Transcriptome deep-sequencing and clustering of expressed isoforms from Favia corals

De novo assembly

Holobiont cDNA libraries were synthesized from the RNA of two individual adult Favia sp. collected from the Gulf of Eilat in the Red Sea. Illumina runs performed on each separate, normalized, cDNA pool generated approximately 80 million reads per sample with average quality scores > Q20 at each base. The first step of assembly was carried out with ABySS [18, 19], a de Brujin graph assembler. In order to recover transcripts across a range of expression levels, we carried out assembly across a range of k-mer values. Transcripts with low depth (i.e. weakly expressed) are best recovered with low k-mer values, while high depth (i.e. highly expressed) transcripts are best recovered with high k-mer values [20]. Using a range of k-mer values also allows for the identification of expressed splice variants arising from a single gene. As the Illumina read length was set to 75 bp, we chose initial k-mer values ranging from 29 to 45 bp for each sample run.

We evaluated various assembly parameters (e.g., total number of contigs, contigs longer than 100 bp, N50 length, and average contig length) as a function of k-mer length. The three k-mer values (35, 39, 45 for Fav1 and 31, 35, 39 for Fav2) with the highest N50 length [21] were selected as being most informative. In each sample, we eliminated contigs shorter than 150 bp [20] in the two k-mer assemblies with the shortest median contig length, but kept all the contigs in the assembly with the longest median contig length in order to retain any information useful for bridging in the subsequent assembly steps. Within each sample, the three k-mer assemblies were then combined, and the combined contigs were assembled with CAP3 (using default parameters), which computes overlaps to correct errors in constructing contigs and generates consensus sequences for contigs [22], thus eliminating redundant contigs. It has been suggested that assembly of ABySS followed CAP3 yield better contigs [19]. As a result, the N50 length distribution improved after using CAP3, and the best N50 values increased from 1027 to 1665 in Fav1, and 742 to 1439 in Fav2 (Figure 2). The final assembled datasets, which were used for all subsequent analyses, contained 58,848 sequences in Fav1 and 62,469 sequences in Fav2. The N50 values of these two datasets were higher than previous short-read publications [5, 7, 23] (675 bp, 1438 bp, 506 bp, respectively), suggesting that the quality of our data was comparable to results in other non-model species (For all commands and parameters, see Additional file 1: File S1).

Homologous clustering of expressed coral transcripts

After using the EMBOSS package [24] to generate all possible open reading frames (ORFs) from stop to stop for each assembled contig, the resulting predicted ORFs were searched for sequence similarity against the N. vectensis proteome [25], using reciprocal BlastP (E-value ≤2e^-30) [26] (Script 1). For the 519,766 predicted ORFs longer than 150 bp, 12,141 unique ORFs in Fav1 showed considerable sequence similarity to 7,186 existing protein sequences in N. vectensis. Similarly, 12,425 unique ORFs in Fav2 showed similarity to 6,862 N. vectensis protein sequences (Additional file 2: File S2, Additional file 3: File S3). The top Blast hits for each sample were saved in a pre-clustering list using a Perl script (Script 2; Output files reported in Additional file 4: File S4, Additional file 5: File S5). These lists were then used in TRIBE-MCL [27] to identify homologous protein family clusters in a comprehensive and uniform way (Additional file 6: File S6, Additional file 7: File S7). The main clustering parameter, inflation value (r), was selected as default (r = 2.5). Fav1 and Fav2 had similar numbers (7,186 and 6,862, respectively) of protein family clusters homologous to unique N. vectensis proteins. These clusters were subjected to further functional annotation.

In order to evaluate the completeness of our annotation using N. vectensis as the reference as opposed to using another available Cnidarian non-annotated proteome (A. digitifera), we applied a newly-developed completeness metric [28] (In prep.) to determine the proportion of the reference proteome covered by our sets of assembled transcripts. Only those ORFs with length coverage ≥80% of the matched protein from the N. vectensis or A. digitifera proteome were included. Completeness measurements in Fav1 and Fav2 compared to N. vectensis were 29.54% and 28.20%, respectively; when the same procedure was carried out using the unannotated proteome of A. digitifera as a reference (23,677 ORFs downloaded from http://marinegenomics.oist.jp/genomes/downloads?project_id=3. This showed an improvement of only 1-3%, thus validating our usage of N. vectensis as a reference proteome (Additional file 8: Table S1).

Functional annotation and characterization of the isoform clusters in Fav1

To identify the putative function of 7,187 isoform clusters, Gene Ontology (GO) and protein domain (KOG, InterPro) searches were performed using the functional annotation of the N. vectensis. (Data downloaded from the JGI genome project http://genome.jgi-psf.org/Nemve1/Nemve1.download.html). The clusters were assigned gene names based on the gene name annotation of the best Blast match for the sequences (Additional file 9: File S8). This process successfully assigned gene names for 6,632 (92.27%) clusters using GO term, KOG description, and InterPro description. Among 12,141 annotated best hits, 11,411 (93.98%) gene names were assigned to sequences. These provide a rough estimate of the number of different genes expressed in Fav1 libraries. Broadly, the putative homologs of genes involved in various cellular processes and pathways found to be functionally conserved.

Annotation of Symbiodinium- derived contigs

Holobiont coral tissues also contain eukaryotic dinoflagellate endosymbionts of the genus Symbiodinium[30, 31]. We therefore determined the contribution of symbiont-derived transcripts in our analysis. First, we extracted the regions of cDNA contigs that corresponded to each individual annotated ORF in two datasets (For commands, see Additional file 1: File S1). Furthermore their similarity search against two Symbiodinium transcriptomes (http://medinalab.org/zoox/) was performed using BlastN. In order to define an E-value as a cutoff threshold, a reciprocal BlastN search between the N. vectensis genome and the two Symbiodinium transcriptomes showed an average E-value of e^-80. Thus all contigs with similarity higher than this threshold to Symbiodinium were defined as likely to be symbiont-derived. Based on these results, 9% of the annotated ORFs (1.34% of the total assembled contigs) of Fav1 were labeled as symbiont sequences, and 8.7% (1.61% of total assembled contigs) of Fav2. FASTA files of these non-symbiont transcripts are reported (Additional file 11: File S9, Additional file 12: File S10). Finally, we performed BlastX (E-value equal to at least e^-30) on the non-symbiont derived cDNA fragments against the N. vectensis proteome to confirm correct initial annotation by BlastP. All the cDNA sequences matched to the same N.vectensis IDs that were predicted using BlastP.

Phylogenetic assessment

Molecular markers are essential tools for population genetic studies. Typically, combination of mitochondrial and nuclear markers are used to examine the species relationships. In order to generate a Favia molecular marker dataset, we downloaded Favia related sequences from NCBI. Similarity searches were carried out against this Favia dataset. Among various molecular markers, we chose COI, Cytb and 28S. Individual sequence regions were identified and extracted from the cDNA contig files in both samples. DNA alignments for each locus were generated using ClustalW2 with default parameters [32] (Additional file 13: File S11, Additional file 14: File S12, Additional file 15: File S13). Consequently, a matrix of these three loci was generated using FASconCAT [33]. A Maximum likelihood phylogenetic analysis (RaxML) was carried out [34]. Maximum likelihood phylogenetic analysis using three loci (COI, Cytb, 28S) suggests that these Favia samples belong to Faviids (Additional file 16: Figure S2). Morphological analysis places them as F. albidus[17], a species that is not yet represented in NCBI. For example, out of 18 Favia species that have been described morphologically, only 15 of them have molecular data in NCBI. F. albidus, F. helianthoides, and F. marshae lack molecular markers in NCBI. Based on geological distribution [17] and morphology, we suggest these two species belong to F. albidus. In fact, F. helianthoides has no morphological similarities with our samples, and F. marshae habitat has never been reported in Red Sea [17]. However, further skeletal samplings are required for final validation [35, 36]. Regardless, this study increases the protein information of the Faviids from 496 proteins to over 12,000 proteins in NCBI.

Characterization of one exemplary homologous protein cluster

From the protein clustering results, we chose to characterize a protein family with a natural fluorescent property. One of the benefits of utilizing scleractinian corals as our model organism is that they posses genes for fluorescent proteins (FPs), a rare characteristic in most other Phyla’ besides Cnidaria [37–40]. In N. vectensis, six protein IDs encode for FPs [41]. A search among the homologous sequence clusters with E values of at least 2e^-30 in each transcriptome led to the identification of one protein cluster group per sample representing potential fluorescent proteins (FPs). A total of 11 new potential FPs were identified, six belonging to the Fav1 sample and four belonging to the Fav2 sample. One additional sequence, s23Contig9635-2 was found by increasing the E-value to 2e^-10 in Fav1. The alignment of these sequences with N. vectensis fluorescent protein sequences (JGI ID:205348, ID:206334), Branchiostoma GFPa1[42] and GFP of Aequorea victoria (GI:17943301) showed a considerable homology (Figure 3). The conserved chromophore region is located at the residues 303 to 305 based on the top sequence. Our data shows that one of the newly identified potential fluorescent protein sequences (Fav1 s23Contig16657-5) is 185 amino acids longer at the N-terminus (416 amino acids in total) and two of them were shown to be 49 (Fav2 s62Contig19888-6) and 41 amino acids (Fav2 s62Contig41210-3) longer than the consensus length of reported sequences in NCBI (wild-type GFP from Aequorea victoria is 236 amino acids) (Additional file 17: Figure S3). This extended region does not seem to interfere with the proper folding and expression of FP, however further studies are required to reveal the function of these upstream domains.

Furthermore, the maximum likelihood trees were generated from the alignment of 156 fluorescent sequences, including the 11 newly identified sequences (Additional file 18: Figure S4, Additional file 19: File S14 contains all the accession numbers). There was a strong bootstrap support for basal clade relationships within tree. This includes the order Ceriantipatharia, and Pennatulacea, although low bootstrap support for FPs within order Scleractinia. Ctenophore FP clustered with hydrozoan FP, therefore the cnidarian clade was not monophyletic. Others have shown that incongruence with taxonomy is not unusual in fluorescent proteins [43]. For better visualization, a smaller maximum likelihood sub-tree was generated from 46 scleractinian FP sequences (Figure 4). Although the bootstrap values improved compared to Additional file 18: Figure S4, some branches still exhibited low bootstrap values. Nonetheless, using RaXML [27], we categorized the newly identified sequences into four clades and using ProtTest [44] we identified “PROTGAMMAWAGF” as the best-fit model.

In order to evaluate our assembly method and the possible impact of ABySS-specific errors on the annotation accuracy of the long candidate FP sequence, we performed both Trans-ABySS [20] and Trinity [45] on reads from Fav1. Both assembly programs led to the generation of sequences identical to Fav1 s23Contig16657-5 as predicted using ABySS and CAP3. (Additional file 20: File S15).

Validation of the identified protein clusters as fluorescent proteins

The intrinsic fluorescence of FPs includes a unique chromophore that is formed post-translationally within the protein upon autocatalytic cyclization and oxidation of residues X-Tyr-Gly [46]. The fluorophore is located almost at the center of the cylinder and is inaccessible to outside enzymes [46, 47]. The GFP fluorophore is capable of forming under a wide range of conditions and once formed is highly stable. The entire structure is very resistant to denaturation by heat and denaturants. The three sequences with longer N-terminal domains (s23Contig16657-5, s62Contig19888-6 and s62Contig41210-3) were cloned into mammalian expression vectors. We used Kozak analysis [48] to pick the best potential start codon, and reading frames were generated using gene synthesis. The start codons are underlined in red in Additional file 21: Figure S5. The synthesized sequences were optimized for expression in mammalian cell lines. The synthesized sequences showed fluorescence when expressed in HEK-293 mammalian cells, thus validating them as genuine FPs (Figure 5).

In Silico quantification of Faviids transcripts

In order to rule out the possibility of promiscuous domain assembly, we assessed the quality of the de novo assembly of FP sequences, as well as all other transcripts, by mapping reads on assembled contigs for each sample. Such read alignment to contigs is necessary to provide support for new transcript identification as well as for determining gene expression levels [49, 50]. In order to measure the Reads Per Kilobase of exon model per Million mapped reads (RPKM) [50], a sub-fasta cDNA region, corresponding to each ORF, within each contig was generated. Reads were aligned to these annotated cDNA regions. Coverage (RPKM) measurements were determined using a Perl script (Script 3). The results are reported (Additional file 22: File S16, Additional file 23: File S17). The mapping of all the reads onto the annotated Faviids transcript showed that the number of reads corresponding to each transcript ranged from 10 to 47,189, with an average of 850 reads per transcript in Fav1, and 10 to 29,222, with an average of 766.37 reads per transcript in Fav2, indicating a wide range of expression level of Faviids transcripts. It also indicates that very low expressed annotated Faviids transcripts were also represented in our assembly. The minimum coverage (RPKM) of an annotated Fav1 transcript was 3.89 and maximum of 6,919.20 with an average of 68.61. The RPKM ranged from 3.60 to 8,576, with an average of 72.64 in Fav2. The average and the range of RPKM per transcript is similar and somewhat higher (25.7) than other whole transcriptome studies [26].

All the cDNA regions annotated for fluorescent property had reasonable coverage, including the long candidate cDNA sequence (Fav1 s23Contig16657-5) (Additional file 21: Figure S5). Based on the calculated RPKMs for each of the identified fluorescent protein in both samples, s23Contig19691-3 in Fav1, and s62Contig57475-7 in Fav2 had the highest coverage level (Figure 6).

Coral collection and total RNA isolation

This study was conducted during the period of May–June 2009 on a coral reef on the northern tip of the Gulf of Eilat, in the northern Red Sea (29º30′N, 34°55′E).

Samples were collected at 65 m, using closed-circuit trimix rebreather system (Megalodon™). The organisms were identified under water to the family level, Faviidae, and brought to the surface in a black mesh bag to avoid sun exposure. The organisms were immediately photographed and vouchered with white light and fluorescent photography as described in [51] and stored in a shaded running-seawater facility. Within 1–2 hours of collection, samples were rinsed in sterile-filtered artificial seawater and processed for RNA and DNA. The tissue of the coral was extracted from the skeleton using QiaShredder (Qiagen). For RNA, the TriZol method was used and stored as an ethanol precipitate for travel back to the US. DNA was extracted using Qiagen DNAeasy kit according to manufacturer’s protocol and stored in at 4°C. The specimens have been photo vouchered and their genomic and transcriptomic raw materials are stored in the American Museum of Natural History Ambrose Monell Cryo Collection.

Preparation and screening of cDNA library

Illumina sequencing using the GAII platform was performed at the Yale University W.M. Keck Biotechnology Resource Laboratory according to manufacturer’s instructions (Illumina, San Diego, CA) (Additional file 24: File S18) and using high quality RNA with a 28S rRNA band at 4.5 kb that is at least twice the intensity of the 18 s rRNA band at 1.9 kb. The cDNA library contained 77,804,306, 75-mer length reads. The sequencing data are deposited in NCBI Sequencing Read Archive [52]. (The BiosampleIDs = SAMN01761696, SAMN01761695).

De novo assembly

De novo assembly was carried out using ABySS with default settings across multiple k-mer values [18]. After assessing different k-mer values, the three best k-mer assemblies (35-mer, 39-mer, 45-mer for Fav1 and 31-mer, 35-mer, 39-mer for Fav2) were selected and concatenated for the second step of assembly. To evaluate the N50 length and the number of assembled contigs using different k-mer values, we used a Perl script. CAP3 [22] was used to remove redundancy across ABySS assemblies and to merge contigs into longer sequences. All assembled contigs were subjected to annotation and further protein homology searches. Trans-ABySS [20] and Trinity [45] were used to confirm the long ORFs, homologous to fluorescent proteins, which were identified with ABySS and CAP3.

Gene annotation and analysis

A set of possible Open Reading Frames (ORF), stop to stop from assembled sequences, was generated using EMBOSS [24]. To annotate the de novo assembled sequences, a similarity search against N. vectensis proteome was conducted using BLASTP with two E values of 2e^-10 and 2e^-30. The resulting data (E-value of 2e^-30) was filtered and clustered using TRIBE-MCL [27]. Each homologous group was annotated using GO and KOG annotated N. vectensis data (http://genome.jgi-psf.org/Nemve1/Nemve1.download.html). For Symbiodinium peptide annotation, a homology search using BLASTN with E values of 2e^-80 against the Symbiodinium transcriptome (http://medinalab.org/zoox/) was carried out. The final non-symbiont FASTA cDNA fragments were reported.

Completeness measurement

The BlastP (E-value of 2e^-30) output list generated from homology search of both samples against N. vectensis[41] and A. digitifera[53] was organized for completeness measurements. The completeness formula according to [28] was implemented into a Perl script (In prep) to determine the percentage of the reference proteome that is covered by each of our sets of assembled transcripts. Length coverage of each of these reference ORFs by a hit from our data set had to be at least 80%.

Phylogenetic analysis of FPs

The maximum likelihood tree of identified fluorescent protein was generated using RaXML [34] under PROTGAMMAWAGF amino acid substitution model, selected based on the results from ProtTest [44]. The alignment was generated using MAFFT [54] and CLUSTALW2 [32] with minor adjustment at the N-terminus region, when long gaps were inconsistent with other isoforms. Bootstrap values were estimated based on 1,000 replicates and were given for all presented branches. The variant sites were visualized with geneious (http://www.geneious.com). Dendroscope was used for visualization [55].

Phylogenetic assessment

Molecular barcodes for all the Favia related sequences were downloaded from NCBI. A similarity search with sequences from our annotation was carried out against this Favia dataset. Cytb, COI and 28S sequences were identified and extracted from the cDNA contig files in both samples. DNA alignments for each locus were generated using ClustalW2 with default parameters [32]. Consequently, a matrix of these three loci was generated using FASconCAT [33]. A Maximum likelihood phylogenetic analysis (RaxML) was carried out [34]. Bootstrap values were estimated based on 10,000 replicates and were given for all presented branches. Dendroscope was used for visualization [55].

Cloning of fluorescent proteins

The three cDNA sequences (Fav1 s23Contig16657, Fav2 s62Contig19888-6 and Fav2 s62Contig41210-3) were synthesized and propagated in pUC57 (GenScript USA Inc.). Kozak [48] analysis was used to determine the location of the potential start codon. The genes were subcloned from pUC57 into the NotI-BamHI site of the mammalian expression vector pcDNA 3.1 (Invitrogen, Inc.) using standard recombinant techniques [56].

In Silico gene coverage measurements

Gene coverage levels were determined using a Perl script (Script 3). This script implements Bowtie [57] to map reads to an annotated reference cDNA, and calculates the RPKM according the formula used in [50]. For visualization, BWA [58] was used to generate the read-to-contig alignment. The annotated cDNA from individual samples were used as the reference contig, and SAMtools [59] was used to generate binary files to be visualized in the IGV [60] genome viewer (For commands, see Additional file 1: File S1).