genes, we identified gametophyte-expressed homologs in Pteridium for over half (52.7%) of the previously characterized Arabidopsis gametophyte specific or enriched genes. This finding suggests that a highly specific suite of genes required for gamete production and syngamy may be conserved over long periods of evolutionary time, despite substantial differences in life cycle and reproductive strategies between Pteridium and Arabidopsis. It should be noted also that these conserved genes are not the genes required for meiosis because the tissues sampled in this study and those used to identify gametophyte specific genes in Arabidopsis were all post-meiotic. An in-depth study of sporophytic genes in Pteridium is needed to better understand the evolution and expression of genes between the sporophyte and gametophyte stages.

Pteridium aquilinum ssp. aquilinum spores (collection number: Wolf 83; sourced from a single sporophyte individual collected in Norwich, UK) were sown onto sterile agar nutrient media containing Bold's macronutrients and Nitch's micronutrients (prepared as described by 81 ) and grown under white light. Whole gametophytes including both vegetative and sexually mature male, female, and hermaphroditic individuals of various ages (up to 9 months from germination) were flash frozen in liquid nitrogen and ground to a fine powder. Total RNA was isolated using the Sigma Spectrum Plant Total RNA Kit, incorporating on-column DNase I (Qiagen) digestion during extraction to remove traces of genomic DNA. Total RNA was concentrated by precipitating in 2.5 M ammonium acetate and 70% ethanol, then resolubilizing the RNA pellet in RNase-free water to approximately 500 ng/μL. Total RNA was quantified and its quality verified using an Agilent Bioanalyzer 2100. Total RNA was sent to the Center for Genomics and Bioinformatics at Indiana University, Bloomington (IU CGB), where a normalized transcriptome (cDNA) library optimized for Roche 454 GS-FLX Titanium sequencing was prepared 82 .

Briefly, full-length enriched cDNA was synthesized with the CloneTech SMART cDNA synthesis kit using modified 454-ready adapter/primer oligos (K Mockaitis, unpublished). The frequency of abundant cDNA species was reduced using the Evrogen Direct Trimmer normalization kit. Normalized cDNA was fragmented by sonication, blunt end repaired, and ligated to custom 454 sequencing adapters. Amplification of the sequencing library incorporated an adapter-mediated PCR suppression effect to preferentially amplify ligation products suitable for 454 sequencing 17 . The final transcriptome library was size selected and 454 sequencing proceeded according to the manufacturers recommended protocol on 3 regions of a four region PicoTiter plate.

Sequence reads generated in this study were deposited in the NCBI sequence read archive (SRA012887). Raw sequence reads that passed instrument software quality filters were trimmed of custom oligonucleotide adapter sequences (Justin Choi, unpublished, IU CGB). The resulting sequences were further processed with SeqClean 83 and SnoWhite v1.0.3 84 to remove low quality, short, and contaminant sequences, and to aggressively trim polyA/T sequences. Cleaned reads were assembled de novo in MIRA v3rc4 55 85 using a minimum percent identity of 94% to align reads, retaining singleton reads in the assembly (-OUT:sssip = yes). This primary assembly was passed through a secondary assembly step in CAP3 (95% identity, 25 bp overlap) 56 to reduce redundancy in the final assembly and join additional contigs. Custom perl scripts were used to extract summary information about the reads and assemblies (Tables 1 and 2; Additional file 7).

We utilized a web-based tool, ESTcalc 57 , to estimate the predicted level of transcriptome coverage for our data set. Input parameters to ESTcalc require that we specify the sequencing technology used (or a combination of technologies) and either the total sequencing level (Mbp), or the number of reads and an estimate of read lengths. We used the best approximation for sequencing technology available (454 GS-FLX) and the empirical values observed for the cleaned sequence data (254 Mbp or 681,722 reads with an average of 372.6 bp/read) to obtain our estimates. The estimates reported were identical whether we parameterized on total sequence or supplied read length information as well.

To determine the number of eukaryotic ultra conserved orthologs (UCOs 58 ) we captured in the Pteridium transcriptome data set, we queried a list of 357 UCO coding sequences from Arabidopsis (sequences available at: http://compgenomics.ucdavis.edu/compositae_reference.php) into the unigene set with an e-value threshold of 1e-10 using NCBI tblastx. These blast results were then parsed to determine then number of UCOs with a positive hit that returned an amino acid alignment greater than 30 residues long.

We assessed the changing rate of new gene detection as a function of sampling effort (unigene accumulation curve, Figure 2) using a bootstrapped random sampling protocol implemented in a custom perl script (Additional file 8). This script uses the empirical distribution of read number per unigene in our final assembly to randomly sample reads one at a time and tracks the total number of unigenes detected at each step. Because the order of sampling can impact the shape of this curve, we computed 1,000 replicate random sample orders and calculated the mean number of unigenes detected after each draw. To evaluate the level of variation in the number of unigenes detected, we also calculated the 95% confidence interval on the number of unigenes. Using this curve, we then estimated the number of reads it took to capture an average of 90%, 95%, and 99% of the unigenes and the average number of additional reads required to detect each of the last 10 unigenes.

To evaluate for the presence of potential contaminating sequences, we examined the taxonomic distribution of blastx hits for the unigene set in the NCBI nr protein database using an e-value threshold of 1e-10. The top 10 blast hits for each unigene were kept and examined in MEGAN v.3.7.2 61 . MEGAN is a tool built for the examination of metagenomic data sets and provides a number of useful functions to explore the information content of large blast results. The blast results for each unigene were mapped onto the NCBI taxonomy tree by examining just the best hit (lowest e-value) or by using the lowest common ancestor (LCA) algorithm 61 . LCA was determined using at least three blast hits with a bitscore greater than 75 and within 10% of the top bitscore for that unigene.

The same blast search used to examine the taxonomic distribution of blast hits was used to identify putative homologous proteins and annotate each sequence with gene ontology (GO) terms using Blast2GO 65 66 67 . Blast2GO was also used to automatically handle InterProScan (IPS) searches to identify conserved protein domains in translations of the longest ORF in each unigene. Any GO terms associated with IPS hits were then merged into the blast-based GO annotation. GO terms were then used to map enzyme codes to each sequence. Enzyme codes were then used to automatically color and retrieve KEGG pathway maps 86 87 . As a final step in examining a broad functional representation of the gametophyte transcriptome, GO terms were mapped to the reduced GO-slim ontology and visualized and explored with directed acyclic graphs (not shown) and summarized with filtered pie charts including GO categories represented by at least 150 sequences (Figure 3)

Unigenes were classified into tribe- and orthoMCL clusters in the PlantTribes2.0 database using a custom Perl pipeline (dePamphilis lab, unpublished) which queries each unigene against the complete inferred protein set from ten plant species that have complete sequenced genomes, using a blastx e-value threshold of 1e-10. Unigenes were assigned to MCL clusters based on the best blast hit. Species used for blast searches and gene clustering in the PlantTribes2.0 database include: Chlamydomonas reinhardtii v3.0, Physcomitrella patens v1.1, Selaginella moellendorffii v1.0, Oryza sativa v5.0, Sorghum bicolor v1.0, Vitis vinifera v1.0, Populus trichocarpa v1.0, Medicago truncatula v1.0, Carica papaya v1.0, and Arabidopsis thaliana (TAIR7). Meta-information about each assigned cluster was extracted from the database for each unigene and was output to a file. Simple text-based searches examined this information to retrieve gene family names and putative gene family functional data. The shared single copy tribes 59 for Arabidopsis, Vitis, Populus, and Oryza were identified in the PlantTribes2.0 database and the number of these tribes detected in the unigene set was determined by examining the pipeline output file. Orthogroup assignments for Pteridium unigenes were examined for cluster membership by Selaginella, Physcomitrella, and Arabidopsis to generate a Venn diagram showing putative gene level overlap (Figure 4A). Unigenes were also directly queried against each of these proteomes using a blastx e-value threshold of 1e-10 to examine the distribution of similar proteins in these three species. Venn diagrams were generated to graphically illustrate the overlap of unigenes for each proteome (Figure 4B). To screen for the presence of putative gametophyte genes, a list of male gametophyte specific genes in Arabidopsis was extracted from the microarray study of Honys and Twell 68 and a combined list of significantly enriched female gametophyte genes was compiled from the studies of Wuest et al. 70 and Yu et al. 69 . These Arabidopsis gametophyte gene lists compiled from the literature were examined for overlap with the list of genes producing significant alignments with Pteridium unigenes in the blastx search against the Arabidopsis genome to produce a Venn diagram of gametophyte genes (Figure 5).