De novo genome and transcriptome assemblies

Annotation of structural genes, repeats and horizontal gene transfer

The genome annotation generated 12,000 gene models, of which 9911 were supported and revised by homologous sequences. Of this highly confident set of 9911 gene models, the average gene length was 5733 bp with an average intron number and length of seven and 547 bp, respectively; introns were shorter than in many published insect genomes (e.g., [23]). The distribution of these gene parameters across the genome is given in Fig. 1. Of the 1066 conserved BUSCO genes, 825 (77.4%) of this gene/protein set were predicted to be full length and 69 (6.5%) partial. Among the complete genes, most of them represent single copies (90.4%). We then annotated the protein models from comparisons with the NCBI Genbank non-redundant (nr) database using BLASTP with a cut-off threshold of e-value 10⁻⁶. The proportion of protein models showing homology to nr records was 82.8% and the two species that the proteins hit most frequently were Zootermopsis nevadensis (6.0%) and Daphnia pulex (3.6%) (Additional file 1). Both of these analyses demonstrate that this set of predicted gene models is largely representative of the H. duospinosa protein coding sequences, and is therefore suitable for subsequent evolutionary and functional studies.

The total level of heterozygosity within the H. duospinosa genome, which is the portion of heterozygous single-nucleotide polymorphisms between the two haploid components in the diploid genome, was estimated to be 1.56 × 10⁻³. Among all called variants, including indels, 20,622 (2.97%) fell within the coding regions of 6150 annotated gene models in 13,162 exons (Additional file 2). The histogram of k-mer copy number was largely uni-modal, reflecting the low level of heterozygosity (Fig. 2).

We identified a total of 59 bacterial and 96 fungal genes as candidates for horizontal gene transfer (HGT) into the H. duospinosa genome (Additional file 3). Compared with the Folsomia candida [16] and Orchesella cincta [10] genomes, we have found fewer candidate HGTs, which is likely due to the H. duospinosa genome being assembled from short Illumina reads and therefore being highly fragmented (Table 2). Nonetheless, the most common blast hits of the HGT candidates are from the two fungal species (Conidiobolus coronatus and Basidiobolus meristosporus). The HGT candidates are involved in a wide variety of metabolic functions, like those identified from Folsomia candida and Orchesella cincta [10, 16]. These include amino acid production, DNA and glycerol metabolic process, ATP synthesis, oxidation-reduction process and cation transport. Our fragmented assembly, along with non-curated genes models, have likely led to an underestimate of the amount of HGT into the H. duospinosa genome. However, our results do confirm that this process is a general one within Collembola.

DNA methylation

In arthropods DNA methylation (the addition of a methyl group to a cytosine residue in a CpG context) occurs predominantly within the exons and introns of genes [29–31]. Methylation of cytosine residues leaves them susceptible to deamination [32] and, over evolutionary time, genes that are highly methylated (in germ-line cells) will have lower than expected CpG content. This affect can be quantified by calculating the normalised CpG content of genes, or CpG[o/e]. In animals where DNA methylation has a demonstrated role in controlling gene expression, such as Apis mellifera (Fig. 3a), the distribution of CpG[o/e] values can be described as, consisting of genes with lower than expected CpG content that are historically methylated and those with higher than expected CpG content that are historically unmethylated. Predictions of historical DNA methylation using this method correlate with measured levels of DNA methylation [33, 34]. In contrast, the same analysis in Drosophila melanogaster, which does not have an intact DNA methylation system and has very low levels of DNA methylation [35, 36], yields a unimodal distribution (Fig. 3b). Analysis of the CpG content of genes predicted from the H. duospinosa genome displayed a unimodal distribution indicating the absence of historical DNA methylation in this species. The mean CpG[o/e] is 0.7, which is lower than the expected value of one but consistent with the relatively low CpG content of the H. duospinosa genome as a whole (mean CpG[o/e] is 0.79).

A full complement of the DNA methyltransferase enzymes, Dnmt1, Dnmt2 (TRDMT1) and Dnmt3, is thought to be required for a fully functional DNA methylation system [37]. Dnmt3 enzymes are the ‘de novo’ methyltransferases and are important in mediating environmentally responsive DNA methylation [37]. Dnmt2 (TRDMT1) methyltransferases are predominantly involved in tRNA methylation and Dnmt1 DNA methyltransferases act as maintenance methyltransferases maintaining methylation marks across cell division [37]. The H. duospinosa genome encodes three orthologs of the tRNA methyltransferase Dnmt2 and an ortholog of the putative DNA demethylation enzyme Tet1 [38, 39], an enzyme that also functions in the modification of mRNA promoting translation in D. melanogaster [40]. However we were unable to identify an ortholog of the de novo methyltransferase, Dnmt3 or the maintenance methyltransferase Dnmt1, within the H. duospinosa genome. The lack of Dnmt1 and Dnmt3 in H. duospinosa is consistent with the absence of any environmentally responsive DNA methylation in this organism predicted from the analysis of CpG[o/e].

Non coding RNA genes

Non-coding RNAs (ncRNAs) form a central role in the catalysis and regulation of key cellular functions such as translation, splicing, transport and the modulation of gene expression. The major RNA families include essential and highly conserved RNAs such as the tRNAs, rRNAs and the RNA components of RNase P and the signal recognition particle [41]. Other ncRNAs, such as the small nucleolar RNAs (snoRNAs), microRNAs (miRNAs) and the long non-coding RNAs (lncRNAs), have a high evolutionary turnover [42, 43]. The ncRNAs pose serious research challenges for genome annotation as they lack many of the strong statistical signals that are associated with protein-coding genes, such as open reading frames and codon-usage biases, and are frequently pseudogenised and duplicated via transposition [44]. Therefore homology-based approaches, as opposed to de-novo prediction, are generally used to find them, although high throughput transcriptomic approaches are increasingly employed [45].

We identified 17 loci with sequence similarity to nine known snoRNA families. These included one scaRNA (SCARNA8), three H/ACA box and 13 C/D box snoRNA associated loci. The snoRNAs are predominantly involved with rRNA maturation. We identified 20 loci with sequence similarity to 14 microRNA families. A number of cis-regulatory elements were also identified. These include 118 histone 3′ UTR stem-loops, three potassium channel RNA editing signal sequences, four selenocysteine insertion sequences (SECIS) and three internal ribosome entry sites (IRES).

The predicted SSU rRNA on “scaffold300_size451797/208930–205,788” matches NCBI-nr sequences from the collembolan species, Morulina verrucosa (Neanuridae: Morulininae) and Crossodonthina koreana (Neanuridae: Neanurinae). However, there is a large, 1454 bp insertion in the SSU rRNA at position 496 to 1949. This region contains a homolog of a reverse transcriptase, suggesting that this rRNA insert is a retrotransposon. We detected eight paralogues of this insertion sequence in the H. duospinosa genome and transcriptome sequences.

Developmental genes

Axis formation genes evolve relatively rapidly in insect lineages [46], and patterns of loss and conservation are well known. Holacanthella duospinosa has no Bicoid, which despite being a key gene in Drosophila, is restricted to dipteran lineages. Holacanthella duospinosa is also missing, like most non-dipteran insects, classical oskar and swallow genes. Genes involved in terminal patterning are well conserved, with a gene related to trunk, as well as a noggin-like gene [47], both present in the H. duospinosa genome. The genes that control segmentation in insects are generally well conserved. The Hox gene complex is an evolutionarily conserved complex of homeobox containing genes derived from the common ancestor of metazoans. The genes in the complex control segmental identity and their duplication and diversification have been instrumental in the evolution of the metazoan body plan [48, 49]. The relationships of genes in the complex, their order along the chromosome and transcriptional direction are all highly conserved. In H. duospinosa, the Hox gene complex is broken over two genome regions. At the 3′ end of the complex, The gene labial is found at the extreme end of scaffold 154. This is likely linked to the next region of the complex, found on scaffold 327, which contains three genes, proboscipedia, hox3 and Deformed. The rest of this scaffold contains multiple genes with strong similarity to evolutionary conserved sequences and evidence of transcription. The 5′ end of the complex (Sex-combs reduced, fushi-taratzu, Antennapedia, Ultrabithorax, abdominal-A and Abdominal-B) are found on scaffold 50. At the 3′ end of this group of genes are multiple genes with strong similarity to evolutionary conserved sequences and evidence for transcription. This genomic structure implies the Hox gene complex is split in H. duospinosa (Fig. 4), which is unusual in insects, but is found especially in Diptera. The alternative possibility is that this is an assembly error, though the placement of conserved, transcribed genes at the ends of the contigs containing both parts of the assembly suggest otherwise. The apparent split in the Hox gene cluster of H. duospinosa is at a different position to those found in Drosophila species [50–52] or in the silkworm, Bombyx mori [53]. The split in the Hox complex described here is also partially consistent with the rearrangement seen in the genome of the collembolan Folsomia candida [16], where a significant complex inversion separates deformed from sex-combs reduced, placing sex combs nearer Abd-A and splitting AbdA from Ubx. In Folsomia, deformed and sex combs reduced are on different scaffolds, but Ubx and AbdA are conventionally arranged next to each other (Fig. 4).

Notch signalling is a highly conserved animal-specific cell-signalling pathway with little change observed over evolutionary time. In H. duospinosa, most of the pathway is conserved. Surprisingly, however, orthologs of Deltex and Serrate were not found in the assembly. These two genes are core components of the Notch pathway conserved in all other insects we have looked at, suggesting either that the genome assembly is incomplete or that there has been lineage-specific loss of these genes. Given that Notch signalling is a pleiotropic pathway with many roles in development and in adult tissues it is not clear what the consequence of the loss of these genes might be.

The Enhancer of split complex is an unusual gene complex found in insects and Crustacea that consists of bHLH-orange domain genes and bearded class genes. This gene complex is Notch signalling responsive [54, 55], and the genes in the complex encode Notch effector proteins [56]. Ancestrally, the complex is made up of four genes, three bHLH-orange genes and a bearded class gene, with this structure conserved (with variation) in insects and Crustacea [57, 58]. In H. duospinosa, the enhancer of split complex is reduced to two bHLH-orange genes (her and bHLH1) on scaffold 36. No bearded class gene is present in the complex; though others may be present in the genome (bearded class genes have little sequence similarity). Reductions in the Enhancer of split complex are common especially in hemipteran insects [57, 58], but it is not clear what effect this reduction might have on Notch-regulated processes.

The runt complex is an insect-specific gene complex [59] comprised of four runt domain encoding genes. The runt complex in H. duospinosa lies on scaffold 154 (upstream from the start of the Hox gene complex) and is identical to those found in other insects [59].

Sex determination genes

In hexapods, a large variety of molecular mechanisms have been described that determine the sexual fate of an individual (for reviews, see [26]). While a remarkable diversity of upstream components of the sex determination cascade have evolved within different hexapod groups, a few key regulatory genes are highly conserved among all taxa investigated to date. One of the key players is doublesex (dsx), a transcription factor belonging to the DM-domain family of genes, which are involved in sex-determination and sexual differentiation in all metazoans [60, 61]. In insects, dsx contains two conserved domains, a DNA binding domain (DM-domain; dsx and mab-3) and a dimerisation domain (Dimer) [62]. Dsx has been described as the master switch gene at the bottom of the sex-determination cascade in all insects. It undergoes sex-specific splicing by transformer (tra), which represents another key player in most, but not all insect sex-determining pathways [63]. Little is currently known about sex determination in Collembola and no molecular mechanisms have been described for this group. Expressed Sequence Tag data [64] provided evidence for dsx in Collembola, identifying both the DM and Dimer domains and a potential alternative splicing of Dimer. However, both domains were only present as singletons on different contigs. Here we identified a putative H. duospinosa dsx transcript of 485 amino acids that contains both, a DM and Dimer domain (Fig. 5a, Additional file 4). No homologues were found for tra, however this gene can be highly divergent among insect lineages [65], which limits our ability to detect tra based on sequence similarity. We did, however, identify a putative transformer 2 transcript, which, in Drosophila, forms a complex with tra to control the sex-specific splicing of dsx pre-mRNA.

Sex-lethal (Sxl) is one of the master regulatory genes in drosophilid sex-determination and is thought to have co-opted its specific function as a result of a gene duplication event in the fly clade [66]. In H. duospinosa we likely detected the Sxl paralogue, CG3056 or sister-of-sex-lethal (Fig. 5b, Additional file 5). It is unclear whether this gene plays a role in sex-determination in insects. Other putative sex-determining genes detected in H. duospinosa are listed in Additional file 6.

Overall, we found many transcripts with sequence similarities to genes that, in insects, are involved in responses to X:Autosome signals, in dosage compensation, and processing of doublesex and Sex-lethal (Additional file 6). It is unclear whether in H. duospinoasa these genes are involved in sex-determination, or if these specific functions were co-opted only later in hexapod evolution. Further experimental work would be needed to determine the exact pathway of sex determination in this species; but our data, together with data from other hexapods, are steadily building our understanding of the core features, and differences of this process, across this startlingly specious taxon.

Chemosensory genes

Collembola are able to respond to odours and tastants, being repelled by bitter, alkaline, acid and salt tastes in food [67, 68]. Thought to have originated early in protostome evolution, ionotropic glutamate receptors (IRs) are involved in chemoreception in insects primarily detecting low volatility acids [69]. Using IRs from Drosophila melanogaster and Dendroctonus ponderosae (mountain pine beetle) as query sequences, at least 15 IRs were identified from the H. duospinosa genome assembly. They included orthologues of the IR co-receptors IR25a and IR8a, suggesting that this collembolan has a functional IR system. Candidate ligands might be feeding cues and pheromones involved in the location of potential mates or conspecifics.

Gustatory receptors (GRs) are involved primarily in taste reception in protostome invertebrates [70], however their cnidarian relatives have a role in pattern formation [71]. In the H. duospinosa genome assembly 18 GRs were identified using louse and termite GRs in BLAST searches with a cut-off of 1e⁻⁰⁵. We also applied the rule that hits needed to contain a C terminal motif of T/SXXXXXXQF, where X = an aliphatic amino acid. No GRs involved in carbon dioxide sensing (Gr21a and Gr63a) were found, consistent with previous findings that carbon dioxide sensing GRs evolved later within insect evolution [72]. It is not clear what tastants the collembola GRs are capable of detecting, however likely candidates include compounds that are indicators of nutritional value and toxins (bitter compounds).

Likely derived from the GRs, the odorant receptors (ORs) are involved in odor reception and have undergone differential gene family expansion and are now a very large gene family in higher insects [73]. Recent genomic studies have suggested that ORs, or at least their obligate co-receptor, Orco, may have evolved early in hexapod evolution. The genomes of crustaceans do not contain Orco or any ORs [74] and ORs were not detected within the genome of a bristletail [75]. Orco has been, however, identified in the genome of a firebrat [75]. Using Orco sequences from locust and firebrat (Thermobia domestica) we could not find any evidence for an Orco orthologue within the Collembola genome or transcriptome. Similarly no other ORs were identified in searches using insect OR sequences with a cut-off of 1e⁻⁰⁵. The lack of any odorant receptors in Collembola is consistent with the hypothesis that the expansion of these genes within insects has been associated with the evolution of insect flight [45].

Chitinase genes

Collembola are members of the Ecdysozoa, a group of protostome metazoans that moult as they grow. The moulting process requires the ability to reshape the chitin that makes up their exoskeleton and chitinases are an important family of enzymes involved in this process. Chitinases may also play an important role in the degradation of fungal hyphae, a major food source for collembolans [76, 77]. Previous research has shown that Collembola display chitinase activity and are therefore able to digest fungal cell walls [77] and the Orchesella cincta genome shows a wide array of chitinase genes [10]. Twenty three chitinase-like genes were identified from the genome of H. duospinosa. Apart from three genes, Cht2, Cht8 and Cht10, which were tandemly arranged on the same scaffold, the chitinase-like genes were identified within discrete scaffolds. Analysis of the transcriptome provided full transcripts for seven of these genes and partial transcripts for five. Evidence for at least one pseudogene was found. Twelve genes did not have any representative sequences in the transcriptome, perhaps indicating these genes could be expressed at different developmental stages not investigated here. Holacanthella duospinosa has a similar number of chitinase-like genes when compared with other crustaceans and insects, including Daphnia, Drosophila, beetles and mosquitos [21, 78].

Insect chitinase and chitinase-like proteins generally contain a combination of single or multiple chitin-binding domains and hydrolase domains, specifically from the glycoside hydrolase 18 (GH18) family. Here we have defined chitinases based on the presence of at least one chitin hydrolase domain. All the H. duospinosa hydrolyase domains fall into the GH18 family, except Cht23 which falls into the GH19 family, predominantly restricted to plants. The pattern of H. duospinosa chitinase protein domain structures includes simple proteins with a single chitin hydrolase domain to the more complex Cht3 that contains three hydrolase domains interspersed with four chitin binding domains. Orthologues of this gene are found in higher insects (eg. Cht10 of Tribolium; [78]), and generally contain 4–5 domains of either type. The crustacean, Daphnia pulex, has an orthologue (Cht3) that contains the same number of domains as H. duospinosa, perhaps indicating an expansion of these domains has occurred in insects.

A phylogeny comprised of extracted GH18 hydrolase domains was used to compare the chitinase-like proteins of H. duospinosa to those of crustaceans and insects (Fig. 6). The phylogeny displays conserved orthologous groups that include GH18 domain sequences from H. duospinosa and a number of collembolan specific expansions. Holacanthella duospinosa has members of a number the conserved chitinase groups, including Group I, II, III, IV and V chitinases and ENGases of [79]. However, no obvious members of the SI-CLPs or IDGFs were identified from H. duospinosa. Four expansions including H. duospinosa sequences were identified (Cht2, 8, 10; Cht9, 11, 12, 13, 15; Cht14, 17, 18 and Cht19, 20, 21, 22) that have yet to be assigned to chitinase groups described in the literature. Of the H. duospinosa chitinase hydrolase domains all except Cht3a, 14, 16b, 19a, 19b, 20 and 21 contain a glutamic acid within the conserved motif II, synonymous with chitinolytic activity.

Chitinase genes have been shown to be differentially expressed throughout development of the insect. In this study, RNAseq data was collected from tissue taken from an adult, hence expression will be exclusive to adult physiology. The H. duospinosa chitinases with evidence for expression from RNAseq include Cht1–8, 10, 22 and 23. Group I and II chitinases have been shown to be involved in degradation of the endocutile during moulting [78]. Certainly H. duospinosa has orthologues within these two groups (Cht1 and Cht3, respectively). Group III are anchored in the plasma membrane by a TM domain and are involved in processes post moulting [78]. Holacanthella duospinosa has an orthologue in this group also, Cht16, however it does not seem to be expressed in the adult. Collembola are members of the Ecdysozoa and therefore moult as they grow so having these conserved chitinases is anticipated. A recent study has implicated a role for Drosophila Cht11 in regulation of cholesterol within mitochondria, impacting pathogen infection [80]. Holacanthella duospinosa Cht6 falls into the same phylogenetic clade as Cht11 from Drosophila, which may indicate a similar role for this chitinase from Collembola. Since fungi are thought to be a major part of collembolan diet it is conceivable that some of these chitinases also may be involved in the digestion of fungal cell wall material.

Phylogenetic analysis based on transcriptome data

We assembled a large set of orthologous genes from which to reconstruct phylogenetic relationships among early diverging hexapod lineages. Of 1478 single copy orthologous genes [2], we found hits for more than one of the nine species for 1470 OGs (Additional file 7). The subsequent outlier check revealed no outlier sequences for any of the nine query species. The identification of protein domains revealed 4026 unannotated regions (so-called voids) and 2841 Pfam-A data blocks. After deleting the ambiguously aligned sections and concatenation, the removal of data blocks (based on gene-boundaries or on protein domains) with an IC = 0 and only keeping partitions having contributing sequences from all nine species, supermatrix A (based on protein domain data blocks) consisted of 370,877 amino acid sites and 1049 data blocks (328 Pfam-A domains, 161 clans, 560 voids) and supermatrix B (based on gene data blocks) comprised 323,917 amino acid sites with 894 data blocks (Additional file 8). PartitionFinder merged input data blocks into 338 meta-partitions for supermatrix A (protein domain-based) and 343 meta-partitions for supermatrix B (gene-based). The best fitting substitution models assigned to the meta-partitions were mostly LG4X and LG + G + F (Additional file 9).

Both datasets yielded a similar optimal tree with Collembola being monophyletic and H. duospinosa consistently placed as closest relative to Anurida maritima (Neanuridae). All clades show maximal support, except for the placement of Sminthurus viridis (Symphypleona, Sminthuridae) (Fig. 7). For supermatrix A (domain-based meta-partitions), we found one unique topology as displayed in Fig. 7. For supermatrix B (gene-based meta-partitions) we found two tree topologies whereas the trees with the better LogLH were similar to the unique topology of supermatrix A (40 out of 50 trees). The alternative topology from supermatrix B (10 out of 50 trees) placed Sminthurus as sister to all other springtails which contributes to the very low bootstrap support for the clade Sminthurus + (Pogonognathellus + Folsomia). Thus, the placement of Sminthurus as representative of Symphypleona remains ambiguous as found previously [2]. This issue will be addressed in future phylogenetic studies from the 1KITE consortium. The phylogenetic analysis of both data sets support the expected sister group relationship between H. duospinosa and Anurida.

DNA and RNA extraction and sequencing

Several individuals of Holacanthella duospinosa were collected from under rotting logs on the slopes of Hauturu-O-Toi (Little Barrier Island, 36.19 °S, 175.11 °E), an island in the Hauraki Gulf, near Auckland, New Zealand. We estimated the size of the genome to be 320 Mbp with flow cytometry using methods described previously [81]. Total genomic DNA was extracted from a single individual with the DNeasy kit (Qiagen) using the animal tissue protocol, and the addition of a 3-min incubation with 0.02 mg of RNase A after the digestion step and then centrifugation for 3 min at 12,100 g to remove any remaining material. We sequenced the genome of H. duospinosa using the Illumina HiSeq 2000 sequencing platform. The sequencing libraries consisted of three paired end (PE) libraries with insert sizes of 188 bp, 200 bp and 470 bp and two mate paired (MP) libraries with 3 kbp and 5 kbp insert sizes. The paired end libraries were prepared using the Illumina TruSeq RNA kit and the mate pair libraries using the Illumina TruSeq DNA kit. These libraries were run on two lanes of an Illumina HiSeq2000 at New Zealand Genomics Ltd., Dunedin. Total RNA was extracted from a separate individual using Trizol (Invitrogen) following manufacturer’s instructions for the TRIzol Plus RNA Purification Kit. Four RNA extractions were made from antennae, head, thorax and abdomen. These RNA extractions were used to prepare four mRNA libraries using the Illumina TruSeq RNA kit and run on a single lane of an Illumina HiSeq2000 at New Zealand Genomics Ltd., Dunedin, New Zealand.

Genome and transcriptome assembly

The paired end (PE) reads were filtered for duplicate pairs, reads containing ambiguities (Ns), and then trimmed of adapter sequences and low quality ends using FastUniq (v1.1), PrinSeq (v0.20.3) and Cutadapt (v1.3), respectively [82–84]. Read pairs with at least one read less than 50 bp and unpaired reads (singletons) were discarded. Reads from the short-insert PE libraries (188, 200 and 470 bp) were decomposed into short sequences of length k (k-mer, with k = 17) using SOAPec (v2.03). These reads were then error corrected using ErrorCorrectReads.pl script from the ALLPATHS-LG (v46436) package [85] and the 188 bp and 200 bp PE libraries were merged into long single reads if a pair was detected with an overlap longer than 11 bp. A similar cleaning procedure was also applied to the two mate pair libraries. However, instead of error correcting, we retained 36 bases from the 5′ end of all cleaned reads in order to avoid disruption from the internal adapter sequences.

We used SOAPdenovo2 (vR223) [86] with a k-mer of 73 to perform the initial de novo assembly on reads from PE libraries. We then filled the gaps of the scaffolds using GapCloser (v1.12-r6) [86] and joined the resulting sequences with a standalone scaffolding program called SSPACE (v2.0) [87] using the same paired information. We then used SSPACE again to join the improved scaffolds into longer sequences with the two MP libraries and finally filled the gaps again with PE data.

We assessed the completeness of the assembled genome through three steps. First, to evaluate if the assembly has covered most of the sequencing reads, we mapped all PE reads back to the assembly using Bowtie2 (v2.2.0) [88]. Second, we mapped transcriptome RNA-Seq read pairs to the genome assembly in order to estimate how well the gene coding regions were assembled. Third, we estimated the completeness of the 1066 highly conserved Arthropoda genes (database: arthropoda_odb9) in the genome assembly using BUSCO (v2.0.1) [28]. Furthermore, the scaffolds were searched for homologues from GenBank nucleotide database (nt) using BLASTN (v2.2.28) [89] to determine whether contaminated sequences derived from microbes were present. The scripts from Assemblage (https://github.com/sujaikumar/assemblage) were used to assign the BLAST matches to different taxonomic categories.

We sequenced the H. duospinosa transcriptome to inform predicted gene models. The collembolan transcriptome included RNAseq libraries from the antennae, head, thorax and abdomen, sequenced together across two lanes of HiSeq 2000. The reads were pooled together for de novo assembly. Before reads were assembled, they were filtered and trimmed using a similar cleaning strategy to that applied to the genomic data, except the RNA-Seq reads were trimmed of 8 bases at the 5′ end before the rest of cleaning steps. The remaining high quality reads were then error corrected, before assembly using Trinity (r20140413p1) [90] with default options. The final transcriptome assembly was achieved after sequence redundancy was removed using CD-HIT (v3.1.1) [91] with a 95% identity threshold.

Genome and transcriptome annotation and comparative analysis

We searched for and classified repeats using RepeatModeler (v1.0.8) [92] and PASTEClassifier (v1.0) [93]. The program RepeatMasker (v4.0.5) [94] was used to mask the genome assembly prior to annotation for protein coding genes. We performed structural gene annotation with MAKER2 (v2.31.3) [95] on the repeat-masked genome assembly, integrating transcripts from the transcriptome assembly and conserved Arthropoda protein sequences to correct the predicted gene models. The whole pipeline was divided into several steps. First, the program Augustus [96] was trained using 248 predicted protein models together with 150 complete protein-coding transcripts determined by TransDecoder from the Trinity transcriptome assembly [90]. The trained gene structure parameters were then used by MAKER2 to predict gene structures. Second, the homology evidence provided to MAKER2 included the assembled transcriptome set, 3028 conserved arthropod protein models, which we downloaded from OrthoDB (v7). For the annotation of specific genes, sequences were identified by BLAST searches on assembled transcriptomes and the genome assembly. Where similar transcripts could not be identified, gene models generated by FGENESH6 [97] were used to identify partial regions of coding sequence. We searched for candidate horizontal transfer events as genes identified from the scaffolds that also contain host (insect) genes. We assigned taxonomic identity to each gene model from the ‘blast_taxonomy_report.sl’ using ASSEMBLAGE (https://github.com/sujaikumar/assemblage).

For annotation of RNA coding genes we used the cmsearch program from INFERNAL (v1.1.1) and corresponding covariance models (CMs) from the Rfam database (v12.0) [98, 99]. All matches above the curated GA threshold were included. INFERNAL was selected as the predictions it makes are the most accurate for ncRNAs that have been identified to date [100]. In order to refine the annotation of tRNA genes we ran tRNA-scan (v1.3.1) [101]. This method also uses CMs to identify tRNAs. However, it also uses some heuristics to increase the search-speed and annotates the isoacceptor type of each prediction. It also has a method to infer whether predictions are likely to be functional or tRNA-derived pseudogenes. Rfam matches and the tRNA-scan results for families belonging to the same clan were then “competed” so that only the best match was retained for any genomic region [102].

Protein sequences of genes known to be associated with sex-determination, particularly in insects and Daphnia (Additional file 6), were collected from UniProtKB and used as queries for TBLASTN searches against the H. duospinosa transcriptome and genome assemblies. The top BLAST hits with an E-value threshold of 1e⁻⁰⁵ were retrieved and used as queries for reciprocal BLAST searches against the NCBI non-redundant protein database to confirm putative orthology. The relationships of the doublesex (dsx) and sex-lethal (sxl) sequences to other known orthologues were tested using phylogenetic approaches. Briefly, the protein sequences (putative ORF identified from the transcript) of dsx and Sxl were aligned against known orthologues of various insects, as well as some Crustacea and Chelicerata (Additional files 4 and 5) using the online version of Mafft (v7) [103] with scoring matrix BLOSUM45 and default settings. The phylogenies of dsx and Sxl were rooted at the Ixodida and Branchiopoda, respectively. Sequence alignments of chitinase homologues were created using the Mafft plugin within Geneious v10.0.3 [104]. Phylogenies were reconstructed from alignments of the sex determination and chitinase proteins with the PhyML plugin [105] within Geneious using the JTT + Γ substitution model and 1000 bootstrap replicates.

Genome-wide heterozygous sites were determined by calling variants between the two haploid components in the diploid genome. Reads used for genome assembly construction were mapped to the H. duospinosa genome assembly using Bowtie2 (v2.2.0) [88]. Paired end reads were maintained if both pairs are concordantly mapped on one scaffold (> 97%). Variant-calling was performed with SAMtools (v1.2) and BCFtools (v1.2) [106, 107]. We filtered variants with low quality (QUAL < 30) and/or with excessive mapping depth (DP > 250). The remaining variants were then assigned to gene coding regions using custom python scripts. Exons of length shorter than 30 bp were filtered due to high occurrence of spurious annotation. All the heterozygous genes were then assigned to Gene Ontology (GO) terms using Blast2GO (v2.8) [108]. Enriched GO terms were calculated with the same program using Fisher’s Exact Test with multiple testing correction of false discovery rate [109] less than 0.05.

Nucleotide and dinucleotide content of gene body sequences (both full-length and uniformly truncated to 1 kb) and 1 kb fragments of the whole genome sequence were calculated using a custom Perl script. CpG[o/e] was calculated using the formula CpG[o/e] = (N*CpG)/(C*G), where N is the length of the genomic region, CpG is the number of CpG sites in the regions, and C and G are the numbers of cytosines and guanines in the region [110]. Calculated CpG[o/e] values were plotted as frequency distributions in R (www.r-project.org). The number of components underlying these distributions was estimated in R using mclust [111] model-based clustering. The best fitting model was identified among several non-nested models using Bayesian information criteria (BIC).

Phylogenetic analyses of whole transcriptomes

We inferred phylogenetic relationships from nine species (one proturan, two diplurans and six collembolans) including transcriptome data of eight species published previously [2] (current assembly version, see Additional file 7 and NCBI, 1KITE-Umbrella Bioproject ID 183205) and the transcriptome of H. duospinosa (Additional file 7). We first assigned assembled transcripts of each species to orthologous single copy genes (OGs) published by [2] using the orthology predicting programme Orthograph (v0.5.11) [112]. The published ortholog set comprises 1478 OGs and is based on 12 arthropod reference species (see [2], Table S3). Briefly, Orthograph generates profile hidden Markov models (pHMMs) from alignments of orthologous genes with a set of reference species. The pHMMs are then used to search transcript assemblies of query species for putative candidate orthologous sequences. Candidate sequences are then validated by a reciprocal BLAST using the official gene sets of reference species included in the ortholog set. Surviving hits are considered as ortholog transcript sequences. We chose a relaxed reciprocal BLAST search to any of the reference species. Other settings than default were “max-blast-searches” and “max-blast-hits” = 50; “extend-orf = 1”, and “substitute-u-with = X”. The latter avoids potential problems in downstream analyses because most programmes cannot handle selenocysteine (U). After summarising the results at the amino acid level, we masked all stop codons with “X”. Since we received no hit for two OGs from any of the query species and for another six OGs only hits for one species, we excluded these genes from further analyses. We then generated amino acid multiple sequence alignments (MSAs) for each OG (1470 OGs) using MAFFT L-INS-i (v7.123b) [103]. As described in [2], we checked each MSA for ambiguously aligned sequences (outliers) with the result of none being identified (see Result). For further downstream analyses we subsequently removed all sequences of the reference species from the alignments, leaving only sequences of the nine query taxa and deleted all columns containing only ‘X’ and/or ‘-’ (gaps).

We proceeded to design two datasets, one with partitions based on gene-boundaries and a second one with partitions based on protein domain-boundaries. For both datasets, we identified randomly similar aligned sites within each MSAs of each orthologous gene using a modified version of Aliscore (v1.2) [2, 9, 113], with the same settings described previously [2]. For the dataset based on gene-boundaries, we removed ambiguously aligned sections with the aid of Alicut (v2.3) [114], replaced terminal gaps by ‘X’, and concatenated masked MSAs into a supermatrix using FasConCat (v1.0) [115]. For the dataset based on protein domain-boundaries, we identified protein domains with the protein family database Pfam (v28, released 5 June 2015) [116], more specifically the Pfam-A pHMM library, following a procedure published previously [2] using the PfamScan software (v1.5, released 26 June 2015) [116] and HMMER (v3.1b1) (http://hmmer.org/). We then deleted ambiguously aligned sections from domain-based data blocks using the results of Aliscore and subsequently concatenated data blocks based on domain-boundaries into a supermatrix using custom Perl scripts. For each supermatrix, one with data blocks based on protein-domains (supermatrix A), the other with data blocks based on gene-boundaries (supermatrix B), we evaluated the information content (IC) of each data block with the software MARE (v0.1.2-rc) [117]. From both supermatrices, we removed data blocks with an IC = 0 and only kept data blocks for which all nine species were present.

For the selection of optimal meta-partitions and the best-fitting amino acid substitution models (see [2], Material and Methods, Section 3.6), we applied PartitionFinder (v2.0.0, prerelease 13) [118, 119] on both supermatrices in combination with RAxML (v8.2.3) [120]. We restricted the estimation of the best-fitting amino acid substitution model to LG [121], WAG [122], DMCUT [123], JTT [124], BLOSUM62 [125], each plus the alpha-shape parameter (+GAMMA) to account for among-site rate variation [126] and, in addition, listed models +Γ and using empirical base frequencies (+F). Moreover, we included the recently published free rate model LG4X [127] resulting in altogether 11 models. For the analyses we chose linked branch lengths and used the corrected Akaike information criterion (AICc, [128]) for final model selection. We applied the rcluster algorithm with the following settings: rcluster-max 10,000, rcluster-percent 100, all-states, min-subset-size 100, weights 1,1,0,1.

For maximum likelihood phylogenetic tree inference from both supermatrices, we applied IQTREE (v1.4.2) [129]. The search settings included 50 tree searches with the best meta-partition scheme and best-fitting model per meta-partition (option -spp), and using random starting trees for tree searches. For statistical support, we applied non-parametric bootstrap analyses (300 bootstrap replicates, partitioned bootstrapping). Finally, we plotted all bootstrap replicates on the ML tree with the best log LH value. We performed a SH-like approximate likelihood ratio test (see [105]) with 10,000 replicates on both data sets. We further checked how many unique topologies were present within the 50 inferred trees using the software Unique Tree (v1.9) (T. Wong, L. Jermiin, available upon request). For visualising and rooting the final tree with Diplura, we used Seaview (v4.2) [130]. We edited the tree graphically using Inkscape (v0.91) (https://inkscape.org).