Transcriptome sequencing was performed by 454 GS FLX Titanium (pyrosequencing) and resulted in 546,903 and 546,846 quality-filtered sequencing reads after adapter trimming with an average size of 333 and 331 nt from B. tuediae and H. digitata, respectively (Table 1). From these reads, 64,442 (Bolocera) and 54,293 (Hormathia) contigs were assembled. Transcripts found in one copy number (single reads) counted for about 20% of all transcripts. The raw sequence data from B. tuediae and H. digitata transcriptomes in this study were archived at NCBI’s Sequence Read Archive (SRA) under the accession number SRP011434.

All contigs together with single reads (182,546 sequences for Bolocera and 159,988 for Hormathia) were analyzed in Blast2GO. From the 64,442 Bolocera contigs, 25,447 (40%) had BLAST hits to known proteins. Furthermore, 17,153 (27%) were assigned with GO terms and 11,666 (18%) were annotated. As expected, a much smaller fraction of the 118,104 single reads had BLAST hits (21,268 reads corresponding to 18%). For the 54,293 Hormathia contigs, 22,210 (48%) had BLAST hits, 13,787 (25%) were assigned with GO terms, and 9514 (18%) were annotated. From 105,695 single reads for Hormathia, 22,864 (22%) had BLAST hits. Additionally, 3674 and 4680 single reads were annotated for Bolocera and Hormathia, respectively, and assigned GO terms added to the final analyses. The GO terms assigned to the contig sequences were then exported to CateGOrizer, and GO slim analyses were run externally and graphs were produced in Microsoft Excel. GO slim terms are higher-level GO ontology categories, which provide a better profile for specie comparison [28]. The sequences were then classified according to three main GO categories, molecular function, biological process and cellular components and visualized in bar charts (Figure 2). After GO slim was performed, there were 8655 GO terms in total for B. tuediae, assignments to the biological process category made up the majority (5,562; 64%), followed by molecular function (2180; 25%) and cellular components (913; 11%). From 9066 GO terms for H. digitata, biological process category was also represented in majority (5799; 64%), compared to molecular function (2257; 25%) and cellular components (1010; 11%). B. tuediae and H. digitata belong to the distantly related families Actiniidae and Hormathiidae, respectively, of the order Actiniaria [29]. Interestingly they possess very similar transcriptome profiles in adult polyps (Figure 2). This may be explained by the identical environmental growth conditions since these sessile individuals were sampled side-by-side at 25 meters depth (Figure 1A). Additionally, both species showed a very high similarity to Nematostella vectensis sea anemone [9]. Here, 17,295 out of 25,447 and 18,277 out of 26,210 TOP-BLAST hits for Bolocera and Hormathia, respectively, were assigned to N. vectensis.

A main objective of this study was to present a workflow for digital bioprospecting. The reported protocol is based on whole transcriptome sequencing of a desired organism, recognizing sequences or motifs by bioinformatical tools, and thereafter expressing the candidate gene to perform further functional characterizations. A flowchart of the approach is presented in Figure 1B.

At least two different approaches can be pursued. First, regular BLAST homology searches can be performed at different stringencies. Stringencies must be evaluated for each query, depending on expected conservation between query sequence, database, and acceptable degree of false positives. Evaluation can be performed by inspection of reciprocal searches when applying different parameters. Sequencing data are arranged into a local database that represents a digital library, and annotated homologs of desired molecules are then employed as the query, either at the nucleotide or amino acid sequence level. The database can be utilized as a collection of crude sequences from the original reads, or sequence assemblies can be produced, representing longer and more complete contigs. One drawback of databases with assembled contigs is the risk of producing false assemblies due to the presence of more than one gene copy or closely related homologs, as well as alternative splicing or RNA editing. This can however be largely avoided by stringent settings for contig assemblies. False assemblies can be a challenge for toxin peptides, which often are expressed together with closely related isoforms [30].

Searches in this study were performed on assembled contigs and on single reads. First, 284 sodium and 268 potassium channel toxins from different species were downloaded from SwissProt/UniProtKB protein database and used as query sequences. After closer examination of possible hits, it was concluded that candidate toxins are more similar to published sea anemone toxins and 78 annotated sea anemone neurotoxins were therefore used as query sequences for follow-up analyses (Table S1).

All possible hits to the query sequences with e-value lower than 1 × 10⁻⁶ were evaluated (after reciprocal searches to the reviewed SwissProt/UniProtKB protein database) and resulted in 15 hits, 4 for Bolocera and 11 for Hormathia (Table S2). Another level of stringency was added by allowing only sequences with e-value lower than 1 × 10⁻⁶ when blasting against the reviewed SwissProt/UniProtKB database. This search finally resulted in four hits, one for Bolocera and three for Hormathia, the alignments are shown in Figure 3A. An additional four hits to potential neurotoxin candidates were also included in this study (Table S2).

The second approach identifies potential sequences of interest in the digital library based on recognition of conserved domains. The domain architecture is sometimes the only indication to derived protein function, and domain analysis will thus increase the probability of discovering novel compounds. Recognition of conserved domains is also based on homology searches, but here multiple sequence alignment models based on experimentally verified structures make up the basis of the search tools. Most prediction pipelines are designed for single sequence analyses, and thus not suited for NGS data. Only a few studies have applied motif recognition on large-scale data set. Kozlov and co-workers [31,32] have developed a motif recognition program called Single Residue Distribution Analysis (SRDA) where predicted motifs, based on conserved amino acid sites in a certain group of proteins or peptides, are used in scanning of translated EST databases. This method was successfully applied on spider and sea anemone EST databases in the identification of potentially novel neurotoxins. Here, conserved domains were recognized using the NCBI’s Conserved Domain Database (CDD) with the Batch CD-Search interface, which can process up to 100,000 sequence predictions at one time [33]. The CDD input data are amino acid sequences, and nucleotide data have to be translated into the correct reading frames prior to analysis. The complete search results are then compiled into a temporary database, which is downloaded or viewed graphically. The output domain footprint is either shown as specific hits, as domain super families, or as multi-domain models.

As a proof-of-concept we applied the protocol in pursuing neurotoxin transcripts in the 454 pyrosequenced transcriptome data from the two cold-water sea anemone species B. tuediae and H. digitata. Unlike many tropical species these sea anemones are non-symbiotic, meaning they are not associated with a zooxanthellae. Although we cannot exclude co-extraction of RNA from protozoa or microbial species, we find it highly likely that transcripts investigated here originate from the sea anemone tissue. Sodium and potassium channel toxins blast searches (blastx) were performed on our local Bolocera and Hormathia transcriptome databases, applying published peptide sequences from other Actiniaria as query sequences (Table S1). Most of these toxins are small peptides, less than a hundred amino acids long. Based on the sequence similarity with published sea anemone neurotoxins, we predicted one sodium channel toxin and three potassium channel toxins, which passed our quite stringent criteria for homology prediction (Figure 3A,C). Reciprocal searches against SwissProt/UniProtKB protein database were performed to assure that the blast hits for neurotoxin genes truly represent the best matches for the sequences. One of the classification parameters of sodium channel toxins is the size of the mature toxins, although the number and positions of cysteine residues seem also to be of greater importance [34,35]. It is well known that the 3D structure of neurotoxins is essential for appropriate binding to the specific ion channel and therefore deletions in loop regions might not be vital. The predicted Hormathia type III sodium channel toxin (HdNa3) was aligned to the type III homolog from Calliactis parasitica (CLX-1). The type III sodium channel toxins are not well defined. The predicted HdNa3 has, however, sequence similarity to the CLX-1 peptide. Surprisingly, no sodium channel toxins were predicted from the Bolocera sequence data. This was unexpected since sodium channel toxins are abundant in other sea anemones, and because a sodium channel toxin has previously been reported from B. tuediae [36].

Transcripts representing type II class of potassium channel toxins were predicted from both B. tuediae and H. digitata (Figure 3A,C). Type II toxins appear to be well conserved both regarding sequence and structure prediction. HdK2a aligns well to the type II toxin from Anemonia sulcata (AsKC3) (Figure 3A), two additional toxins BtK2 and HdK2b align both well to the type II toxin from Anthopleura elegantissima (APEKTx1).

Some toxins are represented as precursors that include an N-terminal signal peptide (Figure 3A). Peptide cleavage is usually initiated at a cleavage tandem site (Lys-Arg) leaving a mature peptide at the C-terminal part [37]. With the exception of a few conserved domains, neurotoxins generally have limited sequence conservation. The predicted neurotoxin sequences were therefore structure determined by SWISS-MODEL in order to establish conserved structural motifs, and thereby support the sequence predictions. SWISS-MODEL 3D predictions resulted in four β-strands configurations of HdNa3 (Figure 3B). This is in agreement with the information stating that most sodium channel toxins seem not to possess α-helix motifs [22]. The potassium channel toxin, HdK2 (Figure 3B), was predicted by SWISS-MODEL to possess a C-terminal, and also small N-terminal α-helix, presence of helices is comparable to other type II potassium channel toxins [27].

An additional 15 toxin candidates were predicted using lower stringency settings, 4 for B. tuediae and 11 for H. digitata (Table S2). These would be also included in possible experimental follow-up studies. However the digital marine bioprospecting approach introduced here serves only for bioinformatic purposes to identify possible homology variants of known proteins by mining next generation sequencing data.

Four of these additional toxin candidates showed high similarity to proteins with Kunitz-type domain (AXPI and Kunitz/BPTI-like toxin). AXPI protein was already previously shown to have sequence similarity to type II potassium channel toxins (AsKC1-3) and was proven to belong to the Kunitz-type family [38]. Proteins from this family have high sequence similarity, concerning especially the position of Cysteine residues, and they share both protease, as well as ion channel inhibitory activity. AXPI protein is also predicted to be structurally similar to these proteins, a sign also very important for neurotoxins, although its possible ion channel inhibitory activity has not been proven yet.

Potential neurotoxin sequences were also predicted by recognizing conserved domains (Table 2). We obtained 144 and 206 neurotoxin candidate transcripts from B. tuediae and H. digitata, respectively. The output result from the Batch CD-Search tool assigned 131 (B. tuediae) and 229 (H. digitata) superfamily queries with 151 and 267 positive domain hits, respectively, and a total of nine unique domains (Table 2, S3 and S4). The most widespread conserved domain was the BPT1/Kunitz superfamily, originally a serine protease inhibitor that has gained a new function as an ion channel blocking toxin [39]. Other serine protease inhibitor domains were also recognized, together with domains characterized by disulfide bridges. A significant fraction of the transcripts was found to contain more than one conserved recognized domain.

The cold-water sea anemones B. tuediae (Order Actiniaria; Family Actiniidae) and H. digitata (Order Actiniaria; Family Hormathiidae) were collected 2009-10-01 in Tromsø, Norway (69°41′ N; 18°56′ E) at 25 m depth using scuba diving (Figure 1A). RNA was extracted by crushing fresh tissue from body wall and tentacles in TRIzol using a Precellys lysis homogenizer (Stretton Scientific, Stretton, UK) to ensure identical sample handling before extraction [40]. 0.2× volume of chloroform was added, incubated on ice for 20 min, centrifuged and the water phase was transferred to a new tube. The RNA was precipitated in isopropanol at 4 °C and the pellet washed with 75% ethanol before the RNA was rehydrated in water. For some of the samples an additional phenol/chloroform extraction was performed, and subsequently RNA was precipitated in ethanol.

Transcriptome sequencing of fragment libraries was performed by the 454 pyrosequencing platform at Eurofins MWG Operon (Germany). Approximately 10 µg total RNA from each species was shipped to Germany. Poly(A)⁺ RNA was prepared by Eurofins MWG Operon, first strand cDNA was synthesized applying random hexamers, with successive ligation of 5′ and 3′ adaptors. PCR amplification was performed with a proof-reading enzyme. Normalization was carried out by denaturation and renaturation of the cDNA, with subsequent removal of ds-cDNA before ss-cDNA PCR amplification. The cDNA was size fractionated (500–700 bp) by elution of preparative agarose gels, subjected to shotgun library preparation and subsequent GS FLX Titanium sequencing. All the handling of the samples after RNA isolation, including cDNA library preparation, was done by Eurofins MWG Operon.

The contig assemblies were performed as a service by the Eurofins MWG Operon. Quality-filtering of the reads was done by Roche/454 sequencer software when performing the base calling. The sequences were additionally trimmed, the key tag (TCAG) at the 5′ end and low quality and adapter sequences were removed from the sequences before assembly by MIRA Assembler software. Only reads ≥40 bp were considered for the assembly by MIRA Assembler. Assembled contigs and single reads were run through Blast2GO [41], they were BLASTed, mapped and annotated. The transcripts were grouped, based on their potential function and visualized in bar charts applying CateGOrizer [42] and Microsoft Excel. Additional statistical data were extracted from Blast2GO. The transcriptome data were collected in two local databases and blastx searches were performed on the contigs and single reads using relevant published anthozoan sequences as queries with a threshold value of e = 1 × 10⁻⁶. Furthermore, CLC Genomic Workbench [43] was applied in the mapping of toxins. Different stringencies were used during the neurotoxin searches and also additional reciprocal searches against second, more comprehensive SwissProt/UniProtKB protein database were performed to verify the candidate toxin hits.

Multiple alignments were made in ClustalW2 [44], and the mature protein length for the aligned known toxins was determined using Protein Knowledgebase (SwissProt/UniProtKB). Cysteine residues in the alignments were highlighted by black colour and the disulphide bridges were marked by lines. 3D structure predictions were performed by SWISS-MODEL [45,46] using translated amino acid sequences. The 3D structures were exported in .pdb format and visualized in PyMol Viewer [47] as a cartoon with α-helices and β-sheets with a colour transition from red to blue, C- to N-terminal. Additionally, disulphide bridges were marked according to the sequence alignments. For better resolution, figures were exported and run in POV-Ray [48].

A collection of sequences were also translated in all six reading frames applying the Six Frame Translation tool from Max-Planck Institute for Developmental Biology [49], and run through the Batch Conserved Domain Database [33] with default settings. For the purpose of this analysis, the transcriptome sequences were first enriched with potential neurotoxin transcripts by performing a low stringency (e = 1 × 10⁻²) blastx homology search applying 78 annotated neurotoxins (Table S1) as queries and these candidate sequences were translated into amino acid sequences in all six reading frames before being introduced to the Batch CD-Search tool.