Gene network analyses support subfunctionalization hypothesis for duplicated hsp70 genes in the Antarctic clam

A computationally predicted gene regulatory network (GRN), generated from mantle-specific gene expression profiles in the Antarctic clam Laternula elliptica, was interrogated to test the regulation and interaction of duplicated inducible hsp70 paralogues. hsp70A and hsp70B were identified in the GRN with each paralogue falling into unique submodules that were linked together by a single shared second neighbour. Annotations associated with the clusters in each submodule suggested that hsp70A primarily shares regulatory relationships with genes encoding ribosomal proteins, where it may have a role in protecting the ribosome under stress. hsp70B, on the other hand, interacted with a suite of genes involved in signalling pathways, including four transcription factors, cellular response to stress and the cytoskeleton. Given the contrasting submodules and associated annotations of the two hsp70 paralogues, the GRN analysis suggests that each gene is carrying out additional separate functions, as well as being involved in the traditional chaperone heat stress response, and therefore supports the hypothesis that subfunctionalization has occurred after gene duplication. The GRN was specifically produced from experiments investigating biomineralization; however, this study shows the utility of such data for investigating multiple questions concerning gene duplications, interactions and putative functions in a non-model species. Electronic supplementary material The online version of this article (10.1007/s12192-020-01118-9) contains supplementary material, which is available to authorized users.


Introduction
Improvement in the understanding of the genomic content of any non-model species is impeded by the lack of functional annotation. Annotation rates, particularly in environmental invertebrates, are often low, and many genes assembled within an experimentally derived transcriptome are either designated as unknown or require more detailed analysis to extract conservation at the domain level. Annotation can be particularly problematical when attempting to differentiate the putative functions and regulation of duplicated genes. With the decrease in sequencing costs and subsequent increase in amount and quality of RNA-seq data being produced, it is becoming feasible to use sophisticated information-theory analyses, such as ARACNe [(Algorithm for the Reconstruction of Accurate Cellular Networks), which require large datasets of around 100 non-replicated samples (Margolin et al. 2006)], to reverse engineer predicted gene regulatory networks for non-model species. Such networks can provide significant added value to gene annotations and putative functional assignments as they are calculated in an unbiased approach using mutual information obtained from quantitative expression profiles, rather than biased prior gene annotation or database-based predictions. Whilst these types of network calculations have been used to infer in vivo transcriptional regulation in the biomedical literature for some years (Theodoris et al. 2015;Walsh et al. 2017), the approach is only beginning to be applied to non-model organisms wellpoised to answer questions of an environmental or evolutionary nature (Antczak et al. 2015). A computationally predicted gene regulatory network (GRN) has recently been produced using ARACNe for the Antarctic clam Laternula elliptica (Sleight et al. 2019). This network was produced from mantle-specific shell damage-repair experimental data and, in addition to its contribution to the study of biomineralization, it also provides a resource to investigate questions related to cellular stress responses in L. elliptica, for example, the identification of putative differentiated functions of paralogous genes through examination of neighbouring genes.
Previous molecular analyses of the heat shock response of L. elliptica had identified a duplication of the inducible form of the 70-kDa heat shock protein [hsp70A and hsp70B, (Clark et al. 2008)]. Although identified as the inducible form of hsp70, both paralogues were constitutively expressed in animals under control conditions, where hsp70A expression was similar between tissues and hsp70B expression varied in a tissue-specific manner (Clark et al. 2008). The expression of both hsp70 paralogues was up-regulated in response to acute heat stress and hypoxia (Clark et al. 2008(Clark et al. , 2013(Clark et al. , 2017. Under each condition, the level of induction was gene-and tissuespecific, with the response to hypoxia revealing additional variation with age (Clark et al. 2013). The retention of duplicated genes within a genome is suggested to occur by the process of subfunctionalization, whereby each of the paralogues evolves additional functions, which are often tissue-or developmental-specific (Force et al. 1999). The production of a GRN for L. elliptica provided an opportunity to analyse the predicted regulation of hsp70 genes and further test the hypothesis that, following duplication, each hsp70 paralogue was retained due to subfunctionalization, and therefore are contained in separate GRN submodules with different putative functions.

Methods
Fragments of the duplicated inducible forms of hsp70 in the Antarctic clam L. elliptica were previously identified via degenerate PCR (Clark et al. 2008) and designated hsp70A (accession number AM293598.1) and hsp70B (accession number AM293600.1). In the present paper, these short fragments were BLAST searched against a recently published in-house mantle transcriptome database for L. elliptica (now available at MolluscDB: https://molluscdb.org/ (Caurcel C (2017)) to identify full-length transcripts for further investigation via the computationally predicted GRN (Sleight et al. 2019).
The two hsp70 paralogues were mapped to network nodes of the GRN that are termed here "clusters" of the GRN, using Cyctoscape (v3.7.1). The term cluster, rather than node, is used due previous methods of data processing. Briefly, 199,321 transcripts generated by a Trinity mantle transcriptome assembly were collapsed into 18,862 expression clusters with shared expression profiles using a selforganizing tree algorithm (SOTA); i.e. transcripts in the same cluster have tightly correlated expression over all the experimental samples in the dataset and were shown to be coexpressed in vivo using mRNA in situ hybridization (Sleight et al. 2019). A brief schematic overview of methods is available in the supplementary material (Supplementary information S1).
Putative divergence of function for each hsp70 paralogue was investigated via annotations of the transcripts in clusters within a network submodule, where submodules are defined here as the first-and second-neighbour connections to a cluster containing the hsp genes of interest. All transcripts in the GRN were previously annotated by Sleight et al. (2019) using tblastx again the NCBI non-redundant dataset (nr). In addition, in the present paper, annotations for each transcript within the hsp70 first-neighbour submodules were re-annotated via BLAST sequence similarity (tblastx) searching and interrogation of putative functional domains via InterPro (https:// www.ebi.ac.uk/interpro/). Finally, the transcripts in the firstneighbour clusters were further BLAST searched (blastx) against the UniProtKB/SwissProt human database (The UniProt Consortium 2019). The resulting first-neighbour human UniProtKB identifiers were entered into STRING-DB, an online server that uses database searches to predict potential protein-protein interactions and enrichment [significance, against a whole human genome background as human annotations were used, was obtained using a statistical overrepresentation Fisher's exact test, corrected for multiple testing using the Benjamin-Hochberg method to control the false discovery rate (FDR) https://string-db.org/ (Szklarczyk et al. 2017)]. The reconstructed GRN and associated submodules, assembled transcripts, and detailed annotations are all available as supplementary files in Sleight et al. (2019). The raw RNA-Seq data are freely available for download from NCBI SRA (accession PRJNA398984).

Results
When using the latest L. elliptica transcriptome as a database (Sleight et al. 2019), the previously designated hsp70A fragment (accession number AM293598.1) matched TRINITY_DN258255_c0_g12 at 99.69% identify, whilst the hsp70B fragment (accession number AM293600.1) matched TRINITY_DN246078_c0_g2 at 100% identify. At present only one full-length hsp70 gene is present for L. elliptica in the public databases (GenBank: ABM92345.1) (Park et al. 2007), which matched the Trinity transcript for hsp70A (TRINITY_DN258255_c0_g12, probability score = 1293.9, 100% amino acid identify, E value 0.0). When using public database searches (tblastx against nr) the Trinity transcript for hsp70B (TRINITY_DN246078_c0_g2) most closely matched the intertidal limpet Cellana toreuma HSP70 protein (probability score = 1031.6, 80.6% amino acid identify; 91.8% amino acid similarity, E value 0.0). Comparison of the translated products of these two full-length hsp70 transcripts in L. elliptica revealed 71.7% identity and 84.7% similarity between them at the protein level. Evaluation of these protein sequences showed that both possess all the signature motifs for the HSP70 family (Supplementary information S2). Each hsp70 paralogue was then mapped onto the GRN, hsp70A was located in cluster7829, whilst hsp70B mapped to cluster9844 (Fig. 1).
hsp70A-first neighbours: Cluster7829 containing hsp70A includes another transcript, annotated as the hypothetical protein CAPTEDRAFT_177551. Domain annotation of this protein revealed that this was a member of the universal ribosomal protein uL3 family, in the cosmopolitan worm Capitella teleta. The hsp70A cluster7829 had eight first-neighbour clusters ( Fig. 1-orange clusters). Analysis of the annotations associated with the genes in these first-neighbour clusters had functions related to ribosome biogenesis and RNA processing (Supplementary information S3). Database-based STRING-DB analysis of first-neighbour annotations suggested that hsp70A directly interacted (with medium confidence or higher) with four of the eighteen genes identified in the first-neighbour clusters (Fig. 2). These comprised MRPL3 (mitochondrial protein L3), PDCD11/RRP5 (essential for generation of mature 18s rRNA), PSMC2 (proteasome component, involved in protein degradation) and PRPF38A (PRP38 pre-mRNA processing factor). hsp70A's first neighbours were functionally enriched in RNA processing, ribosome biogenesis and rRNA processing (Supplementary information S4).
hsp70B-first neighbours: Cluster9844 containing hsp70B also had eight first-neighbour clusters ( Fig. 1-magenta clusters), and the cluster contained a single transcript-hsp70B. The first neighbours of cluster9844 did not show a clear functional pattern, and annotation was generally at the domain level (Supplementary information S3). Many annotations were for PDZ domains and two with GPCR (G-proteincoupled receptor). The most notable first-neighbour annotation was contained in cluster7651 with the predicted protein Hsp7012A in Crassostrea gigas. STRING-DB analysis of the first-neighbour annotations of hsp70B showed no expansive network, with only one protein contained in the firstneighbour clusters identified in the GRN directly interacting with hsp70B, namely HSPA12B. There were no other significant interactions, which could have been partly due to lowquality BLAST mapping between the L. elliptica transcriptome and the human proteome. The median scores, coverage and median identities between the two hsp clusters were notably different: 486, 64%, 55% (7829 and first neighbours) and 129, 21%, 47% (9844 and first neighbours) respectively.
Linkage between hsp70A and hsp70B-second neighbours: Interrogation of the second-neighbour annotations provided further evidence that hsp70A is involved in ribosome indicates all colour and shape coding for the edges and clusters, which are labelled with a representative transcript NCBI nr annotation biogenesis ( Fig. 1-yellow clusters, such as, nucleolar protein 8-like, nucleolar GTP-binding protein 1, nucleolar GTPbinding protein 2). hsp70B's second-neighbour annotation ( Fig. 1-purple clusters) provided an additional layer of functional data with many more annotations than the firstneighbour analysis linking to several signalling pathways (such as MAP kinase kinase 7-like, insulin-like growth factor binding protein 5, noggin-like) and transcription factors (such as homeobox protein six1-like and zinc finger protein squeezelike), in addition to response to cellular stress (such as glutathione peroxidase-like and cdc42 homolog) and cytoskeletal trafficking and remodelling (such as protein bicaudal D-like and tubulin alpha-1 chain). Although in separate submodules, the hsp70 paralogues were linked via a mutual second neighbour within the GRN (Fig. 1-single grey cluster). Annotation of the genes linking the second-neighbour networks of each paralogue (Fig. 1-diamond/square clusters) was related to protein and nucleic acid binding, whilst some evidence of involvement in ribosome biogenesis was also seen, namely in clusters8788 and 7959. The latter of which being the key mutual second neighbour linking the two firstneighbour networks, and contained an SDA1 homolog, which is required for the export of 60S pre-ribosomal subunits (Supplementary information S4).

Discussion
Interrogation of the GRN clearly identified hsp70A and hsp70B in different clusters (7829 and 9844 respectively), with discrete first neighbours. Analysis of the annotations associated with genes within first-and second-neighbour submodules revealed differences in the putative gene regulation and interaction of each hsp70 paralogue. In particular, hsp70A was strongly associated with genes related to ribosome biogenesis and RNA processing. These data were supported by the STRING-DB and enrichment analyses and agree with Truebano et al. (2010), who reported significant changes in expression of ribosomal proteins in L. elliptica under heat stress. Other heat stress studies in the goby Gillichthys mirabilis and the Antarctic fish Trematomus bernacchii reported the same significant variation in genes associated with ribosome biogenesis and protein synthesis (Buckley et al. 2006;Buckley and Somero 2009). Whilst up-regulation of transcriptional machinery, including ribosomal proteins, is often associated with the stress response, another reason for this is thought to be that ribosomal proteins have a stabilizing role in the ribosome (e.g. Beck and De Maio 1994;Cornivelli et al. 2003). Any repression or expression due to heat stress is an effort to protect ribosomal structure and function by replacing damaged regions. Indeed, previous experiments on L. elliptica have shown a strong reaction of hsp70A not only to heat stress but also to hypoxia and ocean acidification (Park et al. 2007;Clark et al. 2008Clark et al. , 2013Clark et al. , 2017Cummings et al. 2011).
In contrast, the analysis of hsp70B provided less defined first-neighbour interactions. Annotation rates were poor for the first-neighbour clusters and were often at the domain level rather than putative orthologous genes in other species. The most numerous domain identified was PDZ (also known as Discs-large homologous regions), which is often found in diverse membrane-associated proteins with signalling functions (Ponting et al. 1997). Two other annotations within these clusters were to GPCRs, which are membrane-associated receptors involved in signal transduction. The first-neighbour cluster annotations included hsp70 12A in Crassostrea gigas. This is an atypical hsp70 gene family, which in vertebrates is not up-regulated in response to stress (e.g. Han et al. 2003); it does, however, appear to have been recruited to the stress response in bivalves. In particular, this gene has been subject to massive duplication in the oyster (Crassostrea) lineage and is up-regulated in response to several stresses, including heat and xenobiotics (Zhang et al. 2012;Luchmann et al. 2015). The second-neighbour connections for hsp70B also showed strong evidence for cellular stress responses, with many genes being involved in signalling, for example the MAPK-signalling, which has previously been shown to interact with Hsp70 during muscle regeneration (Fan et al. 2018). In addition, an oxidative stress response gene, glutathione peroxidase-like, was one of hsp70B's second neighbours, as well as many genes associated with cytoskeletal trafficking and remodelling. It has long been thought that molecular chaperones such as hsp70 recognize cytoskeletal elements and can both modulate assembly and provide protection (Liang and MacRae 1997). Given the large number of cytoskeletal genes we found in the GRN submodule of hsp70B, it is likely that hsp70B is playing a role in protecting elements of the cytoskeleton in L. elliptica.
The two hsp70 paralogues were linked in the GRN by a mutual second neighbour, which also contained transcripts involved in ribosome biogenesis and MAPK-signalling. Combining these GRN data with known molecular responses we find that both paralogues are significantly involved in the L. elliptica stress response, and in addition, it is seems likely that subfunctionalization has partitioned their potential secondary functions towards ribosome protection (hsp70A) and signalling and cytoskeleton protection (hsp70B). Hence, despite the known caveats of using primarily in silico methods, the information from the GRN provides valuable clues as to potential additional functions of these inducible hsp70s in L. elliptica, which can be targeted, and tested more rigorously, in future experiments.
Acknowledgements This work was carried out by AR as part of her WRoCAH Researcher Employability Project (award reference REP0173) with the British Antarctic Survey between 09 May 2018 and 10 May 2018. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.