Combined Transcriptomic and Proteomic Analysis of Perk Toxicity Pathways

In Drosophila, endoplasmic reticulum (ER) stress activates the protein kinase R-like endoplasmic reticulum kinase (dPerk). dPerk can also be activated by defective mitochondria in fly models of Parkinson’s disease caused by mutations in pink1 or parkin. The Perk branch of the unfolded protein response (UPR) has emerged as a major toxic process in neurodegenerative disorders causing a chronic reduction in vital proteins and neuronal death. In this study, we combined microarray analysis and quantitative proteomics analysis in adult flies overexpressing dPerk to investigate the relationship between the transcriptional and translational response to dPerk activation. We identified tribbles and Heat shock protein 22 as two novel Drosophila activating transcription factor 4 (dAtf4) regulated transcripts. Using a combined bioinformatics tool kit, we demonstrated that the activation of dPerk leads to translational repression of mitochondrial proteins associated with glutathione and nucleotide metabolism, calcium signalling and iron-sulphur cluster biosynthesis. Further efforts to enhance these translationally repressed dPerk targets might offer protection against Perk toxicity.


Introduction
As we age, neurodegenerative diseases are becoming a prominent health problem worldwide. Despite large-scale efforts to treat these diseases, current therapies only offer symptomatic relief without the possibility of curative treatments. The unique genetic, pathological and clinical signatures of individual neurodegenerative diseases have focused scientific research on the disease-specific protein-centric mechanism of pathology. However, it is becoming increasingly clear that neurodegenerative diseases share a common subcellular signature. At present, the unfolded protein response (UPR) is perceived to be a generic feature of most neurodegenerative diseases (reviewed in [1]). Modulation of the UPR therefore represents a promising target for therapies designed to delay or prevent neurodegeneration (reviewed in [2]).
The endoplasmic reticulum (ER) represents a dynamic network of interconnected sheets and tube-like structures [3] in eukaryotic cells. The primary function of the ER is to act as a factory for proteins, coordinating the synthesis, folding, maturation and post-translational modifications of proteins. Certain pathological conditions, such as neurodegeneration, are characterised by an overproduction of unfolded or misfolded proteins leading to ER stress. To combat this problem, the ER activates an evolutionarily conserved adaptive response known as the UPR [4], which acts to restrict the protein folding load via translation repression [5] in concert with the transcriptional activation of specific genes encoding ER-resident chaperones and the ER-associated degradation machinery (reviewed in [6]).
The UPR is transduced via three protein kinases, namely, protein kinase RNA (PKR)like ER kinase (PERK), inositol-requiring enzyme 1 (IRE1) and activating transcription factor 6 (ATF6), activated by the presence of unfolded proteins in the ER. While ATF6 and IRE1 primarily act via induction of a transcriptional programme [7][8][9], PERK activation leads to an overall translational shutdown, mediated by the phosphorylation of the eukaryotic translation initiator factor-2 α (eIF2α) [5]. The eIF2 complex is responsible for the methionyl-tRNA delivery, which is necessary for the initiation of protein synthesis by ribosomes. Phosphorylated eIF2α (phospho-eIF2α) can competitively bind eIF2, thereby inhibiting the GTP hydrolysis required for the initiator methionyl tRNA to base pair with the initiation codon. Consequently, the cap-dependent translation is inhibited at the initiation step due to the failure of the 80S ribosome complex to assemble (reviewed in [10]).
During ER stress, reduced protein synthesis serves to conserve energy and resources and thus enable cells to reconfigure gene expression to alleviate stress damage. However, the synthesis of some molecules is required to cope with the perturbed proteostasis, as their protein products are necessary for the resolution of ER stress. Thus, phospho-eIF2α also leads to the preferential translation of specific transcripts that facilitate adaptation to a specific stress condition, suggesting that the UPR requires translational reprogramming via alternative mechanisms of translation initiation (reviewed in [11]). In mammalian cells, phospho-eIF2α is required for the regulated expression of several proteins, such as activating transcription factor 4 (ATF4) and C/EBP homologous protein (CHOP) as well as mRNAs encoding heat shock and UPR proteins, such as ER chaperone binding immunoglobulin protein (BiP) and eIF2α phosphatase growth arrest and DNA damageinducible protein (GADD34) [12][13][14][15]. While these alternative mechanisms of translation have not been fully elucidated to date, studies suggest that 5 untranslated region (UTR) cis-acting sequences, such as upstream open reading frames (uORFs) and internal ribosome entry sites (IRES), confer the 'privileged' translation of these mRNAs in spite of eIF2α inhibition (reviewed in [16]).
Mutations in PINK1 or PARKIN, two genes in mitochondrial quality control cause neurodegeneration in some autosomal recessive forms of Parkinson's (PD) disease (reviewed in [17]). In Drosophila melanogaster, mutations in either pink1 or parkin cause mitochondrial dysfunction and the degeneration of dopaminergic neurons [18].
We have recently shown that in Drosophila pink1 or parkin mutants, defective mitochondria also give rise to ER stress signalling by activating the Perk branch of the UPR. Enhanced ER stress signalling in pink1 and parkin mutants is mediated by contacts between defective mitochondria and the ER [19] and causes the upregulation of transcripts coding for serine hydroxymethyl transferase (Shmt2) and NAD-dependent methylenetetrahydrofolate dehydrogenase (Nmdmc) in both cultured human cells and flies. Shmt2 and Nmdmc are targets of dAtf4 involved in mitochondrial one-carbon metabolism [20]. When overexpressed in pink1 or parkin mutants, either Shmt2 or Nmdmc is neuroprotective [20].
These observations indicate that first, transcripts belonging to one-carbon metabolism that are targets of Perk signalling are upregulated in fly models of neurodegeneration, and second, these transcripts need to be overexpressed to confer protection. Both observations lead to the hypothesis that Drosophila Perk (dPerk) target genes that are upregulated at the transcriptional level but fail to be translated are involved in neuroprotective pathways. To investigate this possibility, we interrogated the full cohort of transcripts and proteins altered by dPerk expression with the aim of identifying the full cohort of genes that might be upregulated at the transcript level but not at the protein level. Such a cohort can comprise potential protective pathways such as the mitochondrial one-carbon pathway.
Therefore, we combined microarray analysis of transcripts and tandem mass tagging (TMT) proteomic analysis, in adult flies overexpressing dPerk. We set to functionally characterised upregulated transcripts and their corresponding changes in proteins via pathway enrichment using a network analysis approach and upstream analysis (Cytoscape with ClueGO, STRING and iRegulon). In this study, we show that dPerk upregulation promotes the translation of a limited set of mRNAs, that function to resolve stress; these mRNAs include transcripts from the mitochondrial chaperone Heat shock protein 22 (Hsp22). Furthermore, we have identified a subpopulation of translationally repressed mitochondrial proteins that are suggested to be involved in nucleotide metabolism, calcium signalling and sulphur compound metabolic processes. Our study provides an insight into the dynamic nature of Perk signalling and identifies potential neuroprotective targets, whose roles in ER stress merit further research.

Drosophila tribbles Is Regulated by dPerk and dAtf4
To study the cohort of transcriptional and protein changes in adult flies expressing dPerk we utilised the GAL4-UAS gene expression system adapted with a temperaturesensitive GAL80 protein (GAL80 ts ), expressed ubiquitously from the tubulin 1α promoter [21]. At the permissive temperature of 18 • C, ubiquitously expressed GAL80 protein is functional, and acts as a repressor of the GAL4 transcriptional activator by binding to the GAL4 protein. At restrictive temperatures above 29 • C, GAL80 can no longer bind GAL4 and its repressive function is therefore lost, thereby enabling the expression of the gene target placed under the control of the upstream activation sequence (UAS). Therefore, we generated a UAS fly line expressing an HA-tagged version of dPerk. Flies were raised at 18 • C and heatshocked for 15 h at 29 • C following eclosion, enabling temporally controlled ubiquitous (daGAL4) expression of dPerk in adult flies, and they were processed for transcriptomics and proteomics analysis ( Figure 1a). We chose to heat-shock the flies for 15 h, as this timepoint provided the highest level of detected dPerk expression by assessment of its mRNA levels ( Figure 1b) and corresponded to a peak of expression of Nmdmc, a dAtf4 target (Figure 1c) [20].
We next confirmed the kinase activity of dPerk by examining the levels of its downstream target eIF2α. We compared the phosphorylation levels of eIF2α by detecting phospho-eIF2α by Western blotting. The overexpression of dPerk caused an increase in phospho-eIF2α levels, when compared to controls expressing either a kinase dead version of dPerk (K671R) or driver alone following a 15 h heat-shock ( Figure 1d).
We have previously characterised a subnetwork of transcriptional changes present in pink1 and parkin mutant flies, associated with the activation of the dPerk signalling pathway [20]. In this analysis we found in silico evidence for the induction of the tribbles pseudokinase 3 (TRB3), a negative regulator of ATF4-dependent transcription [22]. As we detected an upregulation of Drosophila tribbles (trbl) in our transcriptomics analysis (Supplementary Table S1), we next utilised quantitative real-time PCR (qRT-PCR) analysis and confirmed that dPerk expression caused an upregulation of trbl ( Figure 1e). In mammalian cells, TRB3 is an ER stress marker and a target of ATF4 [23,24]. Therefore, we next tested the effects of downregulating dAtf4 in flies expressing dPerk. We observed that downregulation of dAtf4 blocked the increase in the mRNA levels of trbl caused by dPerk expression (Figure 1f). We conclude that trbl is a downstream target of dAtf4 regulated by dPerk in adult flies.  To identify differentially expressed targets, transcripts and proteins were filtered by adjusted false discovery rate (FDR) and fold change (FC) values of 0.05 and ±1.6, respectively. The expression levels of transcripts and proteins were compared and further submitted for enrichment and upstream analysis. (b,c) dPerk or Nmdmc mRNA induction were assessed for different durations of the heat shock treatment (in hours (h)). The figure shows fold-change mRNA levels compared to the control, measured by real-time qPCR. (d) Western blot analysis of total and phospho-eIF2α protein levels in dPerk-HA, kinase-dead dPerk (KD)-HA and control flies following 15 h heat-shock at 29 • C. Whole-fly lysates were analysed using the indicated antibodies. (e) Tribbles (trbl), an endoplasmic reticulum stress marker, was upregulated in dPerk flies following 15 h of heat shock, as measured by real-time qPCR (mean ± SEM; asterisks, unpaired t-test). (f) Trbl overexpression is dAtf4 dependent and requires dPerk kinase function (15 h heat shock at 29 • C), as measured by real-time qPCR (mean ± SEM; asterisks, one-way ANOVA with Tukey's multiple comparison test). Genotypes: Control: tubGal80; daGal4 > +, dPerk: tubGal80; daGal4 > dPerk-HA, dPerk(KD): tubGal80; daGal4 > dPerk-K671R-HA, dPerk, dAtf4 RNAi: tubGal80; daGal4 > dPerk-HA, dAtf4 RNAi.

Divergence between the Upstream Transcriptional Regulators of dPerk-Induced Alterations in Transcripts and Proteins
Next, to assess the global cohort of dPerk-dependent alterations in mRNA and protein levels in adult flies we analysed the transcriptomic and proteomic dataset obtained from flies expressing dPerk using the workflow shown in Figure 1a. Microarray analysis, identified 32,500 distinct mRNA molecules (transcripts) that matched a total of 15,635 genes (Supplementary Table S1). We identified differentially expressed transcripts based on the predetermined thresholds set at ±1.6 fold-change for the FC value and 0.05 for the false discovery rate (FDR) value. This approach identified a total of 977 upregulated and 1022 downregulated transcripts, matching to 517 and 642 genes, respectively (Supplementary Tables S2 and S3). dPerk activation results in global translational repression, modulating protein synthesis under conditions of ER stress [25]. To determine how dPerk altered protein levels in flies we used quantitative proteomic analysis using the TMT technique.
Flies subjected to TMT proteomics were kept in conditions identical to those used for transcriptomics. Quantitative proteomic analysis identified 5795 proteins (Supplementary Table S1). We applied the same cut-off values as in the transcriptomic dataset (±1.6 foldchange for the FC value and 0.05 for the FDR) which yielded a list of 100 upregulated and 145 downregulated proteins (Supplementary Tables S4 and S5).
To understand the main transcriptional drivers of dPerk-mediated upregulation in transcripts and proteins we next used iRegulon, a computational method designed to reverse engineer the transcriptional regulatory network underlying coexpressed genes [26]. By subjecting the cohort of upregulated transcripts to iRegulon analysis we identified ATF4 as the primary transcriptional driver of the upregulation of mRNAs in dPerk expressing flies (Figure 2a,b, Supplementary Table S6). However, when an identical analysis was performed on the cohort of upregulated proteins, ATF4 failed to achieve the top score ( Figure 2a, Supplementary Table S7). We conclude that, as predicted, ATF4 is the primary driver of dPerk-induced transcriptional changes in adult flies. However, the protein increases observed in these flies are induced by more complex regulatory mechanisms.

Pathway Analysis of dPerk-Dependent Alterations in Transcripts and Proteins
We subsequently compared the differentially expressed transcripts and proteins independently. We performed pathway enrichment analysis (PEA) using ClueGo, a Cytoscape plug-in that integrates Gene Ontology (GO) terms as well as KEGG and Reactome pathways creating functionally organised GO/pathway term networks [27]. This analysis involved separate right-sided hypergeometric tests, assessing GO term and pathway enrichment, for the upregulated and downregulated transcripts or proteins.
dPerk overexpression caused a transcriptional enrichment in UPR-related terms such as unfolded protein response, endoplasmic reticulum and recycling of eIF2:GDP that consist of canonical ER stress proteins such as Ire1, dPerk and eIF-2α (Figure 3a, Supplementary Tables S2 and S8). We also detected an enrichment of different aminoacyl-tRNA synthetases and terms closely linked to translation, such as tRNA, cellular amino acid and peptide metabolic processes. The terms related to drug and glutathione metabolism as well as haem binding were also significantly enriched, comprising glutathione S-transferases (GSTs), UDP-glycosyltransferases (UGTs), as well as Cytochrome P450 enzymes (CYPs). We also observed an upregulation of more than 50 mitochondrial transcripts ( Figure 3a).
Functional analysis of upregulated proteins ( Figure 3b, Supplementary Tables S4 and S9) indicated enrichment of two UPR proteins, dPerk and Ire1. We also observed upregulation of proteins that modulate transcriptional and translational elongation. In keeping with the results of the transcriptomic analysis, we observed enrichment in proteins involved in detoxification processes (i.e., drug metabolism, metabolism of xenobiotics by cytochrome p450 and metabolism of steroid hormones). We also observed enrichment of members of the serpin family that act as inhibitors of proteases required for the melanisation reaction, a Drosophila specific immune response [28,29].
Both downregulated transcripts ( Figure 3c, Supplementary Tables S3 and S10) and proteins ( Figure 3d, Supplementary Tables S5 and S11) were enriched in terms related to metabolism, specifically metabolism of carbohydrates and lipids. This result indicates that dPerk activation promotes the inhibition of several metabolic pathways.

Analysis of the dPerk-Induced Correlations between Transcript and Protein Levels
The UPR leads to an induction of a transcriptional programme that does not lead to a corresponding increase in translation, as transcripts are unable to produce their protein products due to a stall in cap-dependent translation. For example, we observed induction of the dAtf4 target gene Nmdmc, which was proposed to be neuroprotective in pink1 or parkin mutants but observed that it was necessary to overexpress this gene to achieve neuroprotection [20]. This finding suggests that an increase in mRNA levels does not necessarily reflect a protein upregulation. To assess the relationship between the induction of mRNA and protein levels we compared the subgroups of transcripts matched to proteins ( Figure 4, Supplementary Table S12). We defined two subgroups within the upregulated transcripts, first, genes that were upregulated at both the transcript and protein levels (group 1, Supplementary Table S13), and second, genes that were upregulated at the transcript level but not at protein level (group 2, Supplementary Table S14), as well as a subgroup of downregulated transcripts and proteins (group 3, Supplementary Table S15).
Next, we analysed the subgroups of upregulated transcripts. Our hypothesis is that genes belonging to group 2, consist of failed protective mechanisms. These factors, such as Nmdmc, might represent target pathways to achieve cellular protection from ER toxicity. The comparison between group 1 and group 2 genes revealed that Nmdmc transcript is in fact upregulated by dPerk, but overexpression of this kinase failed to induce an increase in Nmdmc protein levels ( Figure 4).
We first set to examine group 1 targets. We established 27 targets that demonstrated upregulation in both transcripts and proteins (Table 1, Supplementary Table S13). Our results show that the expression levels of dPerk are upregulated after a 15 h heat-shock, thus validating our model. Resident ER proteins included Ire1, Asparagine synthetase (AsnS) and Sil1 nucleotide exchange factor (Sil1). Both Ire1 and AsnS have previously been reported to be upregulated in response to ER stress [30,31]. Sil1, a GrpE-like protein, has been shown to function as a nucleotide exchange factor for an ER luminal chaperone BiP [32], and is required for protein translocation and folding in the ER [33].  . The DEA shows a proportion of genes consistently regulated between the transcriptome and proteome (Pearson's correlation, r = 0.5, p-value < 2.2 × 10 −16 ) and a number of differentially regulated transcripts and proteins. The two subgroups of genes with upregulated transcripts, representing potentially protective pathways are indicated. Group 1 represents dPerk-dependent targets that show upregulation in both transcripts and proteins; group 2 represents translationally repressed targets that only show an upregulation of transcripts, with the absence of a change in protein levels. Group 3 represents targets that are repressed at both the transcript and protein level. dPerk and Nmdmc are indicated by black lines. Our analysis further identified two GTPases, Dgp-1 and CG2017, homologous to GTPbinding-protein 1 (GTPBP1) and GTPBP2, that are predicted to be involved in translational elongation. In Drosophila, Dgp-1 appears to play a role in the cellular defence against oxidative stress and has been found to be upregulated in parkin mutants [34], as well as in flies challenged with an inducer of oxidative stress, paraquat [35] and hyperoxic conditions [36].
Three mitochondrial proteins were also present, a mitochondrial chaperone Hsp22, Phosphoenolpyruvate carboxykinase 2 (Pepck2) and an uncharacterised protein CG34423. CG34423 is a suggested orthologue of the ATP synthase inhibitory factor subunit 1 (ATP5IF, 66% similarity) and has been found to be upregulated by Hsp22, suggesting that these two proteins might be functionally related [37].
The phosphorylation of eIF2α leads to the preferential translation of specific transcripts via alternative mechanisms of translation initiation such as uORFs and IRES (reviewed in [16]). We next investigated the degree of uORFs present in dPerk targets, detected at both transcript and protein level. We found that 43% of genes (2425 out of 5586) have uORFs (Figure 5a, Supplementary Table S16), consistent with what was previously reported [38]. Furthermore, 7 out of 26 group 1 targets (dPerk was excluded) (Figure 5b), and 40 out of 102 group 2 targets contain at least one uORF (Supplementary Tables S17 and S18).  Using a Monte Carlo-based approach, we determined that the number of genes in group 1 is a true positive whereas the number of genes from group 2 does not significantly differ from random sampling from the total population.
The impact that the uORFs can have on translation depends on many features, including their abundance or length (reviewed in [39]). However, upon inspection of these features we found no difference between the two groups (Figure 5c,d). We further investigated presence of IRES by examining the IRES database [40], but did not find any Drosophila targets with an IRES site within group 1.

Small Mitochondrial Chaperone Hsp22 Is a Novel dPerk/dAtf4 Signalling Target
Studies have shown that ER stress is able to modulate mitochondrial proteostasis and dynamics [37], as well as mitochondrial bioenergetics [41], in order to confer cellular adaptation to stress. Mitochondria and the ER communicate through mitochondrial ER contacts (MERCs), structural microdomains of parts of ER membranes in close apposition to mitochondrial outer membrane domains to regulate cell metabolism (reviewed in [42]). PERK has been identified as a component of MERCs [43]. In addition, ATF4 has been found to modulate mitochondrial stress responses [23]. All of these findings suggest that ER stress can modulate mitochondrial function.
Our analysis of group 1 members identified the mitochondrial Hsp22 to be upregulated at both transcript and protein levels. Hsp22 overexpression has been shown to increase Drosophila lifespan and resistance to oxidative stress, and its role in ageing suggests that Hsp22 helps to maintain mitochondrial integrity and function [44,45]. Hsp22 overexpressing flies display upregulation of genes related to mitochondrial energy production and protein biosynthesis, including genes of the mitochondrial oxidative phosphorylation system (OXPHOS) [46].
Analysis of Hsp22 upstream regulators by iRegulon identified ATF4 as a potential transcriptional regulator of this mitochondrial chaperone (Figure 6a).  Therefore, we subsequently confirmed that dPerk expression caused an upregulation of Hsp22 by qRT-PCR analysis (Figure 6b). We also found that downregulation of dAtf4 blocked the increase in the mRNA levels of Hsp22 caused by dPerk expression (Figure 6b). We conclude that Hsp22 is a downstream target of dAtf4 regulated by dPerk in adult flies. We propose that Hsp22 may function to modulate mitochondrial function in response to ER stress.
We next examined group 2, defined by an upregulation in transcript levels without a corresponding change in protein levels. Group 2 consisted of 102 molecules (Figure 7, left, Supplementary Table S14). The ClueGo GO Cellular Component PEA right-sided hypergeometric test identified a strong enrichment in mitochondrial components, made up of 20 different mitochondrial targets including Nmdmc (Supplementary Table S19). In addition to the enriched mitochondrial terms shown in Figure 7 (left), we observed enrichment in terms related to the metabolism of nucleotides as well as glutathione, drug and xenobiotic metabolism (Supplementary Table S20).
Finally, we looked at the subset of genes with downregulated transcript and proteins (group 3). This included 75 transcripts corresponding to 51 proteins (Supplementary Table S15). The ClueGo PEA right-sided hypergeometric test showed an enrichment in terms related to carbohydrate and triglyceride metabolism (Figure 7, right, and Supplementary Table S21), as observed in Figure 3c,d. Terms hydrolysing O-glycosyl compounds, lysosomal oligosaccharide catabolism and monosaccharide metabolic process display the highest significance, consisting of predominantly lysosomal alpha-mannosidases and maltase enzymes, which catalyze the catabolism of different carbohydrates or carbohydrates of glycoproteins [47,48].

Networks of Translationally Repressed Signalling Pathways Caused by dPerk Expression
To define functional networks within group 2 and visualise the connectivity between molecules (nodes) we next used the STRING database. STRING collects, scores and integrates publicly available sources of protein-protein interaction information, complimenting them with computational predictions to provide comprehensive functional networks [49].
To reduce the complexity of our network we used the Markov cluster algorithm (MCL), a clustering algorithm for graphs. MCL clustering is based on the random walk principle and is used for the large-scale detection of protein clustering [50]. Using MCL we deconstructed the group 2 protein network until we defined functional cliques, clusters in which every node was connected to every other node in that cluster (Figure 8, red circles). Whereas protein molecules provide more information regarding the functionality inside the cell, the techniques currently used for proteome analysis are not at the scale of the methods used for the transcriptome, which are highly suitable for large-scale analyses. To circumvent this problem, we reconstructed our networks by reintroducing the predefined cliques into the background transcriptomics network and investigating the clique first neighbours to enrich for biological information lost by missing proteomics data (Figure 8, grey and green circles). After controlling for MCL cluster pathway specificity with stringApp functional enrichment analysis and through a literature review, we established three subnetworks within group 2 ( Figure 8).  Figure 8. String network analysis of dPerk-dependent targets differentially regulated by the transcriptome and proteome. Genes upregulated at the transcript level but not at the protein level were analysed (group 2, red). Clusters were determined using the inbuilt Cytoscape stringApp Markov Cluster (MCL) algorithm. Clusters were enriched with first neighbours from the background transcriptomic network in order to find additional functionally related targets not recognised by the TMT labelling. The nodes representing molecules with an upregulation in transcript levels but no available protein value (grey) and molecules with an upregulation in transcripts and downregulation in protein levels (green) were added. Molecular pathways involved in glutathione and nucleotide metabolism (a), processing of essential mitochondrial calcium uniporter regulator (EMRE) (b) and iron-sulphur metabolism (c) are shown.
MCL clustering identified a large network consisting of two smaller highly connected clusters (Figure 8a). We found a cluster of GSTs, GstD3, GstD9, GstE7, GstE9, GstE1 and GstT1, further interacting with Ugt36Bc, Cyp6a20, Cyp4p1 and chaperone Hsp67Bb. Notably, Gclc, the glutamate-cysteine ligase catalytic subunit, a transcript encoding for the first rate-limiting enzyme of glutathione synthesis, was also on our list. The tetrahydrofolate dehydrogenase/cyclohydrolase Nmdmc was also present in our other cluster together with other molecules related to nucleotide metabolism, a glycine dehydrogenase (GLDC) orthologue transcript CG3999, a phosphoserine phosphatase transcript astray (aay) of protein that catalyses the last step in the biosynthesis of serine from carbohydrates, GART trifunctional enzyme (Gart), which encodes a protein that exhibits phosphoribosylamine-glycine ligase, phosphoribosylformylglycinamidine cyclo-ligase and phosphoribosylglycinamide formyltransferase activities and Phosphoribosylamidotransferase 2 (Prat2), which encodes type-2 glutamine amidotransferase that is essential in the pathway for de novo synthesis of inosine monophosphate (IMP). Transcripts of enzymes tightly linked to serine and threonine metabolism, amino acids fed into the one-carbon metabolism, Seryl-tRNA synthetase (SerRS) and Threonyl-tRNA synthetase (ThrRS), were also detected in our cluster. Next, using qRT-PCR analysis, we confirmed that the transcripts for Nmdmc, CG3999, aay and Gart were significantly upregulated by dPerk overexpression (Figure 9a).
Our analysis also identified a small cluster of genes involved in the processing of single-pass membrane protein with aspartate rich tail 1 (SMDT1) or essential mitochondrial calcium (Ca 2+ ) uniporter (MCU) regulator (EMRE) (Figure 8b). EMRE is an essential regulatory subunit of the MCU that mediates Ca 2+ uptake into mitochondria and is required to bridge the Ca 2+ sensing proteins mitochondrial calcium uptake 1 (MICU1) and MICU2 with the Ca 2+ conducting subunit MCU [51,52]. Our cluster identified Paraplegin (Spg7) and AFG3 like matrix AAA peptidase subunit 2 (Afg3l2), subunits of m-AAA proteases [53,54] and lethal (2) 37Cc (l(2)37Cc), orthologue of human prohibitin (PHB) and CG8728, PMPCA orthologue and Prohibitin 2 (Phb2), which were transcriptionally upregulated, however, their corresponding protein products were not detected. qRT-PCR analysis further confirmed that dPerk overexpression resulted in transcriptional upregulation of Spg7, Afg3l2 and l(2)37Cc (Figure 9b). Taken together, our results suggested that dPerk mediates mitochondrial Ca 2+ levels by restricting the translation of protein products necessary for EMRE processing.
Finally, we identified a network associated with iron-sulphur (Fe-S) clusters (ISCs) (Figure 8c). ISCs are ancient protein cofactors involved in electron transfer reactions that participate in a variety of cellular processes such as enzymatic and redox reactions, respiration, ribosome biogenesis, regulation of gene expression, and DNA-RNA metabolism (reviewed in [55]). In eukaryotes, known Fe-S proteins are located in the mitochondria, cytosol, and nucleus where they perform these diverse functions. Our network includes Nfs1 cysteine desulphurase (Nfs1) and bcn92 transcripts, whose proteins together form the L-cysteine desulphurase complex, to which proteins encoded by Ferredoxins 1 (Fdx1) and 2 (Fdx2) supply the electrons required for the assembly of the cluster. Quetzalcoatl (Qtzl) is another transcript whose protein product is predicted to have cysteine desulphurase activity.
We further detected upregulation in the expression of CG32500, the protein orthologue of the iron-sulphur cluster scaffold protein NFU1 iron-sulphur cluster scaffold, and nucleotide binding protein 1 (Nubp1), a component of the cytosolic Fe-S cluster assembly (CIA) complex as well as Maroon-like (mal), whose protein is involved in molybdenum cofactor (Moco) production, a process that is tightly connected to ISC synthesis [56]. Finally, the network included transcripts coding for Ferrochelatase (FeCH), mitochondrial inner membrane Fe-S protein [57], CG6115 an electron transfer flavoprotein regulatory factor 1 orthologue and Cytochrome b5 (Cyt-b5), an electron transport haemoprotein. In conclusion, using STRING MCL clustering, we defined dPerk dependent networks of translationally repressed proteins involved in glutathione and nucleotide metabolism, as well as Ca 2+ and Fe-S signalling, suggesting that the overexpression of these targets might provide protection against ER stress-related toxicity, as shown with Nmdmc.

Discussion
PERK activation leads to phosphorylation of eIF2α, a key event in many stress signalling pathways [5]. While this modification leads to global inhibition of protein synthesis, it simultaneously enhances the translation of selected mRNAs, such as that of the transcription factor ATF4 [58], eliciting a gene expression programme designed to confer cellular resistance to stress. We have previously demonstrated that in pink and parkin mutants, characterised by increased levels of phospho-eIF2α [19], dAtf4 leads to the upregulation of transcripts, whose protein products confer neuroprotection [20], leading us to hypothesise that dPerk target genes that are upregulated at the transcriptional level but fail to be translated consist of failed protective mechanisms.
In this study we aimed to further characterise this subset of upregulated transcripts not met at the level of the proteome in order to define other salvage pathways that bestow cellular protection from ER toxicity.
To assess the relationship between the induction of mRNA and protein levels, we established a dPerk overexpression model that combined microarray analysis and TMT proteomic analysis. We showed that dPerk overexpression leads to phosphorylation of eIF2α and upregulation of the dAtf4-dependent ER stress marker Nmdmc, as well as a novel Drosophila target trbl, thereby confirming that our model reflects two key dPerk signalling characteristics, translational inhibition and canonical ATF4 signalling.
Our findings capture a dynamic transcriptional and translational response to dPerk overexpression. While upregulated transcripts suggest an induction of a protective transcriptional programme, our results also show a considerable proportion of downregulated transcripts, as previously reported [59]. These transcripts could be negatively regulated by mRNA turnover or via a specific targeted degradation mechanism, such as IRE1-dependent decay of mRNA (RIDD) [60]. When compared, the differential expression analysis shows a proportion of genes consistently regulated as well as numerous targets differentially regulated, by transcripts and proteins, with the latter indicating that dPerk-dependent translation attenuation creates disparity between transcripts and their protein targets.
We defined two subgroups of upregulated transcripts: genes that were upregulated at both the transcript and protein level (group 1) and genes that were upregulated at the transcript level but not the protein level (group 2). From 27 targets representing group 1, consisting of ER stress, translational, mitochondrial and immune-related molecules, we established Hsp22 as a dPerk-dependent, dAtf4-dependent target. Hsp22 is a mitochondrial matrix chaperone [61] thought to be involved in the mitochondrial unfolded protein response (UPR mt ) (reviewed in [62]), suggesting that ER stress may modulate the UPR mt via Hsp22.
Group 2 ClueGO PEA results showed that dPerk overexpression leads to translational repression of mitochondrial proteins. More detailed STRING network analysis identified that previously established nucleotide metabolism was affected, as well as glutathione and Fe-S signalling, all of which are tightly linked to mitochondrial function. A recent study has reported that mitochondrial S-adenosylmethionine (SAM) (mitoSAM) production depends on one-carbon metabolism, and that mitoSAM depletion leads to dysregulation of ISC biosynthesis, linking nucleotide metabolism and ISC synthesis with mitochondrial function [63]. Glutathione metabolism is also related to the methionine cycle, as its product homocysteine can lead to glutathione production via the trans-sulphuration pathway. In accordance with this finding, ATF4 deficient cells require considerably higher concentrations of cysteine to defend against glutathione depletion [64]. Thus, overexpression of these targets might prove beneficial, as it may help cells adapt to the metabolic consequences of ER stress by modulating mitochondrial function and providing protection against oxidative stress.
Alternatively, the absence of translation itself could represent a mechanism by which a cell can regulate its fate. Our results show that dPerk overexpression leads to overexpression of transcripts, though not proteins, of mitochondrial proteases involved in MCU regulatory subunit EMRE processing. The m-AAA proteases have been found to degrade the non-assembled EMRE, and in this manner prevent the assembly of unregulated MCU. Their loss has been determined to result in accumulation of constitutively active MCU-EMRE in neuronal mitochondria, leading to mitochondrial Ca 2+ overload, mitochondrial permeability transition pore opening, and neuronal death [65]. This finding might help explain the paradox of how activation of UPR and ATF4 might simultaneously lead to cell survival and cell death, as absence of the translational recovery of the UPR, leading to a failure in the production of regulatory proteins might function as an execution signal itself.
Our iRegulon results further support the notion that sustained phosphorylation of eIF2α in dPerk overexpressing flies might not be able to sustain efficient translation of ATF4 targets, as ATF4 was seen as the most strongly associated transcription factor of upregulated transcripts, but not proteins. We demonstrated that Hsp22 is an dAtf4-dependent target that is also upregulated in proteins; however, Hsp mRNAs are efficiently translated [66] due to structural features in their 5 UTRs [67], and translation of Hsp22 was confirmed to be unaffected by rapamycin, a cap-dependent translation inhibitor [68]. This finding suggests that ATF4 leads to upregulation of protective transcripts, however, their translation is dependent on specific translational features of individual transcripts, indicating that the protective response by ATF4 is dependent on the dephosphorylation of eIF2α and thus might be compromised during chronic translational repression. Overexpression of these targets therefore might represent a strategy for combating ER stress related toxicity.
Transcripts like ATF4, which function to relieve stress, are preferentially translated via an uORF-mediated mechanism [58]. Our analysis shows a decrease in uORFs in molecules that are upregulated at the transcript and protein levels (group 1). This suggests that the presence of an uORF is insufficient to predict the resistance to eIF2-dependant translational repression, as previously reported [69,70]. It is known that uORFs can function as both repressors and enhancers of translation (reviewed in [71]). It is therefore reasonable to argue that specific properties of the uORFs are critical for their capacity to regulate translation. However, our analysis did not consider the non-canonical initiation codons (e.g., CUG, UUG, and GUG) for uORF prevalence approximation, which have also been shown to serve as competent sites of translation initiation [72][73][74]. Thus, further work is needed to define the role of group 1 uORFs in translation during ER stress.
Our study has some limitations. First, we used the ubiquitous expression of dPerk, which may mask tissue-specific responses. Second, the data acquired by the proteomics analysis contained fewer molecules than those detected by transcriptomic analysis. This discrepancy might underestimate or overestimate the cellular signatures by the PEA, skewing the interpretation of our results.
A recent study using an astrocytoma cell line reported an increase in the flux of metabolites through one-carbon metabolism [75]. Our study confirms that dPerk regulates one-carbon metabolism genes transcriptionally but indicates that the protein levels of, at least some of these enzymes, are not strongly affected by dPerk upregulation in flies. Nevertheless, it highlights one-carbon metabolism as a major metabolic pathway under the control of dPerk and suggests that its manipulation might have therapeutic potential in diseases associated with PERK activation.

Microarray Acquisition and Analysis
The crosses were kept at 18 • C. Following eclosion, adult flies were transferred to 29 • C for 15 h after which the RNA was extracted using the TRIZOL method (4 biological replicates, 5 replicates for each genotype). dPerk was driven by daGAL4, under control of a temperature sensitive tubGal80 driver. The RNA quality was confirmed using an Agilent 2100 Bioanalyzer (Agilent Technologies, CA, USA). Detailed experimental protocols and raw data were deposited in ArrayExpress under accession E-MTAB-6097. Differential expression was analysed using the Partek Genomics Suite (Partek Inc., St. Louis, MO, USA).
TMT labelling was performed according to the manufacturer's recommended protocol (https://www.thermofisher.com/order/catalogue/product/90110/, accessed on 26 April 2021). One hundred micrograms of each digested protein sample was labelled individually with each of the 10 TMT tags. After labelling, the samples were combined, cleaned on a Sep-Pak C18 cartridge, dried and dissolved in 20 mM ammonium formate (pH 10). TMT peptide fractionation was performed using an Acquity ethylene-bridged hybrid C18 UPLC column (Waters; 2.1 mm i.d. × 150 mm, particle size of 1.7 µm). Dried fractions were separated using the LC-MS/MS method as detailed below. The fractions were combined into pairs (i.e., the first fraction with the middle fraction) and analysed by LC-MS/MS using a Dionex Ultimate 3000 RSLC nanoUPLC (Thermo Fisher Scientific Inc, Waltham, MA, USA) system and a Lumos Orbitrap mass spectrometer (Thermo Fisher Scientific Inc., Waltham, MA, USA).
For data analysis, the raw data files were processed using Proteome Discoverer v2.1 (Thermo Fisher Scientific, Waltham, MA, USA) and Mascot (Matrix Science, Chicago, IL, USA) v2.6. The data were aligned with the UniProt data from Pseudomonas aeruginosa (5584 sequences), which is the common repository of adventitious proteins (cRAP, version 1.0). All comparative analyses were performed with the R statistical language. The R package MSnbase [76] was used for the processing of proteomics data. Briefly, this process entailed the removal of missing values (instances where a protein was identified but not quantified in all channels were rejected from further analysis), log2-transformation of the raw data, and subsequent sample normalisation utilising the 'diff.median' method in MSnbase (this translates all samples columns such that they all match the grand median). The differential abundances of the proteins were evaluated using the limma package, and the differences in protein abundances were statistically analysed using Student's t-test with their variances moderated by the empirical Bayes method in limma. The p-values were adjusted for multiple testing using the Benjamini Hochberg method [77].

RNA Extraction and Quantitative Real-Time PCR
RNA was extracted using TRIzol (Ambion, Waltham, MA, USA) and quantified by Nanodrop. Quantitative real-time PCR with reverse transcription (qRT-PCR) was performed on a real-time cycler (Applied Biosystems, Foster City, CA, USA, 7500 Fast Real-Time PCR Systems) using the SensiFAST SYBER Lo-ROX One-Step Kit (Bioline, London, UK). Drosophila gene specific primers were obtained from QIAGEN (Hilden, Germany): trbl RT 2

Protein Extraction and Western Blotting
Protein extracts from whole flies were prepared by grinding flies in lysis buffer (20 mM Tris at pH 7.5, 150 mM NaCl, 1% (v/v) Nonidet-40, 0.5% (w/v) sodium deoxycholate, 0.1% (w/v) containing phosphatase inhibitor cocktail tablets phosSTOP (Roche, Basel, Switzerland)) and the protease inhibitor cocktail tablets cOmplete ULTRA Tablets, Mini, EASYpack (Roche, Basel, Switzerland) at the manufacturer's recommended dilution. Following a 10 min ice incubation, the suspensions were centrifuged at 21,000 g for 10 min at 4 • C. The protein concentration of the supernatant was measured using the Bradford assay (Bio-Rad, Hemel Hempstead, UK). All the supernatants were mixed in 4x Laemmli loading buffer. For SDS-PAGE, equal concentrations of proteins were run in 12.5% gels and transferred onto a methanol-activated PVDF membrane (Millipore, Watford, UK). The membranes were blocked in TBS/T (0.15 M NaCl and 10 mM Tris-HCl, pH = 7.5) containing either 5% (w/v) dried nonfat milk or 5% (w/v) BSA for 1 h at room temperature. The membranes were probed with the primary antibody at 4 • C overnight, followed by incubation with complementary HRP-conjugated secondary antibody. Antibody complexes were visualised by enhanced chemiluminescence (SuperSignal West Dura Extended Duration Substrate, Thermo Fisher Scientific Inc., Waltham, MA, USA). Primary antibodies were used at a dilution 1:1000, unless specified otherwise. Primary antibodies included: Rabbit Antiα-tubulin (Cell Signalling, 2144), Rabbit Anti-EIF2S1 (phospho S51) (Abcam, ab32157), Rabbit Anti-EIF2S1 (Abcam, ab26197), Rat Anti-HA High Affinity (50 mU/mL, Roche 3F10), together with appropriate secondary antibodies diluted either 1:10,000 or 1:50,000. Stripping of the membrane was performed for 30 min at 37 • C using the Restore Western Blot Stripping Buffer reagent (Thermo Fisher Scientific Inc., Waltham, MA, USA).

Functional Pathway Enrichment and Upstream Analysis
Differential expression transcripts and proteins were obtained based on FC and FDR values of ±1.6 and ≤0.05, respectively, using the R package tidyverse [78]. Pathway enrichment analysis was performed using the Cytoscape v3.8.0 application ClueGo [27]. In brief, a list of differentially expressed targets was submitted for GlueGO functional analysis using Drosophila melanogaster [27] as the organism of choice and enriched for 2021 updated Gene Ontology (GO) Cellular Component, GO Biological function, GO Molecular Function, KEGG and Reactome ontologies. Only pathways with a p-value ≤ 0.05 were considered. Enrichment and right-sided hypergeometric tests were performed using the default Bonferroni step down pV correction. Grouping was based on the leading group term based on the highest significance and using default 0.4 as the Kappa score. Network analysis and MCL clustering were performed using the online STRING database [49] and the Cytoscape stringApp application. In brief, a list of differentially expressed targets was submitted to the online STRING database. Following mapping, nodes and edges were exported to the stringApp Cytoscape application. Markov cluster (MCL) clustering was performed with the inbuilt stringApp MCL clustering option. The MCL inflation parameter was chosen based on the unique STRING functional enrichment analysis of a given cluster, once the functional specificity was reached. The defined subcluster was reintroduced into the full background transcriptomic network, and first neighbours were investigated, in order to find additional functionally related targets not recognised by the proteomics analysis. Upstream analysis was performed using the Cytoscape application iRegulon [26]. In brief, a list of differentially expressed targets was imported into Cytoscape and queried for the most prominent motifs and transcription factors in the Drosophila melanogaster 5 kb upstream and the full transcript putative regulatory region.

uORF Analysis
We used the full list of Drosophila uORFs from a recent study by Zhang and colleagues [69]. Briefly, they queried all Drosophila genes from the FlyBase database [79] and identified 7036 genes with at least 1 uORF, according to their definition (starting with AUG and ending with a stop codon UAA/UAG/UGA). Based on this list, we identified genes that contain uORFs in our total set of genes, group 1 and group 2. To assess whether the number of genes sampled in group 1 and group 2 differ from random sampling from the total set of genes, we generated 1000 Monte Carlo simulations of 100 normally distributed integer values based on the mean of the total group. We used a one-sample t-test to assess whether the tested number of genes in either group 1 or group 2 is different from the mean of the distribution. We considered the alternative hypothesis to be significant if 5% of the simulations were significant at a p value < 0.05. For the analysis of uORF features, the number of uORFs per gene were calculated by counting the number of uORFs in all transcripts of a given gene. The uORF length was calculated by subtracting the uORF sequence end with a uORF sequence start site. The detailed script and verifications of this analysis is deposited on GitHub (https://m1gus.github.io/Perk/, accessed on 26 April 2021).

Statistical Analyses
Statistical analyses were performed using GraphPad Prism v9.0.2 (GraphPad Software, San Diego, CA, USA or R Studio v1.4.1103 (R Studio, Boston, MA, USA). The data presentation and the number of biological replicates per experimental variable (n) is indicated in the respective figure. Significance is indicated as * for p < 0.05, ** for p < 0.01, *** for p < 0.001, **** for p < 0.0001 and ns for p ≥ 0.05.

Supplementary Materials:
The following are available online at https://www.mdpi.com/article/10 .3390/ijms22094598/s1, Table S1: Full list of transcripts and proteins in dPerk overexpressing flies, Table S2: Identification of transcripts positively regulated in dPerk overexpressing flies (FC ≥ 1.6, FDR ≤ 0.05), Table S3: Identification of transcripts negatively regulated in dPerk overexpressing flies (FC ≥−1.6, FDR ≤ 0.05), Table S4: Identification of proteins positively regulated in dPerk overexpressing flies (FC ≥0.7 (log), FDR ≤ 0.05), Table S5: Identification of proteins negatively regulated in dPerk overexpressing flies (FC ≥−0.7 (log), FDR ≤ 0.05), Table S6: iRegulon Results. Putative upstream regulators of transcripts upregulated in dPerk overexpressing flies. This table is related to Figure 2, Table S7: iRegulon Results: Putative upstream regulators of proteins upregulated in dPerk overexpressing flies. This table is related to Figure 2, Table S8: ClueGO Results: Pathway enrichment analysis of upregulated transcripts in dPerk overexpressing flies. This table is related to Figure 3, Table S9: ClueGO Results: Pathway enrichment analysis of upregulated proteins in dPerk overexpressing flies. This table is related to Figure 3, Table S10: ClueGO Results: Pathway enrichment analysis of downregulated transcripts in dPerk overexpressing flies. This table is related to Figure 3, Table S11: ClueGO Results: Pathway enrichment analysis of downregulated proteins in dPerk overexpressing flies. This table is related to Figure 3, Table S12: List of matched transcripts and proteins in dPerk overexpressing flies, Table S13: Genes with upregulated transcripts and proteins in dPerk overexpressing flies (group 1). This table is related to Table 1, Table S14: dPerk-dependent targets differentially regulated by the transcriptome and proteome. Genes that are upregulated at the transcript but not at protein level were analysed (group 2), Table S15: Genes with downregulated transcripts and proteins in dPerk overexpressing flies (group 3), Table S16: uORF analysis of all matched transcripts and proteins in dPerk overexpressing flies. This table is related to Figure 5, Table S17: uORF analysis of group 1 genes. This table is related to Figure 5, Table S18: uORF analysis of group 2 genes. This table is related to Figure 5, Table S19: ClueGO Cellular Component Results: Pathway enrichment analysis of dPerk-dependent targets differentially regulated by the transcriptome and proteome (group 2). This table is related to Figure 7 (left), Table S20: ClueGO Results: Pathway enrichment analysis of dPerk-dependent targets differentially regulated by the transcriptome and proteome (group 2). This table is related to Figure 7 (left), Table S21